Nano Claude Code: A Minimal Python Reimplementation
@@ -13,18 +14,27 @@
+
+
+
+
+
+
+
+---
## ๐ฅ๐ฅ๐ฅ News (Pacific Time)
-- 01:47 PM, Apr 01, 2026: Support VLLM inference (**~2000** lines of Python Code)
-- 11:30 AM, Apr 01, 2026: Support more **closed-source** models and **open-source models**: Claude, GPT, Gemini, Kimi, Qwen, Zhipu, DeepSeek, and local open-source models via Ollama or any OpenAI-compatible endpoint. (**~1700** lines of Python Code)
-- 09:50 AM, Apr 01, 2026: Support more **closed-source** models: Claude, GPT, Gemini. (**~1300** lines of Python Code)
-- 08:23 AM, Apr 01, 2026: Release the initial version of Nano Claude Code (**~900 lines** of Python Code)
+- 12:20 PM, Apr 02, 2026: **v3.0** โ Multi-agent packages (`multi_agent/`), memory package (`memory/`), skill package (`skill/`) with built-in skills, argument substitution, fork/inline execution, AI memory search, git worktree isolation, agent type definitions (**~5000** lines of Python), see [update](https://github.com/SafeRL-Lab/nano-claude-code/blob/main/Update_README.MD).
+- 10:00 AM, Apr 02, 2026: **v2.0** โ Context compression, memory, sub-agents, skills, diff view, tool plugin system (**~3400** lines of Python Code).
+- 01:47 PM, Apr 01, 2026: Support VLLM inference (**~2000** lines of Python Code).
+- 11:30 AM, Apr 01, 2026: Support more **closed-source** models and **open-source models**: Claude, GPT, Gemini, Kimi, Qwen, Zhipu, DeepSeek, and local open-source models via Ollama or any OpenAI-compatible endpoint. (**~1700** lines of Python Code).
+- 09:50 AM, Apr 01, 2026: Support more **closed-source** models: Claude, GPT, Gemini. (**~1300** lines of Python Code).
+- 08:23 AM, Apr 01, 2026: Release the initial version of Nano Claude Code (**~900 lines** of Python Code).
+---
# Nano Claude Code
-
-
A minimal Python implementation of Claude Code in ~900 lines (Initial version), **supporting Claude, GPT, Gemini, Kimi, Qwen, Zhipu, DeepSeek, and local open-source models via Ollama or any OpenAI-compatible endpoint.**
---
@@ -32,30 +42,20 @@ A minimal Python implementation of Claude Code in ~900 lines (Initial version),
## Content
* [Features](#features)
* [Supported Models](#supported-models)
- + [Closed-Source (API)](#closed-source--api-)
- + [Open-Source (Local via Ollama)](#open-source--local-via-ollama-)
* [Installation](#installation)
* [Usage: Closed-Source API Models](#usage--closed-source-api-models)
- + [Anthropic Claude](#anthropic-claude)
- + [OpenAI GPT](#openai-gpt)
- + [Google Gemini](#google-gemini)
- + [Kimi (Moonshot AI)](#kimi--moonshot-ai-)
- + [Qwen (Alibaba DashScope)](#qwen--alibaba-dashscope-)
- + [Zhipu GLM](#zhipu-glm)
- + [DeepSeek](#deepseek)
* [Usage: Open-Source Models (Local)](#usage--open-source-models--local-)
- + [Option A โ Ollama (Recommended)](#option-a---ollama--recommended-)
- + [Option B โ LM Studio](#option-b---lm-studio)
- + [Option C โ vLLM / Self-Hosted OpenAI-Compatible Server](#option-c---vllm---self-hosted-openai-compatible-server)
* [Model Name Format](#model-name-format)
* [CLI Reference](#cli-reference)
* [Slash Commands (REPL)](#slash-commands--repl-)
* [Configuring API Keys](#configuring-api-keys)
- + [Method 1: Environment Variables (recommended)](#method-1--environment-variables--recommended-)
- + [Method 2: Set Inside the REPL (persisted)](#method-2--set-inside-the-repl--persisted-)
- + [Method 3: Edit the Config File Directly](#method-3--edit-the-config-file-directly)
* [Permission System](#permission-system)
* [Built-in Tools](#built-in-tools)
+ * [Memory](#memory)
+ * [Skills](#skills)
+ * [Sub-Agents](#sub-agents)
+ * [Context Compression](#context-compression)
+ * [Diff View](#diff-view)
* [CLAUDE.md Support](#claudemd-support)
* [Session Management](#session-management)
* [Project Structure](#project-structure)
@@ -71,10 +71,16 @@ A minimal Python implementation of Claude Code in ~900 lines (Initial version),
| Multi-provider | Anthropic ยท OpenAI ยท Gemini ยท Kimi ยท Qwen ยท Zhipu ยท DeepSeek ยท Ollama ยท LM Studio ยท Custom endpoint |
| Interactive REPL | readline history, Tab-complete slash commands |
| Agent loop | Streaming API + automatic tool-use loop |
-| 8 built-in tools | Read ยท Write ยท Edit ยท Bash ยท Glob ยท Grep ยท WebFetch ยท WebSearch |
+| 18 built-in tools | Read ยท Write ยท Edit ยท Bash ยท Glob ยท Grep ยท WebFetch ยท WebSearch ยท MemorySave ยท MemoryDelete ยท MemorySearch ยท MemoryList ยท Agent ยท SendMessage ยท CheckAgentResult ยท ListAgentTasks ยท ListAgentTypes ยท Skill ยท SkillList |
+| Diff view | Git-style red/green diff display for Edit and Write |
+| Context compression | Auto-compact long conversations to stay within model limits |
+| Persistent memory | Dual-scope memory (user + project) with 4 types, AI search, staleness warnings |
+| Multi-agent | Spawn typed sub-agents (coder/reviewer/researcher/โฆ), git worktree isolation, background mode |
+| Skills | Built-in `/commit` ยท `/review` + custom markdown skills with argument substitution and fork/inline execution |
+| Plugin tools | Register custom tools via `tool_registry.py` |
| Permission system | `auto` / `accept-all` / `manual` modes |
-| 14 slash commands | `/model` ยท `/config` ยท `/save` ยท `/cost` ยท โฆ |
-| Context injection | Auto-loads `CLAUDE.md`, git status, cwd |
+| 17 slash commands | `/model` ยท `/config` ยท `/save` ยท `/cost` ยท `/memory` ยท `/skills` ยท `/agents` ยท โฆ |
+| Context injection | Auto-loads `CLAUDE.md`, git status, cwd, persistent memory |
| Session persistence | Save / load conversations to `~/.nano_claude/sessions/` |
| Extended Thinking | Toggle on/off (Claude models only) |
| Cost tracking | Token usage + estimated USD cost |
@@ -173,6 +179,7 @@ export OPENAI_API_KEY=sk-...
python nano_claude.py --model gpt-4o
python nano_claude.py --model gpt-4o-mini
+python nano_claude.py --model gpt-4.1-mini
python nano_claude.py --model o3-mini
```
@@ -206,9 +213,9 @@ Get your API key at [dashscope.aliyun.com](https://dashscope.aliyun.com).
```bash
export DASHSCOPE_API_KEY=sk-...
-python nano_claude.py --model qwen/qwen-max
-python nano_claude.py --model qwen/qwq-32b
-python nano_claude.py --model qwen/qwen2.5-coder-32b-instruct
+python nano_claude.py --model qwen/Qwen3.5-Plus
+python nano_claude.py --model qwen/Qwen3-MAX
+python nano_claude.py --model qwen/Qwen3.5-Flash
```
### Zhipu GLM
@@ -478,6 +485,10 @@ Type `/` and press **Tab** to autocomplete.
| `/permissions ` | Set permission mode: `auto` / `accept-all` / `manual` |
| `/cwd` | Show current working directory |
| `/cwd ` | Change working directory |
+| `/memory` | List all persistent memories |
+| `/memory ` | Search memories by keyword |
+| `/skills` | List available skills |
+| `/agents` | Show sub-agent task status |
| `/exit` / `/quit` | Exit |
**Switching models inside a session:**
@@ -573,17 +584,247 @@ Keys are saved to `~/.nano_claude/config.json` and loaded automatically on next
## Built-in Tools
+### Core Tools
+
| Tool | Description | Key Parameters |
|---|---|---|
| `Read` | Read file with line numbers | `file_path`, `limit`, `offset` |
-| `Write` | Create or overwrite file | `file_path`, `content` |
-| `Edit` | Exact string replacement in file | `file_path`, `old_string`, `new_string`, `replace_all` |
+| `Write` | Create or overwrite file (shows diff) | `file_path`, `content` |
+| `Edit` | Exact string replacement (shows diff) | `file_path`, `old_string`, `new_string`, `replace_all` |
| `Bash` | Execute shell command | `command`, `timeout` (default 30s) |
| `Glob` | Find files by glob pattern | `pattern` (e.g. `**/*.py`), `path` |
| `Grep` | Regex search in files (uses ripgrep if available) | `pattern`, `path`, `glob`, `output_mode` |
| `WebFetch` | Fetch and extract text from URL | `url`, `prompt` |
| `WebSearch` | Search the web via DuckDuckGo | `query` |
+### Memory Tools
+
+| Tool | Description | Key Parameters |
+|---|---|---|
+| `MemorySave` | Save or update a persistent memory | `name`, `type`, `description`, `content`, `scope` |
+| `MemoryDelete` | Delete a memory by name | `name`, `scope` |
+| `MemorySearch` | Search memories by keyword (or AI ranking) | `query`, `scope`, `use_ai`, `max_results` |
+| `MemoryList` | List all memories with age and metadata | `scope` |
+
+### Sub-Agent Tools
+
+| Tool | Description | Key Parameters |
+|---|---|---|
+| `Agent` | Spawn a sub-agent for a task | `prompt`, `subagent_type`, `isolation`, `name`, `model`, `wait` |
+| `SendMessage` | Send a message to a named background agent | `name`, `message` |
+| `CheckAgentResult` | Check status/result of a background agent | `task_id` |
+| `ListAgentTasks` | List all active and finished agent tasks | โ |
+| `ListAgentTypes` | List available agent type definitions | โ |
+
+### Skill Tools
+
+| Tool | Description | Key Parameters |
+|---|---|---|
+| `Skill` | Invoke a skill by name from within the conversation | `name`, `args` |
+| `SkillList` | List all available skills with triggers and metadata | โ |
+
+> **Adding custom tools:** See [Architecture Guide](docs/architecture.md#tool-registry) for how to register your own tools.
+
+---
+
+## Memory
+
+The model can remember things across conversations using the built-in memory system.
+
+**How it works:** Memories are stored as markdown files. There are two scopes:
+- **User scope** (`~/.nano_claude/memory/`) โ follows you across all projects
+- **Project scope** (`.nano_claude/memory/` in cwd) โ specific to the current repo
+
+A `MEMORY.md` index (โค 200 lines / 25 KB) is auto-rebuilt on every save or delete and injected into the system prompt so Claude always has an overview.
+
+**Memory types:**
+
+| Type | Use for |
+|---|---|
+| `user` | Your role, preferences, background |
+| `feedback` | How you want the model to behave |
+| `project` | Ongoing work, deadlines, decisions |
+| `reference` | Links to external resources |
+
+**Memory file format** (`~/.nano_claude/memory/coding_style.md`):
+```markdown
+---
+name: coding style
+description: Python formatting preferences
+type: feedback
+created: 2026-04-02
+---
+Prefer 4-space indentation and full type hints in all Python code.
+**Why:** user explicitly stated this preference.
+**How to apply:** apply to every Python file written or edited.
+```
+
+**Example interaction:**
+
+```
+You: Remember that I prefer 4-space indentation and type hints in all Python code.
+AI: [calls MemorySave] Memory saved: coding_style [feedback/user]
+
+You: /memory
+ [feedback/user] coding_style (today): Python formatting preferences
+
+You: /memory python
+ [feedback/user] coding_style: Prefers 4-space indent and type hints in Python
+```
+
+**Staleness warnings:** Memories older than 1 day get a freshness note in `/memory` output so you know when to review or update them.
+
+**AI-ranked search:** `MemorySearch(query="...", use_ai=true)` uses the model to rank results by relevance rather than simple keyword matching.
+
+---
+
+## Skills
+
+Skills are reusable prompt templates that give the model specialized capabilities. Two built-in skills ship out of the box โ no setup required.
+
+**Built-in skills:**
+
+| Trigger | Description |
+|---|---|
+| `/commit` | Review staged changes and create a well-structured git commit |
+| `/review [PR]` | Review code or PR diff with structured feedback |
+
+**Quick start โ custom skill:**
+
+```bash
+mkdir -p ~/.nano_claude/skills
+```
+
+Create `~/.nano_claude/skills/deploy.md`:
+
+```markdown
+---
+name: deploy
+description: Deploy to an environment
+triggers: [/deploy]
+allowed-tools: [Bash, Read]
+when_to_use: Use when the user wants to deploy a version to an environment.
+argument-hint: [env] [version]
+arguments: [env, version]
+context: inline
+---
+
+Deploy $VERSION to the $ENV environment.
+Full args: $ARGUMENTS
+```
+
+Now use it:
+
+```
+You: /deploy staging 2.1.0
+AI: [deploys version 2.1.0 to staging]
+```
+
+**Argument substitution:**
+- `$ARGUMENTS` โ the full raw argument string
+- `$ARG_NAME` โ positional substitution by named argument (first word โ first name)
+- Missing args become empty strings
+
+**Execution modes:**
+- `context: inline` (default) โ runs inside current conversation history
+- `context: fork` โ runs as an isolated sub-agent with fresh history; supports `model` override
+
+**Priority** (highest wins): project-level > user-level > built-in
+
+**List skills:** `/skills` โ shows triggers, argument hint, source, and `when_to_use`
+
+**Skill search paths:**
+
+```
+./.nano_claude/skills/ # project-level (overrides user-level)
+~/.nano_claude/skills/ # user-level
+```
+
+---
+
+## Sub-Agents
+
+The model can spawn independent sub-agents to handle tasks in parallel.
+
+**Specialized agent types** โ built-in:
+
+| Type | Optimized for |
+|---|---|
+| `general-purpose` | Research, exploration, multi-step tasks |
+| `coder` | Writing, reading, and modifying code |
+| `reviewer` | Security, correctness, and code quality analysis |
+| `researcher` | Web search and documentation lookup |
+| `tester` | Writing and running tests |
+
+**Basic usage:**
+```
+You: Search this codebase for all TODO comments and summarize them.
+AI: [calls Agent(prompt="...", subagent_type="researcher")]
+ Sub-agent reads files, greps for TODOs...
+ Result: Found 12 TODOs across 5 files...
+```
+
+**Background mode** โ spawn without waiting, collect result later:
+```
+AI: [calls Agent(prompt="run all tests", name="test-runner", wait=false)]
+AI: [continues other work...]
+AI: [calls CheckAgentResult / SendMessage to follow up]
+```
+
+**Git worktree isolation** โ agents work on an isolated branch with no conflicts:
+```
+Agent(prompt="refactor auth module", isolation="worktree")
+```
+The worktree is auto-cleaned up if no changes were made; otherwise the branch name is reported.
+
+**Custom agent types** โ create `~/.nano_claude/agents/myagent.md`:
+```markdown
+---
+name: myagent
+description: Specialized for X
+model: claude-haiku-4-5-20251001
+tools: [Read, Grep, Bash]
+---
+Extra system prompt for this agent type.
+```
+
+**List running agents:** `/agents`
+
+Sub-agents have independent conversation history, share the file system, and are limited to 3 levels of nesting.
+
+---
+
+## Context Compression
+
+Long conversations are automatically compressed to stay within the model's context window.
+
+**Two layers:**
+
+1. **Snip** โ Old tool outputs (file reads, bash results) are truncated after a few turns. Fast, no API cost.
+2. **Auto-compact** โ When token usage exceeds 70% of the context limit, older messages are summarized by the model into a concise recap.
+
+This happens transparently. You don't need to do anything.
+
+---
+
+## Diff View
+
+When the model edits or overwrites a file, you see a git-style diff:
+
+```diff
+ Changes applied to config.py:
+
+--- a/config.py
++++ b/config.py
+@@ -12,7 +12,7 @@
+ "model": "claude-opus-4-6",
+- "max_tokens": 8192,
++ "max_tokens": 16384,
+ "permission_mode": "auto",
+```
+
+Green lines = added, red lines = removed. New file creations show a summary instead.
+
---
## CLAUDE.md Support
@@ -637,19 +878,49 @@ Sessions are stored as JSON in `~/.nano_claude/sessions/`.
```
nano_claude_code/
-โโโ nano_claude.py # Entry point: REPL + slash commands + output rendering (~580 lines)
-โโโ agent.py # Agent loop: neutral message format + tool dispatch (~160 lines)
-โโโ providers.py # Multi-provider: adapters + message format conversion (~480 lines)
-โโโ tools.py # 8 tool implementations + JSON schemas (~360 lines)
-โโโ context.py # System prompt builder: CLAUDE.md + git + cwd (~100 lines)
-โโโ config.py # Config load/save/defaults (~70 lines)
-โโโ demo.py # Demo script (requires API key)
-โโโ make_demo.py # Generates demo.gif and screenshot.png
-โโโ demo.gif # Animated demo
-โโโ screenshot.png # Static screenshot
-โโโ requirements.txt
+โโโ nano_claude.py # Entry point: REPL + slash commands + diff rendering
+โโโ agent.py # Agent loop: streaming, tool dispatch, compaction
+โโโ providers.py # Multi-provider: Anthropic, OpenAI-compat streaming
+โโโ tools.py # Core tools (Read/Write/Edit/Bash/Glob/Grep/Web) + registry wiring
+โโโ tool_registry.py # Tool plugin registry: register, lookup, execute
+โโโ compaction.py # Context compression: snip + auto-summarize
+โโโ context.py # System prompt builder: CLAUDE.md + git + memory
+โโโ config.py # Config load/save/defaults
+โ
+โโโ multi_agent/ # Multi-agent package
+โ โโโ __init__.py # Re-exports
+โ โโโ subagent.py # AgentDefinition, SubAgentManager, worktree helpers
+โ โโโ tools.py # Agent, SendMessage, CheckAgentResult, ListAgentTasks, ListAgentTypes
+โโโ subagent.py # Backward-compat shim โ multi_agent/
+โ
+โโโ memory/ # Memory package
+โ โโโ __init__.py # Re-exports
+โ โโโ types.py # MEMORY_TYPES and format guidance
+โ โโโ store.py # save/load/delete/search, MEMORY.md index rebuilding
+โ โโโ scan.py # MemoryHeader, age/freshness helpers
+โ โโโ context.py # get_memory_context(), truncation, AI search
+โ โโโ tools.py # MemorySave, MemoryDelete, MemorySearch, MemoryList
+โโโ memory.py # Backward-compat shim โ memory/
+โ
+โโโ skill/ # Skill package
+โ โโโ __init__.py # Re-exports; imports builtin to register built-ins
+โ โโโ loader.py # SkillDef, parse, load_skills, find_skill, substitute_arguments
+โ โโโ builtin.py # Built-in skills: /commit, /review
+โ โโโ executor.py # execute_skill(): inline or forked sub-agent
+โ โโโ tools.py # Skill, SkillList
+โโโ skills.py # Backward-compat shim โ skill/
+โ
+โโโ tests/ # 101 unit tests
+ โโโ test_memory.py
+ โโโ test_skills.py
+ โโโ test_subagent.py
+ โโโ test_tool_registry.py
+ โโโ test_compaction.py
+ โโโ test_diff_view.py
```
+> **For developers:** Each feature package (`multi_agent/`, `memory/`, `skill/`) is self-contained. Add custom tools by calling `register_tool(ToolDef(...))` from any module imported by `tools.py`.
+
---
## FAQ
diff --git a/nano-claude-code/Update_README.MD b/nano-claude-code/Update_README.MD
new file mode 100644
index 0000000..9eb60b8
--- /dev/null
+++ b/nano-claude-code/Update_README.MD
@@ -0,0 +1,526 @@
+# Nano Claude Code โ Update Notes
+
+This document describes three major feature additions to nano-claude-code:
+**Multi-Agent**, **Memory**, and **Skill**. Each feature is organized as a
+self-contained Python package, follows the same architectural pattern, and
+includes a backward-compatibility shim so existing code continues to work.
+
+---
+
+## Architecture Overview
+
+All three packages follow the same pattern:
+
+```
+feature/
+ __init__.py โ public re-exports
+ .py โ data model, loading, business logic
+ tools.py โ registers tools into the central tool_registry
+ ...
+feature.py โ backward-compat shim (re-exports from feature/)
+```
+
+The **tool registry** (`tool_registry.py`) is the central hub. Each feature's
+`tools.py` calls `register_tool(ToolDef(...))` at import time. The top-level
+`tools.py` imports all three feature tool modules, triggering auto-registration.
+
+The **agent loop** (`agent.py`) injects `_depth` and `_system_prompt` into the
+`config` dict on every call, so tool functions can read them via `config.get(...)`.
+
+---
+
+## 1. Multi-Agent (`multi_agent/`)
+
+### What it does
+
+Allows Claude to spawn sub-agents โ nested agent loops that run concurrently
+in background threads. Sub-agents can share the parent's context or run in an
+isolated git worktree. The user can send follow-up messages to named background
+agents and retrieve their results.
+
+### Package structure
+
+```
+multi_agent/
+ __init__.py โ re-exports AgentDefinition, SubAgentTask, SubAgentManager, etc.
+ subagent.py โ core: AgentDefinition, SubAgentTask, SubAgentManager, worktree helpers
+ tools.py โ registers: Agent, SendMessage, CheckAgentResult, ListAgentTasks, ListAgentTypes
+subagent.py โ backward-compat shim
+```
+
+### Key classes and functions
+
+**`AgentDefinition`** (`multi_agent/subagent.py`)
+```python
+@dataclass
+class AgentDefinition:
+ name: str
+ description: str
+ system_prompt: str # prepended to base prompt for this agent type
+ model: str # "" = inherit from parent
+ tools: list # [] = all tools
+ source: str # "built-in" | "user" | "project"
+```
+
+**Built-in agent types**: `general-purpose`, `coder`, `reviewer`, `researcher`, `tester`
+
+**Custom agent definitions** โ place a `.md` file with YAML frontmatter in:
+- `~/.nano_claude/agents/.md` (user-level)
+- `.nano_claude/agents/.md` (project-level, takes priority)
+
+Frontmatter format:
+```markdown
+---
+name: my-agent
+description: What this agent does
+model: claude-opus-4-6
+tools: [Read, Glob, Grep]
+---
+Extra system prompt instructions for this agent.
+```
+
+**`SubAgentManager`** (`multi_agent/subagent.py`)
+- `spawn(prompt, config, agent_def, isolation, name, wait)` โ runs agent in thread pool
+- `send_message(task_id_or_name, message)` โ enqueues message to a running background agent
+- `get_result(task_id)` โ returns final text or status
+- `list_tasks()` โ returns all SubAgentTask objects
+
+**Git worktree isolation**:
+When `isolation="worktree"` is passed to `Agent`, a temporary git worktree is
+created on a fresh branch. The sub-agent works in isolation; if it makes no
+changes the worktree is cleaned up automatically.
+
+### Tools registered
+
+| Tool | Description |
+|------|-------------|
+| `Agent` | Spawn a sub-agent (sync or background with `wait=false`) |
+| `SendMessage` | Send a follow-up message to a named background agent |
+| `CheckAgentResult` | Poll status / retrieve result of a background agent |
+| `ListAgentTasks` | List all active and finished sub-agent tasks |
+| `ListAgentTypes` | List all available agent type definitions |
+
+### Agent tool parameters
+
+```python
+Agent(
+ prompt="...", # required โ task description
+ subagent_type="coder", # optional โ use a specialized agent
+ isolation="worktree", # optional โ isolated git branch
+ name="my-agent", # optional โ name for SendMessage later
+ wait=False, # optional โ run in background
+ model="...", # optional โ model override
+)
+```
+
+### How it was wired in
+
+1. `multi_agent/subagent.py` uses **absolute imports** (`import agent as _agent_mod`)
+ because the project root is in `sys.path` when running from that directory.
+2. `agent.py` was updated to inject `_system_prompt` into `config`:
+ ```python
+ config = {**config, "_depth": depth, "_system_prompt": system_prompt}
+ ```
+3. `tools.py` (top-level) was updated to pass `config` through to the registry:
+ ```python
+ return _registry_execute(name, inputs, cfg)
+ ```
+ and at the bottom:
+ ```python
+ import multi_agent.tools as _multiagent_tools
+ ```
+4. `context.py` system prompt template lists Agent, SendMessage, etc. under
+ `## Multi-Agent`.
+5. `nano_claude.py` `/agents` command calls `get_agent_manager().list_tasks()`
+ and prints status/worktree info. A `_print_background_notifications()` function
+ checks for newly completed background agents before each user prompt.
+
+### Files changed
+
+| File | Change |
+|------|--------|
+| `multi_agent/__init__.py` | Created (re-exports) |
+| `multi_agent/subagent.py` | Created (moved + enhanced from `subagent.py`) |
+| `multi_agent/tools.py` | Created (tool registrations) |
+| `subagent.py` | Converted to backward-compat shim |
+| `agent.py` | Inject `_system_prompt` into config |
+| `tools.py` | Pass config to registry; import `multi_agent.tools` |
+| `context.py` | Add Multi-Agent section to system prompt |
+| `nano_claude.py` | `/agents` command; background notification; `_tool_desc()` |
+| `tests/test_subagent.py` | Update imports to `multi_agent.subagent` |
+
+---
+
+## 2. Memory (`memory/`)
+
+### What it does
+
+Provides persistent, file-based memory across sessions. Memories are stored as
+markdown files with YAML frontmatter. There are two scopes โ **user** (global,
+`~/.nano_claude/memory/`) and **project** (per-repo, `.nano_claude/memory/`).
+A `MEMORY.md` index is auto-rebuilt after every save/delete and injected into
+the system prompt so Claude knows what memories exist.
+
+### Package structure
+
+```
+memory/
+ __init__.py โ re-exports all public symbols
+ types.py โ MEMORY_TYPES, type descriptions, format guidance
+ store.py โ MemoryEntry, save/load/delete/search, index rebuilding
+ scan.py โ MemoryHeader, scan_memory_dir, age/freshness helpers
+ context.py โ get_memory_context(), find_relevant_memories(), truncation
+ tools.py โ registers: MemorySave, MemoryDelete, MemorySearch, MemoryList
+memory.py โ backward-compat shim
+```
+
+### Memory types
+
+Defined in `memory/types.py`, mirrors the four types from Claude Code:
+
+| Type | Purpose |
+|------|---------|
+| `user` | User's role, goals, preferences |
+| `feedback` | Corrections and confirmed approaches |
+| `project` | Ongoing work, decisions, deadlines |
+| `reference` | Pointers to external resources |
+
+### Storage layout
+
+```
+~/.nano_claude/memory/
+ MEMORY.md โ auto-generated index (<=200 lines, <=25 KB)
+ my_note.md
+ feedback_testing.md
+ ...
+
+.nano_claude/memory/ โ project-local (relative to cwd)
+ MEMORY.md
+ ...
+```
+
+Each memory file format:
+```markdown
+---
+name: My Note
+description: one-line description for relevance decisions
+type: user
+created: 2026-04-02
+---
+
+Memory content goes here.
+**Why:** ...
+**How to apply:** ...
+```
+
+### Key API
+
+**`memory/store.py`**
+```python
+save_memory(entry: MemoryEntry, scope="user") # save or update (same name = update)
+delete_memory(name: str, scope="user") # remove entry + rebuild index
+load_entries(scope="user") -> list[MemoryEntry] # load all entries for scope
+load_index(scope="all") -> list[MemoryEntry] # "all" merges user + project
+search_memory(query: str, scope="all") -> list # keyword search across content+name
+get_index_content(scope="all") -> str # raw MEMORY.md text
+```
+
+**`memory/scan.py`**
+```python
+scan_memory_dir(mem_dir, scope) -> list[MemoryHeader] # newest-first, capped at 200
+scan_all_memories() -> list[MemoryHeader] # user + project merged
+memory_age_str(mtime_s) -> str # "today" | "yesterday" | "N days ago"
+memory_freshness_text(mtime_s) -> str # staleness warning for memories >1 day old
+format_memory_manifest(headers) -> str # formatted list for display
+```
+
+**`memory/context.py`**
+```python
+get_memory_context() -> str # injected into system prompt
+truncate_index_content(raw) -> str # enforces <=200 lines / <=25 KB
+find_relevant_memories(query, max_results=5, use_ai=False, config=None)
+```
+
+`find_relevant_memories` supports optional AI ranking: when `use_ai=True` it
+makes a small API call to rank candidates by relevance to the query.
+
+### Tools registered
+
+| Tool | Parameters | Description |
+|------|-----------|-------------|
+| `MemorySave` | `name, description, type, content, scope` | Save or update a memory |
+| `MemoryDelete` | `name, scope` | Delete a memory by name |
+| `MemorySearch` | `query, scope, use_ai, max_results` | Search by keyword (or AI) |
+| `MemoryList` | `scope` | List all memories with age and metadata |
+
+### Index truncation
+
+The `MEMORY.md` index is truncated before being injected into the system prompt:
+- Hard limit: **200 lines** (mirrors Claude Code's limit)
+- Byte limit: **25 000 bytes** (mirrors Claude Code's limit)
+- A `WARNING:` line is appended when either limit is hit
+
+### How it was wired in
+
+1. `memory/store.py` exports `USER_MEMORY_DIR` and `get_project_memory_dir` as
+ module-level names so tests can monkeypatch them cleanly.
+2. `context.py` (system prompt builder) calls `get_memory_context()` at the end
+ of `build_system_prompt()` and appends the result.
+3. `tools.py` (top-level) adds:
+ ```python
+ import memory.tools as _memory_tools
+ ```
+4. `memory.py` (top-level) is now a shim:
+ ```python
+ from memory.store import MemoryEntry, save_memory, ...
+ from memory.context import get_memory_context
+ ```
+5. `nano_claude.py` `/memory` command uses `scan_all_memories()` to display a
+ mtime-sorted list with freshness warnings.
+
+### Files changed
+
+| File | Change |
+|------|--------|
+| `memory/__init__.py` | Created (re-exports) |
+| `memory/types.py` | Created (MEMORY_TYPES, descriptions, format guidance) |
+| `memory/store.py` | Created (replaced top-level `memory.py` logic) |
+| `memory/scan.py` | Created (MemoryHeader, age/freshness, manifest) |
+| `memory/context.py` | Created (context injection, truncation, AI search) |
+| `memory/tools.py` | Created (MemorySave, MemoryDelete, MemorySearch, MemoryList) |
+| `memory.py` | Converted to backward-compat shim |
+| `tools.py` | Import `memory.tools` |
+| `context.py` | Call `get_memory_context()` in `build_system_prompt()` |
+| `nano_claude.py` | `/memory` command uses `scan_all_memories()` |
+| `tests/test_memory.py` | Completely rewritten (101 tests total) |
+
+---
+
+## 3. Skill (`skill/`)
+
+### What it does
+
+Skills are reusable prompt templates stored as markdown files. A user types
+`/commit` or `/review pr-123` in the REPL and the skill's prompt (with
+arguments substituted) is injected into the conversation. Skills can run
+**inline** (current conversation context) or **forked** (isolated sub-agent).
+Two built-in skills (`/commit`, `/review`) are registered programmatically.
+
+### Package structure
+
+```
+skill/
+ __init__.py โ re-exports all public symbols; imports builtin to register them
+ loader.py โ SkillDef dataclass, file parsing, load_skills, find_skill, substitute_arguments
+ builtin.py โ built-in skills: /commit, /review
+ executor.py โ execute_skill() (inline or forked)
+ tools.py โ registers: Skill, SkillList
+skills.py โ backward-compat shim
+```
+
+### Skill file format
+
+Place `.md` files in:
+- `~/.nano_claude/skills/.md` (user-level)
+- `.nano_claude/skills/.md` (project-level, takes priority)
+
+```markdown
+---
+name: deploy
+description: Deploy to an environment
+triggers: [/deploy]
+allowed-tools: [Bash, Read]
+when_to_use: Use when the user wants to deploy. Examples: '/deploy staging v1.2'
+argument-hint: [env] [version]
+arguments: [env, version]
+context: inline
+---
+
+Deploy $VERSION to $ENV.
+
+Full args provided: $ARGUMENTS
+```
+
+### Frontmatter fields
+
+| Field | Default | Description |
+|-------|---------|-------------|
+| `name` | required | Skill identifier |
+| `description` | `""` | One-line description shown in `/skills` |
+| `triggers` | `[/]` | Slash commands or phrases that activate this skill |
+| `allowed-tools` / `tools` | `[]` | Tools the skill is allowed to use |
+| `when_to_use` | `""` | Guidance for when Claude should auto-invoke |
+| `argument-hint` | `""` | Hint shown in `/skills` list, e.g. `[branch] [desc]` |
+| `arguments` | `[]` | Named argument list for `$ARG_NAME` substitution |
+| `model` | `""` | Model override (fork context only) |
+| `user-invocable` | `true` | Show in `/skills` list |
+| `context` | `inline` | `inline` = current conversation, `fork` = isolated sub-agent |
+
+### Argument substitution
+
+`substitute_arguments(prompt, args, arg_names)` in `skill/loader.py`:
+
+- `$ARGUMENTS` โ the full raw args string
+- `$ARG_NAME` โ positional substitution (first word โ first arg name, etc.)
+- Missing args become empty strings
+
+```
+prompt: "Deploy $VERSION to $ENV. Full: $ARGUMENTS"
+args: "1.0 staging"
+arg_names: ["env", "version"]
+
+result: "Deploy staging to 1.0. Full: 1.0 staging"
+```
+
+### Execution modes
+
+**Inline** (`context: inline`, default):
+- Skill prompt is injected into the current `AgentState`
+- History is shared โ the user can see and continue the conversation
+
+**Fork** (`context: fork`):
+- A fresh `AgentState` is created (no shared history)
+- Optional `model` and `allowed-tools` overrides are applied
+- Good for self-contained tasks that don't need mid-process user input
+
+### Built-in skills
+
+Defined in `skill/builtin.py` and registered via `register_builtin_skill()`:
+
+| Trigger | Name | Description |
+|---------|------|-------------|
+| `/commit` | commit | Review staged changes and create a well-structured git commit |
+| `/review`, `/review-pr` | review | Review code or PR diff with structured feedback |
+
+Project-level skill files with the same name override built-ins.
+
+### Tools registered
+
+| Tool | Parameters | Description |
+|------|-----------|-------------|
+| `Skill` | `name, args` | Invoke a skill by name from inside a conversation |
+| `SkillList` | โ | List all available skills with triggers and metadata |
+
+### Priority order
+
+When multiple skill sources define the same name, the highest priority wins:
+
+```
+builtin < user (~/.nano_claude/skills/) < project (.nano_claude/skills/)
+```
+
+### REPL usage
+
+```
+/commit # run built-in commit skill
+/review 123 # review PR #123 (args = "123")
+/deploy staging 2.1.0 # custom skill with named args
+/skills # list all skills
+```
+
+The `/skills` command output includes source label, triggers, argument hint,
+and the first 80 chars of `when_to_use` per skill.
+
+### How it was wired in
+
+1. `skill/__init__.py` imports `skill.builtin` which calls `register_builtin_skill()`
+ for each built-in โ just importing the package registers them.
+2. `tools.py` (top-level) adds:
+ ```python
+ import skill.tools as _skill_tools
+ ```
+3. `skills.py` (top-level) becomes a shim re-exporting from `skill/`.
+4. `context.py` adds a `## Skills` section listing `Skill` and `SkillList`.
+5. `nano_claude.py`:
+ - `cmd_skills` imports from `skill`, shows `when_to_use` and source label
+ - `handle_slash` imports `find_skill` from `skill`; returns `(skill, args)` tuple
+ - REPL loop calls `substitute_arguments` before building the injected message
+
+### Files changed
+
+| File | Change |
+|------|--------|
+| `skill/__init__.py` | Created (re-exports; imports builtin) |
+| `skill/loader.py` | Created (SkillDef, parse, load, find, substitute) |
+| `skill/builtin.py` | Created (/commit, /review built-ins) |
+| `skill/executor.py` | Created (inline + fork execution) |
+| `skill/tools.py` | Created (Skill, SkillList tool registration) |
+| `skills.py` | Converted to backward-compat shim |
+| `tools.py` | Import `skill.tools` |
+| `context.py` | Add Skills section to system prompt |
+| `nano_claude.py` | `cmd_skills`, `handle_slash`, REPL loop updated |
+| `tests/test_skills.py` | Rewritten (22 tests; patches `skill.loader`) |
+
+---
+
+## How to add custom agents, memories, and skills
+
+### Custom agent type
+
+Create `~/.nano_claude/agents/myagent.md`:
+```markdown
+---
+name: myagent
+description: Does specialized work
+model: claude-haiku-4-5-20251001
+tools: [Read, Grep, Bash]
+---
+You are specialized in X. Focus on Y. Never do Z.
+```
+
+Then use: `Agent(prompt="...", subagent_type="myagent")`
+
+### Custom memory
+
+Use the REPL `MemorySave` tool or write a file directly to
+`~/.nano_claude/memory/my_note.md` with frontmatter:
+```markdown
+---
+name: my note
+description: short description
+type: feedback
+created: 2026-04-02
+---
+Memory content here.
+```
+
+### Custom skill
+
+Create `~/.nano_claude/skills/myskill.md` (user-level) or
+`.nano_claude/skills/myskill.md` (project-level):
+```markdown
+---
+name: myskill
+description: Does something useful
+triggers: [/myskill]
+arguments: [target]
+argument-hint: [target]
+when_to_use: Use when the user wants to do X with a target.
+---
+
+Do something useful with $TARGET.
+
+Full context: $ARGUMENTS
+```
+
+Then invoke with `/myskill some-target`.
+
+---
+
+## Running tests
+
+```bash
+cd nano-claude-code
+
+# All tests
+python -m pytest tests/ -v
+
+# Per-feature
+python -m pytest tests/test_subagent.py -v # multi-agent
+python -m pytest tests/test_memory.py -v # memory
+python -m pytest tests/test_skills.py -v # skills
+```
+
+Total: **101 tests**, all passing. Each feature's tests use `monkeypatch` to
+redirect file system paths to `tmp_path` so no real `~/.nano_claude/`
+directories are touched during testing.
diff --git a/nano-claude-code/agent.py b/nano-claude-code/agent.py
index 0cbf708..4964a8a 100644
--- a/nano-claude-code/agent.py
+++ b/nano-claude-code/agent.py
@@ -5,8 +5,11 @@ import uuid
from dataclasses import dataclass, field
from typing import Generator
-from tools import TOOL_SCHEMAS, execute_tool
+from tool_registry import get_tool_schemas
+from tools import execute_tool
+import tools as _tools_init # ensure built-in tools are registered on import
from providers import stream, AssistantTurn, TextChunk, ThinkingChunk, detect_provider
+from compaction import maybe_compact
# โโ Re-export event types (used by nano_claude.py) โโโโโโโโโโโโโโโโโโโโโโโโ
__all__ = [
@@ -54,25 +57,39 @@ def run(
state: AgentState,
config: dict,
system_prompt: str,
+ depth: int = 0,
+ cancel_check=None,
) -> Generator:
"""
Multi-turn agent loop (generator).
Yields: TextChunk | ThinkingChunk | ToolStart | ToolEnd |
PermissionRequest | TurnDone
+
+ Args:
+ depth: sub-agent nesting depth, 0 for top-level
+ cancel_check: callable returning True to abort the loop early
"""
# Append user turn in neutral format
state.messages.append({"role": "user", "content": user_message})
+ # Inject runtime metadata into config so tools (e.g. Agent) can access it
+ config = {**config, "_depth": depth, "_system_prompt": system_prompt}
+
while True:
+ if cancel_check and cancel_check():
+ return
state.turn_count += 1
assistant_turn: AssistantTurn | None = None
+ # Compact context if approaching window limit
+ maybe_compact(state, config)
+
# Stream from provider (auto-detected from model name)
for event in stream(
model=config["model"],
system=system_prompt,
messages=state.messages,
- tool_schemas=TOOL_SCHEMAS,
+ tool_schemas=get_tool_schemas(),
config=config,
):
if isinstance(event, (TextChunk, ThinkingChunk)):
@@ -114,6 +131,7 @@ def run(
result = execute_tool(
tc["name"], tc["input"],
permission_mode="accept-all", # already gate-checked above
+ config=config,
)
yield ToolEnd(tc["name"], result, permitted)
diff --git a/nano-claude-code/compaction.py b/nano-claude-code/compaction.py
new file mode 100644
index 0000000..b92714e
--- /dev/null
+++ b/nano-claude-code/compaction.py
@@ -0,0 +1,196 @@
+"""Context window management: two-layer compression for long conversations."""
+from __future__ import annotations
+
+import providers
+
+
+# โโ Token estimation โโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโ
+
+def estimate_tokens(messages: list) -> int:
+ """Estimate token count by summing content lengths / 3.5.
+
+ Args:
+ messages: list of message dicts with "content" field (str or list of dicts)
+ Returns:
+ approximate token count, int
+ """
+ total_chars = 0
+ for m in messages:
+ content = m.get("content", "")
+ if isinstance(content, str):
+ total_chars += len(content)
+ elif isinstance(content, list):
+ for block in content:
+ if isinstance(block, dict):
+ # Sum all string values in the block
+ for v in block.values():
+ if isinstance(v, str):
+ total_chars += len(v)
+ # Also count tool_calls if present
+ for tc in m.get("tool_calls", []):
+ if isinstance(tc, dict):
+ for v in tc.values():
+ if isinstance(v, str):
+ total_chars += len(v)
+ return int(total_chars / 3.5)
+
+
+def get_context_limit(model: str) -> int:
+ """Look up context window size for a model.
+
+ Args:
+ model: model string (e.g. "claude-opus-4-6", "ollama/llama3.3")
+ Returns:
+ context limit in tokens
+ """
+ provider_name = providers.detect_provider(model)
+ prov = providers.PROVIDERS.get(provider_name, {})
+ return prov.get("context_limit", 128000)
+
+
+# โโ Layer 1: Snip old tool results โโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโ
+
+def snip_old_tool_results(
+ messages: list,
+ max_chars: int = 2000,
+ preserve_last_n_turns: int = 6,
+) -> list:
+ """Truncate tool-role messages older than preserve_last_n_turns from end.
+
+ For old tool messages whose content exceeds max_chars, keep the first half
+ and last quarter, inserting '[... N chars snipped ...]' in between.
+ Mutates in place and returns the same list.
+
+ Args:
+ messages: list of message dicts (mutated in place)
+ max_chars: maximum character length before truncation
+ preserve_last_n_turns: number of messages from end to preserve
+ Returns:
+ the same messages list (mutated)
+ """
+ cutoff = max(0, len(messages) - preserve_last_n_turns)
+ for i in range(cutoff):
+ m = messages[i]
+ if m.get("role") != "tool":
+ continue
+ content = m.get("content", "")
+ if not isinstance(content, str) or len(content) <= max_chars:
+ continue
+ first_half = content[: max_chars // 2]
+ last_quarter = content[-(max_chars // 4):]
+ snipped = len(content) - len(first_half) - len(last_quarter)
+ m["content"] = f"{first_half}\n[... {snipped} chars snipped ...]\n{last_quarter}"
+ return messages
+
+
+# โโ Layer 2: Auto-compact โโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโ
+
+def find_split_point(messages: list, keep_ratio: float = 0.3) -> int:
+ """Find index that splits messages so ~keep_ratio of tokens are in the recent portion.
+
+ Walks backwards from end, accumulating token estimates, and returns the
+ index where the recent portion reaches ~keep_ratio of total tokens.
+
+ Args:
+ messages: list of message dicts
+ keep_ratio: fraction of tokens to keep in the recent portion
+ Returns:
+ split index (messages[:idx] = old, messages[idx:] = recent)
+ """
+ total = estimate_tokens(messages)
+ target = int(total * keep_ratio)
+ running = 0
+ for i in range(len(messages) - 1, -1, -1):
+ running += estimate_tokens([messages[i]])
+ if running >= target:
+ return i
+ return 0
+
+
+def compact_messages(messages: list, config: dict) -> list:
+ """Compress old messages into a summary via LLM call.
+
+ Splits at find_split_point, summarizes old portion, returns
+ [summary_msg, ack_msg, *recent_messages].
+
+ Args:
+ messages: full message list
+ config: agent config dict (must contain "model")
+ Returns:
+ new compacted message list
+ """
+ split = find_split_point(messages)
+ if split <= 0:
+ return messages
+
+ old = messages[:split]
+ recent = messages[split:]
+
+ # Build summary request
+ old_text = ""
+ for m in old:
+ role = m.get("role", "?")
+ content = m.get("content", "")
+ if isinstance(content, str):
+ old_text += f"[{role}]: {content[:500]}\n"
+ elif isinstance(content, list):
+ old_text += f"[{role}]: (structured content)\n"
+
+ summary_prompt = (
+ "Summarize the following conversation history concisely. "
+ "Preserve key decisions, file paths, tool results, and context "
+ "needed to continue the conversation:\n\n" + old_text
+ )
+
+ # Call LLM for summary
+ summary_text = ""
+ for event in providers.stream(
+ model=config["model"],
+ system="You are a concise summarizer.",
+ messages=[{"role": "user", "content": summary_prompt}],
+ tool_schemas=[],
+ config=config,
+ ):
+ if isinstance(event, providers.TextChunk):
+ summary_text += event.text
+
+ summary_msg = {
+ "role": "user",
+ "content": f"[Previous conversation summary]\n{summary_text}",
+ }
+ ack_msg = {
+ "role": "assistant",
+ "content": "Understood. I have the context from the previous conversation. Let's continue.",
+ }
+ return [summary_msg, ack_msg, *recent]
+
+
+# โโ Main entry โโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโ
+
+def maybe_compact(state, config: dict) -> bool:
+ """Check if context window is getting full and compress if needed.
+
+ Runs snip_old_tool_results first, then auto-compact if still over threshold.
+
+ Args:
+ state: AgentState with .messages list
+ config: agent config dict (must contain "model")
+ Returns:
+ True if compaction was performed
+ """
+ model = config.get("model", "")
+ limit = get_context_limit(model)
+ threshold = limit * 0.7
+
+ if estimate_tokens(state.messages) <= threshold:
+ return False
+
+ # Layer 1: snip old tool results
+ snip_old_tool_results(state.messages)
+
+ if estimate_tokens(state.messages) <= threshold:
+ return True
+
+ # Layer 2: auto-compact
+ state.messages = compact_messages(state.messages, config)
+ return True
diff --git a/nano-claude-code/config.py b/nano-claude-code/config.py
index 9008729..1de2244 100644
--- a/nano-claude-code/config.py
+++ b/nano-claude-code/config.py
@@ -16,6 +16,9 @@ DEFAULTS = {
"thinking": False,
"thinking_budget": 10000,
"custom_base_url": "", # for "custom" provider
+ "max_tool_output": 32000,
+ "max_agent_depth": 3,
+ "max_concurrent_agents": 3,
# Per-provider API keys (optional; env vars take priority)
# "anthropic_api_key": "sk-ant-..."
# "openai_api_key": "sk-..."
diff --git a/nano-claude-code/context.py b/nano-claude-code/context.py
index 6b12e20..f6cb435 100644
--- a/nano-claude-code/context.py
+++ b/nano-claude-code/context.py
@@ -4,11 +4,15 @@ import subprocess
from pathlib import Path
from datetime import datetime
+from memory import get_memory_context
+
SYSTEM_PROMPT_TEMPLATE = """\
-You are Nano Claude Code, Created by SAIL Lab (Safe AI and Robot Learning Lab), an AI coding assistant running in the terminal.
+You are Nano Claude Code, Created by SAIL Lab (Safe AI and Robot Learning Lab at UC Berkeley), an AI coding assistant running in the terminal.
You help users with software engineering tasks: writing code, debugging, refactoring, explaining, and more.
# Available Tools
+
+## File & Shell
- **Read**: Read file contents with line numbers
- **Write**: Create or overwrite files
- **Edit**: Replace text in a file (exact string replacement)
@@ -18,6 +22,27 @@ You help users with software engineering tasks: writing code, debugging, refacto
- **WebFetch**: Fetch and extract content from a URL
- **WebSearch**: Search the web via DuckDuckGo
+## Multi-Agent
+- **Agent**: Spawn a sub-agent to handle a task autonomously. Supports:
+ - `subagent_type`: specialized agent types (coder, reviewer, researcher, tester, general-purpose)
+ - `isolation="worktree"`: isolated git branch/worktree for parallel coding
+ - `name`: give the agent a name for later addressing
+ - `wait=false`: run in background, then check result later
+- **SendMessage**: Send a follow-up message to a named background agent
+- **CheckAgentResult**: Check status/result of a background agent by task ID
+- **ListAgentTasks**: List all sub-agent tasks
+- **ListAgentTypes**: List all available agent types and their descriptions
+
+## Memory
+- **MemorySave**: Save a persistent memory entry (user or project scope)
+- **MemoryDelete**: Delete a persistent memory entry by name
+- **MemorySearch**: Search memories by keyword (set use_ai=true for AI ranking)
+- **MemoryList**: List all memories with type, scope, age, and description
+
+## Skills
+- **Skill**: Invoke a named skill (reusable prompt template) by name with optional args
+- **SkillList**: List all available skills with names, triggers, and descriptions
+
# Guidelines
- Be concise and direct. Lead with the answer.
- Prefer editing existing files over creating new ones.
@@ -27,6 +52,12 @@ You help users with software engineering tasks: writing code, debugging, refacto
- For multi-step tasks, work through them systematically.
- If a task is unclear, ask for clarification before proceeding.
+## Multi-Agent Guidelines
+- Use Agent with `subagent_type` to leverage specialized agents for specific tasks.
+- Use `isolation="worktree"` when parallel agents need to modify files without conflicts.
+- Use `wait=false` + `name=...` to run multiple agents in parallel, then collect results.
+- Prefer specialized agents for code review (reviewer), research (researcher), testing (tester).
+
# Environment
- Current date: {date}
- Working directory: {cwd}
@@ -91,10 +122,14 @@ def get_claude_md() -> str:
def build_system_prompt() -> str:
import platform
- return SYSTEM_PROMPT_TEMPLATE.format(
+ prompt = SYSTEM_PROMPT_TEMPLATE.format(
date=datetime.now().strftime("%Y-%m-%d %A"),
cwd=str(Path.cwd()),
platform=platform.system(),
git_info=get_git_info(),
claude_md=get_claude_md(),
)
+ memory_ctx = get_memory_context()
+ if memory_ctx:
+ prompt += f"\n\n# Memory\nYour persistent memories:\n{memory_ctx}\n"
+ return prompt
diff --git a/nano-claude-code/docs/architecture.md b/nano-claude-code/docs/architecture.md
new file mode 100644
index 0000000..c1794e4
--- /dev/null
+++ b/nano-claude-code/docs/architecture.md
@@ -0,0 +1,374 @@
+# Architecture Guide
+
+This document is for developers who want to understand, modify, or extend nano-claude-code.
+For user-facing docs, see [README.md](../README.md).
+
+---
+
+## Overview
+
+Nano-claude-code is a ~3.4K-line Python CLI that lets LLMs (GPT, Gemini, etc.) operate as
+coding agents with tool use, memory, sub-agents, and skills. The architecture is a flat
+module layout designed for readability and future migration to a package structure.
+
+```
+User Input
+ โ
+ โผ
+nano_claude.py โโ REPL, slash commands, rendering
+ โ
+ โโโโบ agent.py โโ multi-turn loop, permission gates
+ โ โ
+ โ โโโโบ providers.py โโ API streaming (Anthropic / OpenAI-compat)
+ โ โโโโบ tool_registry.py โโโบ tools.py โโ 13 tools
+ โ โโโโบ compaction.py โโ context window management
+ โ โโโโบ subagent.py โโ threaded sub-agent lifecycle
+ โ
+ โโโโบ context.py โโ system prompt (git, CLAUDE.md, memory)
+ โ โโโโบ memory.py โโ persistent file-based memory
+ โ
+ โโโโบ skills.py โโ markdown skill loading + execution
+ โโโโบ config.py โโ configuration persistence
+```
+
+**Key invariant:** Dependencies flow downward. No circular imports at the module level
+(subagent.py uses lazy imports to call agent.py).
+
+---
+
+## Module Reference
+
+### `tool_registry.py` โ Tool Plugin System
+
+The central registry that all tools register into. This is the foundation for extensibility.
+
+**Data model:**
+
+```python
+@dataclass
+class ToolDef:
+ name: str # unique identifier (e.g. "Read", "MemorySave")
+ schema: dict # JSON schema sent to the LLM API
+ func: Callable # (params: dict, config: dict) -> str
+ read_only: bool # True = auto-approve in 'auto' permission mode
+ concurrent_safe: bool # True = safe to run in parallel (for sub-agents)
+```
+
+**Public API:**
+
+| Function | Description |
+|---|---|
+| `register_tool(tool_def)` | Add a tool to the registry (overwrites by name) |
+| `get_tool(name)` | Look up by name, returns `None` if not found |
+| `get_all_tools()` | List all registered tools |
+| `get_tool_schemas()` | Return schemas for API calls |
+| `execute_tool(name, params, config, max_output=32000)` | Execute with output truncation |
+| `clear_registry()` | Reset โ for testing only |
+
+**Output truncation:** If a tool returns more than `max_output` chars, the result is
+truncated to `first_half + [... N chars truncated ...] + last_quarter`. This prevents
+a single tool call (e.g. reading a huge file) from blowing up the context window.
+
+**Registering a custom tool:**
+
+```python
+from tool_registry import ToolDef, register_tool
+
+def my_tool(params, config):
+ return f"Hello, {params['name']}!"
+
+register_tool(ToolDef(
+ name="MyTool",
+ schema={
+ "name": "MyTool",
+ "description": "A greeting tool",
+ "input_schema": {
+ "type": "object",
+ "properties": {"name": {"type": "string"}},
+ "required": ["name"],
+ },
+ },
+ func=my_tool,
+ read_only=True,
+ concurrent_safe=True,
+))
+```
+
+### `tools.py` โ Built-in Tool Implementations
+
+Contains the 8 core tools (Read, Write, Edit, Bash, Glob, Grep, WebFetch, WebSearch)
+plus memory tools (MemorySave, MemoryDelete) and sub-agent tools (Agent, CheckAgentResult,
+ListAgentTasks). All register themselves via `tool_registry` at import time.
+
+**Key internals:**
+
+- `_is_safe_bash(cmd)` โ whitelist of safe shell commands for auto-approval
+- `generate_unified_diff(old, new, filename)` โ diff generation for Edit/Write
+- `maybe_truncate_diff(diff_text, max_lines=80)` โ truncate large diffs for display
+- `_get_agent_manager()` โ lazy singleton for SubAgentManager
+- Backward-compatible `execute_tool(name, inputs, permission_mode, ask_permission)` wrapper
+
+### `agent.py` โ Core Agent Loop
+
+The heart of the system. `run()` is a generator that yields events as they happen.
+
+```python
+def run(user_message, state, config, system_prompt,
+ depth=0, cancel_check=None) -> Generator:
+```
+
+**Loop logic:**
+
+```
+1. Append user message
+2. Inject depth into config (for sub-agent depth tracking)
+3. While True:
+ a. Check cancel_check() โ cooperative cancellation for sub-agents
+ b. maybe_compact(state, config) โ compress if near context limit
+ c. Stream from provider โ yield TextChunk / ThinkingChunk
+ d. Record assistant message
+ e. If no tool_calls โ break
+ f. For each tool_call:
+ - Permission check (_check_permission)
+ - If denied โ yield PermissionRequest โ user decides
+ - Execute tool โ yield ToolStart / ToolEnd
+ - Append tool result
+ g. Loop (model sees tool results and responds)
+```
+
+**Event types:**
+
+| Event | Fields | When |
+|---|---|---|
+| `TextChunk` | `text` | Streaming text delta |
+| `ThinkingChunk` | `text` | Extended thinking block |
+| `ToolStart` | `name, inputs` | Before tool execution |
+| `ToolEnd` | `name, result, permitted` | After tool execution |
+| `PermissionRequest` | `description, granted` | Needs user approval |
+| `TurnDone` | `input_tokens, output_tokens` | End of one API turn |
+
+### `compaction.py` โ Context Window Management
+
+Keeps conversations within model context limits using two layers.
+
+**Layer 1: Snip** (`snip_old_tool_results`)
+- Rule-based, no API cost
+- Truncates tool-role messages older than `preserve_last_n_turns` (default 6)
+- Keeps first half + last quarter of the content
+
+**Layer 2: Auto-Compact** (`compact_messages`)
+- Model-driven: calls the current model to summarize old messages
+- Splits messages into [old | recent] at ~70/30 ratio
+- Replaces old messages with a summary + acknowledgment
+
+**Trigger:** `maybe_compact()` checks `estimate_tokens(messages) > context_limit * 0.7`.
+Runs snip first (cheap), then auto-compact if still over.
+
+**Token estimation:** `len(content) / 3.5` โ simple heuristic. Works for most models.
+`get_context_limit(model)` reads from the provider registry.
+
+### `memory.py` โ Persistent Memory
+
+File-based memory system stored in `~/.nano_claude/memory/`.
+
+**Storage format:**
+
+```
+~/.nano_claude/memory/
+โโโ MEMORY.md # Index: one line per memory
+โโโ user_preferences.md # Individual memory file
+โโโ project_auth.md
+```
+
+Each memory file uses markdown with YAML frontmatter:
+
+```markdown
+---
+name: user preferences
+description: coding style preferences
+type: feedback
+created: 2026-04-02
+---
+
+User prefers 4-space indentation and type hints.
+```
+
+**How it integrates:**
+- `get_memory_context()` returns the MEMORY.md index text
+- `context.py` injects this into the system prompt
+- The model reads the index, then uses `Read` tool to access full memory content
+- The model uses `MemorySave` / `MemoryDelete` tools to manage memories
+
+### `subagent.py` โ Threaded Sub-Agents
+
+Sub-agents run in background threads via `ThreadPoolExecutor`.
+
+**Key design decisions:**
+
+1. **Fresh context** โ each sub-agent starts with empty message history + task prompt
+2. **Depth limiting** โ `max_depth=3`, checked at spawn time. Model gets an error message
+ (not silent tool removal) so it can adapt.
+3. **Cooperative cancellation** โ `cancel_check` callable checked each loop iteration.
+ Python threads can't be killed safely, so we set a flag.
+4. **Threading, not asyncio** โ the entire codebase is synchronous generators. Threading
+ via `concurrent.futures` keeps things simple. The SubAgentManager API is designed to
+ be compatible with a future async migration.
+
+**Lifecycle:**
+
+```
+spawn(prompt, config, system_prompt, depth)
+ โ Creates SubAgentTask
+ โ Submits _run to ThreadPoolExecutor
+ โ _run calls agent.run() with depth+1
+
+wait(task_id, timeout) โ blocks until complete
+cancel(task_id) โ sets _cancel_flag
+get_result(task_id) โ returns result string
+```
+
+### `skills.py` โ Reusable Prompt Templates
+
+Skills are markdown files with frontmatter. They are **not code** โ just structured prompts
+that get injected into the agent loop.
+
+**Skill file format:**
+
+```markdown
+---
+name: commit
+description: Create a conventional commit
+triggers: ["/commit"]
+tools: [Bash, Read]
+---
+
+Your prompt instructions here...
+```
+
+**Execution:** `execute_skill()` wraps the skill prompt as a user message and calls
+`agent.run()`. The skill runs through the exact same agent loop as a normal query.
+
+**Search order:** Project-level (`./.nano_claude/skills/`) overrides user-level
+(`~/.nano_claude/skills/`) when skill names collide.
+
+### `providers.py` โ Multi-Provider Abstraction
+
+Two streaming adapters cover all providers:
+
+| Adapter | Providers |
+|---|---|
+| `stream_anthropic()` | Anthropic (native SDK) |
+| `stream_openai_compat()` | OpenAI, Gemini, Kimi, Qwen, Zhipu, DeepSeek, Ollama, LM Studio, Custom |
+
+**Neutral message format** (provider-independent):
+
+```python
+{"role": "user", "content": "..."}
+{"role": "assistant", "content": "...", "tool_calls": [{"id": "...", "name": "...", "input": {...}}]}
+{"role": "tool", "tool_call_id": "...", "name": "...", "content": "..."}
+```
+
+Conversion functions: `messages_to_anthropic()`, `messages_to_openai()`, `tools_to_openai()`.
+
+**Provider-specific handling:**
+- Gemini 3 models require `thought_signature` in tool call responses โ this is transparently
+ captured and passed through via `extra_content` on tool_call dicts.
+
+### `context.py` โ System Prompt Builder
+
+Assembles the system prompt from:
+1. Base template (role, date, cwd, platform)
+2. Git info (branch, status, recent commits)
+3. CLAUDE.md content (project-level + global)
+4. Memory index (from `memory.get_memory_context()`)
+
+### `config.py` โ Configuration
+
+Defaults stored in `~/.nano_claude/config.json`. Key settings:
+
+| Key | Default | Description |
+|---|---|---|
+| `model` | `claude-opus-4-6` | Active model |
+| `max_tokens` | `8192` | Max output tokens |
+| `permission_mode` | `auto` | Permission mode |
+| `max_tool_output` | `32000` | Tool output truncation limit |
+| `max_agent_depth` | `3` | Max sub-agent nesting |
+| `max_concurrent_agents` | `3` | Thread pool size |
+
+---
+
+## Data Flow Example
+
+A user asks "Read config.py and change max_tokens to 16384":
+
+```
+1. nano_claude.py captures input
+2. agent.run() appends user message, calls maybe_compact()
+3. providers.stream() sends to Gemini API with 13 tool schemas
+4. Model responds: text + tool_call[Read(config.py)]
+5. agent.py checks permission (Read = read_only โ auto-approve)
+6. tool_registry.execute_tool("Read", ...) โ file content (truncated if >32K)
+7. Tool result appended to messages, loop back to step 3
+8. Model responds: text + tool_call[Edit(config.py, "8192", "16384")]
+9. agent.py checks permission (Edit = not read_only โ ask user)
+10. User approves โ tools.py._edit() runs, generates diff
+11. nano_claude.py renders diff with ANSI colors (red/green)
+12. Tool result appended, loop back to step 3
+13. Model responds: "Done, max_tokens changed to 16384"
+14. No tool_calls โ loop ends, TurnDone yielded
+```
+
+---
+
+## Testing
+
+```bash
+# Run all 78 tests
+python -m pytest tests/ -v
+
+# Run specific module tests
+python -m pytest tests/test_tool_registry.py -v
+python -m pytest tests/test_compaction.py -v
+python -m pytest tests/test_memory.py -v
+python -m pytest tests/test_subagent.py -v
+python -m pytest tests/test_skills.py -v
+python -m pytest tests/test_diff_view.py -v
+```
+
+Tests use `monkeypatch` and `tmp_path` fixtures to avoid side effects.
+Sub-agent tests mock `_agent_run` to avoid real API calls.
+
+---
+
+## Future: Package Refactoring
+
+When `tools.py` or `agent.py` grow too large, the flat layout can be migrated to:
+
+```
+ncc/
+โโโ __init__.py
+โโโ repl.py # from nano_claude.py
+โโโ agent/
+โ โโโ loop.py # from agent.py
+โ โโโ subagent.py # from subagent.py
+โ โโโ compaction.py # from compaction.py
+โโโ providers/
+โ โโโ base.py
+โ โโโ openai_compat.py
+โ โโโ registry.py
+โโโ tools/
+โ โโโ registry.py # from tool_registry.py
+โ โโโ builtin.py # core 8 tools from tools.py
+โ โโโ memory.py # MemorySave/MemoryDelete from tools.py
+โ โโโ subagent.py # Agent/Check/List from tools.py
+โโโ memory/
+โ โโโ store.py # from memory.py
+โโโ skills/
+โ โโโ loader.py # from skills.py
+โโโ config.py
+```
+
+The current code is structured to make this migration straightforward:
+- Modules communicate via function parameters, not globals
+- Each module has a small public API surface
+- Dependencies are unidirectional
diff --git a/nano-claude-code/docs/logo-v1.png b/nano-claude-code/docs/logo-v1.png
new file mode 100644
index 0000000..627c80e
Binary files /dev/null and b/nano-claude-code/docs/logo-v1.png differ
diff --git a/nano-claude-code/docs/superpowers/plans/2026-04-02-open-cc-enhancement.md b/nano-claude-code/docs/superpowers/plans/2026-04-02-open-cc-enhancement.md
new file mode 100644
index 0000000..5ef0d55
--- /dev/null
+++ b/nano-claude-code/docs/superpowers/plans/2026-04-02-open-cc-enhancement.md
@@ -0,0 +1,2356 @@
+# Open-CC Enhancement Implementation Plan
+
+> **For agentic workers:** REQUIRED SUB-SKILL: Use superpowers:subagent-driven-development (recommended) or superpowers:executing-plans to implement this plan task-by-task. Steps use checkbox (`- [ ]`) syntax for tracking.
+
+**Goal:** Evolve nano-claude-code into a capable AI coding CLI with context management, pluggable tools, sub-agents, memory, skills, and diff view.
+
+**Architecture:** Layered enhancement on existing 6-file structure. New modules (tool_registry, compaction, memory, subagent, skills) added as flat files alongside existing code. Modules communicate via function params and dataclasses, no globals. Agent.py gains depth/cancel_check params for sub-agent support.
+
+**Tech Stack:** Python 3.8-3.10, threading/concurrent.futures, difflib, tiktoken (optional), existing deps (anthropic, openai, httpx, rich)
+
+**Spec:** `docs/superpowers/specs/2026-04-02-open-cc-design.md`
+
+---
+
+## File Structure Overview
+
+**New files:**
+- `tool_registry.py` โ Tool plugin registry (ToolDef dataclass, register/get/execute)
+- `compaction.py` โ Context window management (autoCompact + snip)
+- `memory.py` โ File-based memory system (MEMORY.md index + per-entry markdown files)
+- `subagent.py` โ Sub-agent lifecycle (ThreadPoolExecutor, depth limit, cancel)
+- `skills.py` โ Skill loading and execution (markdown frontmatter, prompt injection)
+- `tests/test_tool_registry.py` โ Tests for tool registry
+- `tests/test_compaction.py` โ Tests for compaction
+- `tests/test_memory.py` โ Tests for memory
+- `tests/test_subagent.py` โ Tests for sub-agent
+- `tests/test_skills.py` โ Tests for skills
+- `tests/test_diff_view.py` โ Tests for diff generation
+
+**Modified files:**
+- `tools.py` โ Refactor to use registry, add diff generation in Edit/Write
+- `agent.py` โ Add compaction call, depth/cancel_check params, sub-agent tool dispatch
+- `context.py` โ Inject memory context into system prompt
+- `config.py` โ Add new config keys (max_tool_output, max_depth, etc.)
+- `nano_claude.py` โ Add /memory, /skill, /agents slash commands, diff rendering
+
+---
+
+## Task 1: Tool Registry (`tool_registry.py`)
+
+**Files:**
+- Create: `tool_registry.py`
+- Create: `tests/test_tool_registry.py`
+
+- [ ] **Step 1: Write failing tests for tool registry**
+
+```python
+# tests/test_tool_registry.py
+import sys
+import os
+sys.path.insert(0, os.path.dirname(os.path.dirname(os.path.abspath(__file__))))
+
+import pytest
+from tool_registry import ToolDef, register_tool, get_tool, get_all_tools, get_tool_schemas, execute_tool, clear_registry
+
+
+@pytest.fixture(autouse=True)
+def clean_registry():
+ clear_registry()
+ yield
+ clear_registry()
+
+
+def _echo_tool(params, config):
+ return f"echo: {params['text']}"
+
+
+ECHO_SCHEMA = {
+ "name": "Echo",
+ "description": "Echo input text",
+ "input_schema": {
+ "type": "object",
+ "properties": {"text": {"type": "string"}},
+ "required": ["text"],
+ },
+}
+
+
+def test_register_and_get():
+ td = ToolDef(name="Echo", schema=ECHO_SCHEMA, func=_echo_tool,
+ read_only=True, concurrent_safe=True)
+ register_tool(td)
+ assert get_tool("Echo") is td
+ assert get_tool("NonExistent") is None
+
+
+def test_get_all_tools():
+ td = ToolDef(name="Echo", schema=ECHO_SCHEMA, func=_echo_tool,
+ read_only=True, concurrent_safe=True)
+ register_tool(td)
+ all_tools = get_all_tools()
+ assert len(all_tools) == 1
+ assert all_tools[0].name == "Echo"
+
+
+def test_get_tool_schemas():
+ td = ToolDef(name="Echo", schema=ECHO_SCHEMA, func=_echo_tool,
+ read_only=True, concurrent_safe=True)
+ register_tool(td)
+ schemas = get_tool_schemas()
+ assert len(schemas) == 1
+ assert schemas[0]["name"] == "Echo"
+
+
+def test_execute_tool():
+ td = ToolDef(name="Echo", schema=ECHO_SCHEMA, func=_echo_tool,
+ read_only=True, concurrent_safe=True)
+ register_tool(td)
+ result = execute_tool("Echo", {"text": "hello"}, {})
+ assert result == "echo: hello"
+
+
+def test_execute_unknown_tool():
+ result = execute_tool("Unknown", {}, {})
+ assert "Unknown tool" in result
+
+
+def test_output_truncation():
+ def big_output(params, config):
+ return "x" * 100_000
+
+ td = ToolDef(name="Big", schema=ECHO_SCHEMA, func=big_output,
+ read_only=True, concurrent_safe=True)
+ register_tool(td)
+ result = execute_tool("Big", {}, {}, max_output=1000)
+ assert len(result) < 2000
+ assert "truncated" in result
+
+
+def test_duplicate_register_overwrites():
+ td1 = ToolDef(name="Echo", schema=ECHO_SCHEMA, func=_echo_tool,
+ read_only=True, concurrent_safe=True)
+ td2 = ToolDef(name="Echo", schema=ECHO_SCHEMA, func=lambda p, c: "v2",
+ read_only=False, concurrent_safe=False)
+ register_tool(td1)
+ register_tool(td2)
+ assert get_tool("Echo").read_only is False
+```
+
+- [ ] **Step 2: Run tests to verify they fail**
+
+Run: `cd D:/git/open-cc/nano-claude-code && python -m pytest tests/test_tool_registry.py -v`
+Expected: FAIL โ `ModuleNotFoundError: No module named 'tool_registry'`
+
+- [ ] **Step 3: Implement tool_registry.py**
+
+```python
+# tool_registry.py
+"""
+Tool plugin registry for nano-claude-code.
+
+Provides a central registry for tool definitions. Tools register themselves
+at import time. The agent loop queries the registry for schemas and dispatches
+execution through it.
+
+Public API:
+ register_tool(tool_def) โ add a tool to the registry
+ get_tool(name) โ look up by name
+ get_all_tools() โ list all registered tools
+ get_tool_schemas() โ schemas for API calls
+ execute_tool(name, ...) โ dispatch execution with output truncation
+ clear_registry() โ reset (for testing)
+"""
+from __future__ import annotations
+
+from dataclasses import dataclass
+from typing import Callable, Dict, List, Optional
+
+
+DEFAULT_MAX_OUTPUT = 32_000 # ~8K tokens
+
+
+@dataclass
+class ToolDef:
+ """Definition of a registered tool."""
+ name: str
+ schema: dict
+ func: Callable[[dict, dict], str]
+ read_only: bool = False
+ concurrent_safe: bool = False
+
+
+_TOOLS: Dict[str, ToolDef] = {}
+
+
+def register_tool(tool_def: ToolDef) -> None:
+ """Register a tool. Overwrites if name already exists."""
+ _TOOLS[tool_def.name] = tool_def
+
+
+def get_tool(name: str) -> Optional[ToolDef]:
+ """Look up a tool by name. Returns None if not found."""
+ return _TOOLS.get(name)
+
+
+def get_all_tools() -> List[ToolDef]:
+ """Return all registered tools in registration order."""
+ return list(_TOOLS.values())
+
+
+def get_tool_schemas() -> List[dict]:
+ """Return JSON schemas for all tools (for API calls)."""
+ return [t.schema for t in _TOOLS.values()]
+
+
+def execute_tool(name: str, params: dict, config: dict,
+ max_output: int = DEFAULT_MAX_OUTPUT) -> str:
+ """
+ Execute a tool by name with output truncation.
+
+ Args:
+ name: tool name, str
+ params: tool input parameters, dict
+ config: global config, dict
+ max_output: max chars before truncation, int
+ Returns:
+ tool result string, possibly truncated
+ """
+ tool = _TOOLS.get(name)
+ if tool is None:
+ return f"Error: Unknown tool '{name}'"
+
+ try:
+ result = tool.func(params, config)
+ except Exception as e:
+ return f"Error executing {name}: {e}"
+
+ # Hard truncation at source
+ if len(result) > max_output:
+ head_size = max_output // 2
+ tail_size = max_output // 4
+ snipped = len(result) - head_size - tail_size
+ result = (
+ f"{result[:head_size]}\n\n"
+ f"[... {snipped} chars truncated ...]\n\n"
+ f"{result[-tail_size:]}"
+ )
+
+ return result
+
+
+def clear_registry() -> None:
+ """Clear all registered tools. For testing only."""
+ _TOOLS.clear()
+```
+
+- [ ] **Step 4: Run tests to verify they pass**
+
+Run: `cd D:/git/open-cc/nano-claude-code && python -m pytest tests/test_tool_registry.py -v`
+Expected: All 7 tests PASS
+
+- [ ] **Step 5: Commit**
+
+```bash
+cd D:/git/open-cc/nano-claude-code
+git add tool_registry.py tests/test_tool_registry.py
+git commit -m "feat: add tool plugin registry with output truncation"
+```
+
+---
+
+## Task 2: Migrate Built-in Tools to Registry (`tools.py` refactor)
+
+**Files:**
+- Modify: `tools.py` (lines 1-360)
+- Modify: `agent.py` (lines 5-6 imports, lines 52-130 run function)
+
+- [ ] **Step 1: Refactor tools.py to register via tool_registry**
+
+The key changes to `tools.py`:
+1. Import `tool_registry` and call `register_tool()` for each built-in tool
+2. Keep `TOOL_SCHEMAS` and `execute_tool` as thin wrappers for backward compat
+3. Keep all `_read`, `_write`, etc. implementations unchanged
+
+Add at the top of `tools.py` after existing imports:
+
+```python
+from tool_registry import ToolDef, register_tool
+from tool_registry import execute_tool as _registry_execute
+from tool_registry import get_tool_schemas as _registry_schemas
+```
+
+At the bottom of `tools.py` (after all function definitions, replacing the existing `execute_tool` function), add the registration block and compatibility layer:
+
+```python
+# --- Tool Registration ---------------------------------------------------
+
+def _register_builtin_tools():
+ """Register all 8 built-in tools with the registry."""
+ _tool_defs = [
+ ToolDef(name="Read", schema=TOOL_SCHEMAS[0], func=lambda p, c: _read(**p),
+ read_only=True, concurrent_safe=True),
+ ToolDef(name="Write", schema=TOOL_SCHEMAS[1], func=lambda p, c: _write(**p),
+ read_only=False, concurrent_safe=False),
+ ToolDef(name="Edit", schema=TOOL_SCHEMAS[2], func=lambda p, c: _edit(**p),
+ read_only=False, concurrent_safe=False),
+ ToolDef(name="Bash", schema=TOOL_SCHEMAS[3], func=lambda p, c: _bash(**p),
+ read_only=False, concurrent_safe=False),
+ ToolDef(name="Glob", schema=TOOL_SCHEMAS[4], func=lambda p, c: _glob(**p),
+ read_only=True, concurrent_safe=True),
+ ToolDef(name="Grep", schema=TOOL_SCHEMAS[5], func=lambda p, c: _grep(**p),
+ read_only=True, concurrent_safe=True),
+ ToolDef(name="WebFetch", schema=TOOL_SCHEMAS[6], func=lambda p, c: _webfetch(**p),
+ read_only=True, concurrent_safe=True),
+ ToolDef(name="WebSearch", schema=TOOL_SCHEMAS[7], func=lambda p, c: _websearch(**p),
+ read_only=True, concurrent_safe=True),
+ ]
+ for td in _tool_defs:
+ register_tool(td)
+
+
+_register_builtin_tools()
+
+
+# --- Backward Compatibility -----------------------------------------------
+
+def execute_tool(name, inputs, permission_mode="auto", ask_permission=None):
+ """
+ Backward-compatible wrapper. Delegates to tool_registry.execute_tool
+ but preserves the permission_mode/ask_permission interface used by agent.py.
+ """
+ # Permission check (same logic as before)
+ from tool_registry import get_tool
+ tool = get_tool(name)
+ if tool is None:
+ return f"Error: unknown tool '{name}'"
+
+ needs_permission = False
+ if permission_mode == "manual":
+ needs_permission = True
+ elif permission_mode == "auto":
+ if not tool.read_only:
+ if name == "Bash" and _is_safe_bash(inputs.get("command", "")):
+ needs_permission = False
+ else:
+ needs_permission = True
+
+ if needs_permission and ask_permission:
+ desc = _permission_desc_for(name, inputs)
+ if not ask_permission(desc):
+ return "[Tool call denied by user]"
+
+ return _registry_execute(name, inputs, {})
+
+
+def _permission_desc_for(name, inputs):
+ """Build a human-readable permission description."""
+ if name == "Bash":
+ return f"Run: {inputs.get('command', '')}"
+ elif name == "Write":
+ return f"Write to: {inputs.get('file_path', '')}"
+ elif name == "Edit":
+ return f"Edit: {inputs.get('file_path', '')}"
+ return f"{name}: {inputs}"
+```
+
+Remove the old `execute_tool` function (lines 304-360 in original tools.py) and replace with the above.
+
+- [ ] **Step 2: Update agent.py imports**
+
+Change `agent.py` lines 5-6 from:
+
+```python
+from tools import TOOL_SCHEMAS, execute_tool
+```
+
+to:
+
+```python
+from tool_registry import get_tool_schemas, get_tool
+from tools import execute_tool # backward-compat wrapper with permissions
+import tools as _tools_init # ensure built-in tools are registered
+```
+
+Update the `run()` function where it references `TOOL_SCHEMAS` (around line 65) to use `get_tool_schemas()` instead.
+
+- [ ] **Step 3: Run the existing code to verify nothing is broken**
+
+Run: `cd D:/git/open-cc/nano-claude-code && python -c "from tools import execute_tool; from tool_registry import get_all_tools; print(f'{len(get_all_tools())} tools registered'); print(execute_tool('Read', {'file_path': 'config.py'}, 'auto'))"`
+Expected: `8 tools registered` + contents of config.py
+
+- [ ] **Step 4: Commit**
+
+```bash
+cd D:/git/open-cc/nano-claude-code
+git add tools.py agent.py
+git commit -m "refactor: migrate built-in tools to plugin registry"
+```
+
+---
+
+## Task 3: Diff View (`tools.py` + `nano_claude.py`)
+
+**Files:**
+- Modify: `tools.py` (the `_edit` and `_write` functions)
+- Modify: `nano_claude.py` (the `print_tool_end` function)
+- Create: `tests/test_diff_view.py`
+
+- [ ] **Step 1: Write failing tests for diff generation**
+
+```python
+# tests/test_diff_view.py
+import sys, os, tempfile
+sys.path.insert(0, os.path.dirname(os.path.dirname(os.path.abspath(__file__))))
+
+import pytest
+
+
+def test_generate_unified_diff():
+ from tools import generate_unified_diff
+ old = "line1\nline2\nline3\n"
+ new = "line1\nline2_modified\nline3\n"
+ diff = generate_unified_diff(old, new, "test.py")
+ assert "--- a/test.py" in diff
+ assert "+++ b/test.py" in diff
+ assert "-line2" in diff
+ assert "+line2_modified" in diff
+
+
+def test_generate_unified_diff_empty_old():
+ from tools import generate_unified_diff
+ diff = generate_unified_diff("", "new content\n", "test.py")
+ assert "+new content" in diff
+
+
+def test_edit_returns_diff(tmp_path):
+ from tools import _edit
+ f = tmp_path / "test.txt"
+ f.write_text("hello world\n")
+ result = _edit(str(f), "hello", "goodbye")
+ assert "-hello world" in result
+ assert "+goodbye world" in result
+
+
+def test_write_existing_returns_diff(tmp_path):
+ from tools import _write
+ f = tmp_path / "test.txt"
+ f.write_text("old content\n")
+ result = _write(str(f), "new content\n")
+ assert "-old content" in result
+ assert "+new content" in result
+
+
+def test_write_new_file_no_diff(tmp_path):
+ from tools import _write
+ f = tmp_path / "new.txt"
+ result = _write(str(f), "content\n")
+ assert "Created" in result
+ assert "---" not in result # no diff for new files
+
+
+def test_diff_truncation():
+ from tools import generate_unified_diff, maybe_truncate_diff
+ old = "\n".join(f"line{i}" for i in range(200))
+ new = "\n".join(f"CHANGED{i}" for i in range(200))
+ diff = generate_unified_diff(old, new, "big.py")
+ truncated = maybe_truncate_diff(diff, max_lines=50)
+ assert "more lines" in truncated
+ assert truncated.count("\n") < 60
+```
+
+- [ ] **Step 2: Run tests to verify they fail**
+
+Run: `cd D:/git/open-cc/nano-claude-code && python -m pytest tests/test_diff_view.py -v`
+Expected: FAIL โ `ImportError: cannot import name 'generate_unified_diff'`
+
+- [ ] **Step 3: Add diff generation functions to tools.py**
+
+Add near the top of `tools.py`, after the existing imports:
+
+```python
+import difflib
+
+
+def generate_unified_diff(old, new, filename, context_lines=3):
+ """
+ Generate a unified diff between two strings.
+
+ Args:
+ old: original content, str
+ new: modified content, str
+ filename: display filename, str
+ context_lines: lines of surrounding context, int
+ Returns:
+ unified diff string, str
+ """
+ old_lines = old.splitlines(keepends=True)
+ new_lines = new.splitlines(keepends=True)
+ diff = difflib.unified_diff(
+ old_lines, new_lines,
+ fromfile=f"a/{filename}",
+ tofile=f"b/{filename}",
+ n=context_lines,
+ )
+ return "".join(diff)
+
+
+def maybe_truncate_diff(diff_text, max_lines=80):
+ """
+ Truncate diff for display if it exceeds max_lines.
+
+ Args:
+ diff_text: unified diff string, str
+ max_lines: max lines to show, int
+ Returns:
+ possibly truncated diff string, str
+ """
+ lines = diff_text.splitlines()
+ if len(lines) <= max_lines:
+ return diff_text
+ shown = lines[:max_lines]
+ remaining = len(lines) - max_lines
+ return "\n".join(shown) + f"\n\n[... {remaining} more lines ...]"
+```
+
+- [ ] **Step 4: Modify _edit to return diff**
+
+Replace the `_edit` function (originally lines 175-192) with:
+
+```python
+def _edit(file_path, old_string, new_string, replace_all=False):
+ p = Path(file_path)
+ if not p.exists():
+ return f"Error: {file_path} not found"
+ content = p.read_text(encoding="utf-8")
+ if old_string not in content:
+ return f"Error: old_string not found in {file_path}"
+ old_content = content
+ if replace_all:
+ new_content = content.replace(old_string, new_string)
+ else:
+ count = content.count(old_string)
+ if count > 1:
+ return f"Error: old_string appears {count} times; use replace_all=true or provide more context"
+ new_content = content.replace(old_string, new_string, 1)
+ p.write_text(new_content, encoding="utf-8")
+ filename = p.name
+ diff = generate_unified_diff(old_content, new_content, filename)
+ return f"Changes applied to {filename}:\n\n{diff}" if diff else f"No changes in {filename}"
+```
+
+- [ ] **Step 5: Modify _write to return diff for existing files**
+
+Replace the `_write` function (originally lines 164-172) with:
+
+```python
+def _write(file_path, content):
+ p = Path(file_path)
+ p.parent.mkdir(parents=True, exist_ok=True)
+ if p.exists():
+ old_content = p.read_text(encoding="utf-8")
+ p.write_text(content, encoding="utf-8")
+ filename = p.name
+ diff = generate_unified_diff(old_content, content, filename)
+ truncated = maybe_truncate_diff(diff)
+ return f"File updated โ {filename}:\n\n{truncated}" if diff else f"No changes in {filename}"
+ else:
+ p.write_text(content, encoding="utf-8")
+ line_count = content.count("\n") + (1 if content and not content.endswith("\n") else 0)
+ return f"Created {file_path} ({line_count} lines)"
+```
+
+- [ ] **Step 6: Add diff rendering to nano_claude.py**
+
+Add a new function after the `flush_response` function (around line 105):
+
+```python
+def render_diff(text):
+ """Render unified diff with ANSI colors."""
+ for line in text.splitlines():
+ if line.startswith("+++") or line.startswith("---"):
+ print(clr(line, "bold"))
+ elif line.startswith("+"):
+ print(clr(line, "green"))
+ elif line.startswith("-"):
+ print(clr(line, "red"))
+ elif line.startswith("@@"):
+ print(clr(line, "cyan"))
+ else:
+ print(line)
+
+
+def _has_diff(text):
+ """Check if text contains a unified diff."""
+ return ("--- a/" in text and "+++ b/" in text) or text.startswith("@@")
+```
+
+Modify the `print_tool_end` function (originally line 115) to detect and render diffs:
+
+```python
+def print_tool_end(name, result, verbose):
+ if name in ("Edit", "Write") and _has_diff(result):
+ # Extract the diff portion and render with colors
+ parts = result.split("\n\n", 1)
+ if len(parts) == 2:
+ info(parts[0]) # "Changes applied to file.py:"
+ render_diff(parts[1]) # colored diff
+ else:
+ render_diff(result)
+ elif verbose:
+ info(f" {C.get('dim','')}Result: {result[:200]}{'...' if len(result)>200 else ''}{C.get('reset','')}")
+```
+
+Also ensure the color dict `C` (line 60-70 of nano_claude.py) includes the needed keys. Check and add if missing:
+
+```python
+# In the C dict, ensure these entries exist:
+"bold": "\033[1m",
+"cyan": "\033[36m",
+```
+
+- [ ] **Step 7: Run tests to verify they pass**
+
+Run: `cd D:/git/open-cc/nano-claude-code && python -m pytest tests/test_diff_view.py -v`
+Expected: All 6 tests PASS
+
+- [ ] **Step 8: Manual smoke test**
+
+Run: `cd D:/git/open-cc/nano-claude-code && python -c "
+from tools import _write, _edit
+import tempfile, os
+f = os.path.join(tempfile.mkdtemp(), 'test.py')
+print(_write(f, 'def hello():\n return 42\n'))
+print('---')
+print(_edit(f, 'return 42', 'return 99'))
+"`
+Expected: First call shows "Created", second call shows colored diff with red `-return 42` and green `+return 99`
+
+- [ ] **Step 9: Commit**
+
+```bash
+cd D:/git/open-cc/nano-claude-code
+git add tools.py nano_claude.py tests/test_diff_view.py
+git commit -m "feat: add git-style diff view for Edit and Write tools"
+```
+
+---
+
+## Task 4: Context Window Management (`compaction.py`)
+
+**Files:**
+- Create: `compaction.py`
+- Create: `tests/test_compaction.py`
+- Modify: `agent.py` (add compaction call in run loop)
+- Modify: `providers.py` (add context_limit to PROVIDERS entries)
+
+- [ ] **Step 1: Write failing tests for compaction**
+
+```python
+# tests/test_compaction.py
+import sys, os
+sys.path.insert(0, os.path.dirname(os.path.dirname(os.path.abspath(__file__))))
+
+import pytest
+from compaction import (
+ estimate_tokens,
+ get_context_limit,
+ snip_old_tool_results,
+ find_split_point,
+)
+
+
+def test_estimate_tokens_simple():
+ msgs = [{"role": "user", "content": "hello world"}]
+ tokens = estimate_tokens(msgs)
+ assert 2 < tokens < 10 # "hello world" ~ 2-3 tokens
+
+
+def test_estimate_tokens_empty():
+ assert estimate_tokens([]) == 0
+
+
+def test_estimate_tokens_with_tool_results():
+ msgs = [
+ {"role": "user", "content": "read file"},
+ {"role": "tool", "content": "x" * 10000, "name": "Read"},
+ ]
+ tokens = estimate_tokens(msgs)
+ assert tokens > 2000 # 10000 / ~3.5
+
+
+def test_get_context_limit_known_model():
+ limit = get_context_limit("gpt-4o")
+ assert limit >= 100_000
+
+
+def test_get_context_limit_unknown_model():
+ limit = get_context_limit("unknown-model-xyz")
+ assert limit == 128_000 # default fallback
+
+
+def test_snip_old_tool_results():
+ msgs = [
+ {"role": "user", "content": "read file"},
+ {"role": "assistant", "content": "ok", "tool_calls": [{"id": "1", "name": "Read", "input": {}}]},
+ {"role": "tool", "content": "x" * 5000, "tool_call_id": "1", "name": "Read"},
+ {"role": "assistant", "content": "got it"},
+ {"role": "user", "content": "now do something else"},
+ {"role": "assistant", "content": "sure"},
+ ]
+ snipped = snip_old_tool_results(msgs, max_chars=100, preserve_last_n_turns=2)
+ tool_msg = snipped[2]
+ assert len(tool_msg["content"]) < 300
+ assert "snipped" in tool_msg["content"].lower()
+
+
+def test_snip_preserves_recent():
+ msgs = [
+ {"role": "user", "content": "recent"},
+ {"role": "tool", "content": "x" * 5000, "tool_call_id": "1", "name": "Read"},
+ ]
+ snipped = snip_old_tool_results(msgs, max_chars=100, preserve_last_n_turns=4)
+ # Everything is recent, nothing should be snipped
+ assert snipped[1]["content"] == "x" * 5000
+
+
+def test_find_split_point():
+ msgs = [{"role": "user", "content": "x" * 1000}] * 10
+ split = find_split_point(msgs, keep_ratio=0.3)
+ assert 5 <= split <= 8 # keep ~30% means split at ~70%
+```
+
+- [ ] **Step 2: Run tests to verify they fail**
+
+Run: `cd D:/git/open-cc/nano-claude-code && python -m pytest tests/test_compaction.py -v`
+Expected: FAIL โ `ModuleNotFoundError: No module named 'compaction'`
+
+- [ ] **Step 3: Add context_limit to providers.py**
+
+In `providers.py`, add a `context_limit` field to each entry in the `PROVIDERS` dict. Add after the `"models"` field in each provider entry:
+
+```python
+# Add to each provider in PROVIDERS dict:
+# anthropic:
+"context_limit": 200_000,
+# openai:
+"context_limit": 128_000,
+# gemini:
+"context_limit": 1_000_000,
+# kimi:
+"context_limit": 128_000,
+# qwen:
+"context_limit": 1_000_000,
+# zhipu:
+"context_limit": 128_000,
+# deepseek:
+"context_limit": 64_000,
+# ollama:
+"context_limit": 128_000,
+# lmstudio:
+"context_limit": 128_000,
+# custom:
+"context_limit": 128_000,
+```
+
+- [ ] **Step 4: Implement compaction.py**
+
+```python
+# compaction.py
+"""
+Context window management for nano-claude-code.
+
+Two-layer compression strategy:
+ Layer 1 (snip): Truncate old tool results โ rule-based, no model call
+ Layer 2 (autoCompact): Summarize old messages โ requires model call
+
+Public API:
+ maybe_compact(state, config) โ run both layers if needed
+ estimate_tokens(messages) โ fast token estimation
+ get_context_limit(model) โ context window size for model
+ snip_old_tool_results(...) โ truncate old tool outputs
+ find_split_point(...) โ find where to split old/recent
+"""
+from __future__ import annotations
+
+from typing import List, Optional
+
+from providers import detect_provider, PROVIDERS, stream
+
+
+# ---------------------------------------------------------------------------
+# Token estimation
+# ---------------------------------------------------------------------------
+
+def estimate_tokens(messages: List[dict]) -> int:
+ """
+ Estimate token count for a message list.
+
+ Uses chars/3.5 as a rough heuristic. tiktoken can be used for GPT
+ models if installed, but is not required.
+
+ Args:
+ messages: list of message dicts with "content" field
+ Returns:
+ estimated token count, int
+ """
+ total_chars = 0
+ for msg in messages:
+ content = msg.get("content", "")
+ if isinstance(content, str):
+ total_chars += len(content)
+ elif isinstance(content, list):
+ for block in content:
+ if isinstance(block, dict):
+ total_chars += len(block.get("text", ""))
+ elif isinstance(block, str):
+ total_chars += len(block)
+ return int(total_chars / 3.5)
+
+
+def get_context_limit(model: str) -> int:
+ """
+ Return context window size for a given model.
+
+ Args:
+ model: model name string
+ Returns:
+ context limit in tokens, int
+ """
+ provider_name = detect_provider(model)
+ provider = PROVIDERS.get(provider_name, {})
+ return provider.get("context_limit", 128_000)
+
+
+# ---------------------------------------------------------------------------
+# Layer 1: Tool-result snipping (rule-based, no model call)
+# ---------------------------------------------------------------------------
+
+def snip_old_tool_results(
+ messages: List[dict],
+ max_chars: int = 2000,
+ preserve_last_n_turns: int = 6,
+) -> List[dict]:
+ """
+ Truncate old tool result contents to max_chars.
+
+ Preserves the most recent preserve_last_n_turns messages untouched.
+ For older tool results, keeps first and last lines with a snip marker.
+
+ Args:
+ messages: message list (mutated in place and returned)
+ max_chars: max chars per tool result, int
+ preserve_last_n_turns: how many recent messages to leave alone, int
+ Returns:
+ the same list, with old tool results truncated
+ """
+ cutoff = max(0, len(messages) - preserve_last_n_turns)
+
+ for i in range(cutoff):
+ msg = messages[i]
+ if msg.get("role") != "tool":
+ continue
+ content = msg.get("content", "")
+ if len(content) <= max_chars:
+ continue
+
+ # Keep first half + last quarter of allowed chars
+ head = content[: max_chars // 2]
+ tail = content[-(max_chars // 4) :]
+ snipped = len(content) - len(head) - len(tail)
+ messages[i] = {
+ **msg,
+ "content": f"{head}\n\n[... {snipped} chars snipped ...]\n\n{tail}",
+ }
+
+ return messages
+
+
+# ---------------------------------------------------------------------------
+# Layer 2: Auto-compact (model-driven summary)
+# ---------------------------------------------------------------------------
+
+def find_split_point(messages: List[dict], keep_ratio: float = 0.3) -> int:
+ """
+ Find the index that splits messages into [old | recent],
+ keeping approximately keep_ratio of total tokens in 'recent'.
+
+ Args:
+ messages: message list
+ keep_ratio: fraction of tokens to keep as recent, float
+ Returns:
+ split index, int
+ """
+ total = estimate_tokens(messages)
+ target_recent = int(total * keep_ratio)
+
+ cumulative = 0
+ for i in range(len(messages) - 1, -1, -1):
+ content = messages[i].get("content", "")
+ if isinstance(content, str):
+ cumulative += int(len(content) / 3.5)
+ if cumulative >= target_recent:
+ return i
+ return 0
+
+
+_SUMMARY_PROMPT = """Summarize the conversation so far. Preserve:
+- Key decisions and their reasoning
+- File paths that were read or modified
+- Code changes made and their purpose
+- Unfinished tasks or pending questions
+- Important error messages or debugging findings
+
+Be concise but complete. Use bullet points."""
+
+
+def compact_messages(messages: List[dict], config: dict) -> List[dict]:
+ """
+ Summarize old messages via model call, keep recent messages verbatim.
+
+ Args:
+ messages: full message list
+ config: global config dict (needs "model" key)
+ Returns:
+ new message list: [summary_msg, ack_msg, *recent]
+ """
+ split = find_split_point(messages, keep_ratio=0.3)
+ if split <= 1:
+ return messages # nothing worth compacting
+
+ old = messages[:split]
+ recent = messages[split:]
+
+ # Build a summary request
+ summary_messages = [
+ {"role": "user", "content": _SUMMARY_PROMPT},
+ ]
+ # Prepend old messages as context
+ for msg in old:
+ role = msg.get("role", "user")
+ content = msg.get("content", "")
+ if role == "tool":
+ summary_messages.append({
+ "role": "user",
+ "content": f"[Tool {msg.get('name', '?')} result]: {content[:1000]}",
+ })
+ else:
+ summary_messages.append({"role": role, "content": str(content)[:2000]})
+ summary_messages.append({"role": "user", "content": _SUMMARY_PROMPT})
+
+ # Call model for summary (non-streaming, collect full response)
+ model = config.get("model", "gpt-4o")
+ summary_text = ""
+ for event in stream(model, "You are a helpful summarizer.", summary_messages, [], config):
+ if hasattr(event, "text") and not hasattr(event, "tool_calls"):
+ summary_text += event.text
+
+ return [
+ {"role": "user", "content": f"[Conversation summary]\n{summary_text}"},
+ {"role": "assistant", "content": "Understood, I have the context from the summary."},
+ *recent,
+ ]
+
+
+# ---------------------------------------------------------------------------
+# Main entry point
+# ---------------------------------------------------------------------------
+
+def maybe_compact(state, config: dict) -> bool:
+ """
+ Run compaction if needed. Called before each API call in agent loop.
+
+ Layer 1 (snip) runs first โ cheap and fast.
+ Layer 2 (autoCompact) runs if still over threshold.
+
+ Args:
+ state: AgentState with .messages
+ config: global config dict
+ Returns:
+ True if compaction occurred, bool
+ """
+ model = config.get("model", "gpt-4o")
+ limit = get_context_limit(model)
+ threshold = int(limit * 0.7)
+
+ current = estimate_tokens(state.messages)
+ if current <= threshold:
+ return False
+
+ # Layer 1: snip old tool results
+ snip_old_tool_results(state.messages)
+ current = estimate_tokens(state.messages)
+ if current <= threshold:
+ return True
+
+ # Layer 2: auto-compact via model summary
+ state.messages = compact_messages(state.messages, config)
+ return True
+```
+
+- [ ] **Step 5: Run tests to verify they pass**
+
+Run: `cd D:/git/open-cc/nano-claude-code && python -m pytest tests/test_compaction.py -v`
+Expected: All 8 tests PASS
+
+- [ ] **Step 6: Integrate compaction into agent.py**
+
+In `agent.py`, add import at the top:
+
+```python
+from compaction import maybe_compact
+```
+
+In the `run()` function, add compaction check before the API call. Inside the `while True` loop, before the `for event in stream(...)` line, add:
+
+```python
+ # --- Context compaction ---
+ maybe_compact(state, config)
+```
+
+- [ ] **Step 7: Commit**
+
+```bash
+cd D:/git/open-cc/nano-claude-code
+git add compaction.py tests/test_compaction.py agent.py providers.py
+git commit -m "feat: add context window management with snip + auto-compact"
+```
+
+---
+
+## Task 5: Memory System (`memory.py`)
+
+**Files:**
+- Create: `memory.py`
+- Create: `tests/test_memory.py`
+- Modify: `context.py` (inject memory into system prompt)
+- Modify: `tools.py` (register MemorySave + MemoryDelete tools)
+
+- [ ] **Step 1: Write failing tests for memory**
+
+```python
+# tests/test_memory.py
+import sys, os, tempfile, shutil
+sys.path.insert(0, os.path.dirname(os.path.dirname(os.path.abspath(__file__))))
+
+import pytest
+from memory import MemoryEntry, save_memory, load_index, delete_memory, search_memory, get_memory_context
+
+
+@pytest.fixture
+def memory_dir(tmp_path, monkeypatch):
+ """Use a temp directory for memory storage."""
+ import memory
+ monkeypatch.setattr(memory, "MEMORY_DIR", tmp_path)
+ monkeypatch.setattr(memory, "INDEX_FILE", tmp_path / "MEMORY.md")
+ return tmp_path
+
+
+def test_save_and_load(memory_dir):
+ entry = MemoryEntry(
+ name="test entry",
+ description="a test memory",
+ type="user",
+ content="User prefers Python.",
+ )
+ save_memory(entry)
+
+ index = load_index()
+ assert len(index) == 1
+ assert index[0].name == "test entry"
+ assert index[0].content == "User prefers Python."
+
+
+def test_save_creates_file(memory_dir):
+ entry = MemoryEntry(name="file check", description="check file", type="feedback",
+ content="Always use type hints.")
+ save_memory(entry)
+ files = list(memory_dir.glob("*.md"))
+ assert len(files) == 2 # MEMORY.md + the entry file
+
+
+def test_delete_memory(memory_dir):
+ entry = MemoryEntry(name="to delete", description="temp", type="project",
+ content="Temporary.")
+ save_memory(entry)
+ assert len(load_index()) == 1
+ delete_memory("to delete")
+ assert len(load_index()) == 0
+
+
+def test_delete_nonexistent(memory_dir):
+ # Should not raise
+ delete_memory("nonexistent")
+
+
+def test_search_memory(memory_dir):
+ save_memory(MemoryEntry(name="python pref", description="user likes python",
+ type="user", content="Python is preferred."))
+ save_memory(MemoryEntry(name="go pref", description="user knows go",
+ type="user", content="User knows Go."))
+ results = search_memory("python")
+ assert len(results) == 1
+ assert results[0].name == "python pref"
+
+
+def test_get_memory_context(memory_dir):
+ save_memory(MemoryEntry(name="ctx test", description="context entry",
+ type="feedback", content="Be concise."))
+ ctx = get_memory_context()
+ assert "ctx test" in ctx
+ assert "context entry" in ctx
+
+
+def test_get_memory_context_empty(memory_dir):
+ ctx = get_memory_context()
+ assert ctx == ""
+
+
+def test_update_existing(memory_dir):
+ save_memory(MemoryEntry(name="evolving", description="v1", type="user", content="Version 1."))
+ save_memory(MemoryEntry(name="evolving", description="v2", type="user", content="Version 2."))
+ index = load_index()
+ assert len(index) == 1
+ assert index[0].description == "v2"
+ assert index[0].content == "Version 2."
+```
+
+- [ ] **Step 2: Run tests to verify they fail**
+
+Run: `cd D:/git/open-cc/nano-claude-code && python -m pytest tests/test_memory.py -v`
+Expected: FAIL โ `ModuleNotFoundError: No module named 'memory'`
+
+- [ ] **Step 3: Implement memory.py**
+
+```python
+# memory.py
+"""
+File-based memory system for nano-claude-code.
+
+Stores memories as individual markdown files with YAML frontmatter.
+MEMORY.md serves as an index. Memories persist across conversations.
+
+Public API:
+ save_memory(entry) โ write/update a memory file + index
+ load_index() โ parse all memory files
+ delete_memory(name) โ remove a memory file + index entry
+ search_memory(query) โ keyword search across memories
+ get_memory_context() โ return index text for system prompt
+"""
+from __future__ import annotations
+
+import re
+from dataclasses import dataclass, field
+from datetime import datetime
+from pathlib import Path
+from typing import List
+
+
+MEMORY_DIR = Path.home() / ".nano_claude" / "memory"
+INDEX_FILE = MEMORY_DIR / "MEMORY.md"
+
+
+@dataclass
+class MemoryEntry:
+ """A single memory record."""
+ name: str
+ description: str
+ type: str # "user" | "feedback" | "project" | "reference"
+ content: str
+ file_path: str = ""
+ created: str = ""
+
+
+def _slugify(name: str) -> str:
+ """Convert a name to a filesystem-safe slug."""
+ slug = re.sub(r"[^\w\s-]", "", name.lower())
+ slug = re.sub(r"[\s]+", "_", slug).strip("_")
+ return slug[:60] or "memory"
+
+
+def _parse_frontmatter(text: str) -> tuple:
+ """
+ Parse YAML-like frontmatter from markdown text.
+
+ Returns:
+ (metadata_dict, body_text)
+ """
+ if not text.startswith("---"):
+ return {}, text
+ parts = text.split("---", 2)
+ if len(parts) < 3:
+ return {}, text
+ meta = {}
+ for line in parts[1].strip().splitlines():
+ if ":" in line:
+ key, val = line.split(":", 1)
+ meta[key.strip()] = val.strip()
+ return meta, parts[2].strip()
+
+
+def _write_entry_file(entry: MemoryEntry) -> Path:
+ """Write a memory entry to its markdown file."""
+ MEMORY_DIR.mkdir(parents=True, exist_ok=True)
+
+ slug = _slugify(entry.name)
+ file_path = MEMORY_DIR / f"{slug}.md"
+
+ created = entry.created or datetime.now().strftime("%Y-%m-%d")
+ content = (
+ f"---\n"
+ f"name: {entry.name}\n"
+ f"description: {entry.description}\n"
+ f"type: {entry.type}\n"
+ f"created: {created}\n"
+ f"---\n\n"
+ f"{entry.content}\n"
+ )
+ file_path.write_text(content, encoding="utf-8")
+ return file_path
+
+
+def _update_index(entries: List[MemoryEntry]) -> None:
+ """Rewrite MEMORY.md index from a list of entries."""
+ MEMORY_DIR.mkdir(parents=True, exist_ok=True)
+ lines = []
+ for e in entries:
+ slug = _slugify(e.name)
+ lines.append(f"- [{e.name}]({slug}.md) โ {e.description}")
+ INDEX_FILE.write_text("\n".join(lines) + "\n" if lines else "", encoding="utf-8")
+
+
+def save_memory(entry: MemoryEntry) -> None:
+ """
+ Save or update a memory entry.
+
+ If a memory with the same name exists, it is overwritten.
+ Updates both the entry file and the MEMORY.md index.
+ """
+ # Load existing entries, filter out same-name
+ existing = [e for e in load_index() if e.name != entry.name]
+ existing.append(entry)
+
+ _write_entry_file(entry)
+ _update_index(existing)
+
+
+def load_index() -> List[MemoryEntry]:
+ """
+ Load all memory entries by scanning markdown files in MEMORY_DIR.
+
+ Returns:
+ list of MemoryEntry, sorted by name
+ """
+ if not MEMORY_DIR.exists():
+ return []
+
+ entries = []
+ for f in sorted(MEMORY_DIR.glob("*.md")):
+ if f.name == "MEMORY.md":
+ continue
+ try:
+ text = f.read_text(encoding="utf-8")
+ meta, body = _parse_frontmatter(text)
+ entries.append(MemoryEntry(
+ name=meta.get("name", f.stem),
+ description=meta.get("description", ""),
+ type=meta.get("type", "user"),
+ content=body,
+ file_path=str(f),
+ created=meta.get("created", ""),
+ ))
+ except Exception:
+ continue
+ return entries
+
+
+def delete_memory(name: str) -> None:
+ """
+ Delete a memory by name.
+
+ Removes the entry file and updates the index.
+ """
+ slug = _slugify(name)
+ file_path = MEMORY_DIR / f"{slug}.md"
+ if file_path.exists():
+ file_path.unlink()
+
+ remaining = [e for e in load_index() if e.name != name]
+ _update_index(remaining)
+
+
+def search_memory(query: str) -> List[MemoryEntry]:
+ """
+ Search memories by keyword match against name, description, and content.
+
+ Args:
+ query: search string (case-insensitive)
+ Returns:
+ list of matching MemoryEntry
+ """
+ query_lower = query.lower()
+ results = []
+ for entry in load_index():
+ searchable = f"{entry.name} {entry.description} {entry.content}".lower()
+ if query_lower in searchable:
+ results.append(entry)
+ return results
+
+
+def get_memory_context() -> str:
+ """
+ Return memory index content for system prompt injection.
+
+ Returns:
+ MEMORY.md content string, or empty string if no memories
+ """
+ if not INDEX_FILE.exists():
+ return ""
+ text = INDEX_FILE.read_text(encoding="utf-8").strip()
+ return text
+```
+
+- [ ] **Step 4: Run tests to verify they pass**
+
+Run: `cd D:/git/open-cc/nano-claude-code && python -m pytest tests/test_memory.py -v`
+Expected: All 9 tests PASS
+
+- [ ] **Step 5: Register MemorySave and MemoryDelete tools**
+
+Add to the bottom of `tools.py`, after the `_register_builtin_tools()` call:
+
+```python
+# --- Memory tools ---------------------------------------------------------
+
+from memory import save_memory, delete_memory, MemoryEntry
+from datetime import datetime as _dt
+
+
+def _memory_save(params, config):
+ entry = MemoryEntry(
+ name=params["name"],
+ description=params["description"],
+ type=params["type"],
+ content=params["content"],
+ created=_dt.now().strftime("%Y-%m-%d"),
+ )
+ save_memory(entry)
+ return f"Memory saved: {entry.name}"
+
+
+def _memory_delete(params, config):
+ delete_memory(params["name"])
+ return f"Memory deleted: {params['name']}"
+
+
+_MEMORY_SAVE_SCHEMA = {
+ "name": "MemorySave",
+ "description": "Save a persistent memory that survives across conversations.",
+ "input_schema": {
+ "type": "object",
+ "properties": {
+ "name": {"type": "string", "description": "Short name for the memory"},
+ "type": {"type": "string", "enum": ["user", "feedback", "project", "reference"],
+ "description": "Memory category"},
+ "description": {"type": "string", "description": "One-line description"},
+ "content": {"type": "string", "description": "Full memory content"},
+ },
+ "required": ["name", "type", "description", "content"],
+ },
+}
+
+_MEMORY_DELETE_SCHEMA = {
+ "name": "MemoryDelete",
+ "description": "Delete a persistent memory by name.",
+ "input_schema": {
+ "type": "object",
+ "properties": {
+ "name": {"type": "string", "description": "Name of memory to delete"},
+ },
+ "required": ["name"],
+ },
+}
+
+register_tool(ToolDef(name="MemorySave", schema=_MEMORY_SAVE_SCHEMA,
+ func=_memory_save, read_only=False, concurrent_safe=True))
+register_tool(ToolDef(name="MemoryDelete", schema=_MEMORY_DELETE_SCHEMA,
+ func=_memory_delete, read_only=False, concurrent_safe=True))
+```
+
+- [ ] **Step 6: Inject memory into context.py**
+
+Modify `context.py`'s `build_system_prompt()` function (line 92). Add import at top:
+
+```python
+from memory import get_memory_context
+```
+
+At the end of `build_system_prompt()`, before the return statement, add:
+
+```python
+ memory_ctx = get_memory_context()
+ if memory_ctx:
+ prompt += f"\n\n# Memory\nYour persistent memories:\n{memory_ctx}\n"
+```
+
+- [ ] **Step 7: Commit**
+
+```bash
+cd D:/git/open-cc/nano-claude-code
+git add memory.py tests/test_memory.py tools.py context.py
+git commit -m "feat: add file-based memory system with MemorySave/MemoryDelete tools"
+```
+
+---
+
+## Task 6: Sub-Agent System (`subagent.py`)
+
+**Files:**
+- Create: `subagent.py`
+- Create: `tests/test_subagent.py`
+- Modify: `agent.py` (add depth, cancel_check params)
+- Modify: `tools.py` (register Agent, CheckAgentResult, ListAgentTasks)
+
+- [ ] **Step 1: Write failing tests for sub-agent**
+
+```python
+# tests/test_subagent.py
+import sys, os, time, threading
+sys.path.insert(0, os.path.dirname(os.path.dirname(os.path.abspath(__file__))))
+
+import pytest
+from subagent import SubAgentTask, SubAgentManager
+
+
+def _mock_agent_run(prompt, state, config, system_prompt, depth=0, cancel_check=None):
+ """Mock agent.run that yields a simple response."""
+ from agent import TurnDone
+ # Simulate work
+ for i in range(3):
+ if cancel_check and cancel_check():
+ return
+ time.sleep(0.05)
+ state.messages.append({"role": "assistant", "content": f"Done: {prompt}"})
+ yield TurnDone(input_tokens=10, output_tokens=20)
+
+
+@pytest.fixture
+def manager(monkeypatch):
+ import subagent
+ monkeypatch.setattr(subagent, "_agent_run", _mock_agent_run)
+ mgr = SubAgentManager(max_concurrent=2, max_depth=3)
+ yield mgr
+ mgr.shutdown()
+
+
+def test_spawn_and_wait(manager):
+ task = manager.spawn("say hello", {}, "system prompt")
+ result = manager.wait(task.id, timeout=5)
+ assert result.status == "completed"
+ assert "Done: say hello" in result.result
+
+
+def test_spawn_returns_immediately(manager):
+ task = manager.spawn("do work", {}, "system prompt")
+ assert task.status in ("pending", "running")
+ assert task.id is not None
+
+
+def test_list_tasks(manager):
+ manager.spawn("task 1", {}, "sys")
+ manager.spawn("task 2", {}, "sys")
+ tasks = manager.list_tasks()
+ assert len(tasks) == 2
+
+
+def test_cancel(manager):
+ # Use a slow mock
+ import subagent
+ original = subagent._agent_run
+
+ def slow_run(prompt, state, config, system_prompt, depth=0, cancel_check=None):
+ from agent import TurnDone
+ for i in range(100):
+ if cancel_check and cancel_check():
+ return
+ time.sleep(0.05)
+ state.messages.append({"role": "assistant", "content": "should not reach"})
+ yield TurnDone(input_tokens=0, output_tokens=0)
+
+ subagent._agent_run = slow_run
+ task = manager.spawn("slow task", {}, "sys")
+ time.sleep(0.1)
+ assert manager.cancel(task.id) is True
+ manager.wait(task.id, timeout=5)
+ assert task.status == "cancelled"
+ subagent._agent_run = original
+
+
+def test_depth_limit(manager):
+ manager.max_depth = 2
+ task = manager.spawn("deep task", {}, "sys", depth=2)
+ assert task.status == "failed"
+ assert "depth" in task.result.lower()
+
+
+def test_get_result(manager):
+ task = manager.spawn("hello", {}, "sys")
+ manager.wait(task.id, timeout=5)
+ result = manager.get_result(task.id)
+ assert "Done: hello" in result
+
+
+def test_get_result_unknown():
+ mgr = SubAgentManager()
+ assert mgr.get_result("nonexistent") is None
+ mgr.shutdown()
+```
+
+- [ ] **Step 2: Run tests to verify they fail**
+
+Run: `cd D:/git/open-cc/nano-claude-code && python -m pytest tests/test_subagent.py -v`
+Expected: FAIL โ `ModuleNotFoundError: No module named 'subagent'`
+
+- [ ] **Step 3: Update agent.py to accept depth and cancel_check**
+
+Modify the `run()` function signature in `agent.py` (line 52) from:
+
+```python
+def run(user_message: str, state: AgentState, config: dict, system_prompt: str) -> Generator:
+```
+
+to:
+
+```python
+def run(user_message: str, state: AgentState, config: dict, system_prompt: str,
+ depth: int = 0, cancel_check=None) -> Generator:
+```
+
+Add a cancellation check inside the `while True` loop, at the very top of the loop body:
+
+```python
+ # --- Cancellation check ---
+ if cancel_check and cancel_check():
+ return
+```
+
+- [ ] **Step 4: Implement subagent.py**
+
+```python
+# subagent.py
+"""
+Sub-agent lifecycle management for nano-claude-code.
+
+Sub-agents run in background threads via ThreadPoolExecutor.
+Each has independent message history and fresh context.
+Depth limiting prevents infinite recursion. Cancellation is cooperative.
+
+Public API:
+ SubAgentManager โ spawn, wait, cancel, list, get_result, shutdown
+"""
+from __future__ import annotations
+
+import uuid
+import threading
+from concurrent.futures import ThreadPoolExecutor, Future
+from dataclasses import dataclass, field
+from typing import Dict, List, Optional, Callable
+
+
+@dataclass
+class SubAgentTask:
+ """A single sub-agent task."""
+ id: str
+ prompt: str
+ status: str = "pending" # pending | running | completed | failed | cancelled
+ result: Optional[str] = None
+ depth: int = 0
+ _cancel_flag: bool = False
+ _future: Optional[Future] = field(default=None, repr=False)
+
+
+def _agent_run(prompt, state, config, system_prompt, depth=0, cancel_check=None):
+ """
+ Reference to agent.run(). Replaced in __init__ or monkeypatched in tests.
+ Importing agent at module level would create circular import,
+ so we do a lazy import.
+ """
+ from agent import run
+ return run(prompt, state, config, system_prompt,
+ depth=depth, cancel_check=cancel_check)
+
+
+def _extract_final_text(messages: list) -> str:
+ """Extract the last assistant message text from a message list."""
+ for msg in reversed(messages):
+ if msg.get("role") == "assistant":
+ content = msg.get("content", "")
+ if isinstance(content, str) and content:
+ return content
+ return "(no response)"
+
+
+class SubAgentManager:
+ """
+ Manages sub-agent lifecycle.
+
+ Args:
+ max_concurrent: max parallel sub-agents, int
+ max_depth: max recursion depth for sub-agent chains, int
+ """
+
+ def __init__(self, max_concurrent: int = 3, max_depth: int = 3):
+ self.tasks: Dict[str, SubAgentTask] = {}
+ self.max_concurrent = max_concurrent
+ self.max_depth = max_depth
+ self._pool = ThreadPoolExecutor(max_workers=max_concurrent)
+
+ def spawn(self, prompt: str, config: dict, system_prompt: str,
+ depth: int = 0) -> SubAgentTask:
+ """
+ Spawn a new sub-agent in a background thread.
+
+ Args:
+ prompt: task description for the sub-agent
+ config: global config dict
+ system_prompt: system prompt for sub-agent
+ depth: current recursion depth
+ Returns:
+ SubAgentTask (status may be "failed" if depth exceeded)
+ """
+ task_id = uuid.uuid4().hex[:8]
+
+ if depth >= self.max_depth:
+ task = SubAgentTask(
+ id=task_id, prompt=prompt, status="failed", depth=depth,
+ result="Error: max sub-agent depth reached. Complete this task directly.",
+ )
+ self.tasks[task_id] = task
+ return task
+
+ task = SubAgentTask(id=task_id, prompt=prompt, status="running", depth=depth)
+ self.tasks[task_id] = task
+
+ def _run():
+ from agent import AgentState, TurnDone
+
+ sub_state = AgentState()
+ sub_config = {**config, "max_tokens": config.get("max_tokens", 8192)}
+
+ sub_sys = (
+ f"You are a sub-agent. Complete the following task:\n\n"
+ f"{prompt}\n\n"
+ f"---\n{system_prompt}"
+ )
+
+ try:
+ for event in _agent_run(
+ prompt, sub_state, sub_config, sub_sys,
+ depth=depth + 1,
+ cancel_check=lambda: task._cancel_flag,
+ ):
+ pass # consume events
+
+ if task._cancel_flag:
+ task.status = "cancelled"
+ task.result = "Task was cancelled."
+ else:
+ task.result = _extract_final_text(sub_state.messages)
+ task.status = "completed"
+ except Exception as e:
+ task.result = f"Error: {e}"
+ task.status = "failed"
+
+ task._future = self._pool.submit(_run)
+ return task
+
+ def wait(self, task_id: str, timeout: float = None) -> Optional[SubAgentTask]:
+ """
+ Block until a task completes or timeout.
+
+ Args:
+ task_id: task identifier
+ timeout: max wait seconds (None = wait forever)
+ Returns:
+ SubAgentTask or None if not found
+ """
+ task = self.tasks.get(task_id)
+ if task is None:
+ return None
+ if task._future is not None:
+ try:
+ task._future.result(timeout=timeout)
+ except Exception:
+ pass
+ return task
+
+ def get_result(self, task_id: str) -> Optional[str]:
+ """Get the result string for a task. None if not found."""
+ task = self.tasks.get(task_id)
+ return task.result if task else None
+
+ def list_tasks(self) -> List[SubAgentTask]:
+ """Return all tasks."""
+ return list(self.tasks.values())
+
+ def cancel(self, task_id: str) -> bool:
+ """
+ Request cancellation of a running task.
+
+ Args:
+ task_id: task identifier
+ Returns:
+ True if cancellation was requested, False if task not found or not running
+ """
+ task = self.tasks.get(task_id)
+ if task is None or task.status != "running":
+ return False
+ task._cancel_flag = True
+ return True
+
+ def shutdown(self) -> None:
+ """Shutdown the thread pool. Cancel all running tasks first."""
+ for task in self.tasks.values():
+ if task.status == "running":
+ task._cancel_flag = True
+ self._pool.shutdown(wait=False)
+```
+
+- [ ] **Step 5: Run tests to verify they pass**
+
+Run: `cd D:/git/open-cc/nano-claude-code && python -m pytest tests/test_subagent.py -v`
+Expected: All 7 tests PASS
+
+- [ ] **Step 6: Register Agent, CheckAgentResult, ListAgentTasks tools**
+
+Add to the bottom of `tools.py`:
+
+```python
+# --- Sub-agent tools ------------------------------------------------------
+
+from subagent import SubAgentManager
+
+# Singleton manager (created lazily)
+_agent_manager = None
+
+
+def _get_agent_manager():
+ global _agent_manager
+ if _agent_manager is None:
+ _agent_manager = SubAgentManager()
+ return _agent_manager
+
+
+def _agent_tool(params, config):
+ from context import build_system_prompt
+ mgr = _get_agent_manager()
+ prompt = params["prompt"]
+ model = params.get("model")
+ wait = params.get("wait", True)
+
+ sub_config = {**config}
+ if model:
+ sub_config["model"] = model
+
+ task = mgr.spawn(prompt, sub_config, build_system_prompt())
+
+ if task.status == "failed":
+ return task.result
+
+ if wait:
+ mgr.wait(task.id, timeout=300)
+ return f"[Agent {task.id}] {task.status}: {task.result}"
+ else:
+ return f"[Agent {task.id}] Spawned in background. Use CheckAgentResult with task_id='{task.id}' to get the result."
+
+
+def _check_agent_result(params, config):
+ mgr = _get_agent_manager()
+ task_id = params["task_id"]
+ task = mgr.tasks.get(task_id)
+ if task is None:
+ return f"Error: no task with id '{task_id}'"
+ if task.status == "running":
+ return f"[Agent {task_id}] Still running..."
+ return f"[Agent {task_id}] {task.status}: {task.result}"
+
+
+def _list_agent_tasks(params, config):
+ mgr = _get_agent_manager()
+ tasks = mgr.list_tasks()
+ if not tasks:
+ return "No sub-agent tasks."
+ lines = ["ID | Status | Prompt"]
+ lines.append("-------- | --------- | ------")
+ for t in tasks:
+ preview = t.prompt[:50] + ("..." if len(t.prompt) > 50 else "")
+ lines.append(f"{t.id} | {t.status:9s} | {preview}")
+ return "\n".join(lines)
+
+
+_AGENT_SCHEMA = {
+ "name": "Agent",
+ "description": "Launch a sub-agent to handle a task independently. "
+ "Use for complex subtasks that can be delegated.",
+ "input_schema": {
+ "type": "object",
+ "properties": {
+ "prompt": {"type": "string", "description": "Task description for the sub-agent"},
+ "model": {"type": "string", "description": "Optional model override"},
+ "wait": {"type": "boolean", "description": "True (default) = wait for result. False = return task_id for later polling."},
+ },
+ "required": ["prompt"],
+ },
+}
+
+_CHECK_AGENT_RESULT_SCHEMA = {
+ "name": "CheckAgentResult",
+ "description": "Check the result of a background sub-agent task.",
+ "input_schema": {
+ "type": "object",
+ "properties": {
+ "task_id": {"type": "string", "description": "Task ID returned by Agent tool"},
+ },
+ "required": ["task_id"],
+ },
+}
+
+_LIST_AGENT_TASKS_SCHEMA = {
+ "name": "ListAgentTasks",
+ "description": "List all sub-agent tasks and their status.",
+ "input_schema": {
+ "type": "object",
+ "properties": {},
+ },
+}
+
+register_tool(ToolDef(name="Agent", schema=_AGENT_SCHEMA,
+ func=_agent_tool, read_only=False, concurrent_safe=False))
+register_tool(ToolDef(name="CheckAgentResult", schema=_CHECK_AGENT_RESULT_SCHEMA,
+ func=_check_agent_result, read_only=True, concurrent_safe=True))
+register_tool(ToolDef(name="ListAgentTasks", schema=_LIST_AGENT_TASKS_SCHEMA,
+ func=_list_agent_tasks, read_only=True, concurrent_safe=True))
+```
+
+- [ ] **Step 7: Commit**
+
+```bash
+cd D:/git/open-cc/nano-claude-code
+git add subagent.py tests/test_subagent.py agent.py tools.py
+git commit -m "feat: add threaded sub-agent system with Agent/CheckAgentResult/ListAgentTasks tools"
+```
+
+---
+
+## Task 7: Skills System (`skills.py`)
+
+**Files:**
+- Create: `skills.py`
+- Create: `tests/test_skills.py`
+- Modify: `nano_claude.py` (skill dispatch in REPL)
+
+- [ ] **Step 1: Write failing tests for skills**
+
+```python
+# tests/test_skills.py
+import sys, os, tempfile
+sys.path.insert(0, os.path.dirname(os.path.dirname(os.path.abspath(__file__))))
+
+import pytest
+from skills import SkillDef, load_skills, find_skill, _parse_skill_file
+
+
+@pytest.fixture
+def skill_dir(tmp_path):
+ """Create a temp skill directory with sample skills."""
+ skill_file = tmp_path / "commit.md"
+ skill_file.write_text(
+ "---\n"
+ "name: commit\n"
+ "description: Create a git commit\n"
+ "triggers: [\"/commit\", \"commit changes\"]\n"
+ "tools: [Bash, Read]\n"
+ "---\n\n"
+ "# Commit Skill\n\n"
+ "Analyze staged changes and create a commit message.\n"
+ )
+
+ review_file = tmp_path / "review.md"
+ review_file.write_text(
+ "---\n"
+ "name: review\n"
+ "description: Review a PR\n"
+ "triggers: [\"/review\", \"/review-pr\"]\n"
+ "tools: [Bash, Read, Grep]\n"
+ "---\n\n"
+ "# Review Skill\n\n"
+ "Review the current PR for issues.\n"
+ )
+ return tmp_path
+
+
+def test_parse_skill_file(skill_dir):
+ skill = _parse_skill_file(skill_dir / "commit.md")
+ assert skill.name == "commit"
+ assert skill.description == "Create a git commit"
+ assert "/commit" in skill.triggers
+ assert "Bash" in skill.tools
+ assert "Analyze staged" in skill.prompt
+
+
+def test_load_skills(skill_dir, monkeypatch):
+ import skills
+ monkeypatch.setattr(skills, "SKILL_PATHS", [skill_dir])
+ loaded = load_skills()
+ assert len(loaded) == 2
+ names = {s.name for s in loaded}
+ assert "commit" in names
+ assert "review" in names
+
+
+def test_find_skill_by_slash_command(skill_dir, monkeypatch):
+ import skills
+ monkeypatch.setattr(skills, "SKILL_PATHS", [skill_dir])
+ skill = find_skill("/commit")
+ assert skill is not None
+ assert skill.name == "commit"
+
+
+def test_find_skill_with_args(skill_dir, monkeypatch):
+ import skills
+ monkeypatch.setattr(skills, "SKILL_PATHS", [skill_dir])
+ skill = find_skill("/review")
+ assert skill is not None
+ assert skill.name == "review"
+
+
+def test_find_skill_not_found(skill_dir, monkeypatch):
+ import skills
+ monkeypatch.setattr(skills, "SKILL_PATHS", [skill_dir])
+ skill = find_skill("/nonexistent")
+ assert skill is None
+
+
+def test_load_skills_empty_dir(tmp_path, monkeypatch):
+ import skills
+ monkeypatch.setattr(skills, "SKILL_PATHS", [tmp_path])
+ loaded = load_skills()
+ assert loaded == []
+
+
+def test_load_skills_nonexistent_dir(monkeypatch):
+ import skills
+ monkeypatch.setattr(skills, "SKILL_PATHS", ["/nonexistent/path"])
+ loaded = load_skills()
+ assert loaded == []
+```
+
+- [ ] **Step 2: Run tests to verify they fail**
+
+Run: `cd D:/git/open-cc/nano-claude-code && python -m pytest tests/test_skills.py -v`
+Expected: FAIL โ `ModuleNotFoundError: No module named 'skills'`
+
+- [ ] **Step 3: Implement skills.py**
+
+```python
+# skills.py
+"""
+Skill loading and execution for nano-claude-code.
+
+Skills are markdown files with YAML frontmatter that define reusable
+prompt templates. They are injected into the agent loop as user messages.
+
+Public API:
+ load_skills() โ scan skill directories
+ find_skill(query) โ match by /command or keyword
+ execute_skill(skill, args, state, ...) โ run skill via agent loop
+"""
+from __future__ import annotations
+
+import re
+from dataclasses import dataclass
+from pathlib import Path
+from typing import Generator, List, Optional
+
+
+SKILL_PATHS = [
+ Path.cwd() / ".nano_claude" / "skills",
+ Path.home() / ".nano_claude" / "skills",
+]
+
+
+@dataclass
+class SkillDef:
+ """Definition of a loaded skill."""
+ name: str
+ description: str
+ triggers: List[str]
+ tools: List[str]
+ prompt: str
+ file_path: str
+
+
+def _parse_list_field(value: str) -> List[str]:
+ """Parse a YAML-like list field: [a, b, c] or \"a, b, c\"."""
+ value = value.strip()
+ if value.startswith("[") and value.endswith("]"):
+ value = value[1:-1]
+ # Handle quoted items
+ items = []
+ for item in value.split(","):
+ item = item.strip().strip('"').strip("'")
+ if item:
+ items.append(item)
+ return items
+
+
+def _parse_skill_file(path: Path) -> Optional[SkillDef]:
+ """
+ Parse a single skill markdown file.
+
+ Args:
+ path: path to .md skill file
+ Returns:
+ SkillDef or None if parsing fails
+ """
+ try:
+ text = path.read_text(encoding="utf-8")
+ except Exception:
+ return None
+
+ if not text.startswith("---"):
+ return None
+
+ parts = text.split("---", 2)
+ if len(parts) < 3:
+ return None
+
+ # Parse frontmatter
+ meta = {}
+ for line in parts[1].strip().splitlines():
+ if ":" in line:
+ key, val = line.split(":", 1)
+ meta[key.strip()] = val.strip()
+
+ name = meta.get("name", "")
+ if not name:
+ return None
+
+ return SkillDef(
+ name=name,
+ description=meta.get("description", ""),
+ triggers=_parse_list_field(meta.get("triggers", "")),
+ tools=_parse_list_field(meta.get("tools", "")),
+ prompt=parts[2].strip(),
+ file_path=str(path),
+ )
+
+
+def load_skills() -> List[SkillDef]:
+ """
+ Scan all skill directories and parse skill files.
+
+ Project-level skills override user-level skills with the same name.
+
+ Returns:
+ list of SkillDef
+ """
+ skills_by_name = {}
+ # Reverse order so project-level (first in list) overrides user-level
+ for skill_dir in reversed(SKILL_PATHS):
+ if not Path(skill_dir).exists():
+ continue
+ for f in sorted(Path(skill_dir).glob("*.md")):
+ skill = _parse_skill_file(f)
+ if skill:
+ skills_by_name[skill.name] = skill
+ return list(skills_by_name.values())
+
+
+def find_skill(query: str) -> Optional[SkillDef]:
+ """
+ Find a skill matching a query string.
+
+ Matches against trigger strings (e.g., "/commit").
+ The query is matched with and without leading arguments.
+
+ Args:
+ query: user input string (e.g., "/commit -m fix bug")
+ Returns:
+ matching SkillDef or None
+ """
+ # Extract the command part (first word)
+ cmd = query.strip().split()[0] if query.strip() else ""
+
+ for skill in load_skills():
+ for trigger in skill.triggers:
+ if cmd == trigger or query.strip() == trigger:
+ return skill
+ return None
+
+
+def execute_skill(skill: SkillDef, args: str, state, config: dict,
+ system_prompt: str) -> Generator:
+ """
+ Execute a skill by injecting its prompt into the agent loop.
+
+ Args:
+ skill: the skill to execute
+ args: additional arguments from the user
+ state: AgentState
+ config: global config dict
+ system_prompt: current system prompt
+ Yields:
+ agent events (TextChunk, ToolStart, etc.)
+ """
+ from agent import run
+
+ prompt = f"[Skill: {skill.name}]\n\n{skill.prompt}"
+ if args:
+ prompt += f"\n\nUser context: {args}"
+
+ for event in run(prompt, state, config, system_prompt):
+ yield event
+```
+
+- [ ] **Step 4: Run tests to verify they pass**
+
+Run: `cd D:/git/open-cc/nano-claude-code && python -m pytest tests/test_skills.py -v`
+Expected: All 7 tests PASS
+
+- [ ] **Step 5: Integrate skills into nano_claude.py REPL**
+
+In `nano_claude.py`, add import near the top:
+
+```python
+from skills import find_skill, execute_skill, load_skills
+```
+
+Modify the `handle_slash` function (line 373) to fall through to skill lookup. At the end of the function, after checking all built-in commands, before the "unknown command" error:
+
+```python
+ # Try skills
+ skill = find_skill(line)
+ if skill:
+ # Extract args: everything after the trigger
+ cmd_parts = line.strip().split(maxsplit=1)
+ args = cmd_parts[1] if len(cmd_parts) > 1 else ""
+ return skill, args # return skill for REPL to execute
+```
+
+This requires adjusting `handle_slash` return type. Currently it returns `bool`. Change it to return `tuple | bool`:
+- `True` / `False` for built-in commands
+- `(SkillDef, args_str)` for skill matches
+
+In the REPL loop (around line 440), where `handle_slash` is called, add skill execution:
+
+```python
+ result = handle_slash(user_input, state, config)
+ if isinstance(result, tuple):
+ # It's a skill
+ skill, args = result
+ system_prompt = build_system_prompt()
+ info(f"Running skill: {skill.name}")
+ for event in execute_skill(skill, args, state, config, system_prompt):
+ # Reuse existing event handling (same as query events)
+ if isinstance(event, TextChunk):
+ stream_text(event.text)
+ elif isinstance(event, ThinkingChunk):
+ stream_thinking(event.text, config.get("verbose"))
+ elif isinstance(event, ToolStart):
+ print_tool_start(event.name, event.inputs, config.get("verbose"))
+ elif isinstance(event, ToolEnd):
+ print_tool_end(event.name, event.result, config.get("verbose"))
+ elif isinstance(event, TurnDone):
+ flush_response()
+ continue
+ elif result:
+ continue
+```
+
+Also add a `/skills` slash command to list available skills:
+
+```python
+def cmd_skills(_args, _state, _config):
+ skills = load_skills()
+ if not skills:
+ info("No skills found. Place .md files in ~/.nano_claude/skills/")
+ return True
+ for s in skills:
+ triggers = ", ".join(s.triggers)
+ info(f" {s.name:15s} {triggers:25s} {s.description}")
+ return True
+```
+
+Add `"skills": cmd_skills` to the `COMMANDS` dict.
+
+- [ ] **Step 6: Commit**
+
+```bash
+cd D:/git/open-cc/nano-claude-code
+git add skills.py tests/test_skills.py nano_claude.py
+git commit -m "feat: add skills system with markdown definitions and /skills command"
+```
+
+---
+
+## Task 8: Config Updates + Remaining Slash Commands
+
+**Files:**
+- Modify: `config.py` (new config keys)
+- Modify: `nano_claude.py` (add /memory, /agents commands)
+
+- [ ] **Step 1: Add new config defaults**
+
+In `config.py`, add to the `DEFAULTS` dict (line 11):
+
+```python
+ "max_tool_output": 32000,
+ "max_agent_depth": 3,
+ "max_concurrent_agents": 3,
+```
+
+- [ ] **Step 2: Add /memory slash command**
+
+In `nano_claude.py`, add:
+
+```python
+def cmd_memory(args, _state, _config):
+ from memory import load_index, search_memory
+ if args:
+ results = search_memory(args)
+ if not results:
+ info(f"No memories matching '{args}'")
+ return True
+ for m in results:
+ info(f" [{m.type}] {m.name}: {m.description}")
+ info(f" {m.content[:100]}{'...' if len(m.content) > 100 else ''}")
+ return True
+ entries = load_index()
+ if not entries:
+ info("No memories stored. The model can save memories via MemorySave tool.")
+ return True
+ info(f" {len(entries)} memories:")
+ for m in entries:
+ info(f" [{m.type:9s}] {m.name}: {m.description}")
+ return True
+```
+
+Add `"memory": cmd_memory` to the `COMMANDS` dict.
+
+- [ ] **Step 3: Add /agents slash command**
+
+In `nano_claude.py`, add:
+
+```python
+def cmd_agents(_args, _state, _config):
+ try:
+ from tools import _get_agent_manager
+ mgr = _get_agent_manager()
+ tasks = mgr.list_tasks()
+ if not tasks:
+ info("No sub-agent tasks.")
+ return True
+ info(f" {len(tasks)} sub-agent tasks:")
+ for t in tasks:
+ preview = t.prompt[:40] + ("..." if len(t.prompt) > 40 else "")
+ info(f" {t.id} [{t.status:9s}] {preview}")
+ except Exception:
+ info("Sub-agent system not initialized.")
+ return True
+```
+
+Add `"agents": cmd_agents` to the `COMMANDS` dict.
+
+- [ ] **Step 4: Update /help to include new commands**
+
+In `cmd_help`, add the new commands to the help text:
+
+```python
+ info(" /memory [query] Show/search persistent memories")
+ info(" /skills List available skills")
+ info(" /agents Show sub-agent tasks")
+```
+
+- [ ] **Step 5: Verify all tests pass**
+
+Run: `cd D:/git/open-cc/nano-claude-code && python -m pytest tests/ -v`
+Expected: All tests across all test files PASS
+
+- [ ] **Step 6: Commit**
+
+```bash
+cd D:/git/open-cc/nano-claude-code
+git add config.py nano_claude.py
+git commit -m "feat: add /memory, /skills, /agents commands and new config defaults"
+```
+
+---
+
+## Task 9: Integration Smoke Test
+
+**Files:** None (testing only)
+
+- [ ] **Step 1: Verify tool registry loads all tools**
+
+Run: `cd D:/git/open-cc/nano-claude-code && python -c "
+from tool_registry import get_all_tools
+tools = get_all_tools()
+print(f'{len(tools)} tools registered:')
+for t in tools:
+ print(f' {t.name:20s} read_only={t.read_only} concurrent_safe={t.concurrent_safe}')
+"`
+Expected: 13 tools (8 built-in + MemorySave + MemoryDelete + Agent + CheckAgentResult + ListAgentTasks)
+
+- [ ] **Step 2: Verify compaction module loads**
+
+Run: `cd D:/git/open-cc/nano-claude-code && python -c "
+from compaction import estimate_tokens, get_context_limit
+msgs = [{'role': 'user', 'content': 'hello ' * 1000}]
+print(f'Tokens: {estimate_tokens(msgs)}')
+print(f'GPT-4o limit: {get_context_limit(\"gpt-4o\")}')
+print(f'Gemini limit: {get_context_limit(\"gemini-2.0-flash\")}')
+"`
+Expected: Token count ~1700, GPT-4o 128000, Gemini 1000000
+
+- [ ] **Step 3: Verify memory roundtrip**
+
+Run: `cd D:/git/open-cc/nano-claude-code && python -c "
+import tempfile
+from pathlib import Path
+import memory
+memory.MEMORY_DIR = Path(tempfile.mkdtemp())
+memory.INDEX_FILE = memory.MEMORY_DIR / 'MEMORY.md'
+from memory import MemoryEntry, save_memory, load_index, get_memory_context
+save_memory(MemoryEntry(name='test', description='integration test', type='user', content='works!'))
+entries = load_index()
+print(f'Entries: {len(entries)}, name={entries[0].name}')
+print(f'Context: {get_memory_context()}')
+"`
+Expected: 1 entry, context contains "test"
+
+- [ ] **Step 4: Verify skills loading**
+
+Run: `cd D:/git/open-cc/nano-claude-code && python -c "
+from skills import load_skills
+skills = load_skills()
+print(f'{len(skills)} skills loaded')
+for s in skills:
+ print(f' {s.name}: {s.triggers}')
+"`
+Expected: 0 skills (none created yet), no errors
+
+- [ ] **Step 5: Run full test suite**
+
+Run: `cd D:/git/open-cc/nano-claude-code && python -m pytest tests/ -v --tb=short`
+Expected: All tests PASS
+
+- [ ] **Step 6: Final commit with all tests passing**
+
+```bash
+cd D:/git/open-cc/nano-claude-code
+git add -A
+git status
+# Only commit if there are uncommitted changes
+git commit -m "chore: integration verification complete"
+```
diff --git a/nano-claude-code/docs/superpowers/specs/2026-04-02-open-cc-design.md b/nano-claude-code/docs/superpowers/specs/2026-04-02-open-cc-design.md
new file mode 100644
index 0000000..c13958b
--- /dev/null
+++ b/nano-claude-code/docs/superpowers/specs/2026-04-02-open-cc-design.md
@@ -0,0 +1,643 @@
+# Open-CC: Nano Claude Code Enhancement Design
+
+**Date:** 2026-04-02
+**Status:** Approved
+**Target:** GPT-5.4, Gemini 3/3.1 Pro (Claude not in scope)
+**Code budget:** ~10K lines total (currently ~2.2K)
+**Constraint:** PR-friendly, mergeable back to nano-claude-code upstream
+
+---
+
+## 1. Overview
+
+Evolve nano-claude-code from a minimal ~2.2K-line reference implementation into a capable AI coding CLI, approaching Claude Code's core functionality while staying lean. Five enhancement areas:
+
+1. **Context Window Management** (`compaction.py`)
+2. **Tool System Enhancement** (`tool_registry.py` + `tools.py` refactor)
+3. **Sub-Agent** (`subagent.py`)
+4. **Memory System** (`memory.py`)
+5. **Skills System** (`skills.py`)
+
+### Strategy
+
+**Approach A: Layered Enhancement** -- add new modules alongside existing files, minimize changes to existing code. When agent.py grows too complex, refactor into Approach B (package structure under `ncc/`).
+
+### Design Principles
+
+- Modules communicate via function parameters / dataclasses, no globals
+- Each new module exposes 2-3 public functions, internals self-contained
+- New logic in agent.py grouped by clear `# --- section ---` comments
+- All code in English (comments, docstrings, commit messages)
+
+---
+
+## 2. File Structure
+
+```
+nano-claude-code/
+โโโ nano_claude.py # REPL -- add /memory, /skill slash commands
+โโโ agent.py # Agent loop -- add compaction call + sub-agent dispatch
+โโโ providers.py # No changes (already solid)
+โโโ tools.py # Refactor: register built-in tools via registry
+โโโ context.py # Extend: inject memory context
+โโโ config.py # Add new config keys
+โ
+โโโ compaction.py # NEW: Context window management
+โโโ subagent.py # NEW: Sub-agent lifecycle
+โโโ memory.py # NEW: File-based memory system
+โโโ skills.py # NEW: Skill loading and execution
+โโโ tool_registry.py # NEW: Tool plugin registry
+```
+
+### Module Dependency Graph (unidirectional)
+
+```
+nano_claude.py
+ โ-> agent.py
+ โ โ-> providers.py
+ โ โ-> tool_registry.py -> tools.py (built-in implementations)
+ โ โ-> compaction.py -> providers.py (for summary model call)
+ โ โ-> subagent.py (calls agent.py:run recursively)
+ โ-> context.py -> memory.py
+ โ-> skills.py -> tool_registry.py
+ โ-> config.py
+```
+
+---
+
+## 3. Context Window Management (`compaction.py`)
+
+Two-layer compression, inspired by Claude Code's three-layer strategy (Layer 3 contextCollapse is experimental, deferred).
+
+### 3.1 Layer 1: Auto-Compact (model-driven summary)
+
+Triggered when estimated token count exceeds 70% of model's context limit.
+
+```python
+def compact_messages(messages: list[dict], config: dict) -> list[dict]:
+ """
+ Split messages into [old | recent].
+ Summarize old via model call.
+ Return [summary_msg, ack_msg, *recent].
+ """
+ split_point = find_split_point(messages, keep_ratio=0.3)
+ old = messages[:split_point]
+ recent = messages[split_point:]
+ summary = call_model_for_summary(old, config)
+ return [
+ {"role": "user", "content": f"[Conversation summary]\n{summary}"},
+ {"role": "assistant", "content": "Understood, I have the context."},
+ *recent
+ ]
+```
+
+### 3.2 Layer 2: Tool-Result Snipping (rule-based)
+
+Truncate old tool outputs without model call. Fast and cheap.
+
+```python
+def snip_old_tool_results(messages: list[dict], max_chars: int = 2000) -> list[dict]:
+ """
+ For tool results older than N turns, truncate to max_chars.
+ Preserve first/last lines, add [snipped N chars] marker.
+ """
+```
+
+### 3.3 Token Estimation
+
+```python
+def estimate_tokens(messages: list[dict]) -> int:
+ """Use tiktoken for GPT models, chars/3.5 fallback."""
+
+def get_context_limit(model: str) -> int:
+ """Return context window size from provider registry."""
+```
+
+### 3.4 Integration Point
+
+```python
+# In agent.py run() loop, before each API call:
+def _maybe_compact(state: AgentState, config: dict) -> bool:
+ token_count = estimate_tokens(state.messages)
+ threshold = get_context_limit(config["model"]) * 0.7
+ if token_count > threshold:
+ state.messages = compact_messages(state.messages, config)
+ return True
+ return False
+```
+
+### 3.5 Public API
+
+```python
+maybe_compact(state: AgentState, config: dict) -> bool
+estimate_tokens(messages: list[dict]) -> int
+get_context_limit(model: str) -> int
+```
+
+---
+
+## 4. Tool System Enhancement (`tool_registry.py` + `tools.py`)
+
+### 4.1 Tool Registry
+
+```python
+@dataclass
+class ToolDef:
+ name: str
+ schema: dict # JSON schema for parameters
+ func: Callable # (params: dict, config: dict) -> str
+ read_only: bool # True = auto-approve in 'auto' mode
+ concurrent_safe: bool # True = safe for parallel sub-agent use
+
+_TOOLS: dict[str, ToolDef] = {}
+
+def register_tool(tool_def: ToolDef) -> None
+def get_tool(name: str) -> ToolDef | None
+def get_all_tools() -> list[ToolDef]
+def get_tool_schemas() -> list[dict]
+def execute_tool(name: str, params: dict, config: dict) -> str
+```
+
+### 4.2 Tool Output Truncation
+
+Prevent oversized tool outputs (e.g., `cat` large file, `ls -R`) from blowing up context
+before compaction even gets a chance to run. Applied at the `execute_tool` boundary:
+
+```python
+MAX_TOOL_OUTPUT = 32_000 # ~8K tokens, configurable per tool
+
+def execute_tool(name, params, config):
+ tool = get_tool(name)
+ result = tool.func(params, config)
+
+ # Immediate truncation at source
+ if len(result) > MAX_TOOL_OUTPUT:
+ head = result[:MAX_TOOL_OUTPUT // 2]
+ tail = result[-MAX_TOOL_OUTPUT // 4:]
+ snipped = len(result) - len(head) - len(tail)
+ result = f"{head}\n\n[... {snipped} chars truncated ...]\n\n{tail}"
+
+ return result
+```
+
+Additionally, `Bash` tool caps `subprocess` stdout reads to prevent unbounded
+output (e.g., `cat /dev/urandom`).
+
+This creates a two-layer defense:
+- **Layer 0 (here):** hard truncation at tool execution time โ prevents oversized messages
+- **Layer 2 (compaction.py snip):** soft truncation of old tool results โ reclaims context space
+
+### 4.3 Built-in Tools Refactor
+
+Existing tools.py implementations unchanged. Wrap each with `register_tool()` at module load:
+
+```python
+register_tool(ToolDef(
+ name="Read", schema=READ_SCHEMA, func=_read_file,
+ read_only=True, concurrent_safe=True
+))
+```
+
+### 4.3 Permission Logic (unified)
+
+```python
+# agent.py
+def _check_permission(tool_name, params, config):
+ tool = get_tool(tool_name)
+ if config["permission_mode"] == "accept-all":
+ return True
+ if tool.read_only:
+ return True
+ if tool_name == "Bash" and _is_safe_command(params["command"]):
+ return True
+ return None # ask user
+```
+
+---
+
+## 5. Sub-Agent (`subagent.py`)
+
+### 5.1 Data Model
+
+```python
+@dataclass
+class SubAgentTask:
+ id: str
+ prompt: str
+ status: str # "pending" | "running" | "completed" | "failed" | "cancelled"
+ messages: list[dict] # independent message history
+ result: str | None
+ model: str | None # optional model override
+ depth: int = 0 # recursion depth counter
+ _cancel_flag: bool = False
+ _future: Future | None = None
+
+@dataclass
+class SubAgentManager:
+ tasks: dict[str, SubAgentTask] = field(default_factory=dict)
+ max_concurrent: int = 3
+ max_depth: int = 3
+ _pool: ThreadPoolExecutor = field(default_factory=
+ lambda: ThreadPoolExecutor(max_workers=3))
+
+ def spawn(self, prompt, config, system_prompt, depth=0) -> SubAgentTask
+ def get_result(self, task_id) -> str | None
+ def list_tasks(self) -> list[SubAgentTask]
+ def cancel(self, task_id) -> bool
+ def wait(self, task_id, timeout=None) -> SubAgentTask
+```
+
+### 5.2 Execution Model โ Threading from Day 1
+
+Sub-agents run in background threads via `ThreadPoolExecutor`. This enables:
+- Non-blocking spawn (main agent continues or waits by choice)
+- Cancellation via cooperative flag
+- Concurrent sub-agents (up to `max_concurrent`)
+
+```python
+def spawn(self, prompt, config, system_prompt, depth=0):
+ if depth >= self.max_depth:
+ return SubAgentTask(status="failed",
+ result="Error: max sub-agent depth reached.")
+
+ task = SubAgentTask(id=uuid4().hex[:8], prompt=prompt,
+ status="running", depth=depth, ...)
+
+ def _run():
+ sub_state = AgentState()
+ try:
+ for event in agent.run(
+ prompt, sub_state, config, system_prompt,
+ depth=depth + 1,
+ cancel_check=lambda: task._cancel_flag
+ ):
+ if isinstance(event, TurnDone):
+ task.result = extract_final_text(sub_state.messages)
+ task.status = "completed"
+ except Exception as e:
+ task.result = f"Error: {e}"
+ task.status = "failed"
+
+ task._future = self._pool.submit(_run)
+ self.tasks[task.id] = task
+ return task
+```
+
+### 5.3 Cooperative Cancellation
+
+Python threads cannot be killed safely. Instead, `agent.run()` checks a
+`cancel_check` callable each loop iteration:
+
+```python
+# agent.py run() โ new parameter
+def run(user_message, state, config, system_prompt,
+ depth=0, cancel_check=None):
+ ...
+ while True:
+ if cancel_check and cancel_check():
+ return # clean exit
+ for event in stream(...):
+ yield event
+ ...
+```
+
+### 5.4 Depth Limiting (No Tool Removal)
+
+Sub-agents CAN call Agent tool (enabling A -> B -> C chains). Depth is
+passed through, and the Agent tool returns an error at `max_depth`:
+
+```python
+def _agent_tool_func(params, config, depth=0):
+ if depth >= manager.max_depth:
+ return ("Error: max sub-agent depth reached. "
+ "Complete this task directly without spawning sub-agents.")
+ return manager.spawn(params["prompt"], config, system_prompt, depth)
+```
+
+The model sees the error and adapts โ no silent capability removal.
+
+### 5.5 Context Strategy
+
+Sub-agent gets **fresh context** (no parent message history):
+
+```python
+sub_system_prompt = f"""You are a sub-agent. Your task:
+{prompt}
+
+Working directory: {cwd}
+{memory_context}
+"""
+```
+
+### 5.6 Tool Registration โ 3 Tools
+
+The sub-agent system registers three tools:
+
+**Agent** โ spawn a sub-agent:
+
+```python
+AGENT_SCHEMA = {
+ "name": "Agent",
+ "description": "Launch a sub-agent to handle a task independently.",
+ "input_schema": {
+ "type": "object",
+ "properties": {
+ "prompt": {"type": "string", "description": "Task description"},
+ "model": {"type": "string", "description": "Optional model override"},
+ "wait": {"type": "boolean", "default": True,
+ "description": "True = block until done (default). "
+ "False = return task_id immediately."}
+ },
+ "required": ["prompt"]
+ }
+}
+```
+
+- `wait=True` (default): spawn + block + return result. Feels synchronous to model.
+- `wait=False`: spawn + return task_id immediately. Model must use CheckAgentResult later.
+
+**CheckAgentResult** โ poll a background sub-agent:
+
+```python
+CHECK_AGENT_RESULT_SCHEMA = {
+ "name": "CheckAgentResult",
+ "description": "Check the result of a background sub-agent task.",
+ "input_schema": {
+ "type": "object",
+ "properties": {
+ "task_id": {"type": "string", "description": "Task ID from Agent tool"}
+ },
+ "required": ["task_id"]
+ }
+}
+```
+
+Returns: status + result (if completed), or status + "still running".
+
+**ListAgentTasks** โ overview of all sub-agents:
+
+```python
+LIST_AGENT_TASKS_SCHEMA = {
+ "name": "ListAgentTasks",
+ "description": "List all sub-agent tasks and their status.",
+ "input_schema": {"type": "object", "properties": {}}
+}
+```
+
+Returns a table of `[id, status, prompt_preview]` for all tasks.
+
+---
+
+## 6. Memory System (`memory.py`)
+
+### 6.1 Storage
+
+```
+~/.nano_claude/memory/
+โโโ MEMORY.md # Index file (max 200 lines)
+โโโ user_role.md # Individual memory files
+โโโ feedback_testing.md
+โโโ ...
+```
+
+Memory file format:
+
+```markdown
+---
+name: user role
+description: user is a data scientist focused on logging
+type: user
+created: 2026-04-02
+---
+
+User is a data scientist, currently investigating observability/logging.
+```
+
+### 6.2 Public API
+
+```python
+@dataclass
+class MemoryEntry:
+ name: str
+ description: str
+ type: str # "user" | "feedback" | "project" | "reference"
+ content: str
+ file_path: str
+ created: str
+
+def load_index() -> list[MemoryEntry]
+def save_memory(entry: MemoryEntry) -> None
+def delete_memory(name: str) -> None
+def search_memory(query: str) -> list[MemoryEntry]
+def get_memory_context() -> str # for system prompt injection
+```
+
+### 6.3 Tool Registration
+
+Two tools for model-driven memory management:
+
+- **MemorySave**: `{name, type, description, content}` -> write file + update index
+- **MemoryDelete**: `{name}` -> remove file + update index
+
+### 6.4 Context Integration
+
+`context.py:build_system_prompt()` appends `memory.get_memory_context()` (the MEMORY.md index). Model uses Read tool to access full memory file content when needed.
+
+---
+
+## 7. Skills System (`skills.py`)
+
+### 7.1 Skill Definition
+
+Markdown files with frontmatter:
+
+```
+~/.nano_claude/skills/commit.md
+```
+
+```markdown
+---
+name: commit
+description: Create a git commit with conventional format
+triggers: ["/commit", "commit changes"]
+tools: [Bash, Read]
+---
+
+# Commit Skill
+
+Analyze staged changes and create a well-formatted commit message.
+...
+```
+
+### 7.2 Search Path
+
+```python
+SKILL_PATHS = [
+ Path.cwd() / ".nano_claude" / "skills", # project-level (priority)
+ Path.home() / ".nano_claude" / "skills", # user-level
+]
+```
+
+### 7.3 Public API
+
+```python
+@dataclass
+class SkillDef:
+ name: str
+ description: str
+ triggers: list[str]
+ tools: list[str]
+ prompt: str
+ file_path: str
+
+def load_skills() -> list[SkillDef]
+def find_skill(query: str) -> SkillDef | None
+def execute_skill(skill, args, state, config) -> Generator
+```
+
+### 7.4 Execution Model
+
+Skills are just prompts injected into the normal agent loop:
+
+```python
+def execute_skill(skill, args, state, config):
+ prompt = f"[Skill: {skill.name}]\n\n{skill.prompt}"
+ if args:
+ prompt += f"\n\nUser context: {args}"
+ system_prompt = build_system_prompt(config)
+ for event in agent.run(prompt, state, config, system_prompt):
+ yield event
+```
+
+### 7.5 REPL Integration
+
+In `nano_claude.py`, unmatched `/` commands fall through to skill lookup:
+
+```python
+if user_input.startswith("/"):
+ # Try built-in slash commands first
+ # If no match -> find_skill(user_input)
+ # If skill found -> execute_skill(...)
+```
+
+---
+
+## 8. Diff View for File Modifications
+
+Core UX improvement: show git-style red/green diff when Edit or Write modifies an existing file.
+
+### 8.1 Diff Generation (in tools.py)
+
+Edit and Write tool implementations capture before/after content and generate unified diff:
+
+```python
+import difflib
+
+def generate_unified_diff(old, new, filename, context_lines=3):
+ """
+ Args:
+ old: original file content, str
+ new: modified file content, str
+ filename: display name, str
+ context_lines: lines of context around changes, int
+ Returns:
+ unified diff string
+ """
+ old_lines = old.splitlines(keepends=True)
+ new_lines = new.splitlines(keepends=True)
+ diff = difflib.unified_diff(
+ old_lines, new_lines,
+ fromfile=f"a/{filename}", tofile=f"b/{filename}",
+ n=context_lines
+ )
+ return "".join(diff)
+```
+
+Tool return values change:
+- **Edit**: `"Changes applied to {filename}:\n\n{diff}"`
+- **Write** (existing file): `"File updated:\n\n{diff}"`
+- **Write** (new file): `"New file created: {filename} ({n} lines)"` (no diff)
+
+### 8.2 REPL Rendering (in nano_claude.py)
+
+Detect diff blocks in tool output and render with ANSI colors:
+
+```python
+def render_diff(diff_text):
+ for line in diff_text.splitlines():
+ if line.startswith("+++") or line.startswith("---"):
+ print(f"\033[1m{line}\033[0m") # bold
+ elif line.startswith("+"):
+ print(f"\033[32m{line}\033[0m") # green
+ elif line.startswith("-"):
+ print(f"\033[31m{line}\033[0m") # red
+ elif line.startswith("@@"):
+ print(f"\033[36m{line}\033[0m") # cyan
+ else:
+ print(line)
+```
+
+### 8.3 Diff Truncation
+
+For large diffs (e.g., Write replaces entire file), cap the diff display:
+
+```python
+MAX_DIFF_LINES = 80
+
+def maybe_truncate_diff(diff_text):
+ lines = diff_text.splitlines()
+ if len(lines) > MAX_DIFF_LINES:
+ shown = lines[:MAX_DIFF_LINES]
+ remaining = len(lines) - MAX_DIFF_LINES
+ return "\n".join(shown) + f"\n\n[... {remaining} more lines ...]"
+ return diff_text
+```
+
+Note: truncation applies to the **display** in REPL only. The full diff is still
+returned to the model so it can verify the change.
+
+---
+
+## 9. Implementation Order
+
+Each step is an independent PR:
+
+| Phase | Module | Depends On | Estimated Lines |
+|-------|--------|-----------|-----------------|
+| 1 | `tool_registry.py` + `tools.py` refactor | None | ~600 |
+| 2 | Diff view in `tools.py` + `nano_claude.py` | Phase 1 | ~100 |
+| 3 | `compaction.py` + agent.py integration | Phase 1 | ~300 |
+| 4 | `memory.py` + context.py integration | Phase 1 | ~200 |
+| 5 | `subagent.py` + agent.py integration (threading) | Phase 1 | ~350 |
+| 6 | `skills.py` + nano_claude.py integration | Phase 1, 4 | ~200 |
+| 7 | Slash commands + config updates | All above | ~300 |
+
+**Total new code: ~2050 lines. Grand total: ~4.2K lines.**
+
+---
+
+## 10. Key Decisions
+
+| Decision | Choice | Rationale |
+|----------|--------|-----------|
+| Compression layers | 2 (autoCompact + snip) | Layer 3 is experimental in Claude Code |
+| Tool output truncation | Hard cap at execute_tool boundary | Prevents oversized outputs before compaction runs |
+| Sub-agent execution | Threading from day 1 | Sync blocks main agent, can't cancel, can't parallelize |
+| Sub-agent depth | Depth counter (max 3), no tool removal | Model sees error and adapts; sub-sub-agents allowed |
+| Sub-agent tools | Agent + CheckAgentResult + ListAgentTasks | Model needs feedback loop for async tasks |
+| Diff view | difflib unified diff + ANSI colors | Core UX, zero dependencies |
+| Memory search | Keyword match, no embeddings | Keep simple, model judges relevance |
+| Skills format | Markdown + frontmatter | Human-readable, git-friendly, no Python needed |
+| Tool registry | Global dict + register function | Simple, extensible, easy to migrate to package |
+| Target models | GPT-5.4, Gemini 3/3.1 Pro | User's primary use case |
+| No Claude support | Intentional | Official Claude Code exists |
+
+---
+
+## 11. Future Considerations (Not in Scope)
+
+- MCP protocol support
+- Remote skill marketplace
+- Voice mode
+- Bridge to desktop apps
+- contextCollapse (Layer 3 compression)
diff --git a/nano-claude-code/memory.py b/nano-claude-code/memory.py
new file mode 100644
index 0000000..dd25231
--- /dev/null
+++ b/nano-claude-code/memory.py
@@ -0,0 +1,11 @@
+"""Backward-compatibility shim โ real implementation is in memory/ package."""
+from memory.store import ( # noqa: F401
+ MemoryEntry,
+ save_memory,
+ delete_memory,
+ load_index,
+ search_memory,
+ get_index_content,
+ parse_frontmatter,
+)
+from memory.context import get_memory_context # noqa: F401
diff --git a/nano-claude-code/memory/__init__.py b/nano-claude-code/memory/__init__.py
new file mode 100644
index 0000000..f9618b4
--- /dev/null
+++ b/nano-claude-code/memory/__init__.py
@@ -0,0 +1,86 @@
+"""Memory package for nano-claude-code.
+
+Provides persistent, file-based memory across conversations.
+
+Storage layout:
+ user scope : ~/.nano_claude/memory/.md (shared across projects)
+ project scope : .nano_claude/memory/.md (local to cwd)
+
+The MEMORY.md index in each directory is auto-maintained and injected
+into the system prompt so Claude has an overview of available memories.
+
+Public API (backward-compatible with the old memory.py module):
+ MemoryEntry โ dataclass for a single memory
+ save_memory() โ write/update a memory file
+ delete_memory() โ remove a memory file
+ load_index() โ load all entries from one or both scopes
+ search_memory() โ keyword search across entries
+ get_memory_context() โ MEMORY.md content for system prompt injection
+"""
+from .store import ( # noqa: F401
+ MemoryEntry,
+ save_memory,
+ delete_memory,
+ load_index,
+ load_entries,
+ search_memory,
+ get_index_content,
+ parse_frontmatter,
+ USER_MEMORY_DIR,
+ INDEX_FILENAME,
+ MAX_INDEX_LINES,
+ MAX_INDEX_BYTES,
+)
+from .scan import ( # noqa: F401
+ MemoryHeader,
+ scan_memory_dir,
+ scan_all_memories,
+ format_memory_manifest,
+ memory_age_days,
+ memory_age_str,
+ memory_freshness_text,
+)
+from .context import ( # noqa: F401
+ get_memory_context,
+ find_relevant_memories,
+ truncate_index_content,
+)
+from .types import ( # noqa: F401
+ MEMORY_TYPES,
+ MEMORY_TYPE_DESCRIPTIONS,
+ MEMORY_SYSTEM_PROMPT,
+ WHAT_NOT_TO_SAVE,
+)
+
+__all__ = [
+ # store
+ "MemoryEntry",
+ "save_memory",
+ "delete_memory",
+ "load_index",
+ "load_entries",
+ "search_memory",
+ "get_index_content",
+ "parse_frontmatter",
+ "USER_MEMORY_DIR",
+ "INDEX_FILENAME",
+ "MAX_INDEX_LINES",
+ "MAX_INDEX_BYTES",
+ # scan
+ "MemoryHeader",
+ "scan_memory_dir",
+ "scan_all_memories",
+ "format_memory_manifest",
+ "memory_age_days",
+ "memory_age_str",
+ "memory_freshness_text",
+ # context
+ "get_memory_context",
+ "find_relevant_memories",
+ "truncate_index_content",
+ # types
+ "MEMORY_TYPES",
+ "MEMORY_TYPE_DESCRIPTIONS",
+ "MEMORY_SYSTEM_PROMPT",
+ "WHAT_NOT_TO_SAVE",
+]
diff --git a/nano-claude-code/memory/context.py b/nano-claude-code/memory/context.py
new file mode 100644
index 0000000..2e14108
--- /dev/null
+++ b/nano-claude-code/memory/context.py
@@ -0,0 +1,221 @@
+"""Memory context building for system prompt injection.
+
+Provides:
+ get_memory_context() โ full context string for system prompt
+ find_relevant_memories() โ keyword (+ optional AI) relevance filtering
+ truncate_index_content() โ line + byte truncation with warning
+"""
+from __future__ import annotations
+
+from pathlib import Path
+
+from .store import (
+ USER_MEMORY_DIR,
+ INDEX_FILENAME,
+ MAX_INDEX_LINES,
+ MAX_INDEX_BYTES,
+ get_memory_dir,
+ get_index_content,
+ load_entries,
+ search_memory,
+)
+from .scan import scan_all_memories, format_memory_manifest, memory_freshness_text
+from .types import MEMORY_SYSTEM_PROMPT
+
+
+# โโ Index truncation โโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโ
+
+def truncate_index_content(raw: str) -> str:
+ """Truncate MEMORY.md content to line AND byte limits, appending a warning.
+
+ Matches Claude Code's truncateEntrypointContent:
+ - Line-truncates first (natural boundary)
+ - Then byte-truncates at the last newline before the cap
+ - Appends which limit fired
+ """
+ trimmed = raw.strip()
+ content_lines = trimmed.split("\n")
+ line_count = len(content_lines)
+ byte_count = len(trimmed.encode())
+
+ was_line_truncated = line_count > MAX_INDEX_LINES
+ was_byte_truncated = byte_count > MAX_INDEX_BYTES
+
+ if not was_line_truncated and not was_byte_truncated:
+ return trimmed
+
+ truncated = "\n".join(content_lines[:MAX_INDEX_LINES]) if was_line_truncated else trimmed
+
+ if len(truncated.encode()) > MAX_INDEX_BYTES:
+ # Cut at last newline before byte limit
+ raw_bytes = truncated.encode()
+ cut = raw_bytes[:MAX_INDEX_BYTES].rfind(b"\n")
+ truncated = raw_bytes[: cut if cut > 0 else MAX_INDEX_BYTES].decode(errors="replace")
+
+ if was_byte_truncated and not was_line_truncated:
+ reason = f"{byte_count:,} bytes (limit: {MAX_INDEX_BYTES:,}) โ index entries are too long"
+ elif was_line_truncated and not was_byte_truncated:
+ reason = f"{line_count} lines (limit: {MAX_INDEX_LINES})"
+ else:
+ reason = f"{line_count} lines and {byte_count:,} bytes"
+
+ warning = (
+ f"\n\n> WARNING: {INDEX_FILENAME} is {reason}. "
+ "Only part of it was loaded. Keep index entries to one line under ~150 chars."
+ )
+ return truncated + warning
+
+
+# โโ System prompt context โโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโ
+
+def get_memory_context(include_guidance: bool = False) -> str:
+ """Return memory context for injection into the system prompt.
+
+ Combines user-level and project-level MEMORY.md content (if present).
+ Returns empty string when no memories exist.
+
+ Args:
+ include_guidance: if True, prepend the full memory system guidance
+ (MEMORY_SYSTEM_PROMPT). Normally False since the
+ system prompt template already includes brief guidance.
+ """
+ parts: list[str] = []
+
+ # User-level index
+ user_content = get_index_content("user")
+ if user_content:
+ truncated = truncate_index_content(user_content)
+ parts.append(truncated)
+
+ # Project-level index (labelled separately)
+ proj_content = get_index_content("project")
+ if proj_content:
+ truncated = truncate_index_content(proj_content)
+ parts.append(f"[Project memories]\n{truncated}")
+
+ if not parts:
+ return ""
+
+ body = "\n\n".join(parts)
+ if include_guidance:
+ return f"{MEMORY_SYSTEM_PROMPT}\n\n## MEMORY.md\n{body}"
+ return body
+
+
+# โโ Relevant memory finder โโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโ
+
+def find_relevant_memories(
+ query: str,
+ max_results: int = 5,
+ use_ai: bool = False,
+ config: dict | None = None,
+) -> list[dict]:
+ """Find memories relevant to a query.
+
+ Strategy:
+ 1. Always: keyword match on name + description + content
+ 2. If use_ai=True and config has a model: use a small AI call to rank
+
+ Returns:
+ List of dicts with keys: name, description, type, scope, content,
+ file_path, mtime_s, freshness_text
+ """
+ # Step 1: Keyword filter
+ keyword_results = search_memory(query)
+ if not keyword_results:
+ return []
+
+ if not use_ai or not config:
+ # Return top max_results by recency (newest first)
+ from .scan import scan_all_memories
+ headers = scan_all_memories()
+ path_to_mtime = {h.file_path: h.mtime_s for h in headers}
+
+ results = []
+ for entry in keyword_results[:max_results]:
+ mtime_s = path_to_mtime.get(entry.file_path, 0)
+ results.append({
+ "name": entry.name,
+ "description": entry.description,
+ "type": entry.type,
+ "scope": entry.scope,
+ "content": entry.content,
+ "file_path": entry.file_path,
+ "mtime_s": mtime_s,
+ "freshness_text": memory_freshness_text(mtime_s),
+ })
+ results.sort(key=lambda r: r["mtime_s"], reverse=True)
+ return results[:max_results]
+
+ # Step 2: AI-powered relevance selection (optional, lightweight)
+ return _ai_select_memories(query, keyword_results, max_results, config)
+
+
+def _ai_select_memories(
+ query: str,
+ candidates: list,
+ max_results: int,
+ config: dict,
+) -> list[dict]:
+ """Use a fast AI call to select the most relevant memories from candidates.
+
+ Falls back to keyword results on any error.
+ """
+ try:
+ from providers import stream, AssistantTurn
+ from .scan import scan_all_memories
+
+ headers = scan_all_memories()
+ path_to_mtime = {h.file_path: h.mtime_s for h in headers}
+
+ # Build manifest of candidates only
+ manifest_lines = []
+ for i, e in enumerate(candidates):
+ manifest_lines.append(f"{i}: [{e.type}] {e.name} โ {e.description}")
+ manifest = "\n".join(manifest_lines)
+
+ system = (
+ "You select memories relevant to a query. "
+ "Return a JSON object with key 'indices' containing a list of integer indices "
+ f"(0-based) from the provided list. Select at most {max_results} entries. "
+ "Only include indices clearly relevant to the query. Return {\"indices\": []} if none."
+ )
+ messages = [{"role": "user", "content": f"Query: {query}\n\nMemories:\n{manifest}"}]
+
+ result_text = ""
+ for event in stream(
+ model=config.get("model", "claude-haiku-4-5-20251001"),
+ system=system,
+ messages=messages,
+ tool_schemas=[],
+ config={**config, "max_tokens": 256, "no_tools": True},
+ ):
+ if isinstance(event, AssistantTurn):
+ result_text = event.text
+ break
+
+ import json as _json
+ parsed = _json.loads(result_text)
+ selected_indices = [int(i) for i in parsed.get("indices", []) if isinstance(i, int)]
+
+ except Exception:
+ # Fall back to keyword results
+ selected_indices = list(range(min(max_results, len(candidates))))
+
+ results = []
+ for i in selected_indices[:max_results]:
+ if i < 0 or i >= len(candidates):
+ continue
+ entry = candidates[i]
+ mtime_s = path_to_mtime.get(entry.file_path, 0) if "path_to_mtime" in dir() else 0
+ results.append({
+ "name": entry.name,
+ "description": entry.description,
+ "type": entry.type,
+ "scope": entry.scope,
+ "content": entry.content,
+ "file_path": entry.file_path,
+ "mtime_s": mtime_s,
+ "freshness_text": memory_freshness_text(mtime_s),
+ })
+ return results
diff --git a/nano-claude-code/memory/scan.py b/nano-claude-code/memory/scan.py
new file mode 100644
index 0000000..b982770
--- /dev/null
+++ b/nano-claude-code/memory/scan.py
@@ -0,0 +1,144 @@
+"""Memory file scanning with mtime tracking and freshness/age helpers.
+
+Mirrors the key ideas from Claude Code's memoryScan.ts and memoryAge.ts:
+ - Scan memory directories, sort newest-first
+ - Format a manifest for display or AI relevance selection
+ - Report memory age in human-readable form ("today", "3 days ago")
+ - Emit a staleness caveat for memories older than 1 day
+"""
+from __future__ import annotations
+
+import math
+import time
+from dataclasses import dataclass
+from pathlib import Path
+
+from .store import get_memory_dir, parse_frontmatter, INDEX_FILENAME
+
+MAX_MEMORY_FILES = 200
+
+
+# โโ Data model โโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโ
+
+@dataclass
+class MemoryHeader:
+ """Lightweight descriptor loaded from a memory file's frontmatter.
+
+ Attributes:
+ filename: basename of the .md file
+ file_path: absolute path
+ mtime_s: modification time (seconds since epoch)
+ description: value from frontmatter `description:` field
+ type: value from frontmatter `type:` field
+ scope: "user" or "project"
+ """
+ filename: str
+ file_path: str
+ mtime_s: float
+ description: str
+ type: str
+ scope: str
+
+
+# โโ Scanning โโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโ
+
+def scan_memory_dir(mem_dir: Path, scope: str) -> list[MemoryHeader]:
+ """Scan a single memory directory and return headers sorted newest-first.
+
+ Reads only the frontmatter (first ~30 lines) for efficiency.
+ Silently skips unreadable files. Caps at MAX_MEMORY_FILES entries.
+ """
+ if not mem_dir.is_dir():
+ return []
+
+ headers: list[MemoryHeader] = []
+ for fp in mem_dir.glob("*.md"):
+ if fp.name == INDEX_FILENAME:
+ continue
+ try:
+ stat = fp.stat()
+ # Read only the first 30 lines for frontmatter
+ lines = fp.read_text(errors="replace").splitlines()[:30]
+ snippet = "\n".join(lines)
+ meta, _ = parse_frontmatter(snippet)
+ headers.append(MemoryHeader(
+ filename=fp.name,
+ file_path=str(fp),
+ mtime_s=stat.st_mtime,
+ description=meta.get("description", ""),
+ type=meta.get("type", ""),
+ scope=scope,
+ ))
+ except Exception:
+ continue
+
+ headers.sort(key=lambda h: h.mtime_s, reverse=True)
+ return headers[:MAX_MEMORY_FILES]
+
+
+def scan_all_memories() -> list[MemoryHeader]:
+ """Scan both user and project memory directories, merged newest-first."""
+ user_dir = get_memory_dir("user")
+ proj_dir = get_memory_dir("project")
+
+ user_headers = scan_memory_dir(user_dir, "user")
+ proj_headers = scan_memory_dir(proj_dir, "project")
+
+ combined = user_headers + proj_headers
+ combined.sort(key=lambda h: h.mtime_s, reverse=True)
+ return combined[:MAX_MEMORY_FILES]
+
+
+# โโ Age / freshness โโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโ
+
+def memory_age_days(mtime_s: float) -> int:
+ """Days since mtime_s (floor-rounded, clamped to 0 for future times)."""
+ return max(0, math.floor((time.time() - mtime_s) / 86_400))
+
+
+def memory_age_str(mtime_s: float) -> str:
+ """Human-readable age: 'today', 'yesterday', or 'N days ago'."""
+ d = memory_age_days(mtime_s)
+ if d == 0:
+ return "today"
+ if d == 1:
+ return "yesterday"
+ return f"{d} days ago"
+
+
+def memory_freshness_text(mtime_s: float) -> str:
+ """Staleness caveat for memories older than 1 day (empty string if fresh).
+
+ Motivated by user reports of stale code-state memories (file:line
+ citations to code that has since changed) being asserted as fact.
+ """
+ d = memory_age_days(mtime_s)
+ if d <= 1:
+ return ""
+ return (
+ f"This memory is {d} days old. "
+ "Memories are point-in-time observations, not live state โ "
+ "claims about code behavior or file:line citations may be outdated. "
+ "Verify against current code before asserting as fact."
+ )
+
+
+# โโ Manifest formatting โโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโ
+
+def format_memory_manifest(headers: list[MemoryHeader]) -> str:
+ """Format a list of MemoryHeader as a text manifest.
+
+ Format per line: [type/scope] filename (age): description
+ Example:
+ [feedback/user] feedback_testing.md (3 days ago): Don't mock DB in tests
+ [project/project] project_freeze.md (today): Merge freeze until 2026-04-10
+ """
+ lines = []
+ for h in headers:
+ tag = f"[{h.type}/{h.scope}]" if h.type else f"[{h.scope}]"
+ age = memory_age_str(h.mtime_s)
+ if h.description:
+ lines.append(f"- {tag} {h.filename} ({age}): {h.description}")
+ else:
+ lines.append(f"- {tag} {h.filename} ({age})")
+ return "\n".join(lines)
diff --git a/nano-claude-code/memory/store.py b/nano-claude-code/memory/store.py
new file mode 100644
index 0000000..c40bf6d
--- /dev/null
+++ b/nano-claude-code/memory/store.py
@@ -0,0 +1,223 @@
+"""File-based memory storage with user-level and project-level scopes.
+
+Storage layout:
+ user scope : ~/.nano_claude/memory/.md
+ project scope : .nano_claude/memory/.md (relative to cwd)
+
+MEMORY.md in each directory is the index file โ rebuilt automatically after
+every save/delete. It is loaded into the system prompt to give Claude an
+overview of available memories.
+"""
+from __future__ import annotations
+
+import re
+from dataclasses import dataclass
+from pathlib import Path
+
+
+# โโ Paths โโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโ
+
+USER_MEMORY_DIR = Path.home() / ".nano_claude" / "memory"
+INDEX_FILENAME = "MEMORY.md"
+
+# Maximum lines/bytes for the index file (mirrors Claude Code limits)
+MAX_INDEX_LINES = 200
+MAX_INDEX_BYTES = 25_000
+
+
+def get_project_memory_dir() -> Path:
+ """Return the project-local memory directory (relative to cwd)."""
+ return Path.cwd() / ".nano_claude" / "memory"
+
+
+def get_memory_dir(scope: str = "user") -> Path:
+ """Return the memory directory for the given scope.
+
+ Args:
+ scope: "user" (global ~/.nano_claude/memory) or
+ "project" (.nano_claude/memory relative to cwd)
+ """
+ if scope == "project":
+ return get_project_memory_dir()
+ return USER_MEMORY_DIR
+
+
+# โโ Data model โโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโ
+
+@dataclass
+class MemoryEntry:
+ """A single memory entry loaded from a .md file.
+
+ Attributes:
+ name: human-readable name (also the display title in the index)
+ description: short one-line description (used for relevance decisions)
+ type: "user" | "feedback" | "project" | "reference"
+ content: body text of the memory
+ file_path: absolute path to the .md file on disk
+ created: date string, e.g. "2026-04-02"
+ scope: "user" | "project" โ which directory this was loaded from
+ """
+ name: str
+ description: str
+ type: str
+ content: str
+ file_path: str = ""
+ created: str = ""
+ scope: str = "user"
+
+
+# โโ Helpers โโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโ
+
+def _slugify(name: str) -> str:
+ """Convert name to a filesystem-safe slug (max 60 chars)."""
+ s = name.lower().strip().replace(" ", "_")
+ s = re.sub(r"[^a-z0-9_]", "", s)
+ return s[:60]
+
+
+def parse_frontmatter(text: str) -> tuple[dict, str]:
+ """Parse ---\\nkey: value\\n---\\nbody format.
+
+ Returns:
+ (meta_dict, body_str)
+ """
+ if not text.startswith("---"):
+ return {}, text
+ parts = text.split("---", 2)
+ if len(parts) < 3:
+ return {}, text
+ meta: dict = {}
+ for line in parts[1].strip().splitlines():
+ if ":" in line:
+ key, _, val = line.partition(":")
+ meta[key.strip()] = val.strip()
+ return meta, parts[2].strip()
+
+
+def _format_entry_md(entry: MemoryEntry) -> str:
+ """Render a MemoryEntry as a markdown file with YAML frontmatter."""
+ return (
+ f"---\n"
+ f"name: {entry.name}\n"
+ f"description: {entry.description}\n"
+ f"type: {entry.type}\n"
+ f"created: {entry.created}\n"
+ f"---\n"
+ f"{entry.content}\n"
+ )
+
+
+# โโ Core storage operations โโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโ
+
+def save_memory(entry: MemoryEntry, scope: str = "user") -> None:
+ """Write/update a memory file and rebuild the index for that scope.
+
+ If a memory with the same name (slug) already exists, it is overwritten.
+
+ Args:
+ entry: MemoryEntry to persist
+ scope: "user" or "project"
+ """
+ mem_dir = get_memory_dir(scope)
+ mem_dir.mkdir(parents=True, exist_ok=True)
+ slug = _slugify(entry.name)
+ fp = mem_dir / f"{slug}.md"
+ fp.write_text(_format_entry_md(entry))
+ entry.file_path = str(fp)
+ entry.scope = scope
+ _rewrite_index(scope)
+
+
+def delete_memory(name: str, scope: str = "user") -> None:
+ """Remove the memory file matching name and rebuild the index.
+
+ No error if not found.
+ """
+ mem_dir = get_memory_dir(scope)
+ slug = _slugify(name)
+ fp = mem_dir / f"{slug}.md"
+ if fp.exists():
+ fp.unlink()
+ _rewrite_index(scope)
+
+
+def load_entries(scope: str = "user") -> list[MemoryEntry]:
+ """Scan all .md files (except MEMORY.md) in a scope and return entries.
+
+ Returns:
+ List of MemoryEntry sorted alphabetically by name.
+ """
+ mem_dir = get_memory_dir(scope)
+ if not mem_dir.exists():
+ return []
+ entries: list[MemoryEntry] = []
+ for fp in sorted(mem_dir.glob("*.md")):
+ if fp.name == INDEX_FILENAME:
+ continue
+ try:
+ text = fp.read_text()
+ except Exception:
+ continue
+ meta, body = parse_frontmatter(text)
+ entries.append(MemoryEntry(
+ name=meta.get("name", fp.stem),
+ description=meta.get("description", ""),
+ type=meta.get("type", "user"),
+ content=body,
+ file_path=str(fp),
+ created=meta.get("created", ""),
+ scope=scope,
+ ))
+ return entries
+
+
+def load_index(scope: str = "all") -> list[MemoryEntry]:
+ """Load memory entries from one or both scopes.
+
+ Args:
+ scope: "user", "project", or "all" (both combined)
+
+ Returns:
+ List of MemoryEntry (user entries first, then project).
+ """
+ if scope == "all":
+ return load_entries("user") + load_entries("project")
+ return load_entries(scope)
+
+
+def search_memory(query: str, scope: str = "all") -> list[MemoryEntry]:
+ """Case-insensitive keyword match on name + description + content.
+
+ Returns:
+ List of matching MemoryEntry objects.
+ """
+ q = query.lower()
+ results = []
+ for entry in load_index(scope):
+ haystack = f"{entry.name} {entry.description} {entry.content}".lower()
+ if q in haystack:
+ results.append(entry)
+ return results
+
+
+def _rewrite_index(scope: str) -> None:
+ """Rebuild MEMORY.md for the given scope from all .md files in that dir."""
+ mem_dir = get_memory_dir(scope)
+ if not mem_dir.exists():
+ return
+ index_path = mem_dir / INDEX_FILENAME
+ entries = load_entries(scope)
+ lines = [
+ f"- [{e.name}]({Path(e.file_path).name}) โ {e.description}"
+ for e in entries
+ ]
+ index_path.write_text("\n".join(lines) + ("\n" if lines else ""))
+
+
+def get_index_content(scope: str = "user") -> str:
+ """Return raw MEMORY.md content for the given scope, or '' if absent."""
+ mem_dir = get_memory_dir(scope)
+ index_path = mem_dir / INDEX_FILENAME
+ if not index_path.exists():
+ return ""
+ return index_path.read_text().strip()
diff --git a/nano-claude-code/memory/tools.py b/nano-claude-code/memory/tools.py
new file mode 100644
index 0000000..3121c68
--- /dev/null
+++ b/nano-claude-code/memory/tools.py
@@ -0,0 +1,216 @@
+"""Memory tool registrations: MemorySave, MemoryDelete, MemorySearch.
+
+Importing this module registers the three tools into the central registry.
+"""
+from __future__ import annotations
+
+from datetime import datetime
+
+from tool_registry import ToolDef, register_tool
+from .store import MemoryEntry, save_memory, delete_memory, load_index
+from .context import find_relevant_memories
+from .scan import scan_all_memories, format_memory_manifest
+
+
+# โโ Tool implementations โโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโ
+
+def _memory_save(params: dict, config: dict) -> str:
+ """Save or update a persistent memory entry."""
+ entry = MemoryEntry(
+ name=params["name"],
+ description=params["description"],
+ type=params["type"],
+ content=params["content"],
+ created=datetime.now().strftime("%Y-%m-%d"),
+ )
+ scope = params.get("scope", "user")
+ save_memory(entry, scope=scope)
+
+ scope_label = "project" if scope == "project" else "user"
+ return f"Memory saved: '{entry.name}' [{entry.type}/{scope_label}]"
+
+
+def _memory_delete(params: dict, config: dict) -> str:
+ """Delete a persistent memory entry by name."""
+ name = params["name"]
+ scope = params.get("scope", "user")
+ delete_memory(name, scope=scope)
+ return f"Memory deleted: '{name}' (scope: {scope})"
+
+
+def _memory_search(params: dict, config: dict) -> str:
+ """Search memories by keyword query with optional AI relevance filtering."""
+ query = params["query"]
+ use_ai = params.get("use_ai", False)
+ max_results = params.get("max_results", 5)
+
+ results = find_relevant_memories(
+ query, max_results=max_results, use_ai=use_ai, config=config
+ )
+
+ if not results:
+ return f"No memories found matching '{query}'."
+
+ lines = [f"Found {len(results)} relevant memory/memories for '{query}':", ""]
+ for r in results:
+ freshness = f" โ {r['freshness_text']}" if r["freshness_text"] else ""
+ lines.append(
+ f"[{r['type']}/{r['scope']}] {r['name']}\n"
+ f" {r['description']}\n"
+ f" {r['content'][:200]}{'...' if len(r['content']) > 200 else ''}"
+ f"{freshness}"
+ )
+ return "\n\n".join(lines)
+
+
+def _memory_list(params: dict, config: dict) -> str:
+ """List all memory entries with their manifest (type, scope, age, description)."""
+ headers = scan_all_memories()
+ if not headers:
+ return "No memories stored."
+
+ scope_filter = params.get("scope", "all")
+ if scope_filter != "all":
+ headers = [h for h in headers if h.scope == scope_filter]
+ if not headers:
+ return f"No {scope_filter} memories stored."
+
+ manifest = format_memory_manifest(headers)
+ return f"{len(headers)} memory/memories:\n\n{manifest}"
+
+
+# โโ Tool registrations โโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโ
+
+register_tool(ToolDef(
+ name="MemorySave",
+ schema={
+ "name": "MemorySave",
+ "description": (
+ "Save a persistent memory entry as a markdown file with frontmatter. "
+ "Use for information that should persist across conversations: "
+ "user preferences, feedback/corrections, project context, or external references. "
+ "Do NOT save: code patterns, architecture, git history, or task state.\n\n"
+ "For feedback/project memories, structure content as: "
+ "rule/fact, then **Why:** and **How to apply:** lines."
+ ),
+ "input_schema": {
+ "type": "object",
+ "properties": {
+ "name": {
+ "type": "string",
+ "description": "Human-readable name (becomes the filename slug)",
+ },
+ "type": {
+ "type": "string",
+ "enum": ["user", "feedback", "project", "reference"],
+ "description": (
+ "user=preferences/role, feedback=guidance on how to work, "
+ "project=ongoing work/decisions, reference=external system pointers"
+ ),
+ },
+ "description": {
+ "type": "string",
+ "description": "Short one-line description (used for relevance decisions โ be specific)",
+ },
+ "content": {
+ "type": "string",
+ "description": "Body text. For feedback/project: rule/fact + **Why:** + **How to apply:**",
+ },
+ "scope": {
+ "type": "string",
+ "enum": ["user", "project"],
+ "description": (
+ "'user' (default) = ~/.nano_claude/memory/ shared across projects; "
+ "'project' = .nano_claude/memory/ local to this project"
+ ),
+ },
+ },
+ "required": ["name", "type", "description", "content"],
+ },
+ },
+ func=_memory_save,
+ read_only=False,
+ concurrent_safe=False,
+))
+
+register_tool(ToolDef(
+ name="MemoryDelete",
+ schema={
+ "name": "MemoryDelete",
+ "description": "Delete a persistent memory entry by name.",
+ "input_schema": {
+ "type": "object",
+ "properties": {
+ "name": {"type": "string", "description": "Name of the memory to delete"},
+ "scope": {
+ "type": "string",
+ "enum": ["user", "project"],
+ "description": "Scope to delete from (default: 'user')",
+ },
+ },
+ "required": ["name"],
+ },
+ },
+ func=_memory_delete,
+ read_only=False,
+ concurrent_safe=False,
+))
+
+register_tool(ToolDef(
+ name="MemorySearch",
+ schema={
+ "name": "MemorySearch",
+ "description": (
+ "Search persistent memories by keyword. Returns matching entries with "
+ "content preview and staleness warning for old memories. "
+ "Set use_ai=true to use AI-powered relevance ranking (costs a small API call)."
+ ),
+ "input_schema": {
+ "type": "object",
+ "properties": {
+ "query": {"type": "string", "description": "Search query"},
+ "max_results": {
+ "type": "integer",
+ "description": "Maximum results to return (default: 5)",
+ },
+ "use_ai": {
+ "type": "boolean",
+ "description": "Use AI relevance ranking (default: false = keyword only)",
+ },
+ "scope": {
+ "type": "string",
+ "enum": ["user", "project", "all"],
+ "description": "Which scope to search (default: 'all')",
+ },
+ },
+ "required": ["query"],
+ },
+ },
+ func=_memory_search,
+ read_only=True,
+ concurrent_safe=True,
+))
+
+register_tool(ToolDef(
+ name="MemoryList",
+ schema={
+ "name": "MemoryList",
+ "description": (
+ "List all memory entries with type, scope, age, and description. "
+ "Useful for reviewing what's been remembered before deciding to save or delete."
+ ),
+ "input_schema": {
+ "type": "object",
+ "properties": {
+ "scope": {
+ "type": "string",
+ "enum": ["user", "project", "all"],
+ "description": "Which scope to list (default: 'all')",
+ },
+ },
+ },
+ },
+ func=_memory_list,
+ read_only=True,
+ concurrent_safe=True,
+))
diff --git a/nano-claude-code/memory/types.py b/nano-claude-code/memory/types.py
new file mode 100644
index 0000000..151d06f
--- /dev/null
+++ b/nano-claude-code/memory/types.py
@@ -0,0 +1,86 @@
+"""Memory type taxonomy and system-prompt guidance text.
+
+Four types capture context NOT derivable from the current project state.
+Code patterns, architecture, git history, and file structure are derivable
+(via grep/git/CLAUDE.md) and should NOT be saved as memories.
+"""
+
+MEMORY_TYPES = ["user", "feedback", "project", "reference"]
+
+# Condensed per-type guidance (used in system prompt injection)
+MEMORY_TYPE_DESCRIPTIONS: dict[str, str] = {
+ "user": (
+ "Information about the user's role, goals, responsibilities, and knowledge. "
+ "Helps tailor future behavior to the user's preferences."
+ ),
+ "feedback": (
+ "Guidance the user has given about how to approach work โ both what to avoid "
+ "and what to keep doing. Lead with the rule, then **Why:** and **How to apply:**."
+ ),
+ "project": (
+ "Ongoing work, goals, bugs, or incidents not derivable from code or git history. "
+ "Lead with the fact/decision, then **Why:** and **How to apply:**. "
+ "Always convert relative dates to absolute dates."
+ ),
+ "reference": (
+ "Pointers to external systems (issue trackers, dashboards, Slack channels, docs)."
+ ),
+}
+
+# What NOT to save (mirrors Claude Code source)
+WHAT_NOT_TO_SAVE = """\
+## What NOT to save in memory
+- Code patterns, conventions, architecture, file paths, or project structure โ derivable from the codebase.
+- Git history, recent changes, who-changed-what โ use `git log` / `git blame`.
+- Debugging solutions or fix recipes โ the fix is in the code; the commit message has context.
+- Anything already documented in CLAUDE.md files.
+- Ephemeral task details: in-progress work, temporary state, current conversation context.
+
+These exclusions apply even when explicitly asked. If asked to save a PR list or activity summary,
+ask what was *surprising* or *non-obvious* โ that is the part worth keeping."""
+
+# Memory format example (frontmatter)
+MEMORY_FORMAT_EXAMPLE = """\
+```markdown
+---
+name: {{memory name}}
+description: {{one-line description โ used to decide relevance, so be specific}}
+type: {{user | feedback | project | reference}}
+---
+
+{{memory content โ for feedback/project types: rule/fact, then **Why:** and **How to apply:** lines}}
+```"""
+
+# Full guidance injected into the system prompt
+MEMORY_SYSTEM_PROMPT = """\
+## Memory system
+
+You have a persistent, file-based memory system. Memories are stored as markdown files with
+YAML frontmatter. Build this up over time so future conversations have context about the user,
+their preferences, and the work you're doing together.
+
+**Types** (save only what cannot be derived from the codebase):
+- **user** โ role, goals, knowledge, preferences
+- **feedback** โ guidance on how to work (corrections AND confirmations of non-obvious approaches)
+- **project** โ ongoing work, decisions, deadlines not in git history
+- **reference** โ pointers to external systems (Linear, Grafana, Slack, etc.)
+
+**When to save**: If the user corrects you, confirms an approach, or shares context that should
+persist beyond this conversation. For feedback: save corrections AND quiet confirmations.
+
+**Body structure for feedback/project**: Lead with the rule/fact, then:
+ **Why:** (reason given) | **How to apply:** (when this guidance kicks in)
+
+**Format**:
+{format_example}
+
+**Saving is two steps**:
+1. Write the memory to its own file (e.g. `feedback_testing.md`) using MemorySave.
+2. The index (MEMORY.md) is updated automatically.
+
+**What NOT to save**: code patterns, architecture, git history, debugging fixes,
+anything already in CLAUDE.md, or ephemeral task state.
+
+**Before recommending from memory**: A memory naming a file, function, or flag may be stale.
+Verify it still exists before acting on it. For current state, prefer `git log` or reading code.
+""".format(format_example=MEMORY_FORMAT_EXAMPLE)
diff --git a/nano-claude-code/multi_agent/__init__.py b/nano-claude-code/multi_agent/__init__.py
new file mode 100644
index 0000000..ffcf516
--- /dev/null
+++ b/nano-claude-code/multi_agent/__init__.py
@@ -0,0 +1,23 @@
+"""Multi-agent package for nano-claude-code.
+
+Provides:
+ - AgentDefinition โ typed agent definition (name, system_prompt, model, tools)
+ - SubAgentTask โ lifecycle-tracked task
+ - SubAgentManager โ thread-pool manager for spawning agents
+ - load_agent_definitions / get_agent_definition โ agent registry
+"""
+from .subagent import (
+ AgentDefinition,
+ SubAgentTask,
+ SubAgentManager,
+ load_agent_definitions,
+ get_agent_definition,
+)
+
+__all__ = [
+ "AgentDefinition",
+ "SubAgentTask",
+ "SubAgentManager",
+ "load_agent_definitions",
+ "get_agent_definition",
+]
diff --git a/nano-claude-code/multi_agent/subagent.py b/nano-claude-code/multi_agent/subagent.py
new file mode 100644
index 0000000..73b563a
--- /dev/null
+++ b/nano-claude-code/multi_agent/subagent.py
@@ -0,0 +1,480 @@
+"""Threaded sub-agent system for spawning nested agent loops."""
+from __future__ import annotations
+
+import os
+import uuid
+import queue
+import subprocess
+import tempfile
+from concurrent.futures import ThreadPoolExecutor, Future
+from dataclasses import dataclass, field
+from pathlib import Path
+from typing import Dict, List, Optional, Any
+
+
+# โโ Agent definition โโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโ
+
+@dataclass
+class AgentDefinition:
+ """Definition for a specialized agent type."""
+ name: str
+ description: str = ""
+ system_prompt: str = "" # extra instructions prepended to the base system prompt
+ model: str = "" # model override; "" = inherit from parent
+ tools: list = field(default_factory=list) # empty list = all tools
+ source: str = "user" # "built-in" | "user" | "project"
+
+
+# โโ Built-in agent definitions โโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโ
+
+_BUILTIN_AGENTS: Dict[str, AgentDefinition] = {
+ "general-purpose": AgentDefinition(
+ name="general-purpose",
+ description=(
+ "General-purpose agent for researching complex questions, "
+ "searching for code, and executing multi-step tasks."
+ ),
+ system_prompt="",
+ source="built-in",
+ ),
+ "coder": AgentDefinition(
+ name="coder",
+ description="Specialized coding agent for writing, reading, and modifying code.",
+ system_prompt=(
+ "You are a specialized coding assistant. Focus on:\n"
+ "- Writing clean, idiomatic code\n"
+ "- Reading and understanding existing code before modifying\n"
+ "- Making minimal targeted changes\n"
+ "- Never adding unnecessary features, comments, or error handling\n"
+ ),
+ source="built-in",
+ ),
+ "reviewer": AgentDefinition(
+ name="reviewer",
+ description="Code review agent analyzing quality, security, and correctness.",
+ system_prompt=(
+ "You are a code reviewer. Analyze code for:\n"
+ "- Correctness and logic errors\n"
+ "- Security vulnerabilities (injection, XSS, auth bypass, etc.)\n"
+ "- Performance issues\n"
+ "- Code quality and maintainability\n"
+ "Be concise and specific. Categorize findings as: Critical | Warning | Suggestion.\n"
+ ),
+ tools=["Read", "Glob", "Grep"],
+ source="built-in",
+ ),
+ "researcher": AgentDefinition(
+ name="researcher",
+ description="Research agent for exploring codebases and answering questions.",
+ system_prompt=(
+ "You are a research assistant focused on understanding codebases.\n"
+ "- Read and analyze code thoroughly before answering\n"
+ "- Provide factual, evidence-based answers\n"
+ "- Cite specific file paths and line numbers\n"
+ "- Be concise and focused\n"
+ ),
+ tools=["Read", "Glob", "Grep", "WebFetch", "WebSearch"],
+ source="built-in",
+ ),
+ "tester": AgentDefinition(
+ name="tester",
+ description="Testing agent that writes and runs tests.",
+ system_prompt=(
+ "You are a testing specialist. Your job:\n"
+ "- Write comprehensive tests for the given code\n"
+ "- Run existing tests and diagnose failures\n"
+ "- Focus on edge cases and error conditions\n"
+ "- Keep tests simple, readable, and fast\n"
+ ),
+ source="built-in",
+ ),
+}
+
+
+# โโ Loading agent definitions from .md files โโโโโโโโโโโโโโโโโโโโโโโโโโโโโโ
+
+def _parse_agent_md(path: Path, source: str = "user") -> AgentDefinition:
+ """Parse a .md file with optional YAML frontmatter into an AgentDefinition.
+
+ File format:
+ ---
+ description: "Short description"
+ model: claude-haiku-4-5-20251001
+ tools: [Read, Write, Edit, Bash]
+ ---
+
+ System prompt body goes here...
+ """
+ content = path.read_text()
+ name = path.stem
+ description = ""
+ model = ""
+ tools: list = []
+ system_prompt_body = content
+
+ if content.startswith("---"):
+ end = content.find("---", 3)
+ if end != -1:
+ fm_text = content[3:end].strip()
+ system_prompt_body = content[end + 3:].strip()
+ try:
+ import yaml as _yaml
+ fm = _yaml.safe_load(fm_text) or {}
+ except ImportError:
+ # Manual key: value parse (no yaml dependency required)
+ fm: dict = {}
+ for line in fm_text.splitlines():
+ if ":" in line:
+ k, _, v = line.partition(":")
+ fm[k.strip()] = v.strip()
+ description = str(fm.get("description", ""))
+ model = str(fm.get("model", ""))
+ raw_tools = fm.get("tools", [])
+ if isinstance(raw_tools, list):
+ tools = [str(t) for t in raw_tools]
+ elif isinstance(raw_tools, str):
+ # Handle "[Read, Write]" or "Read, Write" format
+ s = raw_tools.strip("[]")
+ tools = [t.strip() for t in s.split(",") if t.strip()]
+
+ return AgentDefinition(
+ name=name,
+ description=description,
+ system_prompt=system_prompt_body,
+ model=model,
+ tools=tools,
+ source=source,
+ )
+
+
+def load_agent_definitions() -> Dict[str, AgentDefinition]:
+ """Load all agent definitions: built-ins โ user-level โ project-level.
+
+ Search paths:
+ ~/.nano-claude/agents/*.md (user-level)
+ .nano-claude/agents/*.md (project-level, overrides user)
+ """
+ defs: Dict[str, AgentDefinition] = dict(_BUILTIN_AGENTS)
+
+ # User-level
+ user_dir = Path.home() / ".nano-claude" / "agents"
+ if user_dir.is_dir():
+ for p in sorted(user_dir.glob("*.md")):
+ try:
+ d = _parse_agent_md(p, source="user")
+ defs[d.name] = d
+ except Exception:
+ pass
+
+ # Project-level (overrides user)
+ proj_dir = Path.cwd() / ".nano-claude" / "agents"
+ if proj_dir.is_dir():
+ for p in sorted(proj_dir.glob("*.md")):
+ try:
+ d = _parse_agent_md(p, source="project")
+ defs[d.name] = d
+ except Exception:
+ pass
+
+ return defs
+
+
+def get_agent_definition(name: str) -> Optional[AgentDefinition]:
+ """Look up an agent definition by name. Returns None if not found."""
+ return load_agent_definitions().get(name)
+
+
+# โโ SubAgentTask โโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโ
+
+@dataclass
+class SubAgentTask:
+ """Represents a sub-agent task with lifecycle tracking."""
+ id: str
+ prompt: str
+ status: str = "pending" # pending | running | completed | failed | cancelled
+ result: Optional[str] = None
+ depth: int = 0
+ name: str = "" # optional human-readable name (addressable by SendMessage)
+ worktree_path: str = "" # set if isolation="worktree"
+ worktree_branch: str = "" # set if isolation="worktree"
+ _cancel_flag: bool = False
+ _future: Optional[Future] = field(default=None, repr=False)
+ _inbox: Any = field(default_factory=queue.Queue, repr=False) # for send_message
+
+
+# โโ Worktree helpers โโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโ
+
+def _git_root(cwd: str) -> Optional[str]:
+ """Return the git root directory for cwd, or None if not in a git repo."""
+ try:
+ r = subprocess.run(
+ ["git", "rev-parse", "--show-toplevel"],
+ cwd=cwd, capture_output=True, text=True, check=True,
+ )
+ return r.stdout.strip()
+ except Exception:
+ return None
+
+
+def _create_worktree(base_dir: str) -> tuple:
+ """Create a temporary git worktree.
+
+ Returns:
+ (worktree_path, branch_name)
+ Raises:
+ subprocess.CalledProcessError or OSError on failure.
+ """
+ branch = f"nano-agent-{uuid.uuid4().hex[:8]}"
+ # mkdtemp gives us a path; remove the empty dir so git can create it
+ wt_path = tempfile.mkdtemp(prefix="nano-agent-wt-")
+ os.rmdir(wt_path)
+ subprocess.run(
+ ["git", "worktree", "add", "-b", branch, wt_path],
+ cwd=base_dir, check=True, capture_output=True, text=True,
+ )
+ return wt_path, branch
+
+
+def _remove_worktree(wt_path: str, branch: str, base_dir: str) -> None:
+ """Remove a git worktree and delete its branch (best-effort)."""
+ try:
+ subprocess.run(
+ ["git", "worktree", "remove", "--force", wt_path],
+ cwd=base_dir, capture_output=True,
+ )
+ except Exception:
+ pass
+ try:
+ subprocess.run(
+ ["git", "branch", "-D", branch],
+ cwd=base_dir, capture_output=True,
+ )
+ except Exception:
+ pass
+
+
+# โโ Internal helpers โโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโ
+
+def _agent_run(prompt, state, config, system_prompt, depth=0, cancel_check=None):
+ """Lazy-import wrapper to avoid circular dependency with agent module.
+
+ Uses absolute import so this works whether called from inside or outside
+ the multi_agent package (sys.path includes the project root).
+ """
+ import agent as _agent_mod
+ return _agent_mod.run(prompt, state, config, system_prompt, depth=depth, cancel_check=cancel_check)
+
+
+def _extract_final_text(messages):
+ """Walk backwards through messages, return first assistant content string."""
+ for msg in reversed(messages):
+ if msg.get("role") == "assistant" and msg.get("content"):
+ return msg["content"]
+ return None
+
+
+# โโ SubAgentManager โโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโ
+
+class SubAgentManager:
+ """Manages concurrent sub-agent tasks using a thread pool."""
+
+ def __init__(self, max_concurrent: int = 5, max_depth: int = 5):
+ self.tasks: Dict[str, SubAgentTask] = {}
+ self._by_name: Dict[str, str] = {} # name โ task_id
+ self.max_concurrent = max_concurrent
+ self.max_depth = max_depth
+ self._pool = ThreadPoolExecutor(max_workers=max_concurrent)
+
+ def spawn(
+ self,
+ prompt: str,
+ config: dict,
+ system_prompt: str,
+ depth: int = 0,
+ agent_def: Optional[AgentDefinition] = None,
+ isolation: str = "", # "" | "worktree"
+ name: str = "",
+ ) -> SubAgentTask:
+ """Spawn a new sub-agent task.
+
+ Args:
+ prompt: user message for the sub-agent
+ config: agent configuration dict (copied before modification)
+ system_prompt: base system prompt
+ depth: current nesting depth (prevents infinite recursion)
+ agent_def: optional AgentDefinition with model/system_prompt/tools overrides
+ isolation: "" for normal, "worktree" for isolated git worktree
+ name: optional human-readable name (addressable via SendMessage)
+
+ Returns:
+ SubAgentTask tracking the spawned work.
+ """
+ task_id = uuid.uuid4().hex[:12]
+ short_name = name or task_id[:8]
+ task = SubAgentTask(id=task_id, prompt=prompt, depth=depth, name=short_name)
+ self.tasks[task_id] = task
+ if name:
+ self._by_name[name] = task_id
+
+ if depth >= self.max_depth:
+ task.status = "failed"
+ task.result = f"Max depth ({self.max_depth}) exceeded"
+ return task
+
+ # Build effective config and system prompt for this sub-agent
+ eff_config = dict(config)
+ eff_system = system_prompt
+
+ if agent_def:
+ if agent_def.model:
+ eff_config["model"] = agent_def.model
+ if agent_def.system_prompt:
+ eff_system = agent_def.system_prompt.rstrip() + "\n\n" + system_prompt
+
+ # Handle worktree isolation
+ worktree_path = ""
+ worktree_branch = ""
+ base_dir = os.getcwd()
+
+ if isolation == "worktree":
+ git_root = _git_root(base_dir)
+ if not git_root:
+ task.status = "failed"
+ task.result = "isolation='worktree' requires a git repository"
+ return task
+ try:
+ worktree_path, worktree_branch = _create_worktree(git_root)
+ task.worktree_path = worktree_path
+ task.worktree_branch = worktree_branch
+ notice = (
+ f"\n\n[Note: You are working in an isolated git worktree at "
+ f"{worktree_path} (branch: {worktree_branch}). "
+ f"Your changes are isolated from the main workspace at {git_root}. "
+ f"Commit your changes before finishing so they can be reviewed/merged.]"
+ )
+ prompt = prompt + notice
+ except Exception as e:
+ task.status = "failed"
+ task.result = f"Failed to create worktree: {e}"
+ return task
+
+ def _run():
+ import agent as _agent_mod; AgentState = _agent_mod.AgentState
+ task.status = "running"
+ old_cwd = os.getcwd()
+ try:
+ if worktree_path:
+ os.chdir(worktree_path)
+
+ state = AgentState()
+ gen = _agent_run(
+ prompt, state, eff_config, eff_system,
+ depth=depth + 1,
+ cancel_check=lambda: task._cancel_flag,
+ )
+ for _event in gen:
+ if task._cancel_flag:
+ break
+
+ if task._cancel_flag:
+ task.status = "cancelled"
+ task.result = None
+ else:
+ task.result = _extract_final_text(state.messages)
+ task.status = "completed"
+
+ # Drain inbox: process any messages sent via SendMessage
+ while not task._inbox.empty() and not task._cancel_flag:
+ inbox_msg = task._inbox.get_nowait()
+ task.status = "running"
+ gen2 = _agent_run(
+ inbox_msg, state, eff_config, eff_system,
+ depth=depth + 1,
+ cancel_check=lambda: task._cancel_flag,
+ )
+ for _ev in gen2:
+ if task._cancel_flag:
+ break
+ if not task._cancel_flag:
+ task.result = _extract_final_text(state.messages)
+ task.status = "completed"
+
+ except Exception as e:
+ task.status = "failed"
+ task.result = f"Error: {e}"
+ finally:
+ if worktree_path:
+ os.chdir(old_cwd)
+ _remove_worktree(worktree_path, worktree_branch, old_cwd)
+
+ task._future = self._pool.submit(_run)
+ return task
+
+ def wait(self, task_id: str, timeout: float = None) -> Optional[SubAgentTask]:
+ """Block until a task completes or timeout expires.
+
+ Returns:
+ The task, or None if task_id is unknown.
+ """
+ task = self.tasks.get(task_id)
+ if task is None:
+ return None
+ if task._future is not None:
+ try:
+ task._future.result(timeout=timeout)
+ except Exception:
+ pass
+ return task
+
+ def get_result(self, task_id: str) -> Optional[str]:
+ """Return the result string for a completed task, or None."""
+ task = self.tasks.get(task_id)
+ return task.result if task else None
+
+ def list_tasks(self) -> List[SubAgentTask]:
+ """Return all tracked tasks."""
+ return list(self.tasks.values())
+
+ def send_message(self, task_id_or_name: str, message: str) -> bool:
+ """Send a message to a running background agent.
+
+ The message is queued and the agent will process it after completing
+ its current work.
+
+ Args:
+ task_id_or_name: task ID or the human-readable name passed to spawn()
+ message: message text to send
+
+ Returns:
+ True if the message was queued, False if task not found or already done.
+ """
+ # Resolve name โ task_id
+ task_id = self._by_name.get(task_id_or_name, task_id_or_name)
+ task = self.tasks.get(task_id)
+ if task is None:
+ return False
+ if task.status not in ("running", "pending"):
+ return False
+ task._inbox.put(message)
+ return True
+
+ def cancel(self, task_id: str) -> bool:
+ """Request cancellation of a running task.
+
+ Returns:
+ True if the cancel flag was set, False if task not found or not running.
+ """
+ task = self.tasks.get(task_id)
+ if task is None:
+ return False
+ if task.status == "running":
+ task._cancel_flag = True
+ return True
+ return False
+
+ def shutdown(self) -> None:
+ """Cancel all running tasks and shut down the thread pool."""
+ for task in self.tasks.values():
+ if task.status == "running":
+ task._cancel_flag = True
+ self._pool.shutdown(wait=True)
diff --git a/nano-claude-code/multi_agent/tools.py b/nano-claude-code/multi_agent/tools.py
new file mode 100644
index 0000000..4a8ef74
--- /dev/null
+++ b/nano-claude-code/multi_agent/tools.py
@@ -0,0 +1,295 @@
+"""Multi-agent tool registrations.
+
+Registers the following tools into the central tool_registry:
+ Agent โ spawn a sub-agent (sync or background)
+ SendMessage โ send a message to a named background agent
+ CheckAgentResult โ check status/result of a background agent
+ ListAgentTasks โ list all active/finished agent tasks
+ ListAgentTypes โ list available agent type definitions
+"""
+from __future__ import annotations
+
+from tool_registry import ToolDef, register_tool
+from .subagent import SubAgentManager, get_agent_definition, load_agent_definitions
+
+
+# โโ Singleton manager โโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโ
+
+_agent_manager: SubAgentManager | None = None
+
+
+def get_agent_manager() -> SubAgentManager:
+ """Return (and lazily create) the process-wide SubAgentManager."""
+ global _agent_manager
+ if _agent_manager is None:
+ _agent_manager = SubAgentManager()
+ return _agent_manager
+
+
+# โโ Tool implementations โโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโ
+
+def _agent_tool(params: dict, config: dict) -> str:
+ """Spawn a sub-agent.
+
+ Reads from config:
+ _system_prompt โ injected by agent.py run(), used as base system prompt
+ _depth โ current nesting depth (prevents infinite recursion)
+ """
+ mgr = get_agent_manager()
+
+ prompt = params["prompt"]
+ wait = params.get("wait", True)
+ isolation = params.get("isolation", "")
+ name = params.get("name", "")
+ model_override = params.get("model", "")
+ subagent_type = params.get("subagent_type", "")
+
+ system_prompt = config.get("_system_prompt", "You are a helpful assistant.")
+ depth = config.get("_depth", 0)
+
+ # Strip private keys before passing to sub-agent
+ eff_config = {k: v for k, v in config.items() if not k.startswith("_")}
+ if model_override:
+ eff_config["model"] = model_override
+
+ # Resolve agent definition
+ agent_def = None
+ if subagent_type:
+ agent_def = get_agent_definition(subagent_type)
+ if agent_def is None:
+ return (
+ f"Error: unknown subagent_type '{subagent_type}'. "
+ "Use ListAgentTypes to see available types."
+ )
+
+ task = mgr.spawn(
+ prompt, eff_config, system_prompt,
+ depth=depth,
+ agent_def=agent_def,
+ isolation=isolation,
+ name=name,
+ )
+
+ if task.status == "failed":
+ return f"Error spawning agent: {task.result}"
+
+ if wait:
+ mgr.wait(task.id, timeout=300)
+ result = task.result or f"(no output โ status: {task.status})"
+ header = f"[Agent: {task.name}"
+ if subagent_type:
+ header += f" ({subagent_type})"
+ if task.worktree_branch:
+ header += f", branch: {task.worktree_branch}"
+ header += "]"
+ return f"{header}\n\n{result}"
+ else:
+ info_parts = [f"Task ID: {task.id}", f"Name: {task.name}", f"Status: {task.status}"]
+ if subagent_type:
+ info_parts.append(f"Type: {subagent_type}")
+ if task.worktree_branch:
+ info_parts.append(f"Worktree branch: {task.worktree_branch}")
+ info_parts.append("Use CheckAgentResult or SendMessage to interact with this agent.")
+ return "\n".join(info_parts)
+
+
+def _send_message(params: dict, config: dict) -> str:
+ mgr = get_agent_manager()
+ target = params["to"]
+ message = params["message"]
+ ok = mgr.send_message(target, message)
+ if ok:
+ return f"Message queued for agent '{target}'. It will be processed after current work completes."
+ task_id = mgr._by_name.get(target, target)
+ task = mgr.tasks.get(task_id)
+ if task is None:
+ return f"Error: no agent found with id or name '{target}'"
+ return f"Error: agent '{target}' is not running (status: {task.status}). Cannot send message."
+
+
+def _check_agent_result(params: dict, config: dict) -> str:
+ mgr = get_agent_manager()
+ task_id = params["task_id"]
+ task = mgr.tasks.get(task_id)
+ if task is None:
+ return f"Error: no task with id '{task_id}'"
+ lines = [f"Status: {task.status}", f"Name: {task.name}"]
+ if task.worktree_branch:
+ lines.append(f"Worktree branch: {task.worktree_branch}")
+ if task.result:
+ lines.append(f"\nResult:\n{task.result}")
+ return "\n".join(lines)
+
+
+def _list_agent_tasks(params: dict, config: dict) -> str:
+ mgr = get_agent_manager()
+ tasks = mgr.list_tasks()
+ if not tasks:
+ return "No sub-agent tasks."
+ lines = ["ID | Name | Status | Worktree branch | Prompt"]
+ lines.append("-------------|----------|-----------|-----------------|------")
+ for t in tasks:
+ prompt_short = t.prompt[:50] + ("..." if len(t.prompt) > 50 else "")
+ wt = t.worktree_branch[:15] if t.worktree_branch else "-"
+ lines.append(f"{t.id} | {t.name[:8]:8s} | {t.status:9s} | {wt:15s} | {prompt_short}")
+ return "\n".join(lines)
+
+
+def _list_agent_types(params: dict, config: dict) -> str:
+ defs = load_agent_definitions()
+ if not defs:
+ return "No agent types available."
+ lines = ["Available agent types:", ""]
+ for aname, d in sorted(defs.items()):
+ model_info = f" model: {d.model}" if d.model else ""
+ tools_info = f" tools: {', '.join(d.tools)}" if d.tools else ""
+ lines.append(f" {aname:20s} [{d.source:8s}] {d.description}")
+ if model_info:
+ lines.append(f" {model_info}")
+ if tools_info:
+ lines.append(f" {tools_info}")
+ lines.append("")
+ lines.append(
+ "Create custom agents: place .md files in ~/.nano-claude/agents/ or .nano-claude/agents/"
+ )
+ return "\n".join(lines)
+
+
+# โโ Tool registrations โโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโ
+
+register_tool(ToolDef(
+ name="Agent",
+ schema={
+ "name": "Agent",
+ "description": (
+ "Spawn a sub-agent to handle a task autonomously. The sub-agent runs in a "
+ "separate thread with its own conversation history. Supports specialized agent "
+ "types (coder, reviewer, researcher, tester, or custom from .nano-claude/agents/), "
+ "isolated git worktrees for parallel work, and background execution.\n\n"
+ "When using isolation='worktree', the agent gets its own git branch and "
+ "working copy โ ideal for parallel coding tasks that shouldn't interfere."
+ ),
+ "input_schema": {
+ "type": "object",
+ "properties": {
+ "prompt": {
+ "type": "string",
+ "description": "Task description for the sub-agent",
+ },
+ "subagent_type": {
+ "type": "string",
+ "description": (
+ "Specialized agent type: 'general-purpose', 'coder', 'reviewer', "
+ "'researcher', 'tester', or any custom type. "
+ "Use ListAgentTypes to see all available types."
+ ),
+ },
+ "name": {
+ "type": "string",
+ "description": (
+ "Human-readable name for this agent instance. "
+ "Makes it addressable via SendMessage while running in background."
+ ),
+ },
+ "model": {
+ "type": "string",
+ "description": "Model override for this specific agent (optional)",
+ },
+ "wait": {
+ "type": "boolean",
+ "description": (
+ "Block until complete (default: true). "
+ "Set false to run in background."
+ ),
+ },
+ "isolation": {
+ "type": "string",
+ "enum": ["worktree"],
+ "description": (
+ "'worktree' creates a temporary git worktree so the agent works "
+ "on an isolated copy of the repo. Changes stay on a separate branch "
+ "and can be reviewed/merged after completion."
+ ),
+ },
+ },
+ "required": ["prompt"],
+ },
+ },
+ func=_agent_tool,
+ read_only=False,
+ concurrent_safe=False,
+))
+
+register_tool(ToolDef(
+ name="SendMessage",
+ schema={
+ "name": "SendMessage",
+ "description": (
+ "Send a follow-up message to a running background agent. "
+ "The message is queued and processed after the agent finishes its current work. "
+ "Reference agents by the name set via Agent(name=...) or by task ID."
+ ),
+ "input_schema": {
+ "type": "object",
+ "properties": {
+ "to": {"type": "string", "description": "Agent name or task ID"},
+ "message": {"type": "string", "description": "Message to send to the agent"},
+ },
+ "required": ["to", "message"],
+ },
+ },
+ func=_send_message,
+ read_only=False,
+ concurrent_safe=True,
+))
+
+register_tool(ToolDef(
+ name="CheckAgentResult",
+ schema={
+ "name": "CheckAgentResult",
+ "description": "Check the status and result of a spawned sub-agent task.",
+ "input_schema": {
+ "type": "object",
+ "properties": {
+ "task_id": {"type": "string", "description": "Task ID returned by Agent tool"},
+ },
+ "required": ["task_id"],
+ },
+ },
+ func=_check_agent_result,
+ read_only=True,
+ concurrent_safe=True,
+))
+
+register_tool(ToolDef(
+ name="ListAgentTasks",
+ schema={
+ "name": "ListAgentTasks",
+ "description": "List all sub-agent tasks and their statuses.",
+ "input_schema": {
+ "type": "object",
+ "properties": {},
+ },
+ },
+ func=_list_agent_tasks,
+ read_only=True,
+ concurrent_safe=True,
+))
+
+register_tool(ToolDef(
+ name="ListAgentTypes",
+ schema={
+ "name": "ListAgentTypes",
+ "description": (
+ "List all available agent types (built-in and custom). "
+ "Use the type names as subagent_type when calling Agent."
+ ),
+ "input_schema": {
+ "type": "object",
+ "properties": {},
+ },
+ },
+ func=_list_agent_types,
+ read_only=True,
+ concurrent_safe=True,
+))
diff --git a/nano-claude-code/nano_claude.py b/nano-claude-code/nano_claude.py
old mode 100644
new mode 100755
index 8294b4a..f133594
--- a/nano-claude-code/nano_claude.py
+++ b/nano-claude-code/nano_claude.py
@@ -26,6 +26,9 @@ Slash commands in REPL:
/thinking Toggle extended thinking
/permissions [mode] Set permission mode
/cwd [path] Show or change working directory
+ /memory [query] Show/search persistent memories
+ /skills List available skills
+ /agents Show sub-agent tasks
/exit /quit Exit
"""
from __future__ import annotations
@@ -33,13 +36,16 @@ from __future__ import annotations
import os
import sys
import json
-import readline
+try:
+ import readline
+except ImportError:
+ readline = None # Windows compatibility
import atexit
import argparse
import textwrap
from pathlib import Path
from datetime import datetime
-from typing import Optional
+from typing import Optional, Union
# โโ Optional rich for markdown rendering โโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโ
try:
@@ -78,6 +84,25 @@ def warn(msg: str): print(clr(f"Warning: {msg}", "yellow"))
def err(msg: str): print(clr(f"Error: {msg}", "red"), file=sys.stderr)
+def render_diff(text: str):
+ """Print diff text with ANSI colors: red for removals, green for additions."""
+ for line in text.splitlines():
+ if line.startswith("+++") or line.startswith("---"):
+ print(C["bold"] + line + C["reset"])
+ elif line.startswith("+"):
+ print(C["green"] + line + C["reset"])
+ elif line.startswith("-"):
+ print(C["red"] + line + C["reset"])
+ elif line.startswith("@@"):
+ print(C["cyan"] + line + C["reset"])
+ else:
+ print(line)
+
+def _has_diff(text: str) -> bool:
+ """Check if text contains a unified diff."""
+ return "--- a/" in text and "+++ b/" in text
+
+
# โโ Conversation rendering โโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโ
_accumulated_text: list[str] = [] # buffer text during streaming
@@ -118,6 +143,12 @@ def print_tool_end(name: str, result: str, verbose: bool):
summary = f"โ {lines} lines ({size} chars)"
if not result.startswith("Error") and not result.startswith("Denied"):
print(clr(f" โ {summary}", "dim", "green"), flush=True)
+ # Render diff for Edit/Write results
+ if name in ("Edit", "Write") and _has_diff(result):
+ parts = result.split("\n\n", 1)
+ if len(parts) == 2:
+ print(clr(f" {parts[0]}", "dim"))
+ render_diff(parts[1])
else:
print(clr(f" โ {result[:120]}", "dim", "red"), flush=True)
if verbose and not result.startswith("Denied"):
@@ -131,8 +162,26 @@ def _tool_desc(name: str, inputs: dict) -> str:
if name == "Bash": return f"Bash({inputs.get('command','')[:80]})"
if name == "Glob": return f"Glob({inputs.get('pattern','')})"
if name == "Grep": return f"Grep({inputs.get('pattern','')})"
- if name == "WebFetch": return f"WebFetch({inputs.get('url','')[:60]})"
- if name == "WebSearch": return f"WebSearch({inputs.get('query','')})"
+ if name == "WebFetch": return f"WebFetch({inputs.get('url','')[:60]})"
+ if name == "WebSearch": return f"WebSearch({inputs.get('query','')})"
+ if name == "Agent":
+ atype = inputs.get("subagent_type", "")
+ aname = inputs.get("name", "")
+ iso = inputs.get("isolation", "")
+ bg = not inputs.get("wait", True)
+ parts = []
+ if atype: parts.append(atype)
+ if aname: parts.append(f"name={aname}")
+ if iso: parts.append(f"isolation={iso}")
+ if bg: parts.append("background")
+ suffix = f"({', '.join(parts)})" if parts else ""
+ prompt_short = inputs.get("prompt", "")[:60]
+ return f"Agent{suffix}: {prompt_short}"
+ if name == "SendMessage":
+ return f"SendMessage(to={inputs.get('to','')}: {inputs.get('message','')[:50]})"
+ if name == "CheckAgentResult": return f"CheckAgentResult({inputs.get('task_id','')})"
+ if name == "ListAgentTasks": return "ListAgentTasks()"
+ if name == "ListAgentTypes": return "ListAgentTypes()"
return f"{name}({list(inputs.values())[:1]})"
@@ -351,6 +400,101 @@ def cmd_exit(_args: str, _state, _config) -> bool:
ok("Goodbye!")
sys.exit(0)
+def cmd_memory(args: str, _state, _config) -> bool:
+ from memory import search_memory, load_index
+ from memory.scan import scan_all_memories, format_memory_manifest, memory_freshness_text
+
+ if args.strip():
+ results = search_memory(args.strip())
+ if not results:
+ info(f"No memories matching '{args.strip()}'")
+ return True
+ info(f" {len(results)} result(s) for '{args.strip()}':")
+ for m in results:
+ info(f" [{m.type:9s}|{m.scope:7s}] {m.name}: {m.description}")
+ info(f" {m.content[:120]}{'...' if len(m.content) > 120 else ''}")
+ return True
+
+ # Show manifest with age/freshness
+ headers = scan_all_memories()
+ if not headers:
+ info("No memories stored. The model saves memories via MemorySave.")
+ return True
+ info(f" {len(headers)} memory/memories (newest first):")
+ for h in headers:
+ fresh_warn = " โ stale" if memory_freshness_text(h.mtime_s) else ""
+ tag = f"[{h.type or '?':9s}|{h.scope:7s}]"
+ info(f" {tag} {h.filename}{fresh_warn}")
+ if h.description:
+ info(f" {h.description}")
+ return True
+
+def cmd_agents(_args: str, _state, _config) -> bool:
+ try:
+ from multi_agent.tools import get_agent_manager
+ mgr = get_agent_manager()
+ tasks = mgr.list_tasks()
+ if not tasks:
+ info("No sub-agent tasks.")
+ return True
+ info(f" {len(tasks)} sub-agent task(s):")
+ for t in tasks:
+ preview = t.prompt[:50] + ("..." if len(t.prompt) > 50 else "")
+ wt_info = f" branch:{t.worktree_branch}" if t.worktree_branch else ""
+ info(f" {t.id} [{t.status:9s}] name={t.name}{wt_info} {preview}")
+ except Exception:
+ info("Sub-agent system not initialized.")
+ return True
+
+
+def _print_background_notifications():
+ """Print notifications for newly completed background agent tasks.
+
+ Called before each user prompt so the user sees results without polling.
+ """
+ try:
+ from multi_agent.tools import get_agent_manager
+ mgr = get_agent_manager()
+ except Exception:
+ return
+
+ notified_key = "_notified"
+ if not hasattr(_print_background_notifications, "_seen"):
+ _print_background_notifications._seen = set()
+
+ for task in mgr.list_tasks():
+ if task.id in _print_background_notifications._seen:
+ continue
+ if task.status in ("completed", "failed", "cancelled"):
+ _print_background_notifications._seen.add(task.id)
+ icon = "โ" if task.status == "completed" else "โ"
+ color = "green" if task.status == "completed" else "red"
+ branch_info = f" [branch: {task.worktree_branch}]" if task.worktree_branch else ""
+ print(clr(
+ f"\n {icon} Background agent '{task.name}' {task.status}{branch_info}",
+ color, "bold"
+ ))
+ if task.result:
+ preview = task.result[:200] + ("..." if len(task.result) > 200 else "")
+ print(clr(f" {preview}", "dim"))
+ print()
+
+def cmd_skills(_args: str, _state, _config) -> bool:
+ from skill import load_skills
+ skills = load_skills()
+ if not skills:
+ info("No skills found.")
+ return True
+ info(f"Available skills ({len(skills)}):")
+ for s in skills:
+ triggers = ", ".join(s.triggers)
+ source_label = f"[{s.source}]" if s.source != "builtin" else ""
+ hint = f" args: {s.argument_hint}" if s.argument_hint else ""
+ print(f" {clr(s.name, 'cyan'):24s} {s.description} {clr(triggers, 'dim')}{hint} {clr(source_label, 'yellow')}")
+ if s.when_to_use:
+ print(f" {clr(s.when_to_use[:80], 'dim')}")
+ return True
+
COMMANDS = {
"help": cmd_help,
"clear": cmd_clear,
@@ -365,13 +509,16 @@ COMMANDS = {
"thinking": cmd_thinking,
"permissions": cmd_permissions,
"cwd": cmd_cwd,
+ "skills": cmd_skills,
+ "memory": cmd_memory,
+ "agents": cmd_agents,
"exit": cmd_exit,
"quit": cmd_exit,
}
-def handle_slash(line: str, state, config) -> bool:
- """Handle /command [args]. Returns True if handled."""
+def handle_slash(line: str, state, config) -> Union[bool, tuple]:
+ """Handle /command [args]. Returns True if handled, tuple (skill, args) for skill match."""
if not line.startswith("/"):
return False
parts = line[1:].split(None, 1)
@@ -383,6 +530,15 @@ def handle_slash(line: str, state, config) -> bool:
if handler:
handler(args, state, config)
return True
+
+ # Fall through to skill lookup
+ from skill import find_skill
+ skill = find_skill(line)
+ if skill:
+ cmd_parts = line.strip().split(maxsplit=1)
+ skill_args = cmd_parts[1] if len(cmd_parts) > 1 else ""
+ return (skill, skill_args)
+
err(f"Unknown command: /{cmd} (type /help for commands)")
return True
@@ -390,6 +546,8 @@ def handle_slash(line: str, state, config) -> bool:
# โโ Input history setup โโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโ
def setup_readline(history_file: Path):
+ if readline is None:
+ return
try:
readline.read_history_file(str(history_file))
except FileNotFoundError:
@@ -487,6 +645,8 @@ def repl(config: dict, initial_prompt: str = None):
return
while True:
+ # Show notifications for background agents that finished
+ _print_background_notifications()
try:
cwd_short = Path.cwd().name
prompt = clr(f"\n[{cwd_short}] ", "dim") + clr("โฏ ", "cyan", "bold")
@@ -498,7 +658,19 @@ def repl(config: dict, initial_prompt: str = None):
if not user_input:
continue
- if handle_slash(user_input, state, config):
+
+ result = handle_slash(user_input, state, config)
+ if isinstance(result, tuple):
+ skill, skill_args = result
+ info(f"Running skill: {skill.name}" + (f" [{skill.context}]" if skill.context == "fork" else ""))
+ try:
+ from skill import substitute_arguments
+ rendered = substitute_arguments(skill.prompt, skill_args, skill.arguments)
+ run_query(f"[Skill: {skill.name}]\n\n{rendered}")
+ except KeyboardInterrupt:
+ print(clr("\n (interrupted)", "yellow"))
+ continue
+ if result:
continue
try:
diff --git a/nano-claude-code/providers.py b/nano-claude-code/providers.py
index 1fb4f9a..5e14374 100644
--- a/nano-claude-code/providers.py
+++ b/nano-claude-code/providers.py
@@ -29,6 +29,7 @@ PROVIDERS: dict[str, dict] = {
"anthropic": {
"type": "anthropic",
"api_key_env": "ANTHROPIC_API_KEY",
+ "context_limit": 200000,
"models": [
"claude-opus-4-6", "claude-sonnet-4-6", "claude-haiku-4-5-20251001",
"claude-opus-4-5", "claude-sonnet-4-5",
@@ -39,6 +40,7 @@ PROVIDERS: dict[str, dict] = {
"type": "openai",
"api_key_env": "OPENAI_API_KEY",
"base_url": "https://api.openai.com/v1",
+ "context_limit": 128000,
"models": [
"gpt-4o", "gpt-4o-mini", "gpt-4-turbo",
"o3-mini", "o1", "o1-mini",
@@ -48,6 +50,7 @@ PROVIDERS: dict[str, dict] = {
"type": "openai",
"api_key_env": "GEMINI_API_KEY",
"base_url": "https://generativelanguage.googleapis.com/v1beta/openai/",
+ "context_limit": 1000000,
"models": [
"gemini-2.5-pro-preview-03-25",
"gemini-2.0-flash", "gemini-2.0-flash-lite",
@@ -58,6 +61,7 @@ PROVIDERS: dict[str, dict] = {
"type": "openai",
"api_key_env": "MOONSHOT_API_KEY",
"base_url": "https://api.moonshot.cn/v1",
+ "context_limit": 128000,
"models": [
"moonshot-v1-8k", "moonshot-v1-32k", "moonshot-v1-128k",
"kimi-latest",
@@ -67,6 +71,7 @@ PROVIDERS: dict[str, dict] = {
"type": "openai",
"api_key_env": "DASHSCOPE_API_KEY",
"base_url": "https://dashscope.aliyuncs.com/compatible-mode/v1",
+ "context_limit": 1000000,
"models": [
"qwen-max", "qwen-plus", "qwen-turbo", "qwen-long",
"qwen2.5-72b-instruct", "qwen2.5-coder-32b-instruct",
@@ -77,6 +82,7 @@ PROVIDERS: dict[str, dict] = {
"type": "openai",
"api_key_env": "ZHIPU_API_KEY",
"base_url": "https://open.bigmodel.cn/api/paas/v4/",
+ "context_limit": 128000,
"models": [
"glm-4-plus", "glm-4", "glm-4-flash", "glm-4-air",
"glm-z1-flash",
@@ -86,6 +92,7 @@ PROVIDERS: dict[str, dict] = {
"type": "openai",
"api_key_env": "DEEPSEEK_API_KEY",
"base_url": "https://api.deepseek.com/v1",
+ "context_limit": 64000,
"models": [
"deepseek-chat", "deepseek-coder", "deepseek-reasoner",
],
@@ -95,6 +102,7 @@ PROVIDERS: dict[str, dict] = {
"api_key_env": None,
"base_url": "http://localhost:11434/v1",
"api_key": "ollama",
+ "context_limit": 128000,
"models": [
"llama3.3", "llama3.2", "phi4", "mistral", "mixtral",
"qwen2.5-coder", "deepseek-r1", "gemma3",
@@ -105,12 +113,14 @@ PROVIDERS: dict[str, dict] = {
"api_key_env": None,
"base_url": "http://localhost:1234/v1",
"api_key": "lm-studio",
+ "context_limit": 128000,
"models": [], # dynamic, depends on loaded model
},
"custom": {
"type": "openai",
"api_key_env": "CUSTOM_API_KEY",
"base_url": None, # read from config["custom_base_url"]
+ "context_limit": 128000,
"models": [],
},
}
@@ -277,8 +287,9 @@ def messages_to_openai(messages: list) -> list:
msg: dict = {"role": "assistant", "content": m.get("content") or None}
tcs = m.get("tool_calls", [])
if tcs:
- msg["tool_calls"] = [
- {
+ msg["tool_calls"] = []
+ for tc in tcs:
+ tc_msg = {
"id": tc["id"],
"type": "function",
"function": {
@@ -286,8 +297,10 @@ def messages_to_openai(messages: list) -> list:
"arguments": json.dumps(tc["input"], ensure_ascii=False),
},
}
- for tc in tcs
- ]
+ # Pass through provider-specific fields (e.g. Gemini thought_signature)
+ if tc.get("extra_content"):
+ tc_msg["extra_content"] = tc["extra_content"]
+ msg["tool_calls"].append(tc_msg)
result.append(msg)
elif role == "tool":
@@ -425,7 +438,7 @@ def stream_openai_compat(
for tc in delta.tool_calls:
idx = tc.index
if idx not in tool_buf:
- tool_buf[idx] = {"id": "", "name": "", "args": ""}
+ tool_buf[idx] = {"id": "", "name": "", "args": "", "extra_content": None}
if tc.id:
tool_buf[idx]["id"] = tc.id
if tc.function:
@@ -433,6 +446,10 @@ def stream_openai_compat(
tool_buf[idx]["name"] += tc.function.name
if tc.function.arguments:
tool_buf[idx]["args"] += tc.function.arguments
+ # Capture extra_content (e.g. Gemini thought_signature)
+ extra = getattr(tc, "extra_content", None)
+ if extra:
+ tool_buf[idx]["extra_content"] = extra
# Some providers include usage in the last chunk
if hasattr(chunk, "usage") and chunk.usage:
@@ -446,7 +463,10 @@ def stream_openai_compat(
inp = json.loads(v["args"]) if v["args"] else {}
except json.JSONDecodeError:
inp = {"_raw": v["args"]}
- tool_calls.append({"id": v["id"] or f"call_{idx}", "name": v["name"], "input": inp})
+ tc_entry = {"id": v["id"] or f"call_{idx}", "name": v["name"], "input": inp}
+ if v.get("extra_content"):
+ tc_entry["extra_content"] = v["extra_content"]
+ tool_calls.append(tc_entry)
yield AssistantTurn(text, tool_calls, in_tok, out_tok)
diff --git a/nano-claude-code/skill/__init__.py b/nano-claude-code/skill/__init__.py
new file mode 100644
index 0000000..8ef12f4
--- /dev/null
+++ b/nano-claude-code/skill/__init__.py
@@ -0,0 +1,14 @@
+"""skill package โ reusable prompt templates (skills)."""
+from .loader import ( # noqa: F401
+ SkillDef,
+ load_skills,
+ find_skill,
+ substitute_arguments,
+ register_builtin_skill,
+ _parse_skill_file,
+ _parse_list_field,
+)
+from .executor import execute_skill # noqa: F401
+
+# Importing builtin registers the built-in skills
+from . import builtin as _builtin # noqa: F401
diff --git a/nano-claude-code/skill/builtin.py b/nano-claude-code/skill/builtin.py
new file mode 100644
index 0000000..c6a4b8f
--- /dev/null
+++ b/nano-claude-code/skill/builtin.py
@@ -0,0 +1,100 @@
+"""Built-in skills that ship with nano-claude-code."""
+from __future__ import annotations
+
+from .loader import SkillDef, register_builtin_skill
+
+# โโ /commit โโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโ
+
+_COMMIT_PROMPT = """\
+Review the current git state and create a well-structured commit.
+
+## Steps
+
+1. Run `git status` and `git diff --staged` to see what is staged.
+ - If nothing is staged, run `git diff` to see unstaged changes, then stage relevant files.
+2. Analyze the changes:
+ - Summarize the nature of the change (feature, bug fix, refactor, docs, etc.)
+ - Write a concise commit title (โค72 chars) focusing on *why*, not just *what*.
+ - If multiple logical changes exist, ask the user whether to split them.
+3. Create the commit:
+ ```
+ git commit -m ""
+ ```
+ If additional context is needed, add a body separated by a blank line.
+4. Print the commit hash and summary when done.
+
+**Rules:**
+- Never use `--no-verify`.
+- Never commit files that likely contain secrets (.env, credentials, keys).
+- Prefer imperative mood in the title: "Add X", "Fix Y", "Refactor Z".
+
+User context: $ARGUMENTS
+"""
+
+_REVIEW_PROMPT = """\
+Review the code or pull request and provide structured feedback.
+
+## Steps
+
+1. Understand the scope:
+ - If a PR number or URL is given in $ARGUMENTS, use `gh pr view $ARGUMENTS --patch` to get the diff.
+ - Otherwise, use `git diff main...HEAD` (or `git diff HEAD~1`) for local changes.
+2. Analyze the diff:
+ - Correctness: Are there bugs, edge cases, or logic errors?
+ - Security: Injection, auth issues, exposed secrets, unsafe operations?
+ - Performance: N+1 queries, unnecessary allocations, blocking calls?
+ - Style: Does it follow existing conventions in the codebase?
+ - Tests: Are new behaviors tested? Do existing tests cover the change?
+3. Write a structured review:
+ ```
+ ## Summary
+ One-line overview of what the change does.
+
+ ## Issues
+ - [CRITICAL/MAJOR/MINOR] Description and location
+
+ ## Suggestions
+ - Nice-to-have improvements
+
+ ## Verdict
+ APPROVE / REQUEST CHANGES / COMMENT
+ ```
+4. If changes are needed, list specific file:line references.
+
+User context: $ARGUMENTS
+"""
+
+
+def _register_builtins() -> None:
+ register_builtin_skill(SkillDef(
+ name="commit",
+ description="Review staged changes and create a well-structured git commit",
+ triggers=["/commit"],
+ tools=["Bash", "Read"],
+ prompt=_COMMIT_PROMPT,
+ file_path="",
+ when_to_use="Use when the user wants to commit changes. Triggers: '/commit', 'commit changes', 'make a commit'.",
+ argument_hint="[optional context]",
+ arguments=[],
+ user_invocable=True,
+ context="inline",
+ source="builtin",
+ ))
+
+ register_builtin_skill(SkillDef(
+ name="review",
+ description="Review code changes or a pull request and provide structured feedback",
+ triggers=["/review", "/review-pr"],
+ tools=["Bash", "Read", "Grep"],
+ prompt=_REVIEW_PROMPT,
+ file_path="",
+ when_to_use="Use when the user wants a code review. Triggers: '/review', '/review-pr', 'review this PR'.",
+ argument_hint="[PR number or URL]",
+ arguments=["pr"],
+ user_invocable=True,
+ context="inline",
+ source="builtin",
+ ))
+
+
+_register_builtins()
diff --git a/nano-claude-code/skill/executor.py b/nano-claude-code/skill/executor.py
new file mode 100644
index 0000000..bc15a38
--- /dev/null
+++ b/nano-claude-code/skill/executor.py
@@ -0,0 +1,66 @@
+"""Skill execution: inline (current conversation) or forked (sub-agent)."""
+from __future__ import annotations
+
+from typing import Generator
+
+from .loader import SkillDef, substitute_arguments
+
+
+def execute_skill(
+ skill: SkillDef,
+ args: str,
+ state,
+ config: dict,
+ system_prompt: str,
+) -> Generator:
+ """Execute a skill.
+
+ If skill.context == "fork", runs as an isolated sub-agent and yields its events.
+ Otherwise (inline), injects the rendered prompt into the current agent loop.
+
+ Args:
+ skill: SkillDef to execute
+ args: raw argument string from user (after the trigger word)
+ state: AgentState
+ config: config dict (may contain _depth, model, etc.)
+ system_prompt: current system prompt string
+ Yields:
+ agent events (TextChunk, ToolStart, ToolEnd, TurnDone, โฆ)
+ """
+ rendered = substitute_arguments(skill.prompt, args, skill.arguments)
+ message = f"[Skill: {skill.name}]\n\n{rendered}"
+
+ if skill.context == "fork":
+ yield from _execute_forked(skill, message, config, system_prompt)
+ else:
+ yield from _execute_inline(message, state, config, system_prompt)
+
+
+def _execute_inline(message: str, state, config: dict, system_prompt: str) -> Generator:
+ """Run skill prompt inline in the current conversation."""
+ import agent as _agent
+ yield from _agent.run(message, state, config, system_prompt)
+
+
+def _execute_forked(
+ skill: SkillDef,
+ message: str,
+ config: dict,
+ system_prompt: str,
+) -> Generator:
+ """Run skill as an isolated sub-agent (separate conversation context)."""
+ import agent as _agent
+
+ # Build a sub-agent config with depth tracking
+ depth = config.get("_depth", 0) + 1
+ sub_config = {**config, "_depth": depth, "_system_prompt": system_prompt}
+ if skill.model:
+ sub_config["model"] = skill.model
+
+ # Restrict tools if skill specifies allowed-tools
+ if skill.tools:
+ sub_config["_allowed_tools"] = skill.tools
+
+ # Run in fresh state (no shared history)
+ sub_state = _agent.AgentState()
+ yield from _agent.run(message, sub_state, sub_config, system_prompt)
diff --git a/nano-claude-code/skill/loader.py b/nano-claude-code/skill/loader.py
new file mode 100644
index 0000000..7d21ecb
--- /dev/null
+++ b/nano-claude-code/skill/loader.py
@@ -0,0 +1,184 @@
+"""Skill loading: parse markdown files with YAML frontmatter into SkillDef objects."""
+from __future__ import annotations
+
+from dataclasses import dataclass, field
+from pathlib import Path
+from typing import Optional
+
+
+@dataclass
+class SkillDef:
+ name: str
+ description: str
+ triggers: list[str] # ["/commit", "commit changes"]
+ tools: list[str] # ["Bash", "Read"] (allowed-tools)
+ prompt: str # full prompt body after frontmatter
+ file_path: str
+ # Enhanced fields
+ when_to_use: str = "" # when Claude should auto-invoke this skill
+ argument_hint: str = "" # e.g. "[branch] [description]"
+ arguments: list[str] = field(default_factory=list) # named arg names
+ model: str = "" # model override
+ user_invocable: bool = True # appears in /skills list
+ context: str = "inline" # "inline" or "fork" (fork = sub-agent)
+ source: str = "user" # "user", "project", "builtin"
+
+
+# โโ Directory paths โโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโ
+
+def _get_skill_paths() -> list[Path]:
+ return [
+ Path.cwd() / ".nano_claude" / "skills", # project-level (priority)
+ Path.home() / ".nano_claude" / "skills", # user-level
+ ]
+
+
+# โโ List field parser โโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโ
+
+def _parse_list_field(value: str) -> list[str]:
+ """Parse YAML-like list: ``[a, b, c]`` or ``"a, b, c"``."""
+ value = value.strip()
+ if value.startswith("[") and value.endswith("]"):
+ value = value[1:-1]
+ return [item.strip().strip('"').strip("'") for item in value.split(",") if item.strip()]
+
+
+# โโ Single-file parser โโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโ
+
+def _parse_skill_file(path: Path, source: str = "user") -> Optional[SkillDef]:
+ """Parse a markdown file with ``---`` frontmatter into a SkillDef.
+
+ Frontmatter fields:
+ name, description, triggers, tools / allowed-tools,
+ when_to_use, argument-hint, arguments, model,
+ user-invocable, context
+ """
+ try:
+ text = path.read_text(encoding="utf-8")
+ except Exception:
+ return None
+
+ if not text.startswith("---"):
+ return None
+
+ parts = text.split("---", 2)
+ if len(parts) < 3:
+ return None
+
+ frontmatter_raw = parts[1].strip()
+ prompt = parts[2].strip()
+
+ fields: dict[str, str] = {}
+ for line in frontmatter_raw.splitlines():
+ line = line.strip()
+ if not line or ":" not in line:
+ continue
+ key, _, val = line.partition(":")
+ fields[key.strip().lower()] = val.strip()
+
+ name = fields.get("name", "")
+ if not name:
+ return None
+
+ # allowed-tools wins over tools if present
+ tools_raw = fields.get("allowed-tools", fields.get("tools", ""))
+ tools = _parse_list_field(tools_raw) if tools_raw else []
+
+ triggers_raw = fields.get("triggers", "")
+ triggers = _parse_list_field(triggers_raw) if triggers_raw else [f"/{name}"]
+
+ arguments_raw = fields.get("arguments", "")
+ arguments = _parse_list_field(arguments_raw) if arguments_raw else []
+
+ user_invocable_raw = fields.get("user-invocable", "true")
+ user_invocable = user_invocable_raw.lower() not in ("false", "0", "no")
+
+ context = fields.get("context", "inline").strip().lower()
+ if context not in ("inline", "fork"):
+ context = "inline"
+
+ return SkillDef(
+ name=name,
+ description=fields.get("description", ""),
+ triggers=triggers,
+ tools=tools,
+ prompt=prompt,
+ file_path=str(path),
+ when_to_use=fields.get("when_to_use", ""),
+ argument_hint=fields.get("argument-hint", ""),
+ arguments=arguments,
+ model=fields.get("model", ""),
+ user_invocable=user_invocable,
+ context=context,
+ source=source,
+ )
+
+
+# โโ Registry of built-in skills (registered by builtin.py) โโโโโโโโโโโโโโโโ
+
+_BUILTIN_SKILLS: list[SkillDef] = []
+
+
+def register_builtin_skill(skill: SkillDef) -> None:
+ _BUILTIN_SKILLS.append(skill)
+
+
+# โโ Load all skills โโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโ
+
+def load_skills(include_builtins: bool = True) -> list[SkillDef]:
+ """Return skills from disk + builtins, deduplicated (project > user > builtin)."""
+ seen: dict[str, SkillDef] = {}
+
+ # Builtins go in first (lowest priority)
+ if include_builtins:
+ for sk in _BUILTIN_SKILLS:
+ seen[sk.name] = sk
+
+ # User-level next, project-level last (highest priority)
+ skill_paths = _get_skill_paths()
+ for i, skill_dir in enumerate(reversed(skill_paths)):
+ src = "user" if i == 0 else "project"
+ if not skill_dir.is_dir():
+ continue
+ for md_file in sorted(skill_dir.glob("*.md")):
+ skill = _parse_skill_file(md_file, source=src)
+ if skill:
+ seen[skill.name] = skill
+
+ return list(seen.values())
+
+
+def find_skill(query: str) -> Optional[SkillDef]:
+ """Find a skill whose trigger matches the first word (or whole string) of query."""
+ query = query.strip()
+ if not query:
+ return None
+
+ first_word = query.split()[0]
+ for skill in load_skills():
+ for trigger in skill.triggers:
+ if first_word == trigger:
+ return skill
+ if trigger.startswith(first_word + " "):
+ return skill
+ return None
+
+
+# โโ Argument substitution โโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโ
+
+def substitute_arguments(prompt: str, args: str, arg_names: list[str]) -> str:
+ """Replace $ARGUMENTS (whole args string) and $ARG_NAME placeholders.
+
+ Named args are positional: first word โ first name, etc.
+ """
+ # Always substitute $ARGUMENTS
+ result = prompt.replace("$ARGUMENTS", args)
+
+ # Named args: split by whitespace
+ arg_values = args.split()
+ for i, arg_name in enumerate(arg_names):
+ placeholder = f"${arg_name.upper()}"
+ value = arg_values[i] if i < len(arg_values) else ""
+ result = result.replace(placeholder, value)
+
+ return result
diff --git a/nano-claude-code/skill/tools.py b/nano-claude-code/skill/tools.py
new file mode 100644
index 0000000..6121712
--- /dev/null
+++ b/nano-claude-code/skill/tools.py
@@ -0,0 +1,110 @@
+"""Skill tool: lets the model invoke skills by name via tool call."""
+from __future__ import annotations
+
+from tool_registry import ToolDef, register_tool
+from .loader import find_skill, load_skills, substitute_arguments
+
+
+_SKILL_SCHEMA = {
+ "name": "Skill",
+ "description": (
+ "Invoke a named skill (reusable prompt template). "
+ "Use SkillList to see available skills and their triggers."
+ ),
+ "input_schema": {
+ "type": "object",
+ "properties": {
+ "name": {
+ "type": "string",
+ "description": "Skill name (e.g. 'commit', 'review')",
+ },
+ "args": {
+ "type": "string",
+ "description": "Arguments to pass to the skill (replaces $ARGUMENTS)",
+ "default": "",
+ },
+ },
+ "required": ["name"],
+ },
+}
+
+_SKILL_LIST_SCHEMA = {
+ "name": "SkillList",
+ "description": "List all available skills with their names, triggers, and descriptions.",
+ "input_schema": {
+ "type": "object",
+ "properties": {},
+ "required": [],
+ },
+}
+
+
+def _skill_tool(params: dict, config: dict) -> str:
+ """Execute a skill by name and return its output."""
+ skill_name = params.get("name", "").strip()
+ args = params.get("args", "")
+
+ # Look up by name first, then by trigger
+ skill = None
+ for s in load_skills():
+ if s.name == skill_name:
+ skill = s
+ break
+ if skill is None:
+ skill = find_skill(skill_name)
+ if skill is None:
+ names = [s.name for s in load_skills()]
+ return f"Error: skill '{skill_name}' not found. Available: {', '.join(names)}"
+
+ rendered = substitute_arguments(skill.prompt, args, skill.arguments)
+ message = f"[Skill: {skill.name}]\n\n{rendered}"
+
+ # Run inline via agent and collect text output
+ import agent as _agent
+ system_prompt = config.get("_system_prompt", "")
+
+ # Collect output text
+ output_parts: list[str] = []
+ sub_state = _agent.AgentState()
+ sub_config = {**config, "_depth": config.get("_depth", 0) + 1}
+ try:
+ for event in _agent.run(message, sub_state, sub_config, system_prompt):
+ if hasattr(event, "text"):
+ output_parts.append(event.text)
+ except Exception as e:
+ return f"Skill execution error: {e}"
+
+ return "".join(output_parts) or "(skill completed with no text output)"
+
+
+def _skill_list_tool(params: dict, config: dict) -> str:
+ skills = load_skills()
+ if not skills:
+ return "No skills available."
+ lines = ["Available skills:\n"]
+ for s in skills:
+ triggers = ", ".join(s.triggers)
+ hint = f" args: {s.argument_hint}" if s.argument_hint else ""
+ when = f"\n when: {s.when_to_use}" if s.when_to_use else ""
+ lines.append(f"- **{s.name}** [{triggers}]{hint}\n {s.description}{when}")
+ return "\n".join(lines)
+
+
+def _register() -> None:
+ register_tool(ToolDef(
+ name="Skill",
+ schema=_SKILL_SCHEMA,
+ func=_skill_tool,
+ read_only=False,
+ concurrent_safe=False,
+ ))
+ register_tool(ToolDef(
+ name="SkillList",
+ schema=_SKILL_LIST_SCHEMA,
+ func=_skill_list_tool,
+ read_only=True,
+ concurrent_safe=True,
+ ))
+
+
+_register()
diff --git a/nano-claude-code/skills.py b/nano-claude-code/skills.py
new file mode 100644
index 0000000..da6c370
--- /dev/null
+++ b/nano-claude-code/skills.py
@@ -0,0 +1,14 @@
+"""Backward-compatibility shim โ real implementation is in skill/ package."""
+from skill.loader import ( # noqa: F401
+ SkillDef,
+ load_skills,
+ find_skill,
+ substitute_arguments,
+ _parse_skill_file,
+ _parse_list_field,
+)
+from skill.executor import execute_skill # noqa: F401
+
+# Legacy constant โ kept for tests that patch it
+from skill.loader import _get_skill_paths as _gsp
+SKILL_PATHS = _gsp()
diff --git a/nano-claude-code/subagent.py b/nano-claude-code/subagent.py
new file mode 100644
index 0000000..9ca5d86
--- /dev/null
+++ b/nano-claude-code/subagent.py
@@ -0,0 +1,11 @@
+"""Backward-compatibility shim โ real implementation is in multi_agent/subagent.py."""
+from multi_agent.subagent import ( # noqa: F401
+ AgentDefinition,
+ SubAgentTask,
+ SubAgentManager,
+ load_agent_definitions,
+ get_agent_definition,
+ _extract_final_text,
+ _agent_run,
+ _BUILTIN_AGENTS,
+)
diff --git a/nano-claude-code/tests/__init__.py b/nano-claude-code/tests/__init__.py
new file mode 100644
index 0000000..e69de29
diff --git a/nano-claude-code/tests/test_compaction.py b/nano-claude-code/tests/test_compaction.py
new file mode 100644
index 0000000..88d94e9
--- /dev/null
+++ b/nano-claude-code/tests/test_compaction.py
@@ -0,0 +1,187 @@
+"""Tests for compaction.py โ token estimation, context limits, snipping, split point."""
+from __future__ import annotations
+
+import sys
+import os
+
+# Ensure project root is on sys.path
+sys.path.insert(0, os.path.join(os.path.dirname(__file__), ".."))
+
+from compaction import estimate_tokens, get_context_limit, snip_old_tool_results, find_split_point
+
+
+# โโ estimate_tokens โโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโ
+
+class TestEstimateTokens:
+ def test_simple_messages(self):
+ msgs = [
+ {"role": "user", "content": "Hello world"}, # 11 chars
+ {"role": "assistant", "content": "Hi there!"}, # 9 chars
+ ]
+ result = estimate_tokens(msgs)
+ # (11 + 9) / 3.5 = 5.71 -> 5
+ assert result == int(20 / 3.5)
+
+ def test_empty_messages(self):
+ assert estimate_tokens([]) == 0
+
+ def test_empty_content(self):
+ msgs = [{"role": "user", "content": ""}]
+ assert estimate_tokens(msgs) == 0
+
+ def test_tool_result_messages(self):
+ msgs = [
+ {"role": "tool", "tool_call_id": "abc", "name": "Read", "content": "x" * 350},
+ ]
+ result = estimate_tokens(msgs)
+ assert result == int(350 / 3.5)
+
+ def test_structured_content(self):
+ """Content that is a list of dicts (e.g. Anthropic tool_result blocks)."""
+ msgs = [
+ {
+ "role": "user",
+ "content": [
+ {"type": "tool_result", "tool_use_id": "id1", "content": "A" * 70},
+ ],
+ },
+ ]
+ result = estimate_tokens(msgs)
+ # "tool_result" (11) + "id1" (3) + "A"*70 (70) = 84 -> 84/3.5 = 24
+ assert result == int(84 / 3.5)
+
+ def test_with_tool_calls(self):
+ msgs = [
+ {
+ "role": "assistant",
+ "content": "ok",
+ "tool_calls": [
+ {"id": "c1", "name": "Bash", "input": {"command": "ls"}},
+ ],
+ },
+ ]
+ result = estimate_tokens(msgs)
+ # content "ok" (2) + tool_calls string values: "c1" (2) + "Bash" (4) = 8
+ assert result == int(8 / 3.5)
+
+
+# โโ get_context_limit โโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโ
+
+class TestGetContextLimit:
+ def test_anthropic(self):
+ assert get_context_limit("claude-opus-4-6") == 200000
+
+ def test_gemini(self):
+ assert get_context_limit("gemini-2.0-flash") == 1000000
+
+ def test_deepseek(self):
+ assert get_context_limit("deepseek-chat") == 64000
+
+ def test_openai(self):
+ assert get_context_limit("gpt-4o") == 128000
+
+ def test_qwen(self):
+ assert get_context_limit("qwen-max") == 1000000
+
+ def test_unknown_model_fallback(self):
+ # Unknown models fall back to openai provider which has 128000
+ assert get_context_limit("some-random-model-xyz") == 128000
+
+ def test_explicit_provider_prefix(self):
+ assert get_context_limit("ollama/llama3.3") == 128000
+
+
+# โโ snip_old_tool_results โโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโ
+
+class TestSnipOldToolResults:
+ def test_old_tool_results_get_truncated(self):
+ long_content = "A" * 5000
+ msgs = [
+ {"role": "user", "content": "hello"},
+ {"role": "assistant", "content": "let me check", "tool_calls": []},
+ {"role": "tool", "tool_call_id": "t1", "name": "Read", "content": long_content},
+ {"role": "user", "content": "thanks"},
+ {"role": "assistant", "content": "you're welcome"},
+ {"role": "user", "content": "bye"},
+ {"role": "assistant", "content": "goodbye"},
+ {"role": "user", "content": "wait"},
+ {"role": "assistant", "content": "yes?"},
+ {"role": "user", "content": "never mind"},
+ ]
+ result = snip_old_tool_results(msgs, max_chars=2000, preserve_last_n_turns=6)
+ assert result is msgs # mutated in place
+ tool_msg = msgs[2]
+ assert len(tool_msg["content"]) < 5000
+ assert "snipped" in tool_msg["content"]
+
+ def test_recent_tool_results_preserved(self):
+ long_content = "B" * 5000
+ msgs = [
+ {"role": "user", "content": "hello"},
+ {"role": "assistant", "content": "ok", "tool_calls": []},
+ {"role": "tool", "tool_call_id": "t1", "name": "Read", "content": long_content},
+ ]
+ # All 3 messages are within preserve_last_n_turns=6
+ result = snip_old_tool_results(msgs, max_chars=2000, preserve_last_n_turns=6)
+ assert msgs[2]["content"] == long_content # not truncated
+
+ def test_short_tool_results_not_touched(self):
+ msgs = [
+ {"role": "tool", "tool_call_id": "t1", "name": "Bash", "content": "short"},
+ {"role": "user", "content": "a"},
+ {"role": "user", "content": "b"},
+ {"role": "user", "content": "c"},
+ {"role": "user", "content": "d"},
+ {"role": "user", "content": "e"},
+ {"role": "user", "content": "f"},
+ ]
+ snip_old_tool_results(msgs, max_chars=2000, preserve_last_n_turns=6)
+ assert msgs[0]["content"] == "short"
+
+ def test_non_tool_messages_untouched(self):
+ msgs = [
+ {"role": "user", "content": "X" * 5000},
+ {"role": "user", "content": "a"},
+ {"role": "user", "content": "b"},
+ {"role": "user", "content": "c"},
+ {"role": "user", "content": "d"},
+ {"role": "user", "content": "e"},
+ {"role": "user", "content": "f"},
+ ]
+ snip_old_tool_results(msgs, max_chars=2000, preserve_last_n_turns=6)
+ assert msgs[0]["content"] == "X" * 5000
+
+
+# โโ find_split_point โโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโ
+
+class TestFindSplitPoint:
+ def test_returns_reasonable_index(self):
+ msgs = [
+ {"role": "user", "content": "A" * 1000},
+ {"role": "assistant", "content": "B" * 1000},
+ {"role": "user", "content": "C" * 1000},
+ {"role": "assistant", "content": "D" * 1000},
+ {"role": "user", "content": "E" * 1000},
+ ]
+ idx = find_split_point(msgs, keep_ratio=0.3)
+ # With equal-size messages and keep_ratio=0.3, split should be around index 3-4
+ assert 2 <= idx <= 4
+
+ def test_single_message(self):
+ msgs = [{"role": "user", "content": "hello"}]
+ idx = find_split_point(msgs, keep_ratio=0.3)
+ assert idx == 0
+
+ def test_empty_messages(self):
+ idx = find_split_point([], keep_ratio=0.3)
+ assert idx == 0
+
+ def test_split_preserves_recent(self):
+ # Recent portion should contain ~30% of tokens
+ msgs = [{"role": "user", "content": "X" * 100} for _ in range(10)]
+ idx = find_split_point(msgs, keep_ratio=0.3)
+ total = estimate_tokens(msgs)
+ recent = estimate_tokens(msgs[idx:])
+ # Recent should be roughly 30% of total (allow some tolerance)
+ assert recent >= total * 0.2
+ assert recent <= total * 0.5
diff --git a/nano-claude-code/tests/test_diff_view.py b/nano-claude-code/tests/test_diff_view.py
new file mode 100644
index 0000000..0505c7d
--- /dev/null
+++ b/nano-claude-code/tests/test_diff_view.py
@@ -0,0 +1,50 @@
+import sys, os, tempfile
+sys.path.insert(0, os.path.dirname(os.path.dirname(os.path.abspath(__file__))))
+import pytest
+
+def test_generate_unified_diff():
+ from tools import generate_unified_diff
+ old = "line1\nline2\nline3\n"
+ new = "line1\nline2_modified\nline3\n"
+ diff = generate_unified_diff(old, new, "test.py")
+ assert "--- a/test.py" in diff
+ assert "+++ b/test.py" in diff
+ assert "-line2" in diff
+ assert "+line2_modified" in diff
+
+def test_generate_unified_diff_empty_old():
+ from tools import generate_unified_diff
+ diff = generate_unified_diff("", "new content\n", "test.py")
+ assert "+new content" in diff
+
+def test_edit_returns_diff(tmp_path):
+ from tools import _edit
+ f = tmp_path / "test.txt"
+ f.write_text("hello world\n")
+ result = _edit(str(f), "hello", "goodbye")
+ assert "-hello world" in result
+ assert "+goodbye world" in result
+
+def test_write_existing_returns_diff(tmp_path):
+ from tools import _write
+ f = tmp_path / "test.txt"
+ f.write_text("old content\n")
+ result = _write(str(f), "new content\n")
+ assert "-old content" in result
+ assert "+new content" in result
+
+def test_write_new_file_no_diff(tmp_path):
+ from tools import _write
+ f = tmp_path / "new.txt"
+ result = _write(str(f), "content\n")
+ assert "Created" in result
+ assert "---" not in result
+
+def test_diff_truncation():
+ from tools import generate_unified_diff, maybe_truncate_diff
+ old = "\n".join(f"line{i}" for i in range(200))
+ new = "\n".join(f"CHANGED{i}" for i in range(200))
+ diff = generate_unified_diff(old, new, "big.py")
+ truncated = maybe_truncate_diff(diff, max_lines=50)
+ assert "more lines" in truncated
+ assert truncated.count("\n") < 60
diff --git a/nano-claude-code/tests/test_memory.py b/nano-claude-code/tests/test_memory.py
new file mode 100644
index 0000000..530e9ce
--- /dev/null
+++ b/nano-claude-code/tests/test_memory.py
@@ -0,0 +1,275 @@
+"""Tests for the memory package (memory/)."""
+import pytest
+from pathlib import Path
+
+import memory.store as _store
+from memory.store import (
+ MemoryEntry,
+ save_memory,
+ load_index,
+ load_entries,
+ delete_memory,
+ search_memory,
+ _slugify,
+ parse_frontmatter,
+ get_index_content,
+)
+from memory.context import get_memory_context, truncate_index_content
+from memory.scan import (
+ scan_memory_dir,
+ format_memory_manifest,
+ memory_age_days,
+ memory_age_str,
+ memory_freshness_text,
+ MemoryHeader,
+)
+from memory.types import MEMORY_TYPES
+
+
+# โโ Fixtures โโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโ
+
+@pytest.fixture(autouse=True)
+def redirect_memory_dirs(tmp_path, monkeypatch):
+ """Redirect user and project memory dirs to tmp_path for all tests."""
+ user_mem = tmp_path / "user_memory"
+ user_mem.mkdir()
+ proj_mem = tmp_path / "project_memory"
+ proj_mem.mkdir()
+
+ monkeypatch.setattr(_store, "USER_MEMORY_DIR", user_mem)
+
+ # Patch get_project_memory_dir to return our tmp project dir
+ monkeypatch.setattr(_store, "get_project_memory_dir", lambda: proj_mem)
+
+
+def _make_entry(name="test note", description="a test", type_="user",
+ content="hello world", scope="user"):
+ return MemoryEntry(
+ name=name, description=description, type=type_,
+ content=content, created="2026-04-02", scope=scope,
+ )
+
+
+# โโ Save and Load โโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโ
+
+class TestSaveAndLoad:
+ def test_roundtrip(self):
+ entry = _make_entry()
+ save_memory(entry, scope="user")
+ loaded = load_entries("user")
+ assert len(loaded) == 1
+ assert loaded[0].name == "test note"
+ assert loaded[0].description == "a test"
+ assert loaded[0].type == "user"
+ assert loaded[0].content == "hello world"
+
+ def test_creates_file_on_disk(self):
+ entry = _make_entry()
+ save_memory(entry, scope="user")
+ assert Path(entry.file_path).exists()
+ text = Path(entry.file_path).read_text()
+ assert "hello world" in text
+
+ def test_update_existing(self):
+ """Save same name twice โ only 1 entry with updated content."""
+ save_memory(_make_entry(content="version 1"), scope="user")
+ save_memory(_make_entry(content="version 2"), scope="user")
+ loaded = load_entries("user")
+ assert len(loaded) == 1
+ assert loaded[0].content == "version 2"
+
+ def test_project_scope_stored_separately(self):
+ save_memory(_make_entry(name="user note"), scope="user")
+ save_memory(_make_entry(name="proj note"), scope="project")
+ user_entries = load_entries("user")
+ proj_entries = load_entries("project")
+ assert len(user_entries) == 1
+ assert len(proj_entries) == 1
+ assert user_entries[0].name == "user note"
+ assert proj_entries[0].name == "proj note"
+
+ def test_load_index_all_combines_scopes(self):
+ save_memory(_make_entry(name="user note"), scope="user")
+ save_memory(_make_entry(name="proj note"), scope="project")
+ all_entries = load_index("all")
+ names = {e.name for e in all_entries}
+ assert "user note" in names
+ assert "proj note" in names
+
+
+# โโ Delete โโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโ
+
+class TestDelete:
+ def test_delete_removes_file_and_index(self):
+ entry = _make_entry()
+ save_memory(entry, scope="user")
+ delete_memory("test note", scope="user")
+ assert load_entries("user") == []
+ assert not Path(entry.file_path).exists()
+
+ def test_delete_nonexistent_no_error(self):
+ delete_memory("nonexistent", scope="user")
+
+ def test_delete_from_project_scope(self):
+ save_memory(_make_entry(name="proj note"), scope="project")
+ delete_memory("proj note", scope="project")
+ assert load_entries("project") == []
+
+
+# โโ Search โโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโ
+
+class TestSearch:
+ def test_search_by_keyword(self):
+ save_memory(_make_entry(name="python tips", content="use list comprehension"), scope="user")
+ save_memory(_make_entry(name="rust tips", content="use iterators"), scope="user")
+ results = search_memory("python")
+ assert len(results) == 1
+ assert results[0].name == "python tips"
+
+ def test_search_case_insensitive(self):
+ save_memory(_make_entry(name="Important Note", content="something"), scope="user")
+ results = search_memory("important")
+ assert len(results) == 1
+
+ def test_search_in_content(self):
+ save_memory(_make_entry(name="misc", content="the quick brown fox"), scope="user")
+ results = search_memory("brown fox")
+ assert len(results) == 1
+
+ def test_search_across_scopes(self):
+ save_memory(_make_entry(name="user note", content="alpha"), scope="user")
+ save_memory(_make_entry(name="proj note", content="alpha"), scope="project")
+ results = search_memory("alpha", scope="all")
+ assert len(results) == 2
+
+
+# โโ Memory context โโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโ
+
+class TestGetMemoryContext:
+ def test_returns_index_text(self):
+ save_memory(_make_entry(name="my note", description="desc here"), scope="user")
+ ctx = get_memory_context()
+ assert "my note" in ctx
+ assert "desc here" in ctx
+
+ def test_empty_when_no_memories(self):
+ ctx = get_memory_context()
+ assert ctx == ""
+
+ def test_project_memories_labelled(self):
+ save_memory(_make_entry(name="proj note", description="project context"), scope="project")
+ ctx = get_memory_context()
+ assert "Project memories" in ctx
+ assert "proj note" in ctx
+
+
+# โโ Truncation โโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโ
+
+class TestTruncation:
+ def test_no_truncation_within_limits(self):
+ text = "- line\n" * 10
+ result = truncate_index_content(text)
+ assert "WARNING" not in result
+
+ def test_line_truncation(self):
+ text = "\n".join(f"- line {i}" for i in range(300))
+ result = truncate_index_content(text)
+ assert "WARNING" in result
+ assert "lines" in result
+
+ def test_byte_truncation(self):
+ # 25001 bytes of content
+ text = "x" * 25001
+ result = truncate_index_content(text)
+ assert "WARNING" in result
+
+
+# โโ Slugify โโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโ
+
+class TestSlugify:
+ def test_basic(self):
+ assert _slugify("Hello World") == "hello_world"
+
+ def test_special_chars(self):
+ assert _slugify("foo@bar!baz") == "foobarbaz"
+
+ def test_max_length(self):
+ assert len(_slugify("a" * 100)) == 60
+
+
+# โโ parse_frontmatter โโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโ
+
+class TestParseFrontmatter:
+ def test_parse(self):
+ text = "---\nname: foo\ntype: user\n---\nbody text"
+ meta, body = parse_frontmatter(text)
+ assert meta["name"] == "foo"
+ assert meta["type"] == "user"
+ assert body == "body text"
+
+ def test_no_frontmatter(self):
+ meta, body = parse_frontmatter("just plain text")
+ assert meta == {}
+ assert body == "just plain text"
+
+
+# โโ scan / age / freshness โโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโ
+
+class TestScanAndAge:
+ def test_scan_memory_dir(self):
+ save_memory(_make_entry(name="note a"), scope="user")
+ save_memory(_make_entry(name="note b"), scope="user")
+ user_dir = _store.USER_MEMORY_DIR
+ headers = scan_memory_dir(user_dir, "user")
+ assert len(headers) == 2
+ assert all(isinstance(h, MemoryHeader) for h in headers)
+
+ def test_format_manifest(self):
+ import time
+ headers = [
+ MemoryHeader(
+ filename="foo.md",
+ file_path="/tmp/foo.md",
+ mtime_s=time.time(),
+ description="test desc",
+ type="user",
+ scope="user",
+ )
+ ]
+ manifest = format_memory_manifest(headers)
+ assert "foo.md" in manifest
+ assert "test desc" in manifest
+ assert "today" in manifest
+
+ def test_memory_age_days_today(self):
+ import time
+ assert memory_age_days(time.time()) == 0
+
+ def test_memory_age_days_old(self):
+ import time
+ old = time.time() - 5 * 86400 # 5 days ago
+ assert memory_age_days(old) == 5
+
+ def test_memory_age_str(self):
+ import time
+ assert memory_age_str(time.time()) == "today"
+ assert memory_age_str(time.time() - 86400) == "yesterday"
+ assert memory_age_str(time.time() - 3 * 86400) == "3 days ago"
+
+ def test_freshness_text_fresh(self):
+ import time
+ assert memory_freshness_text(time.time()) == ""
+
+ def test_freshness_text_stale(self):
+ import time
+ old = time.time() - 10 * 86400
+ text = memory_freshness_text(old)
+ assert "10 days old" in text
+ assert "stale" in text.lower() or "outdated" in text.lower()
+
+
+# โโ Memory types โโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโ
+
+class TestMemoryTypes:
+ def test_types_list(self):
+ assert set(MEMORY_TYPES) == {"user", "feedback", "project", "reference"}
diff --git a/nano-claude-code/tests/test_skills.py b/nano-claude-code/tests/test_skills.py
new file mode 100644
index 0000000..6dbc887
--- /dev/null
+++ b/nano-claude-code/tests/test_skills.py
@@ -0,0 +1,234 @@
+from __future__ import annotations
+
+import pytest
+from pathlib import Path
+
+import skill.loader as _loader
+from skill.loader import _parse_skill_file, _parse_list_field, find_skill, SkillDef
+from skill import load_skills, substitute_arguments
+
+
+COMMIT_MD = """\
+---
+name: commit
+description: Create a git commit
+triggers: [/commit, commit changes]
+tools: [Bash, Read]
+---
+Review staged changes and create a commit with a descriptive message.
+"""
+
+REVIEW_MD = """\
+---
+name: review
+description: Review a pull request
+triggers: [/review, /review-pr]
+tools: [Bash, Read, Grep]
+---
+Analyze the PR diff and provide constructive feedback.
+"""
+
+ARGS_MD = """\
+---
+name: deploy
+description: Deploy to an environment
+triggers: [/deploy]
+tools: [Bash]
+argument-hint: [env] [version]
+arguments: [env, version]
+---
+Deploy $VERSION to $ENV environment. Full args: $ARGUMENTS
+"""
+
+
+@pytest.fixture()
+def skill_dir(tmp_path, monkeypatch):
+ """Create a temp skill directory with sample skills and patch _get_skill_paths."""
+ skills_dir = tmp_path / "skills"
+ skills_dir.mkdir()
+ (skills_dir / "commit.md").write_text(COMMIT_MD, encoding="utf-8")
+ (skills_dir / "review.md").write_text(REVIEW_MD, encoding="utf-8")
+
+ monkeypatch.setattr(_loader, "_get_skill_paths", lambda: [skills_dir])
+ # Also patch the builtin list to be empty so tests are predictable
+ monkeypatch.setattr(_loader, "_BUILTIN_SKILLS", [])
+ return skills_dir
+
+
+# ------------------------------------------------------------------
+# _parse_list_field
+# ------------------------------------------------------------------
+
+def test_parse_list_field_bracket():
+ assert _parse_list_field("[a, b, c]") == ["a", "b", "c"]
+
+
+def test_parse_list_field_plain():
+ assert _parse_list_field("a, b, c") == ["a", "b", "c"]
+
+
+def test_parse_list_field_single():
+ assert _parse_list_field("solo") == ["solo"]
+
+
+# ------------------------------------------------------------------
+# _parse_skill_file
+# ------------------------------------------------------------------
+
+def test_parse_skill_file(skill_dir):
+ path = skill_dir / "commit.md"
+ skill = _parse_skill_file(path)
+ assert skill is not None
+ assert skill.name == "commit"
+ assert skill.description == "Create a git commit"
+ assert "/commit" in skill.triggers
+ assert "commit changes" in skill.triggers
+ assert "Bash" in skill.tools
+ assert "Read" in skill.tools
+ assert "commit" in skill.prompt.lower()
+ assert skill.file_path == str(path)
+
+
+def test_parse_skill_file_review(skill_dir):
+ path = skill_dir / "review.md"
+ skill = _parse_skill_file(path)
+ assert skill is not None
+ assert skill.name == "review"
+ assert "/review" in skill.triggers
+ assert "/review-pr" in skill.triggers
+
+
+def test_parse_skill_file_invalid(tmp_path):
+ bad = tmp_path / "bad.md"
+ bad.write_text("no frontmatter here", encoding="utf-8")
+ assert _parse_skill_file(bad) is None
+
+
+def test_parse_skill_file_no_name(tmp_path):
+ no_name = tmp_path / "noname.md"
+ no_name.write_text("---\ndescription: test\n---\nbody\n", encoding="utf-8")
+ assert _parse_skill_file(no_name) is None
+
+
+def test_parse_skill_file_context_fork(tmp_path):
+ fork_md = tmp_path / "fork.md"
+ fork_md.write_text("---\nname: fork-task\ndescription: test\ncontext: fork\n---\nbody\n")
+ skill = _parse_skill_file(fork_md)
+ assert skill is not None
+ assert skill.context == "fork"
+
+
+def test_parse_skill_file_allowed_tools(tmp_path):
+ md = tmp_path / "t.md"
+ md.write_text("---\nname: myskill\ndescription: d\nallowed-tools: [Bash, Read]\n---\nbody\n")
+ skill = _parse_skill_file(md)
+ assert skill is not None
+ assert "Bash" in skill.tools
+ assert "Read" in skill.tools
+
+
+# ------------------------------------------------------------------
+# load_skills
+# ------------------------------------------------------------------
+
+def test_load_skills(skill_dir):
+ skills = load_skills()
+ assert len(skills) == 2
+ names = {s.name for s in skills}
+ assert names == {"commit", "review"}
+
+
+def test_load_skills_empty_dir(tmp_path, monkeypatch):
+ empty = tmp_path / "empty_skills"
+ empty.mkdir()
+ monkeypatch.setattr(_loader, "_get_skill_paths", lambda: [empty])
+ monkeypatch.setattr(_loader, "_BUILTIN_SKILLS", [])
+ assert load_skills() == []
+
+
+def test_load_skills_nonexistent_dir(tmp_path, monkeypatch):
+ monkeypatch.setattr(_loader, "_get_skill_paths", lambda: [tmp_path / "does_not_exist"])
+ monkeypatch.setattr(_loader, "_BUILTIN_SKILLS", [])
+ assert load_skills() == []
+
+
+def test_load_skills_builtins_present(monkeypatch):
+ """Without patching, builtins (commit, review) should be present."""
+ monkeypatch.setattr(_loader, "_get_skill_paths", lambda: [])
+ skills = load_skills()
+ names = {s.name for s in skills}
+ assert "commit" in names
+ assert "review" in names
+
+
+def test_load_skills_project_overrides_builtin(tmp_path, monkeypatch):
+ """A project skill with the same name overrides the builtin."""
+ skills_dir = tmp_path / "skills"
+ skills_dir.mkdir()
+ # project-level "commit" with different description
+ (skills_dir / "commit.md").write_text(
+ "---\nname: commit\ndescription: OVERRIDDEN\n---\ncustom commit prompt\n"
+ )
+ monkeypatch.setattr(_loader, "_get_skill_paths", lambda: [skills_dir])
+ skills = load_skills()
+ commit = next(s for s in skills if s.name == "commit")
+ assert commit.description == "OVERRIDDEN"
+
+
+# ------------------------------------------------------------------
+# find_skill
+# ------------------------------------------------------------------
+
+def test_find_skill_commit(skill_dir):
+ skill = find_skill("/commit")
+ assert skill is not None
+ assert skill.name == "commit"
+
+
+def test_find_skill_review(skill_dir):
+ skill = find_skill("/review")
+ assert skill is not None
+ assert skill.name == "review"
+
+
+def test_find_skill_review_pr(skill_dir):
+ skill = find_skill("/review-pr some-pr-url")
+ assert skill is not None
+ assert skill.name == "review"
+
+
+def test_find_skill_nonexistent(skill_dir):
+ result = find_skill("/nonexistent")
+ assert result is None
+
+
+# ------------------------------------------------------------------
+# substitute_arguments
+# ------------------------------------------------------------------
+
+def test_substitute_arguments_placeholder():
+ result = substitute_arguments("Deploy $ARGUMENTS please", "v1.2 prod", [])
+ assert result == "Deploy v1.2 prod please"
+
+
+def test_substitute_named_args(tmp_path):
+ result = substitute_arguments(
+ "Deploy $VERSION to $ENV. Full args: $ARGUMENTS",
+ "1.0 staging",
+ ["env", "version"],
+ )
+ # arg_names are positional: env=1.0, version=staging
+ assert "$VERSION" not in result
+ assert "$ENV" not in result
+ assert "$ARGUMENTS" not in result
+
+
+def test_substitute_missing_arg():
+ # If user provides fewer args than named slots, missing ones become ""
+ result = substitute_arguments("Hello $NAME!", "", ["name"])
+ assert result == "Hello !"
+
+
+def test_substitute_no_placeholders():
+ result = substitute_arguments("just a plain prompt", "some args", [])
+ assert result == "just a plain prompt"
diff --git a/nano-claude-code/tests/test_subagent.py b/nano-claude-code/tests/test_subagent.py
new file mode 100644
index 0000000..d768235
--- /dev/null
+++ b/nano-claude-code/tests/test_subagent.py
@@ -0,0 +1,136 @@
+"""Tests for the sub-agent system (subagent.py)."""
+import time
+import threading
+
+import pytest
+
+from multi_agent.subagent import SubAgentManager, SubAgentTask, _extract_final_text
+
+
+# โโ Mock for _agent_run โโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโ
+
+def _make_mock_agent_run(sleep_per_iter=0.05, iters=3):
+ """Return a mock _agent_run that simulates work and checks cancellation."""
+
+ def mock_agent_run(prompt, state, config, system_prompt, depth=0, cancel_check=None):
+ for i in range(iters):
+ if cancel_check and cancel_check():
+ return
+ time.sleep(sleep_per_iter)
+ # Append an assistant message to state
+ state.messages.append({
+ "role": "assistant",
+ "content": f"Result for: {prompt}",
+ "tool_calls": [],
+ })
+ # Yield a TurnDone-like event (generator protocol)
+ yield None
+
+ return mock_agent_run
+
+
+def _make_slow_mock(sleep_per_iter=0.2, iters=10):
+ """Return a slow mock for cancellation testing."""
+ return _make_mock_agent_run(sleep_per_iter=sleep_per_iter, iters=iters)
+
+
+@pytest.fixture
+def manager(monkeypatch):
+ """Create a SubAgentManager with mocked _agent_run."""
+ mock = _make_mock_agent_run()
+ monkeypatch.setattr("multi_agent.subagent._agent_run", mock)
+ mgr = SubAgentManager(max_concurrent=3, max_depth=3)
+ yield mgr
+ mgr.shutdown()
+
+
+@pytest.fixture
+def slow_manager(monkeypatch):
+ """Create a SubAgentManager with a slow mock for cancel testing."""
+ mock = _make_slow_mock()
+ monkeypatch.setattr("multi_agent.subagent._agent_run", mock)
+ mgr = SubAgentManager(max_concurrent=3, max_depth=3)
+ yield mgr
+ mgr.shutdown()
+
+
+# โโ Tests โโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโ
+
+class TestSpawnAndWait:
+ def test_spawn_and_wait_completes(self, manager):
+ task = manager.spawn("hello", {}, "system")
+ result_task = manager.wait(task.id, timeout=5)
+ assert result_task is not None
+ assert result_task.status == "completed"
+ assert result_task.result == "Result for: hello"
+
+ def test_spawn_returns_immediately(self, manager):
+ task = manager.spawn("hello", {}, "system")
+ # Task should be pending or running, not yet completed
+ assert task.status in ("pending", "running")
+
+
+class TestListTasks:
+ def test_list_tasks(self, manager):
+ t1 = manager.spawn("task1", {}, "system")
+ t2 = manager.spawn("task2", {}, "system")
+ tasks = manager.list_tasks()
+ task_ids = [t.id for t in tasks]
+ assert t1.id in task_ids
+ assert t2.id in task_ids
+ assert len(tasks) == 2
+
+
+class TestCancel:
+ def test_cancel_running_task(self, slow_manager):
+ task = slow_manager.spawn("slow task", {}, "system")
+ # Wait briefly to ensure the task starts running
+ time.sleep(0.1)
+ assert task.status == "running"
+ success = slow_manager.cancel(task.id)
+ assert success is True
+ # Wait for the task to actually finish
+ slow_manager.wait(task.id, timeout=5)
+ assert task.status == "cancelled"
+
+
+class TestDepthLimit:
+ def test_spawn_at_max_depth_fails(self, manager):
+ task = manager.spawn("deep", {}, "system", depth=3)
+ assert task.status == "failed"
+ assert "Max depth" in task.result
+
+
+class TestGetResult:
+ def test_get_result_completed(self, manager):
+ task = manager.spawn("hello", {}, "system")
+ manager.wait(task.id, timeout=5)
+ result = manager.get_result(task.id)
+ assert result == "Result for: hello"
+
+ def test_get_result_unknown_id(self, manager):
+ result = manager.get_result("nonexistent_id")
+ assert result is None
+
+
+class TestExtractFinalText:
+ def test_extracts_last_assistant(self):
+ messages = [
+ {"role": "user", "content": "hi"},
+ {"role": "assistant", "content": "first"},
+ {"role": "user", "content": "more"},
+ {"role": "assistant", "content": "second"},
+ ]
+ assert _extract_final_text(messages) == "second"
+
+ def test_returns_none_for_empty(self):
+ assert _extract_final_text([]) is None
+
+ def test_returns_none_no_assistant(self):
+ messages = [{"role": "user", "content": "hi"}]
+ assert _extract_final_text(messages) is None
+
+
+class TestWaitUnknown:
+ def test_wait_unknown_returns_none(self, manager):
+ assert manager.wait("nonexistent") is None
diff --git a/nano-claude-code/tests/test_tool_registry.py b/nano-claude-code/tests/test_tool_registry.py
new file mode 100644
index 0000000..2a7a8d4
--- /dev/null
+++ b/nano-claude-code/tests/test_tool_registry.py
@@ -0,0 +1,160 @@
+from __future__ import annotations
+
+import pytest
+
+from tool_registry import (
+ ToolDef,
+ clear_registry,
+ execute_tool,
+ get_all_tools,
+ get_tool,
+ get_tool_schemas,
+ register_tool,
+)
+
+
+@pytest.fixture(autouse=True)
+def _clean_registry():
+ """Reset registry before each test."""
+ clear_registry()
+ yield
+ clear_registry()
+
+
+def _make_echo_tool(name: str = "echo", read_only: bool = False) -> ToolDef:
+ """Helper to build a simple echo tool."""
+ schema = {
+ "name": name,
+ "description": f"Echo tool ({name})",
+ "input_schema": {
+ "type": "object",
+ "properties": {
+ "text": {"type": "string", "description": "text to echo"},
+ },
+ "required": ["text"],
+ },
+ }
+
+ def func(params: dict, config: dict) -> str:
+ return params["text"]
+
+ return ToolDef(
+ name=name,
+ schema=schema,
+ func=func,
+ read_only=read_only,
+ concurrent_safe=True,
+ )
+
+
+# ------------------------------------------------------------------
+# register and get
+# ------------------------------------------------------------------
+
+def test_register_and_get():
+ tool = _make_echo_tool()
+ register_tool(tool)
+ result = get_tool("echo")
+ assert result is not None
+ assert result.name == "echo"
+
+
+def test_get_unknown_returns_none():
+ assert get_tool("no_such_tool") is None
+
+
+# ------------------------------------------------------------------
+# get_all_tools
+# ------------------------------------------------------------------
+
+def test_get_all_tools_empty():
+ assert get_all_tools() == []
+
+
+def test_get_all_tools():
+ register_tool(_make_echo_tool("a"))
+ register_tool(_make_echo_tool("b"))
+ names = [t.name for t in get_all_tools()]
+ assert sorted(names) == ["a", "b"]
+
+
+# ------------------------------------------------------------------
+# get_tool_schemas
+# ------------------------------------------------------------------
+
+def test_get_tool_schemas():
+ register_tool(_make_echo_tool("echo"))
+ schemas = get_tool_schemas()
+ assert len(schemas) == 1
+ assert schemas[0]["name"] == "echo"
+
+
+# ------------------------------------------------------------------
+# execute_tool
+# ------------------------------------------------------------------
+
+def test_execute_tool():
+ register_tool(_make_echo_tool())
+ result = execute_tool("echo", {"text": "hello"}, config={})
+ assert result == "hello"
+
+
+def test_execute_unknown_tool():
+ result = execute_tool("missing", {}, config={})
+ assert "unknown" in result.lower() or "not found" in result.lower()
+
+
+# ------------------------------------------------------------------
+# output truncation
+# ------------------------------------------------------------------
+
+def test_output_truncation():
+ def big_func(params: dict, config: dict) -> str:
+ return "x" * 100
+
+ tool = ToolDef(
+ name="big",
+ schema={"name": "big", "description": "big", "input_schema": {"type": "object", "properties": {}}},
+ func=big_func,
+ read_only=True,
+ concurrent_safe=True,
+ )
+ register_tool(tool)
+
+ result = execute_tool("big", {}, config={}, max_output=40)
+ # first half = 20 chars, last quarter = 10 chars, marker in between
+ assert len(result) < 100
+ assert "truncated" in result
+ # The kept portion: first 20 + last 10 should be present
+ assert result.startswith("x" * 20)
+ assert result.endswith("x" * 10)
+
+
+def test_no_truncation_when_within_limit():
+ register_tool(_make_echo_tool())
+ result = execute_tool("echo", {"text": "short"}, config={})
+ assert result == "short"
+
+
+# ------------------------------------------------------------------
+# duplicate register overwrites
+# ------------------------------------------------------------------
+
+def test_duplicate_register_overwrites():
+ register_tool(_make_echo_tool("dup"))
+
+ def new_func(params: dict, config: dict) -> str:
+ return "new"
+
+ replacement = ToolDef(
+ name="dup",
+ schema={"name": "dup", "description": "new", "input_schema": {"type": "object", "properties": {}}},
+ func=new_func,
+ read_only=False,
+ concurrent_safe=False,
+ )
+ register_tool(replacement)
+
+ assert len(get_all_tools()) == 1
+ result = execute_tool("dup", {}, config={})
+ assert result == "new"
diff --git a/nano-claude-code/tool_registry.py b/nano-claude-code/tool_registry.py
new file mode 100644
index 0000000..c68dc29
--- /dev/null
+++ b/nano-claude-code/tool_registry.py
@@ -0,0 +1,98 @@
+"""Tool plugin registry for nano-claude-code.
+
+Provides a central registry for tool definitions, lookup, schema export,
+and dispatch with output truncation.
+"""
+from __future__ import annotations
+
+from dataclasses import dataclass
+from typing import Any, Callable, Dict, List, Optional
+
+
+@dataclass
+class ToolDef:
+ """Definition of a single tool plugin.
+
+ Attributes:
+ name: unique tool identifier
+ schema: JSON-schema dict sent to the API (name, description, input_schema)
+ func: callable(params: dict, config: dict) -> str
+ read_only: True if the tool never mutates state
+ concurrent_safe: True if safe to run in parallel with other tools
+ """
+ name: str
+ schema: Dict[str, Any]
+ func: Callable[[Dict[str, Any], Dict[str, Any]], str]
+ read_only: bool = False
+ concurrent_safe: bool = False
+
+
+# --------------- internal state ---------------
+
+_registry: Dict[str, ToolDef] = {}
+
+
+# --------------- public API ---------------
+
+def register_tool(tool_def: ToolDef) -> None:
+ """Register a tool, overwriting any existing tool with the same name."""
+ _registry[tool_def.name] = tool_def
+
+
+def get_tool(name: str) -> Optional[ToolDef]:
+ """Look up a tool by name. Returns None if not found."""
+ return _registry.get(name)
+
+
+def get_all_tools() -> List[ToolDef]:
+ """Return all registered tools (insertion order)."""
+ return list(_registry.values())
+
+
+def get_tool_schemas() -> List[Dict[str, Any]]:
+ """Return the schemas of all registered tools (for API tool parameter)."""
+ return [t.schema for t in _registry.values()]
+
+
+def execute_tool(
+ name: str,
+ params: Dict[str, Any],
+ config: Dict[str, Any],
+ max_output: int = 32000,
+) -> str:
+ """Dispatch a tool call by name.
+
+ Args:
+ name: tool name
+ params: tool input parameters dict
+ config: runtime configuration dict
+ max_output: maximum allowed output length in characters
+
+ Returns:
+ Tool result string, possibly truncated.
+ """
+ tool = get_tool(name)
+ if tool is None:
+ return f"Error: tool '{name}' not found."
+
+ try:
+ result = tool.func(params, config)
+ except Exception as e:
+ return f"Error executing {name}: {e}"
+
+ if len(result) > max_output:
+ first_half = max_output // 2
+ last_quarter = max_output // 4
+ truncated = len(result) - first_half - last_quarter
+ result = (
+ result[:first_half]
+ + f"\n[... {truncated} chars truncated ...]\n"
+ + result[-last_quarter:]
+ )
+
+ return result
+
+
+def clear_registry() -> None:
+ """Remove all registered tools. Intended for testing."""
+ _registry.clear()
diff --git a/nano-claude-code/tools.py b/nano-claude-code/tools.py
index 0efd516..08842a7 100644
--- a/nano-claude-code/tools.py
+++ b/nano-claude-code/tools.py
@@ -2,10 +2,14 @@
import os
import re
import glob as _glob
+import difflib
import subprocess
from pathlib import Path
from typing import Callable, Optional
+from tool_registry import ToolDef, register_tool
+from tool_registry import execute_tool as _registry_execute
+
# โโ Tool JSON schemas (sent to Claude API) โโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโ
TOOL_SCHEMAS = [
@@ -142,6 +146,24 @@ def _is_safe_bash(cmd: str) -> bool:
return any(c.startswith(p) for p in _SAFE_PREFIXES)
+# โโ Diff helpers โโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโ
+
+def generate_unified_diff(old, new, filename, context_lines=3):
+ old_lines = old.splitlines(keepends=True)
+ new_lines = new.splitlines(keepends=True)
+ diff = difflib.unified_diff(old_lines, new_lines,
+ fromfile=f"a/{filename}", tofile=f"b/{filename}", n=context_lines)
+ return "".join(diff)
+
+def maybe_truncate_diff(diff_text, max_lines=80):
+ lines = diff_text.splitlines()
+ if len(lines) <= max_lines:
+ return diff_text
+ shown = lines[:max_lines]
+ remaining = len(lines) - max_lines
+ return "\n".join(shown) + f"\n\n[... {remaining} more lines ...]"
+
+
# โโ Tool implementations โโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโ
def _read(file_path: str, limit: int = None, offset: int = None) -> str:
@@ -164,10 +186,19 @@ def _read(file_path: str, limit: int = None, offset: int = None) -> str:
def _write(file_path: str, content: str) -> str:
p = Path(file_path)
try:
+ is_new = not p.exists()
+ old_content = "" if is_new else p.read_text(errors="replace")
p.parent.mkdir(parents=True, exist_ok=True)
p.write_text(content)
- lc = content.count("\n") + (1 if content and not content.endswith("\n") else 0)
- return f"Wrote {lc} lines to {file_path}"
+ if is_new:
+ lc = content.count("\n") + (1 if content and not content.endswith("\n") else 0)
+ return f"Created {file_path} ({lc} lines)"
+ filename = p.name
+ diff = generate_unified_diff(old_content, content, filename)
+ if not diff:
+ return f"No changes in {file_path}"
+ truncated = maybe_truncate_diff(diff)
+ return f"File updated โ {file_path}:\n\n{truncated}"
except Exception as e:
return f"Error: {e}"
@@ -184,10 +215,13 @@ def _edit(file_path: str, old_string: str, new_string: str, replace_all: bool =
if count > 1 and not replace_all:
return (f"Error: old_string appears {count} times. "
"Provide more context to make it unique, or use replace_all=true.")
+ old_content = content
new_content = content.replace(old_string, new_string) if replace_all else \
content.replace(old_string, new_string, 1)
p.write_text(new_content)
- return f"Replaced {'all ' + str(count) if replace_all else '1'} occurrence(s) in {file_path}"
+ filename = p.name
+ diff = generate_unified_diff(old_content, new_content, filename)
+ return f"Changes applied to {filename}:\n\n{diff}"
except Exception as e:
return f"Error: {e}"
@@ -299,15 +333,22 @@ def _websearch(query: str) -> str:
return f"Error: {e}"
-# โโ Dispatcher โโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโ
+# โโ Dispatcher (backward-compatible wrapper) โโโโโโโโโโโโโโโโโโโโโโโโโโโโโโ
def execute_tool(
name: str,
inputs: dict,
permission_mode: str = "auto",
ask_permission: Optional[Callable[[str], bool]] = None,
+ config: dict = None,
) -> str:
- """Dispatch tool execution; ask permission for write/destructive ops."""
+ """Dispatch tool execution; ask permission for write/destructive ops.
+
+ Permission checking is done here, then delegation goes to the registry.
+ The config dict is forwarded to tool functions so they can access
+ runtime context like _depth, _system_prompt, model, etc.
+ """
+ cfg = config or {}
def _check(desc: str) -> bool:
"""Return True if action is allowed."""
@@ -317,43 +358,110 @@ def execute_tool(
return ask_permission(desc)
return True # headless: allow everything
- if name == "Read":
- return _read(inputs["file_path"], inputs.get("limit"), inputs.get("offset"))
-
- elif name == "Write":
+ # --- permission gate ---
+ if name == "Write":
if not _check(f"Write to {inputs['file_path']}"):
return "Denied: user rejected write operation"
- return _write(inputs["file_path"], inputs["content"])
-
elif name == "Edit":
if not _check(f"Edit {inputs['file_path']}"):
return "Denied: user rejected edit operation"
- return _edit(inputs["file_path"], inputs["old_string"],
- inputs["new_string"], inputs.get("replace_all", False))
-
elif name == "Bash":
cmd = inputs["command"]
if permission_mode != "accept-all" and not _is_safe_bash(cmd):
if not _check(f"Bash: {cmd}"):
return "Denied: user rejected bash command"
- return _bash(cmd, inputs.get("timeout", 30))
- elif name == "Glob":
- return _glob(inputs["pattern"], inputs.get("path"))
+ return _registry_execute(name, inputs, cfg)
- elif name == "Grep":
- return _grep(
- inputs["pattern"], inputs.get("path"), inputs.get("glob"),
- inputs.get("output_mode", "files_with_matches"),
- inputs.get("case_insensitive", False),
- inputs.get("context", 0),
- )
- elif name == "WebFetch":
- return _webfetch(inputs["url"], inputs.get("prompt"))
+# โโ Register built-in tools with the plugin registry โโโโโโโโโโโโโโโโโโโโโ
- elif name == "WebSearch":
- return _websearch(inputs["query"])
+def _register_builtins() -> None:
+ """Register all 8 built-in tools into the central registry."""
+ _tool_defs = [
+ ToolDef(
+ name="Read",
+ schema=TOOL_SCHEMAS[0],
+ func=lambda p, c: _read(**p),
+ read_only=True,
+ concurrent_safe=True,
+ ),
+ ToolDef(
+ name="Write",
+ schema=TOOL_SCHEMAS[1],
+ func=lambda p, c: _write(**p),
+ read_only=False,
+ concurrent_safe=False,
+ ),
+ ToolDef(
+ name="Edit",
+ schema=TOOL_SCHEMAS[2],
+ func=lambda p, c: _edit(**p),
+ read_only=False,
+ concurrent_safe=False,
+ ),
+ ToolDef(
+ name="Bash",
+ schema=TOOL_SCHEMAS[3],
+ func=lambda p, c: _bash(p["command"], p.get("timeout", 30)),
+ read_only=False,
+ concurrent_safe=False,
+ ),
+ ToolDef(
+ name="Glob",
+ schema=TOOL_SCHEMAS[4],
+ func=lambda p, c: _glob(p["pattern"], p.get("path")),
+ read_only=True,
+ concurrent_safe=True,
+ ),
+ ToolDef(
+ name="Grep",
+ schema=TOOL_SCHEMAS[5],
+ func=lambda p, c: _grep(
+ p["pattern"], p.get("path"), p.get("glob"),
+ p.get("output_mode", "files_with_matches"),
+ p.get("case_insensitive", False),
+ p.get("context", 0),
+ ),
+ read_only=True,
+ concurrent_safe=True,
+ ),
+ ToolDef(
+ name="WebFetch",
+ schema=TOOL_SCHEMAS[6],
+ func=lambda p, c: _webfetch(p["url"], p.get("prompt")),
+ read_only=True,
+ concurrent_safe=True,
+ ),
+ ToolDef(
+ name="WebSearch",
+ schema=TOOL_SCHEMAS[7],
+ func=lambda p, c: _websearch(p["query"]),
+ read_only=True,
+ concurrent_safe=True,
+ ),
+ ]
+ for td in _tool_defs:
+ register_tool(td)
- else:
- return f"Unknown tool: {name}"
+
+_register_builtins()
+
+
+# โโ Memory tools (MemorySave, MemoryDelete, MemorySearch, MemoryList) โโโโโโโโ
+# Defined in memory/tools.py; importing registers them automatically.
+import memory.tools as _memory_tools # noqa: F401
+
+
+
+# โโ Multi-agent tools (Agent, SendMessage, CheckAgentResult, ListAgentTasks, ListAgentTypes) โโ
+# Defined in multi_agent/tools.py; importing registers them automatically.
+import multi_agent.tools as _multiagent_tools # noqa: F401
+
+# Expose get_agent_manager at module level for backward compatibility
+from multi_agent.tools import get_agent_manager as _get_agent_manager # noqa: F401
+
+
+# โโ Skill tools (Skill, SkillList) โโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโ
+# Defined in skill/tools.py; importing registers them automatically.
+import skill.tools as _skill_tools # noqa: F401
diff --git a/original-source-code/src/QueryEngine.ts b/original-source-code/src/QueryEngine.ts
new file mode 100644
index 0000000..0a80c61
--- /dev/null
+++ b/original-source-code/src/QueryEngine.ts
@@ -0,0 +1,1295 @@
+import { feature } from 'bun:bundle'
+import type { ContentBlockParam } from '@anthropic-ai/sdk/resources/messages.mjs'
+import { randomUUID } from 'crypto'
+import last from 'lodash-es/last.js'
+import {
+ getSessionId,
+ isSessionPersistenceDisabled,
+} from 'src/bootstrap/state.js'
+import type {
+ PermissionMode,
+ SDKCompactBoundaryMessage,
+ SDKMessage,
+ SDKPermissionDenial,
+ SDKStatus,
+ SDKUserMessageReplay,
+} from 'src/entrypoints/agentSdkTypes.js'
+import { accumulateUsage, updateUsage } from 'src/services/api/claude.js'
+import type { NonNullableUsage } from 'src/services/api/logging.js'
+import { EMPTY_USAGE } from 'src/services/api/logging.js'
+import stripAnsi from 'strip-ansi'
+import type { Command } from './commands.js'
+import { getSlashCommandToolSkills } from './commands.js'
+import {
+ LOCAL_COMMAND_STDERR_TAG,
+ LOCAL_COMMAND_STDOUT_TAG,
+} from './constants/xml.js'
+import {
+ getModelUsage,
+ getTotalAPIDuration,
+ getTotalCost,
+} from './cost-tracker.js'
+import type { CanUseToolFn } from './hooks/useCanUseTool.js'
+import { loadMemoryPrompt } from './memdir/memdir.js'
+import { hasAutoMemPathOverride } from './memdir/paths.js'
+import { query } from './query.js'
+import { categorizeRetryableAPIError } from './services/api/errors.js'
+import type { MCPServerConnection } from './services/mcp/types.js'
+import type { AppState } from './state/AppState.js'
+import { type Tools, type ToolUseContext, toolMatchesName } from './Tool.js'
+import type { AgentDefinition } from './tools/AgentTool/loadAgentsDir.js'
+import { SYNTHETIC_OUTPUT_TOOL_NAME } from './tools/SyntheticOutputTool/SyntheticOutputTool.js'
+import type { Message } from './types/message.js'
+import type { OrphanedPermission } from './types/textInputTypes.js'
+import { createAbortController } from './utils/abortController.js'
+import type { AttributionState } from './utils/commitAttribution.js'
+import { getGlobalConfig } from './utils/config.js'
+import { getCwd } from './utils/cwd.js'
+import { isBareMode, isEnvTruthy } from './utils/envUtils.js'
+import { getFastModeState } from './utils/fastMode.js'
+import {
+ type FileHistoryState,
+ fileHistoryEnabled,
+ fileHistoryMakeSnapshot,
+} from './utils/fileHistory.js'
+import {
+ cloneFileStateCache,
+ type FileStateCache,
+} from './utils/fileStateCache.js'
+import { headlessProfilerCheckpoint } from './utils/headlessProfiler.js'
+import { registerStructuredOutputEnforcement } from './utils/hooks/hookHelpers.js'
+import { getInMemoryErrors } from './utils/log.js'
+import { countToolCalls, SYNTHETIC_MESSAGES } from './utils/messages.js'
+import {
+ getMainLoopModel,
+ parseUserSpecifiedModel,
+} from './utils/model/model.js'
+import { loadAllPluginsCacheOnly } from './utils/plugins/pluginLoader.js'
+import {
+ type ProcessUserInputContext,
+ processUserInput,
+} from './utils/processUserInput/processUserInput.js'
+import { fetchSystemPromptParts } from './utils/queryContext.js'
+import { setCwd } from './utils/Shell.js'
+import {
+ flushSessionStorage,
+ recordTranscript,
+} from './utils/sessionStorage.js'
+import { asSystemPrompt } from './utils/systemPromptType.js'
+import { resolveThemeSetting } from './utils/systemTheme.js'
+import {
+ shouldEnableThinkingByDefault,
+ type ThinkingConfig,
+} from './utils/thinking.js'
+
+// Lazy: MessageSelector.tsx pulls React/ink; only needed for message filtering at query time
+/* eslint-disable @typescript-eslint/no-require-imports */
+const messageSelector =
+ (): typeof import('src/components/MessageSelector.js') =>
+ require('src/components/MessageSelector.js')
+
+import {
+ localCommandOutputToSDKAssistantMessage,
+ toSDKCompactMetadata,
+} from './utils/messages/mappers.js'
+import {
+ buildSystemInitMessage,
+ sdkCompatToolName,
+} from './utils/messages/systemInit.js'
+import {
+ getScratchpadDir,
+ isScratchpadEnabled,
+} from './utils/permissions/filesystem.js'
+/* eslint-enable @typescript-eslint/no-require-imports */
+import {
+ handleOrphanedPermission,
+ isResultSuccessful,
+ normalizeMessage,
+} from './utils/queryHelpers.js'
+
+// Dead code elimination: conditional import for coordinator mode
+/* eslint-disable @typescript-eslint/no-require-imports */
+const getCoordinatorUserContext: (
+ mcpClients: ReadonlyArray<{ name: string }>,
+ scratchpadDir?: string,
+) => { [k: string]: string } = feature('COORDINATOR_MODE')
+ ? require('./coordinator/coordinatorMode.js').getCoordinatorUserContext
+ : () => ({})
+/* eslint-enable @typescript-eslint/no-require-imports */
+
+// Dead code elimination: conditional import for snip compaction
+/* eslint-disable @typescript-eslint/no-require-imports */
+const snipModule = feature('HISTORY_SNIP')
+ ? (require('./services/compact/snipCompact.js') as typeof import('./services/compact/snipCompact.js'))
+ : null
+const snipProjection = feature('HISTORY_SNIP')
+ ? (require('./services/compact/snipProjection.js') as typeof import('./services/compact/snipProjection.js'))
+ : null
+/* eslint-enable @typescript-eslint/no-require-imports */
+
+export type QueryEngineConfig = {
+ cwd: string
+ tools: Tools
+ commands: Command[]
+ mcpClients: MCPServerConnection[]
+ agents: AgentDefinition[]
+ canUseTool: CanUseToolFn
+ getAppState: () => AppState
+ setAppState: (f: (prev: AppState) => AppState) => void
+ initialMessages?: Message[]
+ readFileCache: FileStateCache
+ customSystemPrompt?: string
+ appendSystemPrompt?: string
+ userSpecifiedModel?: string
+ fallbackModel?: string
+ thinkingConfig?: ThinkingConfig
+ maxTurns?: number
+ maxBudgetUsd?: number
+ taskBudget?: { total: number }
+ jsonSchema?: Record
+ verbose?: boolean
+ replayUserMessages?: boolean
+ /** Handler for URL elicitations triggered by MCP tool -32042 errors. */
+ handleElicitation?: ToolUseContext['handleElicitation']
+ includePartialMessages?: boolean
+ setSDKStatus?: (status: SDKStatus) => void
+ abortController?: AbortController
+ orphanedPermission?: OrphanedPermission
+ /**
+ * Snip-boundary handler: receives each yielded system message plus the
+ * current mutableMessages store. Returns undefined if the message is not a
+ * snip boundary; otherwise returns the replayed snip result. Injected by
+ * ask() when HISTORY_SNIP is enabled so feature-gated strings stay inside
+ * the gated module (keeps QueryEngine free of excluded strings and testable
+ * despite feature() returning false under bun test). SDK-only: the REPL
+ * keeps full history for UI scrollback and projects on demand via
+ * projectSnippedView; QueryEngine truncates here to bound memory in long
+ * headless sessions (no UI to preserve).
+ */
+ snipReplay?: (
+ yieldedSystemMsg: Message,
+ store: Message[],
+ ) => { messages: Message[]; executed: boolean } | undefined
+}
+
+/**
+ * QueryEngine owns the query lifecycle and session state for a conversation.
+ * It extracts the core logic from ask() into a standalone class that can be
+ * used by both the headless/SDK path and (in a future phase) the REPL.
+ *
+ * One QueryEngine per conversation. Each submitMessage() call starts a new
+ * turn within the same conversation. State (messages, file cache, usage, etc.)
+ * persists across turns.
+ */
+export class QueryEngine {
+ private config: QueryEngineConfig
+ private mutableMessages: Message[]
+ private abortController: AbortController
+ private permissionDenials: SDKPermissionDenial[]
+ private totalUsage: NonNullableUsage
+ private hasHandledOrphanedPermission = false
+ private readFileState: FileStateCache
+ // Turn-scoped skill discovery tracking (feeds was_discovered on
+ // tengu_skill_tool_invocation). Must persist across the two
+ // processUserInputContext rebuilds inside submitMessage, but is cleared
+ // at the start of each submitMessage to avoid unbounded growth across
+ // many turns in SDK mode.
+ private discoveredSkillNames = new Set()
+ private loadedNestedMemoryPaths = new Set()
+
+ constructor(config: QueryEngineConfig) {
+ this.config = config
+ this.mutableMessages = config.initialMessages ?? []
+ this.abortController = config.abortController ?? createAbortController()
+ this.permissionDenials = []
+ this.readFileState = config.readFileCache
+ this.totalUsage = EMPTY_USAGE
+ }
+
+ async *submitMessage(
+ prompt: string | ContentBlockParam[],
+ options?: { uuid?: string; isMeta?: boolean },
+ ): AsyncGenerator {
+ const {
+ cwd,
+ commands,
+ tools,
+ mcpClients,
+ verbose = false,
+ thinkingConfig,
+ maxTurns,
+ maxBudgetUsd,
+ taskBudget,
+ canUseTool,
+ customSystemPrompt,
+ appendSystemPrompt,
+ userSpecifiedModel,
+ fallbackModel,
+ jsonSchema,
+ getAppState,
+ setAppState,
+ replayUserMessages = false,
+ includePartialMessages = false,
+ agents = [],
+ setSDKStatus,
+ orphanedPermission,
+ } = this.config
+
+ this.discoveredSkillNames.clear()
+ setCwd(cwd)
+ const persistSession = !isSessionPersistenceDisabled()
+ const startTime = Date.now()
+
+ // Wrap canUseTool to track permission denials
+ const wrappedCanUseTool: CanUseToolFn = async (
+ tool,
+ input,
+ toolUseContext,
+ assistantMessage,
+ toolUseID,
+ forceDecision,
+ ) => {
+ const result = await canUseTool(
+ tool,
+ input,
+ toolUseContext,
+ assistantMessage,
+ toolUseID,
+ forceDecision,
+ )
+
+ // Track denials for SDK reporting
+ if (result.behavior !== 'allow') {
+ this.permissionDenials.push({
+ tool_name: sdkCompatToolName(tool.name),
+ tool_use_id: toolUseID,
+ tool_input: input,
+ })
+ }
+
+ return result
+ }
+
+ const initialAppState = getAppState()
+ const initialMainLoopModel = userSpecifiedModel
+ ? parseUserSpecifiedModel(userSpecifiedModel)
+ : getMainLoopModel()
+
+ const initialThinkingConfig: ThinkingConfig = thinkingConfig
+ ? thinkingConfig
+ : shouldEnableThinkingByDefault() !== false
+ ? { type: 'adaptive' }
+ : { type: 'disabled' }
+
+ headlessProfilerCheckpoint('before_getSystemPrompt')
+ // Narrow once so TS tracks the type through the conditionals below.
+ const customPrompt =
+ typeof customSystemPrompt === 'string' ? customSystemPrompt : undefined
+ const {
+ defaultSystemPrompt,
+ userContext: baseUserContext,
+ systemContext,
+ } = await fetchSystemPromptParts({
+ tools,
+ mainLoopModel: initialMainLoopModel,
+ additionalWorkingDirectories: Array.from(
+ initialAppState.toolPermissionContext.additionalWorkingDirectories.keys(),
+ ),
+ mcpClients,
+ customSystemPrompt: customPrompt,
+ })
+ headlessProfilerCheckpoint('after_getSystemPrompt')
+ const userContext = {
+ ...baseUserContext,
+ ...getCoordinatorUserContext(
+ mcpClients,
+ isScratchpadEnabled() ? getScratchpadDir() : undefined,
+ ),
+ }
+
+ // When an SDK caller provides a custom system prompt AND has set
+ // CLAUDE_COWORK_MEMORY_PATH_OVERRIDE, inject the memory-mechanics prompt.
+ // The env var is an explicit opt-in signal โ the caller has wired up
+ // a memory directory and needs Claude to know how to use it (which
+ // Write/Edit tools to call, MEMORY.md filename, loading semantics).
+ // The caller can layer their own policy text via appendSystemPrompt.
+ const memoryMechanicsPrompt =
+ customPrompt !== undefined && hasAutoMemPathOverride()
+ ? await loadMemoryPrompt()
+ : null
+
+ const systemPrompt = asSystemPrompt([
+ ...(customPrompt !== undefined ? [customPrompt] : defaultSystemPrompt),
+ ...(memoryMechanicsPrompt ? [memoryMechanicsPrompt] : []),
+ ...(appendSystemPrompt ? [appendSystemPrompt] : []),
+ ])
+
+ // Register function hook for structured output enforcement
+ const hasStructuredOutputTool = tools.some(t =>
+ toolMatchesName(t, SYNTHETIC_OUTPUT_TOOL_NAME),
+ )
+ if (jsonSchema && hasStructuredOutputTool) {
+ registerStructuredOutputEnforcement(setAppState, getSessionId())
+ }
+
+ let processUserInputContext: ProcessUserInputContext = {
+ messages: this.mutableMessages,
+ // Slash commands that mutate the message array (e.g. /force-snip)
+ // call setMessages(fn). In interactive mode this writes back to
+ // AppState; in print mode we write back to mutableMessages so the
+ // rest of the query loop (push at :389, snapshot at :392) sees
+ // the result. The second processUserInputContext below (after
+ // slash-command processing) keeps the no-op โ nothing else calls
+ // setMessages past that point.
+ setMessages: fn => {
+ this.mutableMessages = fn(this.mutableMessages)
+ },
+ onChangeAPIKey: () => {},
+ handleElicitation: this.config.handleElicitation,
+ options: {
+ commands,
+ debug: false, // we use stdout, so don't want to clobber it
+ tools,
+ verbose,
+ mainLoopModel: initialMainLoopModel,
+ thinkingConfig: initialThinkingConfig,
+ mcpClients,
+ mcpResources: {},
+ ideInstallationStatus: null,
+ isNonInteractiveSession: true,
+ customSystemPrompt,
+ appendSystemPrompt,
+ agentDefinitions: { activeAgents: agents, allAgents: [] },
+ theme: resolveThemeSetting(getGlobalConfig().theme),
+ maxBudgetUsd,
+ },
+ getAppState,
+ setAppState,
+ abortController: this.abortController,
+ readFileState: this.readFileState,
+ nestedMemoryAttachmentTriggers: new Set(),
+ loadedNestedMemoryPaths: this.loadedNestedMemoryPaths,
+ dynamicSkillDirTriggers: new Set(),
+ discoveredSkillNames: this.discoveredSkillNames,
+ setInProgressToolUseIDs: () => {},
+ setResponseLength: () => {},
+ updateFileHistoryState: (
+ updater: (prev: FileHistoryState) => FileHistoryState,
+ ) => {
+ setAppState(prev => {
+ const updated = updater(prev.fileHistory)
+ if (updated === prev.fileHistory) return prev
+ return { ...prev, fileHistory: updated }
+ })
+ },
+ updateAttributionState: (
+ updater: (prev: AttributionState) => AttributionState,
+ ) => {
+ setAppState(prev => {
+ const updated = updater(prev.attribution)
+ if (updated === prev.attribution) return prev
+ return { ...prev, attribution: updated }
+ })
+ },
+ setSDKStatus,
+ }
+
+ // Handle orphaned permission (only once per engine lifetime)
+ if (orphanedPermission && !this.hasHandledOrphanedPermission) {
+ this.hasHandledOrphanedPermission = true
+ for await (const message of handleOrphanedPermission(
+ orphanedPermission,
+ tools,
+ this.mutableMessages,
+ processUserInputContext,
+ )) {
+ yield message
+ }
+ }
+
+ const {
+ messages: messagesFromUserInput,
+ shouldQuery,
+ allowedTools,
+ model: modelFromUserInput,
+ resultText,
+ } = await processUserInput({
+ input: prompt,
+ mode: 'prompt',
+ setToolJSX: () => {},
+ context: {
+ ...processUserInputContext,
+ messages: this.mutableMessages,
+ },
+ messages: this.mutableMessages,
+ uuid: options?.uuid,
+ isMeta: options?.isMeta,
+ querySource: 'sdk',
+ })
+
+ // Push new messages, including user input and any attachments
+ this.mutableMessages.push(...messagesFromUserInput)
+
+ // Update params to reflect updates from processing /slash commands
+ const messages = [...this.mutableMessages]
+
+ // Persist the user's message(s) to transcript BEFORE entering the query
+ // loop. The for-await below only calls recordTranscript when ask() yields
+ // an assistant/user/compact_boundary message โ which doesn't happen until
+ // the API responds. If the process is killed before that (e.g. user clicks
+ // Stop in cowork seconds after send), the transcript is left with only
+ // queue-operation entries; getLastSessionLog filters those out, returns
+ // null, and --resume fails with "No conversation found". Writing now makes
+ // the transcript resumable from the point the user message was accepted,
+ // even if no API response ever arrives.
+ //
+ // --bare / SIMPLE: fire-and-forget. Scripted calls don't --resume after
+ // kill-mid-request. The await is ~4ms on SSD, ~30ms under disk contention
+ // โ the single largest controllable critical-path cost after module eval.
+ // Transcript is still written (for post-hoc debugging); just not blocking.
+ if (persistSession && messagesFromUserInput.length > 0) {
+ const transcriptPromise = recordTranscript(messages)
+ if (isBareMode()) {
+ void transcriptPromise
+ } else {
+ await transcriptPromise
+ if (
+ isEnvTruthy(process.env.CLAUDE_CODE_EAGER_FLUSH) ||
+ isEnvTruthy(process.env.CLAUDE_CODE_IS_COWORK)
+ ) {
+ await flushSessionStorage()
+ }
+ }
+ }
+
+ // Filter messages that should be acknowledged after transcript
+ const replayableMessages = messagesFromUserInput.filter(
+ msg =>
+ (msg.type === 'user' &&
+ !msg.isMeta && // Skip synthetic caveat messages
+ !msg.toolUseResult && // Skip tool results (they'll be acked from query)
+ messageSelector().selectableUserMessagesFilter(msg)) || // Skip non-user-authored messages (task notifications, etc.)
+ (msg.type === 'system' && msg.subtype === 'compact_boundary'), // Always ack compact boundaries
+ )
+ const messagesToAck = replayUserMessages ? replayableMessages : []
+
+ // Update the ToolPermissionContext based on user input processing (as necessary)
+ setAppState(prev => ({
+ ...prev,
+ toolPermissionContext: {
+ ...prev.toolPermissionContext,
+ alwaysAllowRules: {
+ ...prev.toolPermissionContext.alwaysAllowRules,
+ command: allowedTools,
+ },
+ },
+ }))
+
+ const mainLoopModel = modelFromUserInput ?? initialMainLoopModel
+
+ // Recreate after processing the prompt to pick up updated messages and
+ // model (from slash commands).
+ processUserInputContext = {
+ messages,
+ setMessages: () => {},
+ onChangeAPIKey: () => {},
+ handleElicitation: this.config.handleElicitation,
+ options: {
+ commands,
+ debug: false,
+ tools,
+ verbose,
+ mainLoopModel,
+ thinkingConfig: initialThinkingConfig,
+ mcpClients,
+ mcpResources: {},
+ ideInstallationStatus: null,
+ isNonInteractiveSession: true,
+ customSystemPrompt,
+ appendSystemPrompt,
+ theme: resolveThemeSetting(getGlobalConfig().theme),
+ agentDefinitions: { activeAgents: agents, allAgents: [] },
+ maxBudgetUsd,
+ },
+ getAppState,
+ setAppState,
+ abortController: this.abortController,
+ readFileState: this.readFileState,
+ nestedMemoryAttachmentTriggers: new Set(),
+ loadedNestedMemoryPaths: this.loadedNestedMemoryPaths,
+ dynamicSkillDirTriggers: new Set(),
+ discoveredSkillNames: this.discoveredSkillNames,
+ setInProgressToolUseIDs: () => {},
+ setResponseLength: () => {},
+ updateFileHistoryState: processUserInputContext.updateFileHistoryState,
+ updateAttributionState: processUserInputContext.updateAttributionState,
+ setSDKStatus,
+ }
+
+ headlessProfilerCheckpoint('before_skills_plugins')
+ // Cache-only: headless/SDK/CCR startup must not block on network for
+ // ref-tracked plugins. CCR populates the cache via CLAUDE_CODE_SYNC_PLUGIN_INSTALL
+ // (headlessPluginInstall) or CLAUDE_CODE_PLUGIN_SEED_DIR before this runs;
+ // SDK callers that need fresh source can call /reload-plugins.
+ const [skills, { enabled: enabledPlugins }] = await Promise.all([
+ getSlashCommandToolSkills(getCwd()),
+ loadAllPluginsCacheOnly(),
+ ])
+ headlessProfilerCheckpoint('after_skills_plugins')
+
+ yield buildSystemInitMessage({
+ tools,
+ mcpClients,
+ model: mainLoopModel,
+ permissionMode: initialAppState.toolPermissionContext
+ .mode as PermissionMode, // TODO: avoid the cast
+ commands,
+ agents,
+ skills,
+ plugins: enabledPlugins,
+ fastMode: initialAppState.fastMode,
+ })
+
+ // Record when system message is yielded for headless latency tracking
+ headlessProfilerCheckpoint('system_message_yielded')
+
+ if (!shouldQuery) {
+ // Return the results of local slash commands.
+ // Use messagesFromUserInput (not replayableMessages) for command output
+ // because selectableUserMessagesFilter excludes local-command-stdout tags.
+ for (const msg of messagesFromUserInput) {
+ if (
+ msg.type === 'user' &&
+ typeof msg.message.content === 'string' &&
+ (msg.message.content.includes(`<${LOCAL_COMMAND_STDOUT_TAG}>`) ||
+ msg.message.content.includes(`<${LOCAL_COMMAND_STDERR_TAG}>`) ||
+ msg.isCompactSummary)
+ ) {
+ yield {
+ type: 'user',
+ message: {
+ ...msg.message,
+ content: stripAnsi(msg.message.content),
+ },
+ session_id: getSessionId(),
+ parent_tool_use_id: null,
+ uuid: msg.uuid,
+ timestamp: msg.timestamp,
+ isReplay: !msg.isCompactSummary,
+ isSynthetic: msg.isMeta || msg.isVisibleInTranscriptOnly,
+ } as SDKUserMessageReplay
+ }
+
+ // Local command output โ yield as a synthetic assistant message so
+ // RC renders it as assistant-style text rather than a user bubble.
+ // Emitted as assistant (not the dedicated SDKLocalCommandOutputMessage
+ // system subtype) so mobile clients + session-ingress can parse it.
+ if (
+ msg.type === 'system' &&
+ msg.subtype === 'local_command' &&
+ typeof msg.content === 'string' &&
+ (msg.content.includes(`<${LOCAL_COMMAND_STDOUT_TAG}>`) ||
+ msg.content.includes(`<${LOCAL_COMMAND_STDERR_TAG}>`))
+ ) {
+ yield localCommandOutputToSDKAssistantMessage(msg.content, msg.uuid)
+ }
+
+ if (msg.type === 'system' && msg.subtype === 'compact_boundary') {
+ yield {
+ type: 'system',
+ subtype: 'compact_boundary' as const,
+ session_id: getSessionId(),
+ uuid: msg.uuid,
+ compact_metadata: toSDKCompactMetadata(msg.compactMetadata),
+ } as SDKCompactBoundaryMessage
+ }
+ }
+
+ if (persistSession) {
+ await recordTranscript(messages)
+ if (
+ isEnvTruthy(process.env.CLAUDE_CODE_EAGER_FLUSH) ||
+ isEnvTruthy(process.env.CLAUDE_CODE_IS_COWORK)
+ ) {
+ await flushSessionStorage()
+ }
+ }
+
+ yield {
+ type: 'result',
+ subtype: 'success',
+ is_error: false,
+ duration_ms: Date.now() - startTime,
+ duration_api_ms: getTotalAPIDuration(),
+ num_turns: messages.length - 1,
+ result: resultText ?? '',
+ stop_reason: null,
+ session_id: getSessionId(),
+ total_cost_usd: getTotalCost(),
+ usage: this.totalUsage,
+ modelUsage: getModelUsage(),
+ permission_denials: this.permissionDenials,
+ fast_mode_state: getFastModeState(
+ mainLoopModel,
+ initialAppState.fastMode,
+ ),
+ uuid: randomUUID(),
+ }
+ return
+ }
+
+ if (fileHistoryEnabled() && persistSession) {
+ messagesFromUserInput
+ .filter(messageSelector().selectableUserMessagesFilter)
+ .forEach(message => {
+ void fileHistoryMakeSnapshot(
+ (updater: (prev: FileHistoryState) => FileHistoryState) => {
+ setAppState(prev => ({
+ ...prev,
+ fileHistory: updater(prev.fileHistory),
+ }))
+ },
+ message.uuid,
+ )
+ })
+ }
+
+ // Track current message usage (reset on each message_start)
+ let currentMessageUsage: NonNullableUsage = EMPTY_USAGE
+ let turnCount = 1
+ let hasAcknowledgedInitialMessages = false
+ // Track structured output from StructuredOutput tool calls
+ let structuredOutputFromTool: unknown
+ // Track the last stop_reason from assistant messages
+ let lastStopReason: string | null = null
+ // Reference-based watermark so error_during_execution's errors[] is
+ // turn-scoped. A length-based index breaks when the 100-entry ring buffer
+ // shift()s during the turn โ the index slides. If this entry is rotated
+ // out, lastIndexOf returns -1 and we include everything (safe fallback).
+ const errorLogWatermark = getInMemoryErrors().at(-1)
+ // Snapshot count before this query for delta-based retry limiting
+ const initialStructuredOutputCalls = jsonSchema
+ ? countToolCalls(this.mutableMessages, SYNTHETIC_OUTPUT_TOOL_NAME)
+ : 0
+
+ for await (const message of query({
+ messages,
+ systemPrompt,
+ userContext,
+ systemContext,
+ canUseTool: wrappedCanUseTool,
+ toolUseContext: processUserInputContext,
+ fallbackModel,
+ querySource: 'sdk',
+ maxTurns,
+ taskBudget,
+ })) {
+ // Record assistant, user, and compact boundary messages
+ if (
+ message.type === 'assistant' ||
+ message.type === 'user' ||
+ (message.type === 'system' && message.subtype === 'compact_boundary')
+ ) {
+ // Before writing a compact boundary, flush any in-memory-only
+ // messages up through the preservedSegment tail. Attachments and
+ // progress are now recorded inline (their switch cases below), but
+ // this flush still matters for the preservedSegment tail walk.
+ // If the SDK subprocess restarts before then (claude-desktop kills
+ // between turns), tailUuid points to a never-written message โ
+ // applyPreservedSegmentRelinks fails its tailโhead walk โ returns
+ // without pruning โ resume loads full pre-compact history.
+ if (
+ persistSession &&
+ message.type === 'system' &&
+ message.subtype === 'compact_boundary'
+ ) {
+ const tailUuid = message.compactMetadata?.preservedSegment?.tailUuid
+ if (tailUuid) {
+ const tailIdx = this.mutableMessages.findLastIndex(
+ m => m.uuid === tailUuid,
+ )
+ if (tailIdx !== -1) {
+ await recordTranscript(this.mutableMessages.slice(0, tailIdx + 1))
+ }
+ }
+ }
+ messages.push(message)
+ if (persistSession) {
+ // Fire-and-forget for assistant messages. claude.ts yields one
+ // assistant message per content block, then mutates the last
+ // one's message.usage/stop_reason on message_delta โ relying on
+ // the write queue's 100ms lazy jsonStringify. Awaiting here
+ // blocks ask()'s generator, so message_delta can't run until
+ // every block is consumed; the drain timer (started at block 1)
+ // elapses first. Interactive CC doesn't hit this because
+ // useLogMessages.ts fire-and-forgets. enqueueWrite is
+ // order-preserving so fire-and-forget here is safe.
+ if (message.type === 'assistant') {
+ void recordTranscript(messages)
+ } else {
+ await recordTranscript(messages)
+ }
+ }
+
+ // Acknowledge initial user messages after first transcript recording
+ if (!hasAcknowledgedInitialMessages && messagesToAck.length > 0) {
+ hasAcknowledgedInitialMessages = true
+ for (const msgToAck of messagesToAck) {
+ if (msgToAck.type === 'user') {
+ yield {
+ type: 'user',
+ message: msgToAck.message,
+ session_id: getSessionId(),
+ parent_tool_use_id: null,
+ uuid: msgToAck.uuid,
+ timestamp: msgToAck.timestamp,
+ isReplay: true,
+ } as SDKUserMessageReplay
+ }
+ }
+ }
+ }
+
+ if (message.type === 'user') {
+ turnCount++
+ }
+
+ switch (message.type) {
+ case 'tombstone':
+ // Tombstone messages are control signals for removing messages, skip them
+ break
+ case 'assistant':
+ // Capture stop_reason if already set (synthetic messages). For
+ // streamed responses, this is null at content_block_stop time;
+ // the real value arrives via message_delta (handled below).
+ if (message.message.stop_reason != null) {
+ lastStopReason = message.message.stop_reason
+ }
+ this.mutableMessages.push(message)
+ yield* normalizeMessage(message)
+ break
+ case 'progress':
+ this.mutableMessages.push(message)
+ // Record inline so the dedup loop in the next ask() call sees it
+ // as already-recorded. Without this, deferred progress interleaves
+ // with already-recorded tool_results in mutableMessages, and the
+ // dedup walk freezes startingParentUuid at the wrong message โ
+ // forking the chain and orphaning the conversation on resume.
+ if (persistSession) {
+ messages.push(message)
+ void recordTranscript(messages)
+ }
+ yield* normalizeMessage(message)
+ break
+ case 'user':
+ this.mutableMessages.push(message)
+ yield* normalizeMessage(message)
+ break
+ case 'stream_event':
+ if (message.event.type === 'message_start') {
+ // Reset current message usage for new message
+ currentMessageUsage = EMPTY_USAGE
+ currentMessageUsage = updateUsage(
+ currentMessageUsage,
+ message.event.message.usage,
+ )
+ }
+ if (message.event.type === 'message_delta') {
+ currentMessageUsage = updateUsage(
+ currentMessageUsage,
+ message.event.usage,
+ )
+ // Capture stop_reason from message_delta. The assistant message
+ // is yielded at content_block_stop with stop_reason=null; the
+ // real value only arrives here (see claude.ts message_delta
+ // handler). Without this, result.stop_reason is always null.
+ if (message.event.delta.stop_reason != null) {
+ lastStopReason = message.event.delta.stop_reason
+ }
+ }
+ if (message.event.type === 'message_stop') {
+ // Accumulate current message usage into total
+ this.totalUsage = accumulateUsage(
+ this.totalUsage,
+ currentMessageUsage,
+ )
+ }
+
+ if (includePartialMessages) {
+ yield {
+ type: 'stream_event' as const,
+ event: message.event,
+ session_id: getSessionId(),
+ parent_tool_use_id: null,
+ uuid: randomUUID(),
+ }
+ }
+
+ break
+ case 'attachment':
+ this.mutableMessages.push(message)
+ // Record inline (same reason as progress above).
+ if (persistSession) {
+ messages.push(message)
+ void recordTranscript(messages)
+ }
+
+ // Extract structured output from StructuredOutput tool calls
+ if (message.attachment.type === 'structured_output') {
+ structuredOutputFromTool = message.attachment.data
+ }
+ // Handle max turns reached signal from query.ts
+ else if (message.attachment.type === 'max_turns_reached') {
+ if (persistSession) {
+ if (
+ isEnvTruthy(process.env.CLAUDE_CODE_EAGER_FLUSH) ||
+ isEnvTruthy(process.env.CLAUDE_CODE_IS_COWORK)
+ ) {
+ await flushSessionStorage()
+ }
+ }
+ yield {
+ type: 'result',
+ subtype: 'error_max_turns',
+ duration_ms: Date.now() - startTime,
+ duration_api_ms: getTotalAPIDuration(),
+ is_error: true,
+ num_turns: message.attachment.turnCount,
+ stop_reason: lastStopReason,
+ session_id: getSessionId(),
+ total_cost_usd: getTotalCost(),
+ usage: this.totalUsage,
+ modelUsage: getModelUsage(),
+ permission_denials: this.permissionDenials,
+ fast_mode_state: getFastModeState(
+ mainLoopModel,
+ initialAppState.fastMode,
+ ),
+ uuid: randomUUID(),
+ errors: [
+ `Reached maximum number of turns (${message.attachment.maxTurns})`,
+ ],
+ }
+ return
+ }
+ // Yield queued_command attachments as SDK user message replays
+ else if (
+ replayUserMessages &&
+ message.attachment.type === 'queued_command'
+ ) {
+ yield {
+ type: 'user',
+ message: {
+ role: 'user' as const,
+ content: message.attachment.prompt,
+ },
+ session_id: getSessionId(),
+ parent_tool_use_id: null,
+ uuid: message.attachment.source_uuid || message.uuid,
+ timestamp: message.timestamp,
+ isReplay: true,
+ } as SDKUserMessageReplay
+ }
+ break
+ case 'stream_request_start':
+ // Don't yield stream request start messages
+ break
+ case 'system': {
+ // Snip boundary: replay on our store to remove zombie messages and
+ // stale markers. The yielded boundary is a signal, not data to push โ
+ // the replay produces its own equivalent boundary. Without this,
+ // markers persist and re-trigger on every turn, and mutableMessages
+ // never shrinks (memory leak in long SDK sessions). The subtype
+ // check lives inside the injected callback so feature-gated strings
+ // stay out of this file (excluded-strings check).
+ const snipResult = this.config.snipReplay?.(
+ message,
+ this.mutableMessages,
+ )
+ if (snipResult !== undefined) {
+ if (snipResult.executed) {
+ this.mutableMessages.length = 0
+ this.mutableMessages.push(...snipResult.messages)
+ }
+ break
+ }
+ this.mutableMessages.push(message)
+ // Yield compact boundary messages to SDK
+ if (
+ message.subtype === 'compact_boundary' &&
+ message.compactMetadata
+ ) {
+ // Release pre-compaction messages for GC. The boundary was just
+ // pushed so it's the last element. query.ts already uses
+ // getMessagesAfterCompactBoundary() internally, so only
+ // post-boundary messages are needed going forward.
+ const mutableBoundaryIdx = this.mutableMessages.length - 1
+ if (mutableBoundaryIdx > 0) {
+ this.mutableMessages.splice(0, mutableBoundaryIdx)
+ }
+ const localBoundaryIdx = messages.length - 1
+ if (localBoundaryIdx > 0) {
+ messages.splice(0, localBoundaryIdx)
+ }
+
+ yield {
+ type: 'system',
+ subtype: 'compact_boundary' as const,
+ session_id: getSessionId(),
+ uuid: message.uuid,
+ compact_metadata: toSDKCompactMetadata(message.compactMetadata),
+ }
+ }
+ if (message.subtype === 'api_error') {
+ yield {
+ type: 'system',
+ subtype: 'api_retry' as const,
+ attempt: message.retryAttempt,
+ max_retries: message.maxRetries,
+ retry_delay_ms: message.retryInMs,
+ error_status: message.error.status ?? null,
+ error: categorizeRetryableAPIError(message.error),
+ session_id: getSessionId(),
+ uuid: message.uuid,
+ }
+ }
+ // Don't yield other system messages in headless mode
+ break
+ }
+ case 'tool_use_summary':
+ // Yield tool use summary messages to SDK
+ yield {
+ type: 'tool_use_summary' as const,
+ summary: message.summary,
+ preceding_tool_use_ids: message.precedingToolUseIds,
+ session_id: getSessionId(),
+ uuid: message.uuid,
+ }
+ break
+ }
+
+ // Check if USD budget has been exceeded
+ if (maxBudgetUsd !== undefined && getTotalCost() >= maxBudgetUsd) {
+ if (persistSession) {
+ if (
+ isEnvTruthy(process.env.CLAUDE_CODE_EAGER_FLUSH) ||
+ isEnvTruthy(process.env.CLAUDE_CODE_IS_COWORK)
+ ) {
+ await flushSessionStorage()
+ }
+ }
+ yield {
+ type: 'result',
+ subtype: 'error_max_budget_usd',
+ duration_ms: Date.now() - startTime,
+ duration_api_ms: getTotalAPIDuration(),
+ is_error: true,
+ num_turns: turnCount,
+ stop_reason: lastStopReason,
+ session_id: getSessionId(),
+ total_cost_usd: getTotalCost(),
+ usage: this.totalUsage,
+ modelUsage: getModelUsage(),
+ permission_denials: this.permissionDenials,
+ fast_mode_state: getFastModeState(
+ mainLoopModel,
+ initialAppState.fastMode,
+ ),
+ uuid: randomUUID(),
+ errors: [`Reached maximum budget ($${maxBudgetUsd})`],
+ }
+ return
+ }
+
+ // Check if structured output retry limit exceeded (only on user messages)
+ if (message.type === 'user' && jsonSchema) {
+ const currentCalls = countToolCalls(
+ this.mutableMessages,
+ SYNTHETIC_OUTPUT_TOOL_NAME,
+ )
+ const callsThisQuery = currentCalls - initialStructuredOutputCalls
+ const maxRetries = parseInt(
+ process.env.MAX_STRUCTURED_OUTPUT_RETRIES || '5',
+ 10,
+ )
+ if (callsThisQuery >= maxRetries) {
+ if (persistSession) {
+ if (
+ isEnvTruthy(process.env.CLAUDE_CODE_EAGER_FLUSH) ||
+ isEnvTruthy(process.env.CLAUDE_CODE_IS_COWORK)
+ ) {
+ await flushSessionStorage()
+ }
+ }
+ yield {
+ type: 'result',
+ subtype: 'error_max_structured_output_retries',
+ duration_ms: Date.now() - startTime,
+ duration_api_ms: getTotalAPIDuration(),
+ is_error: true,
+ num_turns: turnCount,
+ stop_reason: lastStopReason,
+ session_id: getSessionId(),
+ total_cost_usd: getTotalCost(),
+ usage: this.totalUsage,
+ modelUsage: getModelUsage(),
+ permission_denials: this.permissionDenials,
+ fast_mode_state: getFastModeState(
+ mainLoopModel,
+ initialAppState.fastMode,
+ ),
+ uuid: randomUUID(),
+ errors: [
+ `Failed to provide valid structured output after ${maxRetries} attempts`,
+ ],
+ }
+ return
+ }
+ }
+ }
+
+ // Stop hooks yield progress/attachment messages AFTER the assistant
+ // response (via yield* handleStopHooks in query.ts). Since #23537 pushes
+ // those to `messages` inline, last(messages) can be a progress/attachment
+ // instead of the assistant โ which makes textResult extraction below
+ // return '' and -p mode emit a blank line. Allowlist to assistant|user:
+ // isResultSuccessful handles both (user with all tool_result blocks is a
+ // valid successful terminal state).
+ const result = messages.findLast(
+ m => m.type === 'assistant' || m.type === 'user',
+ )
+ // Capture for the error_during_execution diagnostic โ isResultSuccessful
+ // is a type predicate (message is Message), so inside the false branch
+ // `result` narrows to never and these accesses don't typecheck.
+ const edeResultType = result?.type ?? 'undefined'
+ const edeLastContentType =
+ result?.type === 'assistant'
+ ? (last(result.message.content)?.type ?? 'none')
+ : 'n/a'
+
+ // Flush buffered transcript writes before yielding result.
+ // The desktop app kills the CLI process immediately after receiving the
+ // result message, so any unflushed writes would be lost.
+ if (persistSession) {
+ if (
+ isEnvTruthy(process.env.CLAUDE_CODE_EAGER_FLUSH) ||
+ isEnvTruthy(process.env.CLAUDE_CODE_IS_COWORK)
+ ) {
+ await flushSessionStorage()
+ }
+ }
+
+ if (!isResultSuccessful(result, lastStopReason)) {
+ yield {
+ type: 'result',
+ subtype: 'error_during_execution',
+ duration_ms: Date.now() - startTime,
+ duration_api_ms: getTotalAPIDuration(),
+ is_error: true,
+ num_turns: turnCount,
+ stop_reason: lastStopReason,
+ session_id: getSessionId(),
+ total_cost_usd: getTotalCost(),
+ usage: this.totalUsage,
+ modelUsage: getModelUsage(),
+ permission_denials: this.permissionDenials,
+ fast_mode_state: getFastModeState(
+ mainLoopModel,
+ initialAppState.fastMode,
+ ),
+ uuid: randomUUID(),
+ // Diagnostic prefix: these are what isResultSuccessful() checks โ if
+ // the result type isn't assistant-with-text/thinking or user-with-
+ // tool_result, and stop_reason isn't end_turn, that's why this fired.
+ // errors[] is turn-scoped via the watermark; previously it dumped the
+ // entire process's logError buffer (ripgrep timeouts, ENOENT, etc).
+ errors: (() => {
+ const all = getInMemoryErrors()
+ const start = errorLogWatermark
+ ? all.lastIndexOf(errorLogWatermark) + 1
+ : 0
+ return [
+ `[ede_diagnostic] result_type=${edeResultType} last_content_type=${edeLastContentType} stop_reason=${lastStopReason}`,
+ ...all.slice(start).map(_ => _.error),
+ ]
+ })(),
+ }
+ return
+ }
+
+ // Extract the text result based on message type
+ let textResult = ''
+ let isApiError = false
+
+ if (result.type === 'assistant') {
+ const lastContent = last(result.message.content)
+ if (
+ lastContent?.type === 'text' &&
+ !SYNTHETIC_MESSAGES.has(lastContent.text)
+ ) {
+ textResult = lastContent.text
+ }
+ isApiError = Boolean(result.isApiErrorMessage)
+ }
+
+ yield {
+ type: 'result',
+ subtype: 'success',
+ is_error: isApiError,
+ duration_ms: Date.now() - startTime,
+ duration_api_ms: getTotalAPIDuration(),
+ num_turns: turnCount,
+ result: textResult,
+ stop_reason: lastStopReason,
+ session_id: getSessionId(),
+ total_cost_usd: getTotalCost(),
+ usage: this.totalUsage,
+ modelUsage: getModelUsage(),
+ permission_denials: this.permissionDenials,
+ structured_output: structuredOutputFromTool,
+ fast_mode_state: getFastModeState(
+ mainLoopModel,
+ initialAppState.fastMode,
+ ),
+ uuid: randomUUID(),
+ }
+ }
+
+ interrupt(): void {
+ this.abortController.abort()
+ }
+
+ getMessages(): readonly Message[] {
+ return this.mutableMessages
+ }
+
+ getReadFileState(): FileStateCache {
+ return this.readFileState
+ }
+
+ getSessionId(): string {
+ return getSessionId()
+ }
+
+ setModel(model: string): void {
+ this.config.userSpecifiedModel = model
+ }
+}
+
+/**
+ * Sends a single prompt to the Claude API and returns the response.
+ * Assumes that claude is being used non-interactively -- will not
+ * ask the user for permissions or further input.
+ *
+ * Convenience wrapper around QueryEngine for one-shot usage.
+ */
+export async function* ask({
+ commands,
+ prompt,
+ promptUuid,
+ isMeta,
+ cwd,
+ tools,
+ mcpClients,
+ verbose = false,
+ thinkingConfig,
+ maxTurns,
+ maxBudgetUsd,
+ taskBudget,
+ canUseTool,
+ mutableMessages = [],
+ getReadFileCache,
+ setReadFileCache,
+ customSystemPrompt,
+ appendSystemPrompt,
+ userSpecifiedModel,
+ fallbackModel,
+ jsonSchema,
+ getAppState,
+ setAppState,
+ abortController,
+ replayUserMessages = false,
+ includePartialMessages = false,
+ handleElicitation,
+ agents = [],
+ setSDKStatus,
+ orphanedPermission,
+}: {
+ commands: Command[]
+ prompt: string | Array
+ promptUuid?: string
+ isMeta?: boolean
+ cwd: string
+ tools: Tools
+ verbose?: boolean
+ mcpClients: MCPServerConnection[]
+ thinkingConfig?: ThinkingConfig
+ maxTurns?: number
+ maxBudgetUsd?: number
+ taskBudget?: { total: number }
+ canUseTool: CanUseToolFn
+ mutableMessages?: Message[]
+ customSystemPrompt?: string
+ appendSystemPrompt?: string
+ userSpecifiedModel?: string
+ fallbackModel?: string
+ jsonSchema?: Record
+ getAppState: () => AppState
+ setAppState: (f: (prev: AppState) => AppState) => void
+ getReadFileCache: () => FileStateCache
+ setReadFileCache: (cache: FileStateCache) => void
+ abortController?: AbortController
+ replayUserMessages?: boolean
+ includePartialMessages?: boolean
+ handleElicitation?: ToolUseContext['handleElicitation']
+ agents?: AgentDefinition[]
+ setSDKStatus?: (status: SDKStatus) => void
+ orphanedPermission?: OrphanedPermission
+}): AsyncGenerator {
+ const engine = new QueryEngine({
+ cwd,
+ tools,
+ commands,
+ mcpClients,
+ agents,
+ canUseTool,
+ getAppState,
+ setAppState,
+ initialMessages: mutableMessages,
+ readFileCache: cloneFileStateCache(getReadFileCache()),
+ customSystemPrompt,
+ appendSystemPrompt,
+ userSpecifiedModel,
+ fallbackModel,
+ thinkingConfig,
+ maxTurns,
+ maxBudgetUsd,
+ taskBudget,
+ jsonSchema,
+ verbose,
+ handleElicitation,
+ replayUserMessages,
+ includePartialMessages,
+ setSDKStatus,
+ abortController,
+ orphanedPermission,
+ ...(feature('HISTORY_SNIP')
+ ? {
+ snipReplay: (yielded: Message, store: Message[]) => {
+ if (!snipProjection!.isSnipBoundaryMessage(yielded))
+ return undefined
+ return snipModule!.snipCompactIfNeeded(store, { force: true })
+ },
+ }
+ : {}),
+ })
+
+ try {
+ yield* engine.submitMessage(prompt, {
+ uuid: promptUuid,
+ isMeta,
+ })
+ } finally {
+ setReadFileCache(engine.getReadFileState())
+ }
+}
diff --git a/original-source-code/src/Task.ts b/original-source-code/src/Task.ts
new file mode 100644
index 0000000..196caf3
--- /dev/null
+++ b/original-source-code/src/Task.ts
@@ -0,0 +1,125 @@
+import { randomBytes } from 'crypto'
+import type { AppState } from './state/AppState.js'
+import type { AgentId } from './types/ids.js'
+import { getTaskOutputPath } from './utils/task/diskOutput.js'
+
+export type TaskType =
+ | 'local_bash'
+ | 'local_agent'
+ | 'remote_agent'
+ | 'in_process_teammate'
+ | 'local_workflow'
+ | 'monitor_mcp'
+ | 'dream'
+
+export type TaskStatus =
+ | 'pending'
+ | 'running'
+ | 'completed'
+ | 'failed'
+ | 'killed'
+
+/**
+ * True when a task is in a terminal state and will not transition further.
+ * Used to guard against injecting messages into dead teammates, evicting
+ * finished tasks from AppState, and orphan-cleanup paths.
+ */
+export function isTerminalTaskStatus(status: TaskStatus): boolean {
+ return status === 'completed' || status === 'failed' || status === 'killed'
+}
+
+export type TaskHandle = {
+ taskId: string
+ cleanup?: () => void
+}
+
+export type SetAppState = (f: (prev: AppState) => AppState) => void
+
+export type TaskContext = {
+ abortController: AbortController
+ getAppState: () => AppState
+ setAppState: SetAppState
+}
+
+// Base fields shared by all task states
+export type TaskStateBase = {
+ id: string
+ type: TaskType
+ status: TaskStatus
+ description: string
+ toolUseId?: string
+ startTime: number
+ endTime?: number
+ totalPausedMs?: number
+ outputFile: string
+ outputOffset: number
+ notified: boolean
+}
+
+export type LocalShellSpawnInput = {
+ command: string
+ description: string
+ timeout?: number
+ toolUseId?: string
+ agentId?: AgentId
+ /** UI display variant: description-as-label, dialog title, status bar pill. */
+ kind?: 'bash' | 'monitor'
+}
+
+// What getTaskByType dispatches for: kill. spawn/render were never
+// called polymorphically (removed in #22546). All six kill implementations
+// use only setAppState โ getAppState/abortController were dead weight.
+export type Task = {
+ name: string
+ type: TaskType
+ kill(taskId: string, setAppState: SetAppState): Promise
+}
+
+// Task ID prefixes
+const TASK_ID_PREFIXES: Record = {
+ local_bash: 'b', // Keep as 'b' for backward compatibility
+ local_agent: 'a',
+ remote_agent: 'r',
+ in_process_teammate: 't',
+ local_workflow: 'w',
+ monitor_mcp: 'm',
+ dream: 'd',
+}
+
+// Get task ID prefix
+function getTaskIdPrefix(type: TaskType): string {
+ return TASK_ID_PREFIXES[type] ?? 'x'
+}
+
+// Case-insensitive-safe alphabet (digits + lowercase) for task IDs.
+// 36^8 โ 2.8 trillion combinations, sufficient to resist brute-force symlink attacks.
+const TASK_ID_ALPHABET = '0123456789abcdefghijklmnopqrstuvwxyz'
+
+export function generateTaskId(type: TaskType): string {
+ const prefix = getTaskIdPrefix(type)
+ const bytes = randomBytes(8)
+ let id = prefix
+ for (let i = 0; i < 8; i++) {
+ id += TASK_ID_ALPHABET[bytes[i]! % TASK_ID_ALPHABET.length]
+ }
+ return id
+}
+
+export function createTaskStateBase(
+ id: string,
+ type: TaskType,
+ description: string,
+ toolUseId?: string,
+): TaskStateBase {
+ return {
+ id,
+ type,
+ status: 'pending',
+ description,
+ toolUseId,
+ startTime: Date.now(),
+ outputFile: getTaskOutputPath(id),
+ outputOffset: 0,
+ notified: false,
+ }
+}
diff --git a/original-source-code/src/Tool.ts b/original-source-code/src/Tool.ts
new file mode 100644
index 0000000..205cac0
--- /dev/null
+++ b/original-source-code/src/Tool.ts
@@ -0,0 +1,792 @@
+import type {
+ ToolResultBlockParam,
+ ToolUseBlockParam,
+} from '@anthropic-ai/sdk/resources/index.mjs'
+import type {
+ ElicitRequestURLParams,
+ ElicitResult,
+} from '@modelcontextprotocol/sdk/types.js'
+import type { UUID } from 'crypto'
+import type { z } from 'zod/v4'
+import type { Command } from './commands.js'
+import type { CanUseToolFn } from './hooks/useCanUseTool.js'
+import type { ThinkingConfig } from './utils/thinking.js'
+
+export type ToolInputJSONSchema = {
+ [x: string]: unknown
+ type: 'object'
+ properties?: {
+ [x: string]: unknown
+ }
+}
+
+import type { Notification } from './context/notifications.js'
+import type {
+ MCPServerConnection,
+ ServerResource,
+} from './services/mcp/types.js'
+import type {
+ AgentDefinition,
+ AgentDefinitionsResult,
+} from './tools/AgentTool/loadAgentsDir.js'
+import type {
+ AssistantMessage,
+ AttachmentMessage,
+ Message,
+ ProgressMessage,
+ SystemLocalCommandMessage,
+ SystemMessage,
+ UserMessage,
+} from './types/message.js'
+// Import permission types from centralized location to break import cycles
+// Import PermissionResult from centralized location to break import cycles
+import type {
+ AdditionalWorkingDirectory,
+ PermissionMode,
+ PermissionResult,
+} from './types/permissions.js'
+// Import tool progress types from centralized location to break import cycles
+import type {
+ AgentToolProgress,
+ BashProgress,
+ MCPProgress,
+ REPLToolProgress,
+ SkillToolProgress,
+ TaskOutputProgress,
+ ToolProgressData,
+ WebSearchProgress,
+} from './types/tools.js'
+import type { FileStateCache } from './utils/fileStateCache.js'
+import type { DenialTrackingState } from './utils/permissions/denialTracking.js'
+import type { SystemPrompt } from './utils/systemPromptType.js'
+import type { ContentReplacementState } from './utils/toolResultStorage.js'
+
+// Re-export progress types for backwards compatibility
+export type {
+ AgentToolProgress,
+ BashProgress,
+ MCPProgress,
+ REPLToolProgress,
+ SkillToolProgress,
+ TaskOutputProgress,
+ WebSearchProgress,
+}
+
+import type { SpinnerMode } from './components/Spinner.js'
+import type { QuerySource } from './constants/querySource.js'
+import type { SDKStatus } from './entrypoints/agentSdkTypes.js'
+import type { AppState } from './state/AppState.js'
+import type {
+ HookProgress,
+ PromptRequest,
+ PromptResponse,
+} from './types/hooks.js'
+import type { AgentId } from './types/ids.js'
+import type { DeepImmutable } from './types/utils.js'
+import type { AttributionState } from './utils/commitAttribution.js'
+import type { FileHistoryState } from './utils/fileHistory.js'
+import type { Theme, ThemeName } from './utils/theme.js'
+
+export type QueryChainTracking = {
+ chainId: string
+ depth: number
+}
+
+export type ValidationResult =
+ | { result: true }
+ | {
+ result: false
+ message: string
+ errorCode: number
+ }
+
+export type SetToolJSXFn = (
+ args: {
+ jsx: React.ReactNode | null
+ shouldHidePromptInput: boolean
+ shouldContinueAnimation?: true
+ showSpinner?: boolean
+ isLocalJSXCommand?: boolean
+ isImmediate?: boolean
+ /** Set to true to clear a local JSX command (e.g., from its onDone callback) */
+ clearLocalJSX?: boolean
+ } | null,
+) => void
+
+// Import tool permission types from centralized location to break import cycles
+import type { ToolPermissionRulesBySource } from './types/permissions.js'
+
+// Re-export for backwards compatibility
+export type { ToolPermissionRulesBySource }
+
+// Apply DeepImmutable to the imported type
+export type ToolPermissionContext = DeepImmutable<{
+ mode: PermissionMode
+ additionalWorkingDirectories: Map
+ alwaysAllowRules: ToolPermissionRulesBySource
+ alwaysDenyRules: ToolPermissionRulesBySource
+ alwaysAskRules: ToolPermissionRulesBySource
+ isBypassPermissionsModeAvailable: boolean
+ isAutoModeAvailable?: boolean
+ strippedDangerousRules?: ToolPermissionRulesBySource
+ /** When true, permission prompts are auto-denied (e.g., background agents that can't show UI) */
+ shouldAvoidPermissionPrompts?: boolean
+ /** When true, automated checks (classifier, hooks) are awaited before showing the permission dialog (coordinator workers) */
+ awaitAutomatedChecksBeforeDialog?: boolean
+ /** Stores the permission mode before model-initiated plan mode entry, so it can be restored on exit */
+ prePlanMode?: PermissionMode
+}>
+
+export const getEmptyToolPermissionContext: () => ToolPermissionContext =
+ () => ({
+ mode: 'default',
+ additionalWorkingDirectories: new Map(),
+ alwaysAllowRules: {},
+ alwaysDenyRules: {},
+ alwaysAskRules: {},
+ isBypassPermissionsModeAvailable: false,
+ })
+
+export type CompactProgressEvent =
+ | {
+ type: 'hooks_start'
+ hookType: 'pre_compact' | 'post_compact' | 'session_start'
+ }
+ | { type: 'compact_start' }
+ | { type: 'compact_end' }
+
+export type ToolUseContext = {
+ options: {
+ commands: Command[]
+ debug: boolean
+ mainLoopModel: string
+ tools: Tools
+ verbose: boolean
+ thinkingConfig: ThinkingConfig
+ mcpClients: MCPServerConnection[]
+ mcpResources: Record
+ isNonInteractiveSession: boolean
+ agentDefinitions: AgentDefinitionsResult
+ maxBudgetUsd?: number
+ /** Custom system prompt that replaces the default system prompt */
+ customSystemPrompt?: string
+ /** Additional system prompt appended after the main system prompt */
+ appendSystemPrompt?: string
+ /** Override querySource for analytics tracking */
+ querySource?: QuerySource
+ /** Optional callback to get the latest tools (e.g., after MCP servers connect mid-query) */
+ refreshTools?: () => Tools
+ }
+ abortController: AbortController
+ readFileState: FileStateCache
+ getAppState(): AppState
+ setAppState(f: (prev: AppState) => AppState): void
+ /**
+ * Always-shared setAppState for session-scoped infrastructure (background
+ * tasks, session hooks). Unlike setAppState, which is no-op for async agents
+ * (see createSubagentContext), this always reaches the root store so agents
+ * at any nesting depth can register/clean up infrastructure that outlives
+ * a single turn. Only set by createSubagentContext; main-thread contexts
+ * fall back to setAppState.
+ */
+ setAppStateForTasks?: (f: (prev: AppState) => AppState) => void
+ /**
+ * Optional handler for URL elicitations triggered by tool call errors (-32042).
+ * In print/SDK mode, this delegates to structuredIO.handleElicitation.
+ * In REPL mode, this is undefined and the queue-based UI path is used.
+ */
+ handleElicitation?: (
+ serverName: string,
+ params: ElicitRequestURLParams,
+ signal: AbortSignal,
+ ) => Promise
+ setToolJSX?: SetToolJSXFn
+ addNotification?: (notif: Notification) => void
+ /** Append a UI-only system message to the REPL message list. Stripped at the
+ * normalizeMessagesForAPI boundary โ the Exclude<> makes that type-enforced. */
+ appendSystemMessage?: (
+ msg: Exclude,
+ ) => void
+ /** Send an OS-level notification (iTerm2, Kitty, Ghostty, bell, etc.) */
+ sendOSNotification?: (opts: {
+ message: string
+ notificationType: string
+ }) => void
+ nestedMemoryAttachmentTriggers?: Set
+ /**
+ * CLAUDE.md paths already injected as nested_memory attachments this
+ * session. Dedup for memoryFilesToAttachments โ readFileState is an LRU
+ * that evicts entries in busy sessions, so its .has() check alone can
+ * re-inject the same CLAUDE.md dozens of times.
+ */
+ loadedNestedMemoryPaths?: Set
+ dynamicSkillDirTriggers?: Set
+ /** Skill names surfaced via skill_discovery this session. Telemetry only (feeds was_discovered). */
+ discoveredSkillNames?: Set
+ userModified?: boolean
+ setInProgressToolUseIDs: (f: (prev: Set) => Set) => void
+ /** Only wired in interactive (REPL) contexts; SDK/QueryEngine don't set this. */
+ setHasInterruptibleToolInProgress?: (v: boolean) => void
+ setResponseLength: (f: (prev: number) => number) => void
+ /** Ant-only: push a new API metrics entry for OTPS tracking.
+ * Called by subagent streaming when a new API request starts. */
+ pushApiMetricsEntry?: (ttftMs: number) => void
+ setStreamMode?: (mode: SpinnerMode) => void
+ onCompactProgress?: (event: CompactProgressEvent) => void
+ setSDKStatus?: (status: SDKStatus) => void
+ openMessageSelector?: () => void
+ updateFileHistoryState: (
+ updater: (prev: FileHistoryState) => FileHistoryState,
+ ) => void
+ updateAttributionState: (
+ updater: (prev: AttributionState) => AttributionState,
+ ) => void
+ setConversationId?: (id: UUID) => void
+ agentId?: AgentId // Only set for subagents; use getSessionId() for session ID. Hooks use this to distinguish subagent calls.
+ agentType?: string // Subagent type name. For the main thread's --agent type, hooks fall back to getMainThreadAgentType().
+ /** When true, canUseTool must always be called even when hooks auto-approve.
+ * Used by speculation for overlay file path rewriting. */
+ requireCanUseTool?: boolean
+ messages: Message[]
+ fileReadingLimits?: {
+ maxTokens?: number
+ maxSizeBytes?: number
+ }
+ globLimits?: {
+ maxResults?: number
+ }
+ toolDecisions?: Map<
+ string,
+ {
+ source: string
+ decision: 'accept' | 'reject'
+ timestamp: number
+ }
+ >
+ queryTracking?: QueryChainTracking
+ /** Callback factory for requesting interactive prompts from the user.
+ * Returns a prompt callback bound to the given source name.
+ * Only available in interactive (REPL) contexts. */
+ requestPrompt?: (
+ sourceName: string,
+ toolInputSummary?: string | null,
+ ) => (request: PromptRequest) => Promise
+ toolUseId?: string
+ criticalSystemReminder_EXPERIMENTAL?: string
+ /** When true, preserve toolUseResult on messages even for subagents.
+ * Used by in-process teammates whose transcripts are viewable by the user. */
+ preserveToolUseResults?: boolean
+ /** Local denial tracking state for async subagents whose setAppState is a
+ * no-op. Without this, the denial counter never accumulates and the
+ * fallback-to-prompting threshold is never reached. Mutable โ the
+ * permissions code updates it in place. */
+ localDenialTracking?: DenialTrackingState
+ /**
+ * Per-conversation-thread content replacement state for the tool result
+ * budget. When present, query.ts applies the aggregate tool result budget.
+ * Main thread: REPL provisions once (never resets โ stale UUID keys
+ * are inert). Subagents: createSubagentContext clones the parent's state
+ * by default (cache-sharing forks need identical decisions), or
+ * resumeAgentBackground threads one reconstructed from sidechain records.
+ */
+ contentReplacementState?: ContentReplacementState
+ /**
+ * Parent's rendered system prompt bytes, frozen at turn start.
+ * Used by fork subagents to share the parent's prompt cache โ re-calling
+ * getSystemPrompt() at fork-spawn time can diverge (GrowthBook coldโwarm)
+ * and bust the cache. See forkSubagent.ts.
+ */
+ renderedSystemPrompt?: SystemPrompt
+}
+
+// Re-export ToolProgressData from centralized location
+export type { ToolProgressData }
+
+export type Progress = ToolProgressData | HookProgress
+
+export type ToolProgress
= {
+ toolUseID: string
+ data: P
+}
+
+export function filterToolProgressMessages(
+ progressMessagesForMessage: ProgressMessage[],
+): ProgressMessage[] {
+ return progressMessagesForMessage.filter(
+ (msg): msg is ProgressMessage =>
+ msg.data?.type !== 'hook_progress',
+ )
+}
+
+export type ToolResult = {
+ data: T
+ newMessages?: (
+ | UserMessage
+ | AssistantMessage
+ | AttachmentMessage
+ | SystemMessage
+ )[]
+ // contextModifier is only honored for tools that aren't concurrency safe.
+ contextModifier?: (context: ToolUseContext) => ToolUseContext
+ /** MCP protocol metadata (structuredContent, _meta) to pass through to SDK consumers */
+ mcpMeta?: {
+ _meta?: Record
+ structuredContent?: Record
+ }
+}
+
+export type ToolCallProgress
= (
+ progress: ToolProgress
,
+) => void
+
+// Type for any schema that outputs an object with string keys
+export type AnyObject = z.ZodType<{ [key: string]: unknown }>
+
+/**
+ * Checks if a tool matches the given name (primary name or alias).
+ */
+export function toolMatchesName(
+ tool: { name: string; aliases?: string[] },
+ name: string,
+): boolean {
+ return tool.name === name || (tool.aliases?.includes(name) ?? false)
+}
+
+/**
+ * Finds a tool by name or alias from a list of tools.
+ */
+export function findToolByName(tools: Tools, name: string): Tool | undefined {
+ return tools.find(t => toolMatchesName(t, name))
+}
+
+export type Tool<
+ Input extends AnyObject = AnyObject,
+ Output = unknown,
+ P extends ToolProgressData = ToolProgressData,
+> = {
+ /**
+ * Optional aliases for backwards compatibility when a tool is renamed.
+ * The tool can be looked up by any of these names in addition to its primary name.
+ */
+ aliases?: string[]
+ /**
+ * One-line capability phrase used by ToolSearch for keyword matching.
+ * Helps the model find this tool via keyword search when it's deferred.
+ * 3โ10 words, no trailing period.
+ * Prefer terms not already in the tool name (e.g. 'jupyter' for NotebookEdit).
+ */
+ searchHint?: string
+ call(
+ args: z.infer,
+ context: ToolUseContext,
+ canUseTool: CanUseToolFn,
+ parentMessage: AssistantMessage,
+ onProgress?: ToolCallProgress
,
+ ): Promise>
+ description(
+ input: z.infer,
+ options: {
+ isNonInteractiveSession: boolean
+ toolPermissionContext: ToolPermissionContext
+ tools: Tools
+ },
+ ): Promise
+ readonly inputSchema: Input
+ // Type for MCP tools that can specify their input schema directly in JSON Schema format
+ // rather than converting from Zod schema
+ readonly inputJSONSchema?: ToolInputJSONSchema
+ // Optional because TungstenTool doesn't define this. TODO: Make it required.
+ // When we do that, we can also go through and make this a bit more type-safe.
+ outputSchema?: z.ZodType
+ inputsEquivalent?(a: z.infer, b: z.infer): boolean
+ isConcurrencySafe(input: z.infer): boolean
+ isEnabled(): boolean
+ isReadOnly(input: z.infer): boolean
+ /** Defaults to false. Only set when the tool performs irreversible operations (delete, overwrite, send). */
+ isDestructive?(input: z.infer): boolean
+ /**
+ * What should happen when the user submits a new message while this tool
+ * is running.
+ *
+ * - `'cancel'` โ stop the tool and discard its result
+ * - `'block'` โ keep running; the new message waits
+ *
+ * Defaults to `'block'` when not implemented.
+ */
+ interruptBehavior?(): 'cancel' | 'block'
+ /**
+ * Returns information about whether this tool use is a search or read operation
+ * that should be collapsed into a condensed display in the UI. Examples include
+ * file searching (Grep, Glob), file reading (Read), and bash commands like find,
+ * grep, wc, etc.
+ *
+ * Returns an object indicating whether the operation is a search or read operation:
+ * - `isSearch: true` for search operations (grep, find, glob patterns)
+ * - `isRead: true` for read operations (cat, head, tail, file read)
+ * - `isList: true` for directory-listing operations (ls, tree, du)
+ * - All can be false if the operation shouldn't be collapsed
+ */
+ isSearchOrReadCommand?(input: z.infer): {
+ isSearch: boolean
+ isRead: boolean
+ isList?: boolean
+ }
+ isOpenWorld?(input: z.infer): boolean
+ requiresUserInteraction?(): boolean
+ isMcp?: boolean
+ isLsp?: boolean
+ /**
+ * When true, this tool is deferred (sent with defer_loading: true) and requires
+ * ToolSearch to be used before it can be called.
+ */
+ readonly shouldDefer?: boolean
+ /**
+ * When true, this tool is never deferred โ its full schema appears in the
+ * initial prompt even when ToolSearch is enabled. For MCP tools, set via
+ * `_meta['anthropic/alwaysLoad']`. Use for tools the model must see on
+ * turn 1 without a ToolSearch round-trip.
+ */
+ readonly alwaysLoad?: boolean
+ /**
+ * For MCP tools: the server and tool names as received from the MCP server (unnormalized).
+ * Present on all MCP tools regardless of whether `name` is prefixed (mcp__server__tool)
+ * or unprefixed (CLAUDE_AGENT_SDK_MCP_NO_PREFIX mode).
+ */
+ mcpInfo?: { serverName: string; toolName: string }
+ readonly name: string
+ /**
+ * Maximum size in characters for tool result before it gets persisted to disk.
+ * When exceeded, the result is saved to a file and Claude receives a preview
+ * with the file path instead of the full content.
+ *
+ * Set to Infinity for tools whose output must never be persisted (e.g. Read,
+ * where persisting creates a circular ReadโfileโRead loop and the tool
+ * already self-bounds via its own limits).
+ */
+ maxResultSizeChars: number
+ /**
+ * When true, enables strict mode for this tool, which causes the API to
+ * more strictly adhere to tool instructions and parameter schemas.
+ * Only applied when the tengu_tool_pear is enabled.
+ */
+ readonly strict?: boolean
+
+ /**
+ * Called on copies of tool_use input before observers see it (SDK stream,
+ * transcript, canUseTool, PreToolUse/PostToolUse hooks). Mutate in place
+ * to add legacy/derived fields. Must be idempotent. The original API-bound
+ * input is never mutated (preserves prompt cache). Not re-applied when a
+ * hook/permission returns a fresh updatedInput โ those own their shape.
+ */
+ backfillObservableInput?(input: Record): void
+
+ /**
+ * Determines if this tool is allowed to run with this input in the current context.
+ * It informs the model of why the tool use failed, and does not directly display any UI.
+ * @param input
+ * @param context
+ */
+ validateInput?(
+ input: z.infer,
+ context: ToolUseContext,
+ ): Promise
+
+ /**
+ * Determines if the user is asked for permission. Only called after validateInput() passes.
+ * General permission logic is in permissions.ts. This method contains tool-specific logic.
+ * @param input
+ * @param context
+ */
+ checkPermissions(
+ input: z.infer,
+ context: ToolUseContext,
+ ): Promise
+
+ // Optional method for tools that operate on a file path
+ getPath?(input: z.infer): string
+
+ /**
+ * Prepare a matcher for hook `if` conditions (permission-rule patterns like
+ * "git *" from "Bash(git *)"). Called once per hook-input pair; any
+ * expensive parsing happens here. Returns a closure that is called per
+ * hook pattern. If not implemented, only tool-name-level matching works.
+ */
+ preparePermissionMatcher?(
+ input: z.infer,
+ ): Promise<(pattern: string) => boolean>
+
+ prompt(options: {
+ getToolPermissionContext: () => Promise
+ tools: Tools
+ agents: AgentDefinition[]
+ allowedAgentTypes?: string[]
+ }): Promise
+ userFacingName(input: Partial> | undefined): string
+ userFacingNameBackgroundColor?(
+ input: Partial> | undefined,
+ ): keyof Theme | undefined
+ /**
+ * Transparent wrappers (e.g. REPL) delegate all rendering to their progress
+ * handler, which emits native-looking blocks for each inner tool call.
+ * The wrapper itself shows nothing.
+ */
+ isTransparentWrapper?(): boolean
+ /**
+ * Returns a short string summary of this tool use for display in compact views.
+ * @param input The tool input
+ * @returns A short string summary, or null to not display
+ */
+ getToolUseSummary?(input: Partial> | undefined): string | null
+ /**
+ * Returns a human-readable present-tense activity description for spinner display.
+ * Example: "Reading src/foo.ts", "Running bun test", "Searching for pattern"
+ * @param input The tool input
+ * @returns Activity description string, or null to fall back to tool name
+ */
+ getActivityDescription?(
+ input: Partial> | undefined,
+ ): string | null
+ /**
+ * Returns a compact representation of this tool use for the auto-mode
+ * security classifier. Examples: `ls -la` for Bash, `/tmp/x: new content`
+ * for Edit. Return '' to skip this tool in the classifier transcript
+ * (e.g. tools with no security relevance). May return an object to avoid
+ * double-encoding when the caller JSON-wraps the value.
+ */
+ toAutoClassifierInput(input: z.infer): unknown
+ mapToolResultToToolResultBlockParam(
+ content: Output,
+ toolUseID: string,
+ ): ToolResultBlockParam
+ /**
+ * Optional. When omitted, the tool result renders nothing (same as returning
+ * null). Omit for tools whose results are surfaced elsewhere (e.g., TodoWrite
+ * updates the todo panel, not the transcript).
+ */
+ renderToolResultMessage?(
+ content: Output,
+ progressMessagesForMessage: ProgressMessage
[],
+ options: {
+ style?: 'condensed'
+ theme: ThemeName
+ tools: Tools
+ verbose: boolean
+ isTranscriptMode?: boolean
+ isBriefOnly?: boolean
+ /** Original tool_use input, when available. Useful for compact result
+ * summaries that reference what was requested (e.g. "Sent to #foo"). */
+ input?: unknown
+ },
+ ): React.ReactNode
+ /**
+ * Flattened text of what renderToolResultMessage shows IN TRANSCRIPT
+ * MODE (verbose=true, isTranscriptMode=true). For transcript search
+ * indexing: the index counts occurrences in this string, the highlight
+ * overlay scans the actual screen buffer. For count โก highlight, this
+ * must return the text that ends up visible โ not the model-facing
+ * serialization from mapToolResultToToolResultBlockParam (which adds
+ * system-reminders, persisted-output wrappers).
+ *
+ * Chrome can be skipped (under-count is fine). "Found 3 files in 12ms"
+ * isn't worth indexing. Phantoms are not fine โ text that's claimed
+ * here but doesn't render is a countโ highlight bug.
+ *
+ * Optional: omitted โ field-name heuristic in transcriptSearch.ts.
+ * Drift caught by test/utils/transcriptSearch.renderFidelity.test.tsx
+ * which renders sample outputs and flags text that's indexed-but-not-
+ * rendered (phantom) or rendered-but-not-indexed (under-count warning).
+ */
+ extractSearchText?(out: Output): string
+ /**
+ * Render the tool use message. Note that `input` is partial because we render
+ * the message as soon as possible, possibly before tool parameters have fully
+ * streamed in.
+ */
+ renderToolUseMessage(
+ input: Partial>,
+ options: { theme: ThemeName; verbose: boolean; commands?: Command[] },
+ ): React.ReactNode
+ /**
+ * Returns true when the non-verbose rendering of this output is truncated
+ * (i.e., clicking to expand would reveal more content). Gates
+ * click-to-expand in fullscreen โ only messages where verbose actually
+ * shows more get a hover/click affordance. Unset means never truncated.
+ */
+ isResultTruncated?(output: Output): boolean
+ /**
+ * Renders an optional tag to display after the tool use message.
+ * Used for additional metadata like timeout, model, resume ID, etc.
+ * Returns null to not display anything.
+ */
+ renderToolUseTag?(input: Partial>): React.ReactNode
+ /**
+ * Optional. When omitted, no progress UI is shown while the tool runs.
+ */
+ renderToolUseProgressMessage?(
+ progressMessagesForMessage: ProgressMessage
[],
+ options: {
+ tools: Tools
+ verbose: boolean
+ terminalSize?: { columns: number; rows: number }
+ inProgressToolCallCount?: number
+ isTranscriptMode?: boolean
+ },
+ ): React.ReactNode
+ renderToolUseQueuedMessage?(): React.ReactNode
+ /**
+ * Optional. When omitted, falls back to .
+ * Only define this for tools that need custom rejection UI (e.g., file edits
+ * that show the rejected diff).
+ */
+ renderToolUseRejectedMessage?(
+ input: z.infer,
+ options: {
+ columns: number
+ messages: Message[]
+ style?: 'condensed'
+ theme: ThemeName
+ tools: Tools
+ verbose: boolean
+ progressMessagesForMessage: ProgressMessage
[]
+ isTranscriptMode?: boolean
+ },
+ ): React.ReactNode
+ /**
+ * Optional. When omitted, falls back to .
+ * Only define this for tools that need custom error UI (e.g., search tools
+ * that show "File not found" instead of the raw error).
+ */
+ renderToolUseErrorMessage?(
+ result: ToolResultBlockParam['content'],
+ options: {
+ progressMessagesForMessage: ProgressMessage
[]
+ tools: Tools
+ verbose: boolean
+ isTranscriptMode?: boolean
+ },
+ ): React.ReactNode
+
+ /**
+ * Renders multiple parallel instances of this tool as a group.
+ * @returns React node to render, or null to fall back to individual rendering
+ */
+ /**
+ * Renders multiple tool uses as a group (non-verbose mode only).
+ * In verbose mode, individual tool uses render at their original positions.
+ * @returns React node to render, or null to fall back to individual rendering
+ */
+ renderGroupedToolUse?(
+ toolUses: Array<{
+ param: ToolUseBlockParam
+ isResolved: boolean
+ isError: boolean
+ isInProgress: boolean
+ progressMessages: ProgressMessage
[]
+ result?: {
+ param: ToolResultBlockParam
+ output: unknown
+ }
+ }>,
+ options: {
+ shouldAnimate: boolean
+ tools: Tools
+ },
+ ): React.ReactNode | null
+}
+
+/**
+ * A collection of tools. Use this type instead of `Tool[]` to make it easier
+ * to track where tool sets are assembled, passed, and filtered across the codebase.
+ */
+export type Tools = readonly Tool[]
+
+/**
+ * Methods that `buildTool` supplies a default for. A `ToolDef` may omit these;
+ * the resulting `Tool` always has them.
+ */
+type DefaultableToolKeys =
+ | 'isEnabled'
+ | 'isConcurrencySafe'
+ | 'isReadOnly'
+ | 'isDestructive'
+ | 'checkPermissions'
+ | 'toAutoClassifierInput'
+ | 'userFacingName'
+
+/**
+ * Tool definition accepted by `buildTool`. Same shape as `Tool` but with the
+ * defaultable methods optional โ `buildTool` fills them in so callers always
+ * see a complete `Tool`.
+ */
+export type ToolDef<
+ Input extends AnyObject = AnyObject,
+ Output = unknown,
+ P extends ToolProgressData = ToolProgressData,
+> = Omit, DefaultableToolKeys> &
+ Partial, DefaultableToolKeys>>
+
+/**
+ * Type-level spread mirroring `{ ...TOOL_DEFAULTS, ...def }`. For each
+ * defaultable key: if D provides it (required), D's type wins; if D omits
+ * it or has it optional (inherited from Partial<> in the constraint), the
+ * default fills in. All other keys come from D verbatim โ preserving arity,
+ * optional presence, and literal types exactly as `satisfies Tool` did.
+ */
+type BuiltTool = Omit & {
+ [K in DefaultableToolKeys]-?: K extends keyof D
+ ? undefined extends D[K]
+ ? ToolDefaults[K]
+ : D[K]
+ : ToolDefaults[K]
+}
+
+/**
+ * Build a complete `Tool` from a partial definition, filling in safe defaults
+ * for the commonly-stubbed methods. All tool exports should go through this so
+ * that defaults live in one place and callers never need `?.() ?? default`.
+ *
+ * Defaults (fail-closed where it matters):
+ * - `isEnabled` โ `true`
+ * - `isConcurrencySafe` โ `false` (assume not safe)
+ * - `isReadOnly` โ `false` (assume writes)
+ * - `isDestructive` โ `false`
+ * - `checkPermissions` โ `{ behavior: 'allow', updatedInput }` (defer to general permission system)
+ * - `toAutoClassifierInput` โ `''` (skip classifier โ security-relevant tools must override)
+ * - `userFacingName` โ `name`
+ */
+const TOOL_DEFAULTS = {
+ isEnabled: () => true,
+ isConcurrencySafe: (_input?: unknown) => false,
+ isReadOnly: (_input?: unknown) => false,
+ isDestructive: (_input?: unknown) => false,
+ checkPermissions: (
+ input: { [key: string]: unknown },
+ _ctx?: ToolUseContext,
+ ): Promise =>
+ Promise.resolve({ behavior: 'allow', updatedInput: input }),
+ toAutoClassifierInput: (_input?: unknown) => '',
+ userFacingName: (_input?: unknown) => '',
+}
+
+// The defaults type is the ACTUAL shape of TOOL_DEFAULTS (optional params so
+// both 0-arg and full-arg call sites type-check โ stubs varied in arity and
+// tests relied on that), not the interface's strict signatures.
+type ToolDefaults = typeof TOOL_DEFAULTS
+
+// D infers the concrete object-literal type from the call site. The
+// constraint provides contextual typing for method parameters; `any` in
+// constraint position is structural and never leaks into the return type.
+// BuiltTool mirrors runtime `{...TOOL_DEFAULTS, ...def}` at the type level.
+// eslint-disable-next-line @typescript-eslint/no-explicit-any
+type AnyToolDef = ToolDef
+
+export function buildTool(def: D): BuiltTool {
+ // The runtime spread is straightforward; the `as` bridges the gap between
+ // the structural-any constraint and the precise BuiltTool return. The
+ // type semantics are proven by the 0-error typecheck across all 60+ tools.
+ return {
+ ...TOOL_DEFAULTS,
+ userFacingName: () => def.name,
+ ...def,
+ } as BuiltTool
+}
diff --git a/original-source-code/src/assistant/sessionHistory.ts b/original-source-code/src/assistant/sessionHistory.ts
new file mode 100644
index 0000000..9e1ddc5
--- /dev/null
+++ b/original-source-code/src/assistant/sessionHistory.ts
@@ -0,0 +1,87 @@
+import axios from 'axios'
+import { getOauthConfig } from '../constants/oauth.js'
+import type { SDKMessage } from '../entrypoints/agentSdkTypes.js'
+import { logForDebugging } from '../utils/debug.js'
+import { getOAuthHeaders, prepareApiRequest } from '../utils/teleport/api.js'
+
+export const HISTORY_PAGE_SIZE = 100
+
+export type HistoryPage = {
+ /** Chronological order within the page. */
+ events: SDKMessage[]
+ /** Oldest event ID in this page โ before_id cursor for next-older page. */
+ firstId: string | null
+ /** true = older events exist. */
+ hasMore: boolean
+}
+
+type SessionEventsResponse = {
+ data: SDKMessage[]
+ has_more: boolean
+ first_id: string | null
+ last_id: string | null
+}
+
+export type HistoryAuthCtx = {
+ baseUrl: string
+ headers: Record
+}
+
+/** Prepare auth + headers + base URL once, reuse across pages. */
+export async function createHistoryAuthCtx(
+ sessionId: string,
+): Promise {
+ const { accessToken, orgUUID } = await prepareApiRequest()
+ return {
+ baseUrl: `${getOauthConfig().BASE_API_URL}/v1/sessions/${sessionId}/events`,
+ headers: {
+ ...getOAuthHeaders(accessToken),
+ 'anthropic-beta': 'ccr-byoc-2025-07-29',
+ 'x-organization-uuid': orgUUID,
+ },
+ }
+}
+
+async function fetchPage(
+ ctx: HistoryAuthCtx,
+ params: Record,
+ label: string,
+): Promise {
+ const resp = await axios
+ .get(ctx.baseUrl, {
+ headers: ctx.headers,
+ params,
+ timeout: 15000,
+ validateStatus: () => true,
+ })
+ .catch(() => null)
+ if (!resp || resp.status !== 200) {
+ logForDebugging(`[${label}] HTTP ${resp?.status ?? 'error'}`)
+ return null
+ }
+ return {
+ events: Array.isArray(resp.data.data) ? resp.data.data : [],
+ firstId: resp.data.first_id,
+ hasMore: resp.data.has_more,
+ }
+}
+
+/**
+ * Newest page: last `limit` events, chronological, via anchor_to_latest.
+ * has_more=true means older events exist.
+ */
+export async function fetchLatestEvents(
+ ctx: HistoryAuthCtx,
+ limit = HISTORY_PAGE_SIZE,
+): Promise {
+ return fetchPage(ctx, { limit, anchor_to_latest: true }, 'fetchLatestEvents')
+}
+
+/** Older page: events immediately before `beforeId` cursor. */
+export async function fetchOlderEvents(
+ ctx: HistoryAuthCtx,
+ beforeId: string,
+ limit = HISTORY_PAGE_SIZE,
+): Promise {
+ return fetchPage(ctx, { limit, before_id: beforeId }, 'fetchOlderEvents')
+}
diff --git a/original-source-code/src/bootstrap/state.ts b/original-source-code/src/bootstrap/state.ts
new file mode 100644
index 0000000..d7199e5
--- /dev/null
+++ b/original-source-code/src/bootstrap/state.ts
@@ -0,0 +1,1758 @@
+import type { BetaMessageStreamParams } from '@anthropic-ai/sdk/resources/beta/messages/messages.mjs'
+import type { Attributes, Meter, MetricOptions } from '@opentelemetry/api'
+import type { logs } from '@opentelemetry/api-logs'
+import type { LoggerProvider } from '@opentelemetry/sdk-logs'
+import type { MeterProvider } from '@opentelemetry/sdk-metrics'
+import type { BasicTracerProvider } from '@opentelemetry/sdk-trace-base'
+import { realpathSync } from 'fs'
+import sumBy from 'lodash-es/sumBy.js'
+import { cwd } from 'process'
+import type { HookEvent, ModelUsage } from 'src/entrypoints/agentSdkTypes.js'
+import type { AgentColorName } from 'src/tools/AgentTool/agentColorManager.js'
+import type { HookCallbackMatcher } from 'src/types/hooks.js'
+// Indirection for browser-sdk build (package.json "browser" field swaps
+// crypto.ts for crypto.browser.ts). Pure leaf re-export of node:crypto โ
+// zero circular-dep risk. Path-alias import bypasses bootstrap-isolation
+// (rule only checks ./ and / prefixes); explicit disable documents intent.
+// eslint-disable-next-line custom-rules/bootstrap-isolation
+import { randomUUID } from 'src/utils/crypto.js'
+import type { ModelSetting } from 'src/utils/model/model.js'
+import type { ModelStrings } from 'src/utils/model/modelStrings.js'
+import type { SettingSource } from 'src/utils/settings/constants.js'
+import { resetSettingsCache } from 'src/utils/settings/settingsCache.js'
+import type { PluginHookMatcher } from 'src/utils/settings/types.js'
+import { createSignal } from 'src/utils/signal.js'
+
+// Union type for registered hooks - can be SDK callbacks or native plugin hooks
+type RegisteredHookMatcher = HookCallbackMatcher | PluginHookMatcher
+
+import type { SessionId } from 'src/types/ids.js'
+
+// DO NOT ADD MORE STATE HERE - BE JUDICIOUS WITH GLOBAL STATE
+
+// dev: true on entries that came via --dangerously-load-development-channels.
+// The allowlist gate checks this per-entry (not the session-wide
+// hasDevChannels bit) so passing both flags doesn't let the dev dialog's
+// acceptance leak allowlist-bypass to the --channels entries.
+export type ChannelEntry =
+ | { kind: 'plugin'; name: string; marketplace: string; dev?: boolean }
+ | { kind: 'server'; name: string; dev?: boolean }
+
+export type AttributedCounter = {
+ add(value: number, additionalAttributes?: Attributes): void
+}
+
+type State = {
+ originalCwd: string
+ // Stable project root - set once at startup (including by --worktree flag),
+ // never updated by mid-session EnterWorktreeTool.
+ // Use for project identity (history, skills, sessions) not file operations.
+ projectRoot: string
+ totalCostUSD: number
+ totalAPIDuration: number
+ totalAPIDurationWithoutRetries: number
+ totalToolDuration: number
+ turnHookDurationMs: number
+ turnToolDurationMs: number
+ turnClassifierDurationMs: number
+ turnToolCount: number
+ turnHookCount: number
+ turnClassifierCount: number
+ startTime: number
+ lastInteractionTime: number
+ totalLinesAdded: number
+ totalLinesRemoved: number
+ hasUnknownModelCost: boolean
+ cwd: string
+ modelUsage: { [modelName: string]: ModelUsage }
+ mainLoopModelOverride: ModelSetting | undefined
+ initialMainLoopModel: ModelSetting
+ modelStrings: ModelStrings | null
+ isInteractive: boolean
+ kairosActive: boolean
+ // When true, ensureToolResultPairing throws on mismatch instead of
+ // repairing with synthetic placeholders. HFI opts in at startup so
+ // trajectories fail fast rather than conditioning the model on fake
+ // tool_results.
+ strictToolResultPairing: boolean
+ sdkAgentProgressSummariesEnabled: boolean
+ userMsgOptIn: boolean
+ clientType: string
+ sessionSource: string | undefined
+ questionPreviewFormat: 'markdown' | 'html' | undefined
+ flagSettingsPath: string | undefined
+ flagSettingsInline: Record | null
+ allowedSettingSources: SettingSource[]
+ sessionIngressToken: string | null | undefined
+ oauthTokenFromFd: string | null | undefined
+ apiKeyFromFd: string | null | undefined
+ // Telemetry state
+ meter: Meter | null
+ sessionCounter: AttributedCounter | null
+ locCounter: AttributedCounter | null
+ prCounter: AttributedCounter | null
+ commitCounter: AttributedCounter | null
+ costCounter: AttributedCounter | null
+ tokenCounter: AttributedCounter | null
+ codeEditToolDecisionCounter: AttributedCounter | null
+ activeTimeCounter: AttributedCounter | null
+ statsStore: { observe(name: string, value: number): void } | null
+ sessionId: SessionId
+ // Parent session ID for tracking session lineage (e.g., plan mode -> implementation)
+ parentSessionId: SessionId | undefined
+ // Logger state
+ loggerProvider: LoggerProvider | null
+ eventLogger: ReturnType | null
+ // Meter provider state
+ meterProvider: MeterProvider | null
+ // Tracer provider state
+ tracerProvider: BasicTracerProvider | null
+ // Agent color state
+ agentColorMap: Map
+ agentColorIndex: number
+ // Last API request for bug reports
+ lastAPIRequest: Omit | null
+ // Messages from the last API request (ant-only; reference, not clone).
+ // Captures the exact post-compaction, CLAUDE.md-injected message set sent
+ // to the API so /share's serialized_conversation.json reflects reality.
+ lastAPIRequestMessages: BetaMessageStreamParams['messages'] | null
+ // Last auto-mode classifier request(s) for /share transcript
+ lastClassifierRequests: unknown[] | null
+ // CLAUDE.md content cached by context.ts for the auto-mode classifier.
+ // Breaks the yoloClassifier โ claudemd โ filesystem โ permissions cycle.
+ cachedClaudeMdContent: string | null
+ // In-memory error log for recent errors
+ inMemoryErrorLog: Array<{ error: string; timestamp: string }>
+ // Session-only plugins from --plugin-dir flag
+ inlinePlugins: Array
+ // Explicit --chrome / --no-chrome flag value (undefined = not set on CLI)
+ chromeFlagOverride: boolean | undefined
+ // Use cowork_plugins directory instead of plugins (--cowork flag or env var)
+ useCoworkPlugins: boolean
+ // Session-only bypass permissions mode flag (not persisted)
+ sessionBypassPermissionsMode: boolean
+ // Session-only flag gating the .claude/scheduled_tasks.json watcher
+ // (useScheduledTasks). Set by cronScheduler.start() when the JSON has
+ // entries, or by CronCreateTool. Not persisted.
+ scheduledTasksEnabled: boolean
+ // Session-only cron tasks created via CronCreate with durable: false.
+ // Fire on schedule like file-backed tasks but are never written to
+ // .claude/scheduled_tasks.json โ they die with the process. Typed via
+ // SessionCronTask below (not importing from cronTasks.ts keeps
+ // bootstrap a leaf of the import DAG).
+ sessionCronTasks: SessionCronTask[]
+ // Teams created this session via TeamCreate. cleanupSessionTeams()
+ // removes these on gracefulShutdown so subagent-created teams don't
+ // persist on disk forever (gh-32730). TeamDelete removes entries to
+ // avoid double-cleanup. Lives here (not teamHelpers.ts) so
+ // resetStateForTests() clears it between tests.
+ sessionCreatedTeams: Set
+ // Session-only trust flag for home directory (not persisted to disk)
+ // When running from home dir, trust dialog is shown but not saved to disk.
+ // This flag allows features requiring trust to work during the session.
+ sessionTrustAccepted: boolean
+ // Session-only flag to disable session persistence to disk
+ sessionPersistenceDisabled: boolean
+ // Track if user has exited plan mode in this session (for re-entry guidance)
+ hasExitedPlanMode: boolean
+ // Track if we need to show the plan mode exit attachment (one-time notification)
+ needsPlanModeExitAttachment: boolean
+ // Track if we need to show the auto mode exit attachment (one-time notification)
+ needsAutoModeExitAttachment: boolean
+ // Track if LSP plugin recommendation has been shown this session (only show once)
+ lspRecommendationShownThisSession: boolean
+ // SDK init event state - jsonSchema for structured output
+ initJsonSchema: Record | null
+ // Registered hooks - SDK callbacks and plugin native hooks
+ registeredHooks: Partial> | null
+ // Cache for plan slugs: sessionId -> wordSlug
+ planSlugCache: Map
+ // Track teleported session for reliability logging
+ teleportedSessionInfo: {
+ isTeleported: boolean
+ hasLoggedFirstMessage: boolean
+ sessionId: string | null
+ } | null
+ // Track invoked skills for preservation across compaction
+ // Keys are composite: `${agentId ?? ''}:${skillName}` to prevent cross-agent overwrites
+ invokedSkills: Map<
+ string,
+ {
+ skillName: string
+ skillPath: string
+ content: string
+ invokedAt: number
+ agentId: string | null
+ }
+ >
+ // Track slow operations for dev bar display (ant-only)
+ slowOperations: Array<{
+ operation: string
+ durationMs: number
+ timestamp: number
+ }>
+ // SDK-provided betas (e.g., context-1m-2025-08-07)
+ sdkBetas: string[] | undefined
+ // Main thread agent type (from --agent flag or settings)
+ mainThreadAgentType: string | undefined
+ // Remote mode (--remote flag)
+ isRemoteMode: boolean
+ // Direct connect server URL (for display in header)
+ directConnectServerUrl: string | undefined
+ // System prompt section cache state
+ systemPromptSectionCache: Map
+ // Last date emitted to the model (for detecting midnight date changes)
+ lastEmittedDate: string | null
+ // Additional directories from --add-dir flag (for CLAUDE.md loading)
+ additionalDirectoriesForClaudeMd: string[]
+ // Channel server allowlist from --channels flag (servers whose channel
+ // notifications should register this session). Parsed once in main.tsx โ
+ // the tag decides trust model: 'plugin' โ marketplace verification +
+ // allowlist, 'server' โ allowlist always fails (schema is plugin-only).
+ // Either kind needs entry.dev to bypass allowlist.
+ allowedChannels: ChannelEntry[]
+ // True if any entry in allowedChannels came from
+ // --dangerously-load-development-channels (so ChannelsNotice can name the
+ // right flag in policy-blocked messages)
+ hasDevChannels: boolean
+ // Dir containing the session's `.jsonl`; null = derive from originalCwd.
+ sessionProjectDir: string | null
+ // Cached prompt cache 1h TTL allowlist from GrowthBook (session-stable)
+ promptCache1hAllowlist: string[] | null
+ // Cached 1h TTL user eligibility (session-stable). Latched on first
+ // evaluation so mid-session overage flips don't change the cache_control
+ // TTL, which would bust the server-side prompt cache.
+ promptCache1hEligible: boolean | null
+ // Sticky-on latch for AFK_MODE_BETA_HEADER. Once auto mode is first
+ // activated, keep sending the header for the rest of the session so
+ // Shift+Tab toggles don't bust the ~50-70K token prompt cache.
+ afkModeHeaderLatched: boolean | null
+ // Sticky-on latch for FAST_MODE_BETA_HEADER. Once fast mode is first
+ // enabled, keep sending the header so cooldown enter/exit doesn't
+ // double-bust the prompt cache. The `speed` body param stays dynamic.
+ fastModeHeaderLatched: boolean | null
+ // Sticky-on latch for the cache-editing beta header. Once cached
+ // microcompact is first enabled, keep sending the header so mid-session
+ // GrowthBook/settings toggles don't bust the prompt cache.
+ cacheEditingHeaderLatched: boolean | null
+ // Sticky-on latch for clearing thinking from prior tool loops. Triggered
+ // when >1h since last API call (confirmed cache miss โ no cache-hit
+ // benefit to keeping thinking). Once latched, stays on so the newly-warmed
+ // thinking-cleared cache isn't busted by flipping back to keep:'all'.
+ thinkingClearLatched: boolean | null
+ // Current prompt ID (UUID) correlating a user prompt with subsequent OTel events
+ promptId: string | null
+ // Last API requestId for the main conversation chain (not subagents).
+ // Updated after each successful API response for main-session queries.
+ // Read at shutdown to send cache eviction hints to inference.
+ lastMainRequestId: string | undefined
+ // Timestamp (Date.now()) of the last successful API call completion.
+ // Used to compute timeSinceLastApiCallMs in tengu_api_success for
+ // correlating cache misses with idle time (cache TTL is ~5min).
+ lastApiCompletionTimestamp: number | null
+ // Set to true after compaction (auto or manual /compact). Consumed by
+ // logAPISuccess to tag the first post-compaction API call so we can
+ // distinguish compaction-induced cache misses from TTL expiry.
+ pendingPostCompaction: boolean
+}
+
+// ALSO HERE - THINK THRICE BEFORE MODIFYING
+function getInitialState(): State {
+ // Resolve symlinks in cwd to match behavior of shell.ts setCwd
+ // This ensures consistency with how paths are sanitized for session storage
+ let resolvedCwd = ''
+ if (
+ typeof process !== 'undefined' &&
+ typeof process.cwd === 'function' &&
+ typeof realpathSync === 'function'
+ ) {
+ const rawCwd = cwd()
+ try {
+ resolvedCwd = realpathSync(rawCwd).normalize('NFC')
+ } catch {
+ // File Provider EPERM on CloudStorage mounts (lstat per path component).
+ resolvedCwd = rawCwd.normalize('NFC')
+ }
+ }
+ const state: State = {
+ originalCwd: resolvedCwd,
+ projectRoot: resolvedCwd,
+ totalCostUSD: 0,
+ totalAPIDuration: 0,
+ totalAPIDurationWithoutRetries: 0,
+ totalToolDuration: 0,
+ turnHookDurationMs: 0,
+ turnToolDurationMs: 0,
+ turnClassifierDurationMs: 0,
+ turnToolCount: 0,
+ turnHookCount: 0,
+ turnClassifierCount: 0,
+ startTime: Date.now(),
+ lastInteractionTime: Date.now(),
+ totalLinesAdded: 0,
+ totalLinesRemoved: 0,
+ hasUnknownModelCost: false,
+ cwd: resolvedCwd,
+ modelUsage: {},
+ mainLoopModelOverride: undefined,
+ initialMainLoopModel: null,
+ modelStrings: null,
+ isInteractive: false,
+ kairosActive: false,
+ strictToolResultPairing: false,
+ sdkAgentProgressSummariesEnabled: false,
+ userMsgOptIn: false,
+ clientType: 'cli',
+ sessionSource: undefined,
+ questionPreviewFormat: undefined,
+ sessionIngressToken: undefined,
+ oauthTokenFromFd: undefined,
+ apiKeyFromFd: undefined,
+ flagSettingsPath: undefined,
+ flagSettingsInline: null,
+ allowedSettingSources: [
+ 'userSettings',
+ 'projectSettings',
+ 'localSettings',
+ 'flagSettings',
+ 'policySettings',
+ ],
+ // Telemetry state
+ meter: null,
+ sessionCounter: null,
+ locCounter: null,
+ prCounter: null,
+ commitCounter: null,
+ costCounter: null,
+ tokenCounter: null,
+ codeEditToolDecisionCounter: null,
+ activeTimeCounter: null,
+ statsStore: null,
+ sessionId: randomUUID() as SessionId,
+ parentSessionId: undefined,
+ // Logger state
+ loggerProvider: null,
+ eventLogger: null,
+ // Meter provider state
+ meterProvider: null,
+ tracerProvider: null,
+ // Agent color state
+ agentColorMap: new Map(),
+ agentColorIndex: 0,
+ // Last API request for bug reports
+ lastAPIRequest: null,
+ lastAPIRequestMessages: null,
+ // Last auto-mode classifier request(s) for /share transcript
+ lastClassifierRequests: null,
+ cachedClaudeMdContent: null,
+ // In-memory error log for recent errors
+ inMemoryErrorLog: [],
+ // Session-only plugins from --plugin-dir flag
+ inlinePlugins: [],
+ // Explicit --chrome / --no-chrome flag value (undefined = not set on CLI)
+ chromeFlagOverride: undefined,
+ // Use cowork_plugins directory instead of plugins
+ useCoworkPlugins: false,
+ // Session-only bypass permissions mode flag (not persisted)
+ sessionBypassPermissionsMode: false,
+ // Scheduled tasks disabled until flag or dialog enables them
+ scheduledTasksEnabled: false,
+ sessionCronTasks: [],
+ sessionCreatedTeams: new Set(),
+ // Session-only trust flag (not persisted to disk)
+ sessionTrustAccepted: false,
+ // Session-only flag to disable session persistence to disk
+ sessionPersistenceDisabled: false,
+ // Track if user has exited plan mode in this session
+ hasExitedPlanMode: false,
+ // Track if we need to show the plan mode exit attachment
+ needsPlanModeExitAttachment: false,
+ // Track if we need to show the auto mode exit attachment
+ needsAutoModeExitAttachment: false,
+ // Track if LSP plugin recommendation has been shown this session
+ lspRecommendationShownThisSession: false,
+ // SDK init event state
+ initJsonSchema: null,
+ registeredHooks: null,
+ // Cache for plan slugs
+ planSlugCache: new Map(),
+ // Track teleported session for reliability logging
+ teleportedSessionInfo: null,
+ // Track invoked skills for preservation across compaction
+ invokedSkills: new Map(),
+ // Track slow operations for dev bar display
+ slowOperations: [],
+ // SDK-provided betas
+ sdkBetas: undefined,
+ // Main thread agent type
+ mainThreadAgentType: undefined,
+ // Remote mode
+ isRemoteMode: false,
+ ...(process.env.USER_TYPE === 'ant'
+ ? {
+ replBridgeActive: false,
+ }
+ : {}),
+ // Direct connect server URL
+ directConnectServerUrl: undefined,
+ // System prompt section cache state
+ systemPromptSectionCache: new Map(),
+ // Last date emitted to the model
+ lastEmittedDate: null,
+ // Additional directories from --add-dir flag (for CLAUDE.md loading)
+ additionalDirectoriesForClaudeMd: [],
+ // Channel server allowlist from --channels flag
+ allowedChannels: [],
+ hasDevChannels: false,
+ // Session project dir (null = derive from originalCwd)
+ sessionProjectDir: null,
+ // Prompt cache 1h allowlist (null = not yet fetched from GrowthBook)
+ promptCache1hAllowlist: null,
+ // Prompt cache 1h eligibility (null = not yet evaluated)
+ promptCache1hEligible: null,
+ // Beta header latches (null = not yet triggered)
+ afkModeHeaderLatched: null,
+ fastModeHeaderLatched: null,
+ cacheEditingHeaderLatched: null,
+ thinkingClearLatched: null,
+ // Current prompt ID
+ promptId: null,
+ lastMainRequestId: undefined,
+ lastApiCompletionTimestamp: null,
+ pendingPostCompaction: false,
+ }
+
+ return state
+}
+
+// AND ESPECIALLY HERE
+const STATE: State = getInitialState()
+
+export function getSessionId(): SessionId {
+ return STATE.sessionId
+}
+
+export function regenerateSessionId(
+ options: { setCurrentAsParent?: boolean } = {},
+): SessionId {
+ if (options.setCurrentAsParent) {
+ STATE.parentSessionId = STATE.sessionId
+ }
+ // Drop the outgoing session's plan-slug entry so the Map doesn't
+ // accumulate stale keys. Callers that need to carry the slug across
+ // (REPL.tsx clearContext) read it before calling clearConversation.
+ STATE.planSlugCache.delete(STATE.sessionId)
+ // Regenerated sessions live in the current project: reset projectDir to
+ // null so getTranscriptPath() derives from originalCwd.
+ STATE.sessionId = randomUUID() as SessionId
+ STATE.sessionProjectDir = null
+ return STATE.sessionId
+}
+
+export function getParentSessionId(): SessionId | undefined {
+ return STATE.parentSessionId
+}
+
+/**
+ * Atomically switch the active session. `sessionId` and `sessionProjectDir`
+ * always change together โ there is no separate setter for either, so they
+ * cannot drift out of sync (CC-34).
+ *
+ * @param projectDir โ directory containing `.jsonl`. Omit (or
+ * pass `null`) for sessions in the current project โ the path will derive
+ * from originalCwd at read time. Pass `dirname(transcriptPath)` when the
+ * session lives in a different project directory (git worktrees,
+ * cross-project resume). Every call resets the project dir; it never
+ * carries over from the previous session.
+ */
+export function switchSession(
+ sessionId: SessionId,
+ projectDir: string | null = null,
+): void {
+ // Drop the outgoing session's plan-slug entry so the Map stays bounded
+ // across repeated /resume. Only the current session's slug is ever read
+ // (plans.ts getPlanSlug defaults to getSessionId()).
+ STATE.planSlugCache.delete(STATE.sessionId)
+ STATE.sessionId = sessionId
+ STATE.sessionProjectDir = projectDir
+ sessionSwitched.emit(sessionId)
+}
+
+const sessionSwitched = createSignal<[id: SessionId]>()
+
+/**
+ * Register a callback that fires when switchSession changes the active
+ * sessionId. bootstrap can't import listeners directly (DAG leaf), so
+ * callers register themselves. concurrentSessions.ts uses this to keep the
+ * PID file's sessionId in sync with --resume.
+ */
+export const onSessionSwitch = sessionSwitched.subscribe
+
+/**
+ * Project directory the current session's transcript lives in, or `null` if
+ * the session was created in the current project (common case โ derive from
+ * originalCwd). See `switchSession()`.
+ */
+export function getSessionProjectDir(): string | null {
+ return STATE.sessionProjectDir
+}
+
+export function getOriginalCwd(): string {
+ return STATE.originalCwd
+}
+
+/**
+ * Get the stable project root directory.
+ * Unlike getOriginalCwd(), this is never updated by mid-session EnterWorktreeTool
+ * (so skills/history stay stable when entering a throwaway worktree).
+ * It IS set at startup by --worktree, since that worktree is the session's project.
+ * Use for project identity (history, skills, sessions) not file operations.
+ */
+export function getProjectRoot(): string {
+ return STATE.projectRoot
+}
+
+export function setOriginalCwd(cwd: string): void {
+ STATE.originalCwd = cwd.normalize('NFC')
+}
+
+/**
+ * Only for --worktree startup flag. Mid-session EnterWorktreeTool must NOT
+ * call this โ skills/history should stay anchored to where the session started.
+ */
+export function setProjectRoot(cwd: string): void {
+ STATE.projectRoot = cwd.normalize('NFC')
+}
+
+export function getCwdState(): string {
+ return STATE.cwd
+}
+
+export function setCwdState(cwd: string): void {
+ STATE.cwd = cwd.normalize('NFC')
+}
+
+export function getDirectConnectServerUrl(): string | undefined {
+ return STATE.directConnectServerUrl
+}
+
+export function setDirectConnectServerUrl(url: string): void {
+ STATE.directConnectServerUrl = url
+}
+
+export function addToTotalDurationState(
+ duration: number,
+ durationWithoutRetries: number,
+): void {
+ STATE.totalAPIDuration += duration
+ STATE.totalAPIDurationWithoutRetries += durationWithoutRetries
+}
+
+export function resetTotalDurationStateAndCost_FOR_TESTS_ONLY(): void {
+ STATE.totalAPIDuration = 0
+ STATE.totalAPIDurationWithoutRetries = 0
+ STATE.totalCostUSD = 0
+}
+
+export function addToTotalCostState(
+ cost: number,
+ modelUsage: ModelUsage,
+ model: string,
+): void {
+ STATE.modelUsage[model] = modelUsage
+ STATE.totalCostUSD += cost
+}
+
+export function getTotalCostUSD(): number {
+ return STATE.totalCostUSD
+}
+
+export function getTotalAPIDuration(): number {
+ return STATE.totalAPIDuration
+}
+
+export function getTotalDuration(): number {
+ return Date.now() - STATE.startTime
+}
+
+export function getTotalAPIDurationWithoutRetries(): number {
+ return STATE.totalAPIDurationWithoutRetries
+}
+
+export function getTotalToolDuration(): number {
+ return STATE.totalToolDuration
+}
+
+export function addToToolDuration(duration: number): void {
+ STATE.totalToolDuration += duration
+ STATE.turnToolDurationMs += duration
+ STATE.turnToolCount++
+}
+
+export function getTurnHookDurationMs(): number {
+ return STATE.turnHookDurationMs
+}
+
+export function addToTurnHookDuration(duration: number): void {
+ STATE.turnHookDurationMs += duration
+ STATE.turnHookCount++
+}
+
+export function resetTurnHookDuration(): void {
+ STATE.turnHookDurationMs = 0
+ STATE.turnHookCount = 0
+}
+
+export function getTurnHookCount(): number {
+ return STATE.turnHookCount
+}
+
+export function getTurnToolDurationMs(): number {
+ return STATE.turnToolDurationMs
+}
+
+export function resetTurnToolDuration(): void {
+ STATE.turnToolDurationMs = 0
+ STATE.turnToolCount = 0
+}
+
+export function getTurnToolCount(): number {
+ return STATE.turnToolCount
+}
+
+export function getTurnClassifierDurationMs(): number {
+ return STATE.turnClassifierDurationMs
+}
+
+export function addToTurnClassifierDuration(duration: number): void {
+ STATE.turnClassifierDurationMs += duration
+ STATE.turnClassifierCount++
+}
+
+export function resetTurnClassifierDuration(): void {
+ STATE.turnClassifierDurationMs = 0
+ STATE.turnClassifierCount = 0
+}
+
+export function getTurnClassifierCount(): number {
+ return STATE.turnClassifierCount
+}
+
+export function getStatsStore(): {
+ observe(name: string, value: number): void
+} | null {
+ return STATE.statsStore
+}
+
+export function setStatsStore(
+ store: { observe(name: string, value: number): void } | null,
+): void {
+ STATE.statsStore = store
+}
+
+/**
+ * Marks that an interaction occurred.
+ *
+ * By default the actual Date.now() call is deferred until the next Ink render
+ * frame (via flushInteractionTime()) so we avoid calling Date.now() on every
+ * single keypress.
+ *
+ * Pass `immediate = true` when calling from React useEffect callbacks or
+ * other code that runs *after* the Ink render cycle has already flushed.
+ * Without it the timestamp stays stale until the next render, which may never
+ * come if the user is idle (e.g. permission dialog waiting for input).
+ */
+let interactionTimeDirty = false
+
+export function updateLastInteractionTime(immediate?: boolean): void {
+ if (immediate) {
+ flushInteractionTime_inner()
+ } else {
+ interactionTimeDirty = true
+ }
+}
+
+/**
+ * If an interaction was recorded since the last flush, update the timestamp
+ * now. Called by Ink before each render cycle so we batch many keypresses into
+ * a single Date.now() call.
+ */
+export function flushInteractionTime(): void {
+ if (interactionTimeDirty) {
+ flushInteractionTime_inner()
+ }
+}
+
+function flushInteractionTime_inner(): void {
+ STATE.lastInteractionTime = Date.now()
+ interactionTimeDirty = false
+}
+
+export function addToTotalLinesChanged(added: number, removed: number): void {
+ STATE.totalLinesAdded += added
+ STATE.totalLinesRemoved += removed
+}
+
+export function getTotalLinesAdded(): number {
+ return STATE.totalLinesAdded
+}
+
+export function getTotalLinesRemoved(): number {
+ return STATE.totalLinesRemoved
+}
+
+export function getTotalInputTokens(): number {
+ return sumBy(Object.values(STATE.modelUsage), 'inputTokens')
+}
+
+export function getTotalOutputTokens(): number {
+ return sumBy(Object.values(STATE.modelUsage), 'outputTokens')
+}
+
+export function getTotalCacheReadInputTokens(): number {
+ return sumBy(Object.values(STATE.modelUsage), 'cacheReadInputTokens')
+}
+
+export function getTotalCacheCreationInputTokens(): number {
+ return sumBy(Object.values(STATE.modelUsage), 'cacheCreationInputTokens')
+}
+
+export function getTotalWebSearchRequests(): number {
+ return sumBy(Object.values(STATE.modelUsage), 'webSearchRequests')
+}
+
+let outputTokensAtTurnStart = 0
+let currentTurnTokenBudget: number | null = null
+export function getTurnOutputTokens(): number {
+ return getTotalOutputTokens() - outputTokensAtTurnStart
+}
+export function getCurrentTurnTokenBudget(): number | null {
+ return currentTurnTokenBudget
+}
+let budgetContinuationCount = 0
+export function snapshotOutputTokensForTurn(budget: number | null): void {
+ outputTokensAtTurnStart = getTotalOutputTokens()
+ currentTurnTokenBudget = budget
+ budgetContinuationCount = 0
+}
+export function getBudgetContinuationCount(): number {
+ return budgetContinuationCount
+}
+export function incrementBudgetContinuationCount(): void {
+ budgetContinuationCount++
+}
+
+export function setHasUnknownModelCost(): void {
+ STATE.hasUnknownModelCost = true
+}
+
+export function hasUnknownModelCost(): boolean {
+ return STATE.hasUnknownModelCost
+}
+
+export function getLastMainRequestId(): string | undefined {
+ return STATE.lastMainRequestId
+}
+
+export function setLastMainRequestId(requestId: string): void {
+ STATE.lastMainRequestId = requestId
+}
+
+export function getLastApiCompletionTimestamp(): number | null {
+ return STATE.lastApiCompletionTimestamp
+}
+
+export function setLastApiCompletionTimestamp(timestamp: number): void {
+ STATE.lastApiCompletionTimestamp = timestamp
+}
+
+/** Mark that a compaction just occurred. The next API success event will
+ * include isPostCompaction=true, then the flag auto-resets. */
+export function markPostCompaction(): void {
+ STATE.pendingPostCompaction = true
+}
+
+/** Consume the post-compaction flag. Returns true once after compaction,
+ * then returns false until the next compaction. */
+export function consumePostCompaction(): boolean {
+ const was = STATE.pendingPostCompaction
+ STATE.pendingPostCompaction = false
+ return was
+}
+
+export function getLastInteractionTime(): number {
+ return STATE.lastInteractionTime
+}
+
+// Scroll drain suspension โ background intervals check this before doing work
+// so they don't compete with scroll frames for the event loop. Set by
+// ScrollBox scrollBy/scrollTo, cleared SCROLL_DRAIN_IDLE_MS after the last
+// scroll event. Module-scope (not in STATE) โ ephemeral hot-path flag, no
+// test-reset needed since the debounce timer self-clears.
+let scrollDraining = false
+let scrollDrainTimer: ReturnType | undefined
+const SCROLL_DRAIN_IDLE_MS = 150
+
+/** Mark that a scroll event just happened. Background intervals gate on
+ * getIsScrollDraining() and skip their work until the debounce clears. */
+export function markScrollActivity(): void {
+ scrollDraining = true
+ if (scrollDrainTimer) clearTimeout(scrollDrainTimer)
+ scrollDrainTimer = setTimeout(() => {
+ scrollDraining = false
+ scrollDrainTimer = undefined
+ }, SCROLL_DRAIN_IDLE_MS)
+ scrollDrainTimer.unref?.()
+}
+
+/** True while scroll is actively draining (within 150ms of last event).
+ * Intervals should early-return when this is set โ the work picks up next
+ * tick after scroll settles. */
+export function getIsScrollDraining(): boolean {
+ return scrollDraining
+}
+
+/** Await this before expensive one-shot work (network, subprocess) that could
+ * coincide with scroll. Resolves immediately if not scrolling; otherwise
+ * polls at the idle interval until the flag clears. */
+export async function waitForScrollIdle(): Promise {
+ while (scrollDraining) {
+ // bootstrap-isolation forbids importing sleep() from src/utils/
+ // eslint-disable-next-line no-restricted-syntax
+ await new Promise(r => setTimeout(r, SCROLL_DRAIN_IDLE_MS).unref?.())
+ }
+}
+
+export function getModelUsage(): { [modelName: string]: ModelUsage } {
+ return STATE.modelUsage
+}
+
+export function getUsageForModel(model: string): ModelUsage | undefined {
+ return STATE.modelUsage[model]
+}
+
+/**
+ * Gets the model override set from the --model CLI flag or after the user
+ * updates their configured model.
+ */
+export function getMainLoopModelOverride(): ModelSetting | undefined {
+ return STATE.mainLoopModelOverride
+}
+
+export function getInitialMainLoopModel(): ModelSetting {
+ return STATE.initialMainLoopModel
+}
+
+export function setMainLoopModelOverride(
+ model: ModelSetting | undefined,
+): void {
+ STATE.mainLoopModelOverride = model
+}
+
+export function setInitialMainLoopModel(model: ModelSetting): void {
+ STATE.initialMainLoopModel = model
+}
+
+export function getSdkBetas(): string[] | undefined {
+ return STATE.sdkBetas
+}
+
+export function setSdkBetas(betas: string[] | undefined): void {
+ STATE.sdkBetas = betas
+}
+
+export function resetCostState(): void {
+ STATE.totalCostUSD = 0
+ STATE.totalAPIDuration = 0
+ STATE.totalAPIDurationWithoutRetries = 0
+ STATE.totalToolDuration = 0
+ STATE.startTime = Date.now()
+ STATE.totalLinesAdded = 0
+ STATE.totalLinesRemoved = 0
+ STATE.hasUnknownModelCost = false
+ STATE.modelUsage = {}
+ STATE.promptId = null
+}
+
+/**
+ * Sets cost state values for session restore.
+ * Called by restoreCostStateForSession in cost-tracker.ts.
+ */
+export function setCostStateForRestore({
+ totalCostUSD,
+ totalAPIDuration,
+ totalAPIDurationWithoutRetries,
+ totalToolDuration,
+ totalLinesAdded,
+ totalLinesRemoved,
+ lastDuration,
+ modelUsage,
+}: {
+ totalCostUSD: number
+ totalAPIDuration: number
+ totalAPIDurationWithoutRetries: number
+ totalToolDuration: number
+ totalLinesAdded: number
+ totalLinesRemoved: number
+ lastDuration: number | undefined
+ modelUsage: { [modelName: string]: ModelUsage } | undefined
+}): void {
+ STATE.totalCostUSD = totalCostUSD
+ STATE.totalAPIDuration = totalAPIDuration
+ STATE.totalAPIDurationWithoutRetries = totalAPIDurationWithoutRetries
+ STATE.totalToolDuration = totalToolDuration
+ STATE.totalLinesAdded = totalLinesAdded
+ STATE.totalLinesRemoved = totalLinesRemoved
+
+ // Restore per-model usage breakdown
+ if (modelUsage) {
+ STATE.modelUsage = modelUsage
+ }
+
+ // Adjust startTime to make wall duration accumulate
+ if (lastDuration) {
+ STATE.startTime = Date.now() - lastDuration
+ }
+}
+
+// Only used in tests
+export function resetStateForTests(): void {
+ if (process.env.NODE_ENV !== 'test') {
+ throw new Error('resetStateForTests can only be called in tests')
+ }
+ Object.entries(getInitialState()).forEach(([key, value]) => {
+ STATE[key as keyof State] = value as never
+ })
+ outputTokensAtTurnStart = 0
+ currentTurnTokenBudget = null
+ budgetContinuationCount = 0
+ sessionSwitched.clear()
+}
+
+// You shouldn't use this directly. See src/utils/model/modelStrings.ts::getModelStrings()
+export function getModelStrings(): ModelStrings | null {
+ return STATE.modelStrings
+}
+
+// You shouldn't use this directly. See src/utils/model/modelStrings.ts
+export function setModelStrings(modelStrings: ModelStrings): void {
+ STATE.modelStrings = modelStrings
+}
+
+// Test utility function to reset model strings for re-initialization.
+// Separate from setModelStrings because we only want to accept 'null' in tests.
+export function resetModelStringsForTestingOnly() {
+ STATE.modelStrings = null
+}
+
+export function setMeter(
+ meter: Meter,
+ createCounter: (name: string, options: MetricOptions) => AttributedCounter,
+): void {
+ STATE.meter = meter
+
+ // Initialize all counters using the provided factory
+ STATE.sessionCounter = createCounter('claude_code.session.count', {
+ description: 'Count of CLI sessions started',
+ })
+ STATE.locCounter = createCounter('claude_code.lines_of_code.count', {
+ description:
+ "Count of lines of code modified, with the 'type' attribute indicating whether lines were added or removed",
+ })
+ STATE.prCounter = createCounter('claude_code.pull_request.count', {
+ description: 'Number of pull requests created',
+ })
+ STATE.commitCounter = createCounter('claude_code.commit.count', {
+ description: 'Number of git commits created',
+ })
+ STATE.costCounter = createCounter('claude_code.cost.usage', {
+ description: 'Cost of the Claude Code session',
+ unit: 'USD',
+ })
+ STATE.tokenCounter = createCounter('claude_code.token.usage', {
+ description: 'Number of tokens used',
+ unit: 'tokens',
+ })
+ STATE.codeEditToolDecisionCounter = createCounter(
+ 'claude_code.code_edit_tool.decision',
+ {
+ description:
+ 'Count of code editing tool permission decisions (accept/reject) for Edit, Write, and NotebookEdit tools',
+ },
+ )
+ STATE.activeTimeCounter = createCounter('claude_code.active_time.total', {
+ description: 'Total active time in seconds',
+ unit: 's',
+ })
+}
+
+export function getMeter(): Meter | null {
+ return STATE.meter
+}
+
+export function getSessionCounter(): AttributedCounter | null {
+ return STATE.sessionCounter
+}
+
+export function getLocCounter(): AttributedCounter | null {
+ return STATE.locCounter
+}
+
+export function getPrCounter(): AttributedCounter | null {
+ return STATE.prCounter
+}
+
+export function getCommitCounter(): AttributedCounter | null {
+ return STATE.commitCounter
+}
+
+export function getCostCounter(): AttributedCounter | null {
+ return STATE.costCounter
+}
+
+export function getTokenCounter(): AttributedCounter | null {
+ return STATE.tokenCounter
+}
+
+export function getCodeEditToolDecisionCounter(): AttributedCounter | null {
+ return STATE.codeEditToolDecisionCounter
+}
+
+export function getActiveTimeCounter(): AttributedCounter | null {
+ return STATE.activeTimeCounter
+}
+
+export function getLoggerProvider(): LoggerProvider | null {
+ return STATE.loggerProvider
+}
+
+export function setLoggerProvider(provider: LoggerProvider | null): void {
+ STATE.loggerProvider = provider
+}
+
+export function getEventLogger(): ReturnType | null {
+ return STATE.eventLogger
+}
+
+export function setEventLogger(
+ logger: ReturnType | null,
+): void {
+ STATE.eventLogger = logger
+}
+
+export function getMeterProvider(): MeterProvider | null {
+ return STATE.meterProvider
+}
+
+export function setMeterProvider(provider: MeterProvider | null): void {
+ STATE.meterProvider = provider
+}
+export function getTracerProvider(): BasicTracerProvider | null {
+ return STATE.tracerProvider
+}
+export function setTracerProvider(provider: BasicTracerProvider | null): void {
+ STATE.tracerProvider = provider
+}
+
+export function getIsNonInteractiveSession(): boolean {
+ return !STATE.isInteractive
+}
+
+export function getIsInteractive(): boolean {
+ return STATE.isInteractive
+}
+
+export function setIsInteractive(value: boolean): void {
+ STATE.isInteractive = value
+}
+
+export function getClientType(): string {
+ return STATE.clientType
+}
+
+export function setClientType(type: string): void {
+ STATE.clientType = type
+}
+
+export function getSdkAgentProgressSummariesEnabled(): boolean {
+ return STATE.sdkAgentProgressSummariesEnabled
+}
+
+export function setSdkAgentProgressSummariesEnabled(value: boolean): void {
+ STATE.sdkAgentProgressSummariesEnabled = value
+}
+
+export function getKairosActive(): boolean {
+ return STATE.kairosActive
+}
+
+export function setKairosActive(value: boolean): void {
+ STATE.kairosActive = value
+}
+
+export function getStrictToolResultPairing(): boolean {
+ return STATE.strictToolResultPairing
+}
+
+export function setStrictToolResultPairing(value: boolean): void {
+ STATE.strictToolResultPairing = value
+}
+
+// Field name 'userMsgOptIn' avoids excluded-string substrings ('BriefTool',
+// 'SendUserMessage' โ case-insensitive). All callers are inside feature()
+// guards so these accessors don't need their own (matches getKairosActive).
+export function getUserMsgOptIn(): boolean {
+ return STATE.userMsgOptIn
+}
+
+export function setUserMsgOptIn(value: boolean): void {
+ STATE.userMsgOptIn = value
+}
+
+export function getSessionSource(): string | undefined {
+ return STATE.sessionSource
+}
+
+export function setSessionSource(source: string): void {
+ STATE.sessionSource = source
+}
+
+export function getQuestionPreviewFormat(): 'markdown' | 'html' | undefined {
+ return STATE.questionPreviewFormat
+}
+
+export function setQuestionPreviewFormat(format: 'markdown' | 'html'): void {
+ STATE.questionPreviewFormat = format
+}
+
+export function getAgentColorMap(): Map {
+ return STATE.agentColorMap
+}
+
+export function getFlagSettingsPath(): string | undefined {
+ return STATE.flagSettingsPath
+}
+
+export function setFlagSettingsPath(path: string | undefined): void {
+ STATE.flagSettingsPath = path
+}
+
+export function getFlagSettingsInline(): Record | null {
+ return STATE.flagSettingsInline
+}
+
+export function setFlagSettingsInline(
+ settings: Record | null,
+): void {
+ STATE.flagSettingsInline = settings
+}
+
+export function getSessionIngressToken(): string | null | undefined {
+ return STATE.sessionIngressToken
+}
+
+export function setSessionIngressToken(token: string | null): void {
+ STATE.sessionIngressToken = token
+}
+
+export function getOauthTokenFromFd(): string | null | undefined {
+ return STATE.oauthTokenFromFd
+}
+
+export function setOauthTokenFromFd(token: string | null): void {
+ STATE.oauthTokenFromFd = token
+}
+
+export function getApiKeyFromFd(): string | null | undefined {
+ return STATE.apiKeyFromFd
+}
+
+export function setApiKeyFromFd(key: string | null): void {
+ STATE.apiKeyFromFd = key
+}
+
+export function setLastAPIRequest(
+ params: Omit | null,
+): void {
+ STATE.lastAPIRequest = params
+}
+
+export function getLastAPIRequest(): Omit<
+ BetaMessageStreamParams,
+ 'messages'
+> | null {
+ return STATE.lastAPIRequest
+}
+
+export function setLastAPIRequestMessages(
+ messages: BetaMessageStreamParams['messages'] | null,
+): void {
+ STATE.lastAPIRequestMessages = messages
+}
+
+export function getLastAPIRequestMessages():
+ | BetaMessageStreamParams['messages']
+ | null {
+ return STATE.lastAPIRequestMessages
+}
+
+export function setLastClassifierRequests(requests: unknown[] | null): void {
+ STATE.lastClassifierRequests = requests
+}
+
+export function getLastClassifierRequests(): unknown[] | null {
+ return STATE.lastClassifierRequests
+}
+
+export function setCachedClaudeMdContent(content: string | null): void {
+ STATE.cachedClaudeMdContent = content
+}
+
+export function getCachedClaudeMdContent(): string | null {
+ return STATE.cachedClaudeMdContent
+}
+
+export function addToInMemoryErrorLog(errorInfo: {
+ error: string
+ timestamp: string
+}): void {
+ const MAX_IN_MEMORY_ERRORS = 100
+ if (STATE.inMemoryErrorLog.length >= MAX_IN_MEMORY_ERRORS) {
+ STATE.inMemoryErrorLog.shift() // Remove oldest error
+ }
+ STATE.inMemoryErrorLog.push(errorInfo)
+}
+
+export function getAllowedSettingSources(): SettingSource[] {
+ return STATE.allowedSettingSources
+}
+
+export function setAllowedSettingSources(sources: SettingSource[]): void {
+ STATE.allowedSettingSources = sources
+}
+
+export function preferThirdPartyAuthentication(): boolean {
+ // IDE extension should behave as 1P for authentication reasons.
+ return getIsNonInteractiveSession() && STATE.clientType !== 'claude-vscode'
+}
+
+export function setInlinePlugins(plugins: Array): void {
+ STATE.inlinePlugins = plugins
+}
+
+export function getInlinePlugins(): Array {
+ return STATE.inlinePlugins
+}
+
+export function setChromeFlagOverride(value: boolean | undefined): void {
+ STATE.chromeFlagOverride = value
+}
+
+export function getChromeFlagOverride(): boolean | undefined {
+ return STATE.chromeFlagOverride
+}
+
+export function setUseCoworkPlugins(value: boolean): void {
+ STATE.useCoworkPlugins = value
+ resetSettingsCache()
+}
+
+export function getUseCoworkPlugins(): boolean {
+ return STATE.useCoworkPlugins
+}
+
+export function setSessionBypassPermissionsMode(enabled: boolean): void {
+ STATE.sessionBypassPermissionsMode = enabled
+}
+
+export function getSessionBypassPermissionsMode(): boolean {
+ return STATE.sessionBypassPermissionsMode
+}
+
+export function setScheduledTasksEnabled(enabled: boolean): void {
+ STATE.scheduledTasksEnabled = enabled
+}
+
+export function getScheduledTasksEnabled(): boolean {
+ return STATE.scheduledTasksEnabled
+}
+
+export type SessionCronTask = {
+ id: string
+ cron: string
+ prompt: string
+ createdAt: number
+ recurring?: boolean
+ /**
+ * When set, the task was created by an in-process teammate (not the team lead).
+ * The scheduler routes fires to that teammate's pendingUserMessages queue
+ * instead of the main REPL command queue. Session-only โ never written to disk.
+ */
+ agentId?: string
+}
+
+export function getSessionCronTasks(): SessionCronTask[] {
+ return STATE.sessionCronTasks
+}
+
+export function addSessionCronTask(task: SessionCronTask): void {
+ STATE.sessionCronTasks.push(task)
+}
+
+/**
+ * Returns the number of tasks actually removed. Callers use this to skip
+ * downstream work (e.g. the disk read in removeCronTasks) when all ids
+ * were accounted for here.
+ */
+export function removeSessionCronTasks(ids: readonly string[]): number {
+ if (ids.length === 0) return 0
+ const idSet = new Set(ids)
+ const remaining = STATE.sessionCronTasks.filter(t => !idSet.has(t.id))
+ const removed = STATE.sessionCronTasks.length - remaining.length
+ if (removed === 0) return 0
+ STATE.sessionCronTasks = remaining
+ return removed
+}
+
+export function setSessionTrustAccepted(accepted: boolean): void {
+ STATE.sessionTrustAccepted = accepted
+}
+
+export function getSessionTrustAccepted(): boolean {
+ return STATE.sessionTrustAccepted
+}
+
+export function setSessionPersistenceDisabled(disabled: boolean): void {
+ STATE.sessionPersistenceDisabled = disabled
+}
+
+export function isSessionPersistenceDisabled(): boolean {
+ return STATE.sessionPersistenceDisabled
+}
+
+export function hasExitedPlanModeInSession(): boolean {
+ return STATE.hasExitedPlanMode
+}
+
+export function setHasExitedPlanMode(value: boolean): void {
+ STATE.hasExitedPlanMode = value
+}
+
+export function needsPlanModeExitAttachment(): boolean {
+ return STATE.needsPlanModeExitAttachment
+}
+
+export function setNeedsPlanModeExitAttachment(value: boolean): void {
+ STATE.needsPlanModeExitAttachment = value
+}
+
+export function handlePlanModeTransition(
+ fromMode: string,
+ toMode: string,
+): void {
+ // If switching TO plan mode, clear any pending exit attachment
+ // This prevents sending both plan_mode and plan_mode_exit when user toggles quickly
+ if (toMode === 'plan' && fromMode !== 'plan') {
+ STATE.needsPlanModeExitAttachment = false
+ }
+
+ // If switching out of plan mode, trigger the plan_mode_exit attachment
+ if (fromMode === 'plan' && toMode !== 'plan') {
+ STATE.needsPlanModeExitAttachment = true
+ }
+}
+
+export function needsAutoModeExitAttachment(): boolean {
+ return STATE.needsAutoModeExitAttachment
+}
+
+export function setNeedsAutoModeExitAttachment(value: boolean): void {
+ STATE.needsAutoModeExitAttachment = value
+}
+
+export function handleAutoModeTransition(
+ fromMode: string,
+ toMode: string,
+): void {
+ // Autoโplan transitions are handled by prepareContextForPlanMode (auto may
+ // stay active through plan if opted in) and ExitPlanMode (restores mode).
+ // Skip both directions so this function only handles direct auto transitions.
+ if (
+ (fromMode === 'auto' && toMode === 'plan') ||
+ (fromMode === 'plan' && toMode === 'auto')
+ ) {
+ return
+ }
+ const fromIsAuto = fromMode === 'auto'
+ const toIsAuto = toMode === 'auto'
+
+ // If switching TO auto mode, clear any pending exit attachment
+ // This prevents sending both auto_mode and auto_mode_exit when user toggles quickly
+ if (toIsAuto && !fromIsAuto) {
+ STATE.needsAutoModeExitAttachment = false
+ }
+
+ // If switching out of auto mode, trigger the auto_mode_exit attachment
+ if (fromIsAuto && !toIsAuto) {
+ STATE.needsAutoModeExitAttachment = true
+ }
+}
+
+// LSP plugin recommendation session tracking
+export function hasShownLspRecommendationThisSession(): boolean {
+ return STATE.lspRecommendationShownThisSession
+}
+
+export function setLspRecommendationShownThisSession(value: boolean): void {
+ STATE.lspRecommendationShownThisSession = value
+}
+
+// SDK init event state
+export function setInitJsonSchema(schema: Record): void {
+ STATE.initJsonSchema = schema
+}
+
+export function getInitJsonSchema(): Record | null {
+ return STATE.initJsonSchema
+}
+
+export function registerHookCallbacks(
+ hooks: Partial>,
+): void {
+ if (!STATE.registeredHooks) {
+ STATE.registeredHooks = {}
+ }
+
+ // `registerHookCallbacks` may be called multiple times, so we need to merge (not overwrite)
+ for (const [event, matchers] of Object.entries(hooks)) {
+ const eventKey = event as HookEvent
+ if (!STATE.registeredHooks[eventKey]) {
+ STATE.registeredHooks[eventKey] = []
+ }
+ STATE.registeredHooks[eventKey]!.push(...matchers)
+ }
+}
+
+export function getRegisteredHooks(): Partial<
+ Record
+> | null {
+ return STATE.registeredHooks
+}
+
+export function clearRegisteredHooks(): void {
+ STATE.registeredHooks = null
+}
+
+export function clearRegisteredPluginHooks(): void {
+ if (!STATE.registeredHooks) {
+ return
+ }
+
+ const filtered: Partial> = {}
+ for (const [event, matchers] of Object.entries(STATE.registeredHooks)) {
+ // Keep only callback hooks (those without pluginRoot)
+ const callbackHooks = matchers.filter(m => !('pluginRoot' in m))
+ if (callbackHooks.length > 0) {
+ filtered[event as HookEvent] = callbackHooks
+ }
+ }
+
+ STATE.registeredHooks = Object.keys(filtered).length > 0 ? filtered : null
+}
+
+export function resetSdkInitState(): void {
+ STATE.initJsonSchema = null
+ STATE.registeredHooks = null
+}
+
+export function getPlanSlugCache(): Map {
+ return STATE.planSlugCache
+}
+
+export function getSessionCreatedTeams(): Set {
+ return STATE.sessionCreatedTeams
+}
+
+// Teleported session tracking for reliability logging
+export function setTeleportedSessionInfo(info: {
+ sessionId: string | null
+}): void {
+ STATE.teleportedSessionInfo = {
+ isTeleported: true,
+ hasLoggedFirstMessage: false,
+ sessionId: info.sessionId,
+ }
+}
+
+export function getTeleportedSessionInfo(): {
+ isTeleported: boolean
+ hasLoggedFirstMessage: boolean
+ sessionId: string | null
+} | null {
+ return STATE.teleportedSessionInfo
+}
+
+export function markFirstTeleportMessageLogged(): void {
+ if (STATE.teleportedSessionInfo) {
+ STATE.teleportedSessionInfo.hasLoggedFirstMessage = true
+ }
+}
+
+// Invoked skills tracking for preservation across compaction
+export type InvokedSkillInfo = {
+ skillName: string
+ skillPath: string
+ content: string
+ invokedAt: number
+ agentId: string | null
+}
+
+export function addInvokedSkill(
+ skillName: string,
+ skillPath: string,
+ content: string,
+ agentId: string | null = null,
+): void {
+ const key = `${agentId ?? ''}:${skillName}`
+ STATE.invokedSkills.set(key, {
+ skillName,
+ skillPath,
+ content,
+ invokedAt: Date.now(),
+ agentId,
+ })
+}
+
+export function getInvokedSkills(): Map {
+ return STATE.invokedSkills
+}
+
+export function getInvokedSkillsForAgent(
+ agentId: string | undefined | null,
+): Map {
+ const normalizedId = agentId ?? null
+ const filtered = new Map()
+ for (const [key, skill] of STATE.invokedSkills) {
+ if (skill.agentId === normalizedId) {
+ filtered.set(key, skill)
+ }
+ }
+ return filtered
+}
+
+export function clearInvokedSkills(
+ preservedAgentIds?: ReadonlySet,
+): void {
+ if (!preservedAgentIds || preservedAgentIds.size === 0) {
+ STATE.invokedSkills.clear()
+ return
+ }
+ for (const [key, skill] of STATE.invokedSkills) {
+ if (skill.agentId === null || !preservedAgentIds.has(skill.agentId)) {
+ STATE.invokedSkills.delete(key)
+ }
+ }
+}
+
+export function clearInvokedSkillsForAgent(agentId: string): void {
+ for (const [key, skill] of STATE.invokedSkills) {
+ if (skill.agentId === agentId) {
+ STATE.invokedSkills.delete(key)
+ }
+ }
+}
+
+// Slow operations tracking for dev bar
+const MAX_SLOW_OPERATIONS = 10
+const SLOW_OPERATION_TTL_MS = 10000
+
+export function addSlowOperation(operation: string, durationMs: number): void {
+ if (process.env.USER_TYPE !== 'ant') return
+ // Skip tracking for editor sessions (user editing a prompt file in $EDITOR)
+ // These are intentionally slow since the user is drafting text
+ if (operation.includes('exec') && operation.includes('claude-prompt-')) {
+ return
+ }
+ const now = Date.now()
+ // Remove stale operations
+ STATE.slowOperations = STATE.slowOperations.filter(
+ op => now - op.timestamp < SLOW_OPERATION_TTL_MS,
+ )
+ // Add new operation
+ STATE.slowOperations.push({ operation, durationMs, timestamp: now })
+ // Keep only the most recent operations
+ if (STATE.slowOperations.length > MAX_SLOW_OPERATIONS) {
+ STATE.slowOperations = STATE.slowOperations.slice(-MAX_SLOW_OPERATIONS)
+ }
+}
+
+const EMPTY_SLOW_OPERATIONS: ReadonlyArray<{
+ operation: string
+ durationMs: number
+ timestamp: number
+}> = []
+
+export function getSlowOperations(): ReadonlyArray<{
+ operation: string
+ durationMs: number
+ timestamp: number
+}> {
+ // Most common case: nothing tracked. Return a stable reference so the
+ // caller's setState() can bail via Object.is instead of re-rendering at 2fps.
+ if (STATE.slowOperations.length === 0) {
+ return EMPTY_SLOW_OPERATIONS
+ }
+ const now = Date.now()
+ // Only allocate a new array when something actually expired; otherwise keep
+ // the reference stable across polls while ops are still fresh.
+ if (
+ STATE.slowOperations.some(op => now - op.timestamp >= SLOW_OPERATION_TTL_MS)
+ ) {
+ STATE.slowOperations = STATE.slowOperations.filter(
+ op => now - op.timestamp < SLOW_OPERATION_TTL_MS,
+ )
+ if (STATE.slowOperations.length === 0) {
+ return EMPTY_SLOW_OPERATIONS
+ }
+ }
+ // Safe to return directly: addSlowOperation() reassigns STATE.slowOperations
+ // before pushing, so the array held in React state is never mutated.
+ return STATE.slowOperations
+}
+
+export function getMainThreadAgentType(): string | undefined {
+ return STATE.mainThreadAgentType
+}
+
+export function setMainThreadAgentType(agentType: string | undefined): void {
+ STATE.mainThreadAgentType = agentType
+}
+
+export function getIsRemoteMode(): boolean {
+ return STATE.isRemoteMode
+}
+
+export function setIsRemoteMode(value: boolean): void {
+ STATE.isRemoteMode = value
+}
+
+// System prompt section accessors
+
+export function getSystemPromptSectionCache(): Map {
+ return STATE.systemPromptSectionCache
+}
+
+export function setSystemPromptSectionCacheEntry(
+ name: string,
+ value: string | null,
+): void {
+ STATE.systemPromptSectionCache.set(name, value)
+}
+
+export function clearSystemPromptSectionState(): void {
+ STATE.systemPromptSectionCache.clear()
+}
+
+// Last emitted date accessors (for detecting midnight date changes)
+
+export function getLastEmittedDate(): string | null {
+ return STATE.lastEmittedDate
+}
+
+export function setLastEmittedDate(date: string | null): void {
+ STATE.lastEmittedDate = date
+}
+
+export function getAdditionalDirectoriesForClaudeMd(): string[] {
+ return STATE.additionalDirectoriesForClaudeMd
+}
+
+export function setAdditionalDirectoriesForClaudeMd(
+ directories: string[],
+): void {
+ STATE.additionalDirectoriesForClaudeMd = directories
+}
+
+export function getAllowedChannels(): ChannelEntry[] {
+ return STATE.allowedChannels
+}
+
+export function setAllowedChannels(entries: ChannelEntry[]): void {
+ STATE.allowedChannels = entries
+}
+
+export function getHasDevChannels(): boolean {
+ return STATE.hasDevChannels
+}
+
+export function setHasDevChannels(value: boolean): void {
+ STATE.hasDevChannels = value
+}
+
+export function getPromptCache1hAllowlist(): string[] | null {
+ return STATE.promptCache1hAllowlist
+}
+
+export function setPromptCache1hAllowlist(allowlist: string[] | null): void {
+ STATE.promptCache1hAllowlist = allowlist
+}
+
+export function getPromptCache1hEligible(): boolean | null {
+ return STATE.promptCache1hEligible
+}
+
+export function setPromptCache1hEligible(eligible: boolean | null): void {
+ STATE.promptCache1hEligible = eligible
+}
+
+export function getAfkModeHeaderLatched(): boolean | null {
+ return STATE.afkModeHeaderLatched
+}
+
+export function setAfkModeHeaderLatched(v: boolean): void {
+ STATE.afkModeHeaderLatched = v
+}
+
+export function getFastModeHeaderLatched(): boolean | null {
+ return STATE.fastModeHeaderLatched
+}
+
+export function setFastModeHeaderLatched(v: boolean): void {
+ STATE.fastModeHeaderLatched = v
+}
+
+export function getCacheEditingHeaderLatched(): boolean | null {
+ return STATE.cacheEditingHeaderLatched
+}
+
+export function setCacheEditingHeaderLatched(v: boolean): void {
+ STATE.cacheEditingHeaderLatched = v
+}
+
+export function getThinkingClearLatched(): boolean | null {
+ return STATE.thinkingClearLatched
+}
+
+export function setThinkingClearLatched(v: boolean): void {
+ STATE.thinkingClearLatched = v
+}
+
+/**
+ * Reset beta header latches to null. Called on /clear and /compact so a
+ * fresh conversation gets fresh header evaluation.
+ */
+export function clearBetaHeaderLatches(): void {
+ STATE.afkModeHeaderLatched = null
+ STATE.fastModeHeaderLatched = null
+ STATE.cacheEditingHeaderLatched = null
+ STATE.thinkingClearLatched = null
+}
+
+export function getPromptId(): string | null {
+ return STATE.promptId
+}
+
+export function setPromptId(id: string | null): void {
+ STATE.promptId = id
+}
+
diff --git a/original-source-code/src/bridge/bridgeApi.ts b/original-source-code/src/bridge/bridgeApi.ts
new file mode 100644
index 0000000..052bd4f
--- /dev/null
+++ b/original-source-code/src/bridge/bridgeApi.ts
@@ -0,0 +1,539 @@
+import axios from 'axios'
+
+import { debugBody, extractErrorDetail } from './debugUtils.js'
+import {
+ BRIDGE_LOGIN_INSTRUCTION,
+ type BridgeApiClient,
+ type BridgeConfig,
+ type PermissionResponseEvent,
+ type WorkResponse,
+} from './types.js'
+
+type BridgeApiDeps = {
+ baseUrl: string
+ getAccessToken: () => string | undefined
+ runnerVersion: string
+ onDebug?: (msg: string) => void
+ /**
+ * Called on 401 to attempt OAuth token refresh. Returns true if refreshed,
+ * in which case the request is retried once. Injected because
+ * handleOAuth401Error from utils/auth.ts transitively pulls in config.ts โ
+ * file.ts โ permissions/filesystem.ts โ sessionStorage.ts โ commands.ts
+ * (~1300 modules). Daemon callers using env-var tokens omit this โ their
+ * tokens don't refresh, so 401 goes straight to BridgeFatalError.
+ */
+ onAuth401?: (staleAccessToken: string) => Promise
+ /**
+ * Returns the trusted device token to send as X-Trusted-Device-Token on
+ * bridge API calls. Bridge sessions have SecurityTier=ELEVATED on the
+ * server (CCR v2); when the server's enforcement flag is on,
+ * ConnectBridgeWorker requires a trusted device at JWT-issuance.
+ * Optional โ when absent or returning undefined, the header is omitted
+ * and the server falls through to its flag-off/no-op path. The CLI-side
+ * gate is tengu_sessions_elevated_auth_enforcement (see trustedDevice.ts).
+ */
+ getTrustedDeviceToken?: () => string | undefined
+}
+
+const BETA_HEADER = 'environments-2025-11-01'
+
+/** Allowlist pattern for server-provided IDs used in URL path segments. */
+const SAFE_ID_PATTERN = /^[a-zA-Z0-9_-]+$/
+
+/**
+ * Validate that a server-provided ID is safe to interpolate into a URL path.
+ * Prevents path traversal (e.g. `../../admin`) and injection via IDs that
+ * contain slashes, dots, or other special characters.
+ */
+export function validateBridgeId(id: string, label: string): string {
+ if (!id || !SAFE_ID_PATTERN.test(id)) {
+ throw new Error(`Invalid ${label}: contains unsafe characters`)
+ }
+ return id
+}
+
+/** Fatal bridge errors that should not be retried (e.g. auth failures). */
+export class BridgeFatalError extends Error {
+ readonly status: number
+ /** Server-provided error type, e.g. "environment_expired". */
+ readonly errorType: string | undefined
+ constructor(message: string, status: number, errorType?: string) {
+ super(message)
+ this.name = 'BridgeFatalError'
+ this.status = status
+ this.errorType = errorType
+ }
+}
+
+export function createBridgeApiClient(deps: BridgeApiDeps): BridgeApiClient {
+ function debug(msg: string): void {
+ deps.onDebug?.(msg)
+ }
+
+ let consecutiveEmptyPolls = 0
+ const EMPTY_POLL_LOG_INTERVAL = 100
+
+ function getHeaders(accessToken: string): Record {
+ const headers: Record = {
+ Authorization: `Bearer ${accessToken}`,
+ 'Content-Type': 'application/json',
+ 'anthropic-version': '2023-06-01',
+ 'anthropic-beta': BETA_HEADER,
+ 'x-environment-runner-version': deps.runnerVersion,
+ }
+ const deviceToken = deps.getTrustedDeviceToken?.()
+ if (deviceToken) {
+ headers['X-Trusted-Device-Token'] = deviceToken
+ }
+ return headers
+ }
+
+ function resolveAuth(): string {
+ const accessToken = deps.getAccessToken()
+ if (!accessToken) {
+ throw new Error(BRIDGE_LOGIN_INSTRUCTION)
+ }
+ return accessToken
+ }
+
+ /**
+ * Execute an OAuth-authenticated request with a single retry on 401.
+ * On 401, attempts token refresh via handleOAuth401Error (same pattern as
+ * withRetry.ts for v1/messages). If refresh succeeds, retries the request
+ * once with the new token. If refresh fails or the retry also returns 401,
+ * the 401 response is returned for handleErrorStatus to throw BridgeFatalError.
+ */
+ async function withOAuthRetry(
+ fn: (accessToken: string) => Promise<{ status: number; data: T }>,
+ context: string,
+ ): Promise<{ status: number; data: T }> {
+ const accessToken = resolveAuth()
+ const response = await fn(accessToken)
+
+ if (response.status !== 401) {
+ return response
+ }
+
+ if (!deps.onAuth401) {
+ debug(`[bridge:api] ${context}: 401 received, no refresh handler`)
+ return response
+ }
+
+ // Attempt token refresh โ matches the pattern in withRetry.ts
+ debug(`[bridge:api] ${context}: 401 received, attempting token refresh`)
+ const refreshed = await deps.onAuth401(accessToken)
+ if (refreshed) {
+ debug(`[bridge:api] ${context}: Token refreshed, retrying request`)
+ const newToken = resolveAuth()
+ const retryResponse = await fn(newToken)
+ if (retryResponse.status !== 401) {
+ return retryResponse
+ }
+ debug(`[bridge:api] ${context}: Retry after refresh also got 401`)
+ } else {
+ debug(`[bridge:api] ${context}: Token refresh failed`)
+ }
+
+ // Refresh failed โ return 401 for handleErrorStatus to throw
+ return response
+ }
+
+ return {
+ async registerBridgeEnvironment(
+ config: BridgeConfig,
+ ): Promise<{ environment_id: string; environment_secret: string }> {
+ debug(
+ `[bridge:api] POST /v1/environments/bridge bridgeId=${config.bridgeId}`,
+ )
+
+ const response = await withOAuthRetry(
+ (token: string) =>
+ axios.post<{
+ environment_id: string
+ environment_secret: string
+ }>(
+ `${deps.baseUrl}/v1/environments/bridge`,
+ {
+ machine_name: config.machineName,
+ directory: config.dir,
+ branch: config.branch,
+ git_repo_url: config.gitRepoUrl,
+ // Advertise session capacity so claude.ai/code can show
+ // "2/4 sessions" badges and only block the picker when
+ // actually at capacity. Backends that don't yet accept
+ // this field will silently ignore it.
+ max_sessions: config.maxSessions,
+ // worker_type lets claude.ai filter environments by origin
+ // (e.g. assistant picker only shows assistant-mode workers).
+ // Desktop cowork app sends "cowork"; we send a distinct value.
+ metadata: { worker_type: config.workerType },
+ // Idempotent re-registration: if we have a backend-issued
+ // environment_id from a prior session (--session-id resume),
+ // send it back so the backend reattaches instead of creating
+ // a new env. The backend may still hand back a fresh ID if
+ // the old one expired โ callers must compare the response.
+ ...(config.reuseEnvironmentId && {
+ environment_id: config.reuseEnvironmentId,
+ }),
+ },
+ {
+ headers: getHeaders(token),
+ timeout: 15_000,
+ validateStatus: status => status < 500,
+ },
+ ),
+ 'Registration',
+ )
+
+ handleErrorStatus(response.status, response.data, 'Registration')
+ debug(
+ `[bridge:api] POST /v1/environments/bridge -> ${response.status} environment_id=${response.data.environment_id}`,
+ )
+ debug(
+ `[bridge:api] >>> ${debugBody({ machine_name: config.machineName, directory: config.dir, branch: config.branch, git_repo_url: config.gitRepoUrl, max_sessions: config.maxSessions, metadata: { worker_type: config.workerType } })}`,
+ )
+ debug(`[bridge:api] <<< ${debugBody(response.data)}`)
+ return response.data
+ },
+
+ async pollForWork(
+ environmentId: string,
+ environmentSecret: string,
+ signal?: AbortSignal,
+ reclaimOlderThanMs?: number,
+ ): Promise {
+ validateBridgeId(environmentId, 'environmentId')
+
+ // Save and reset so errors break the "consecutive empty" streak.
+ // Restored below when the response is truly empty.
+ const prevEmptyPolls = consecutiveEmptyPolls
+ consecutiveEmptyPolls = 0
+
+ const response = await axios.get(
+ `${deps.baseUrl}/v1/environments/${environmentId}/work/poll`,
+ {
+ headers: getHeaders(environmentSecret),
+ params:
+ reclaimOlderThanMs !== undefined
+ ? { reclaim_older_than_ms: reclaimOlderThanMs }
+ : undefined,
+ timeout: 10_000,
+ signal,
+ validateStatus: status => status < 500,
+ },
+ )
+
+ handleErrorStatus(response.status, response.data, 'Poll')
+
+ // Empty body or null = no work available
+ if (!response.data) {
+ consecutiveEmptyPolls = prevEmptyPolls + 1
+ if (
+ consecutiveEmptyPolls === 1 ||
+ consecutiveEmptyPolls % EMPTY_POLL_LOG_INTERVAL === 0
+ ) {
+ debug(
+ `[bridge:api] GET .../work/poll -> ${response.status} (no work, ${consecutiveEmptyPolls} consecutive empty polls)`,
+ )
+ }
+ return null
+ }
+
+ debug(
+ `[bridge:api] GET .../work/poll -> ${response.status} workId=${response.data.id} type=${response.data.data?.type}${response.data.data?.id ? ` sessionId=${response.data.data.id}` : ''}`,
+ )
+ debug(`[bridge:api] <<< ${debugBody(response.data)}`)
+ return response.data
+ },
+
+ async acknowledgeWork(
+ environmentId: string,
+ workId: string,
+ sessionToken: string,
+ ): Promise {
+ validateBridgeId(environmentId, 'environmentId')
+ validateBridgeId(workId, 'workId')
+
+ debug(`[bridge:api] POST .../work/${workId}/ack`)
+
+ const response = await axios.post(
+ `${deps.baseUrl}/v1/environments/${environmentId}/work/${workId}/ack`,
+ {},
+ {
+ headers: getHeaders(sessionToken),
+ timeout: 10_000,
+ validateStatus: s => s < 500,
+ },
+ )
+
+ handleErrorStatus(response.status, response.data, 'Acknowledge')
+ debug(`[bridge:api] POST .../work/${workId}/ack -> ${response.status}`)
+ },
+
+ async stopWork(
+ environmentId: string,
+ workId: string,
+ force: boolean,
+ ): Promise {
+ validateBridgeId(environmentId, 'environmentId')
+ validateBridgeId(workId, 'workId')
+
+ debug(`[bridge:api] POST .../work/${workId}/stop force=${force}`)
+
+ const response = await withOAuthRetry(
+ (token: string) =>
+ axios.post(
+ `${deps.baseUrl}/v1/environments/${environmentId}/work/${workId}/stop`,
+ { force },
+ {
+ headers: getHeaders(token),
+ timeout: 10_000,
+ validateStatus: s => s < 500,
+ },
+ ),
+ 'StopWork',
+ )
+
+ handleErrorStatus(response.status, response.data, 'StopWork')
+ debug(`[bridge:api] POST .../work/${workId}/stop -> ${response.status}`)
+ },
+
+ async deregisterEnvironment(environmentId: string): Promise {
+ validateBridgeId(environmentId, 'environmentId')
+
+ debug(`[bridge:api] DELETE /v1/environments/bridge/${environmentId}`)
+
+ const response = await withOAuthRetry(
+ (token: string) =>
+ axios.delete(
+ `${deps.baseUrl}/v1/environments/bridge/${environmentId}`,
+ {
+ headers: getHeaders(token),
+ timeout: 10_000,
+ validateStatus: s => s < 500,
+ },
+ ),
+ 'Deregister',
+ )
+
+ handleErrorStatus(response.status, response.data, 'Deregister')
+ debug(
+ `[bridge:api] DELETE /v1/environments/bridge/${environmentId} -> ${response.status}`,
+ )
+ },
+
+ async archiveSession(sessionId: string): Promise {
+ validateBridgeId(sessionId, 'sessionId')
+
+ debug(`[bridge:api] POST /v1/sessions/${sessionId}/archive`)
+
+ const response = await withOAuthRetry(
+ (token: string) =>
+ axios.post(
+ `${deps.baseUrl}/v1/sessions/${sessionId}/archive`,
+ {},
+ {
+ headers: getHeaders(token),
+ timeout: 10_000,
+ validateStatus: s => s < 500,
+ },
+ ),
+ 'ArchiveSession',
+ )
+
+ // 409 = already archived (idempotent, not an error)
+ if (response.status === 409) {
+ debug(
+ `[bridge:api] POST /v1/sessions/${sessionId}/archive -> 409 (already archived)`,
+ )
+ return
+ }
+
+ handleErrorStatus(response.status, response.data, 'ArchiveSession')
+ debug(
+ `[bridge:api] POST /v1/sessions/${sessionId}/archive -> ${response.status}`,
+ )
+ },
+
+ async reconnectSession(
+ environmentId: string,
+ sessionId: string,
+ ): Promise {
+ validateBridgeId(environmentId, 'environmentId')
+ validateBridgeId(sessionId, 'sessionId')
+
+ debug(
+ `[bridge:api] POST /v1/environments/${environmentId}/bridge/reconnect session_id=${sessionId}`,
+ )
+
+ const response = await withOAuthRetry(
+ (token: string) =>
+ axios.post(
+ `${deps.baseUrl}/v1/environments/${environmentId}/bridge/reconnect`,
+ { session_id: sessionId },
+ {
+ headers: getHeaders(token),
+ timeout: 10_000,
+ validateStatus: s => s < 500,
+ },
+ ),
+ 'ReconnectSession',
+ )
+
+ handleErrorStatus(response.status, response.data, 'ReconnectSession')
+ debug(`[bridge:api] POST .../bridge/reconnect -> ${response.status}`)
+ },
+
+ async heartbeatWork(
+ environmentId: string,
+ workId: string,
+ sessionToken: string,
+ ): Promise<{ lease_extended: boolean; state: string }> {
+ validateBridgeId(environmentId, 'environmentId')
+ validateBridgeId(workId, 'workId')
+
+ debug(`[bridge:api] POST .../work/${workId}/heartbeat`)
+
+ const response = await axios.post<{
+ lease_extended: boolean
+ state: string
+ last_heartbeat: string
+ ttl_seconds: number
+ }>(
+ `${deps.baseUrl}/v1/environments/${environmentId}/work/${workId}/heartbeat`,
+ {},
+ {
+ headers: getHeaders(sessionToken),
+ timeout: 10_000,
+ validateStatus: s => s < 500,
+ },
+ )
+
+ handleErrorStatus(response.status, response.data, 'Heartbeat')
+ debug(
+ `[bridge:api] POST .../work/${workId}/heartbeat -> ${response.status} lease_extended=${response.data.lease_extended} state=${response.data.state}`,
+ )
+ return response.data
+ },
+
+ async sendPermissionResponseEvent(
+ sessionId: string,
+ event: PermissionResponseEvent,
+ sessionToken: string,
+ ): Promise {
+ validateBridgeId(sessionId, 'sessionId')
+
+ debug(
+ `[bridge:api] POST /v1/sessions/${sessionId}/events type=${event.type}`,
+ )
+
+ const response = await axios.post(
+ `${deps.baseUrl}/v1/sessions/${sessionId}/events`,
+ { events: [event] },
+ {
+ headers: getHeaders(sessionToken),
+ timeout: 10_000,
+ validateStatus: s => s < 500,
+ },
+ )
+
+ handleErrorStatus(
+ response.status,
+ response.data,
+ 'SendPermissionResponseEvent',
+ )
+ debug(
+ `[bridge:api] POST /v1/sessions/${sessionId}/events -> ${response.status}`,
+ )
+ debug(`[bridge:api] >>> ${debugBody({ events: [event] })}`)
+ debug(`[bridge:api] <<< ${debugBody(response.data)}`)
+ },
+ }
+}
+
+function handleErrorStatus(
+ status: number,
+ data: unknown,
+ context: string,
+): void {
+ if (status === 200 || status === 204) {
+ return
+ }
+ const detail = extractErrorDetail(data)
+ const errorType = extractErrorTypeFromData(data)
+ switch (status) {
+ case 401:
+ throw new BridgeFatalError(
+ `${context}: Authentication failed (401)${detail ? `: ${detail}` : ''}. ${BRIDGE_LOGIN_INSTRUCTION}`,
+ 401,
+ errorType,
+ )
+ case 403:
+ throw new BridgeFatalError(
+ isExpiredErrorType(errorType)
+ ? 'Remote Control session has expired. Please restart with `claude remote-control` or /remote-control.'
+ : `${context}: Access denied (403)${detail ? `: ${detail}` : ''}. Check your organization permissions.`,
+ 403,
+ errorType,
+ )
+ case 404:
+ throw new BridgeFatalError(
+ detail ??
+ `${context}: Not found (404). Remote Control may not be available for this organization.`,
+ 404,
+ errorType,
+ )
+ case 410:
+ throw new BridgeFatalError(
+ detail ??
+ 'Remote Control session has expired. Please restart with `claude remote-control` or /remote-control.',
+ 410,
+ errorType ?? 'environment_expired',
+ )
+ case 429:
+ throw new Error(`${context}: Rate limited (429). Polling too frequently.`)
+ default:
+ throw new Error(
+ `${context}: Failed with status ${status}${detail ? `: ${detail}` : ''}`,
+ )
+ }
+}
+
+/** Check whether an error type string indicates a session/environment expiry. */
+export function isExpiredErrorType(errorType: string | undefined): boolean {
+ if (!errorType) {
+ return false
+ }
+ return errorType.includes('expired') || errorType.includes('lifetime')
+}
+
+/**
+ * Check whether a BridgeFatalError is a suppressible 403 permission error.
+ * These are 403 errors for scopes like 'external_poll_sessions' or operations
+ * like StopWork that fail because the user's role lacks 'environments:manage'.
+ * They don't affect core functionality and shouldn't be shown to users.
+ */
+export function isSuppressible403(err: BridgeFatalError): boolean {
+ if (err.status !== 403) {
+ return false
+ }
+ return (
+ err.message.includes('external_poll_sessions') ||
+ err.message.includes('environments:manage')
+ )
+}
+
+function extractErrorTypeFromData(data: unknown): string | undefined {
+ if (data && typeof data === 'object') {
+ if (
+ 'error' in data &&
+ data.error &&
+ typeof data.error === 'object' &&
+ 'type' in data.error &&
+ typeof data.error.type === 'string'
+ ) {
+ return data.error.type
+ }
+ }
+ return undefined
+}
diff --git a/original-source-code/src/bridge/bridgeConfig.ts b/original-source-code/src/bridge/bridgeConfig.ts
new file mode 100644
index 0000000..02f0876
--- /dev/null
+++ b/original-source-code/src/bridge/bridgeConfig.ts
@@ -0,0 +1,48 @@
+/**
+ * Shared bridge auth/URL resolution. Consolidates the ant-only
+ * CLAUDE_BRIDGE_* dev overrides that were previously copy-pasted across
+ * a dozen files โ inboundAttachments, BriefTool/upload, bridgeMain,
+ * initReplBridge, remoteBridgeCore, daemon workers, /rename,
+ * /remote-control.
+ *
+ * Two layers: *Override() returns the ant-only env var (or undefined);
+ * the non-Override versions fall through to the real OAuth store/config.
+ * Callers that compose with a different auth source (e.g. daemon workers
+ * using IPC auth) use the Override getters directly.
+ */
+
+import { getOauthConfig } from '../constants/oauth.js'
+import { getClaudeAIOAuthTokens } from '../utils/auth.js'
+
+/** Ant-only dev override: CLAUDE_BRIDGE_OAUTH_TOKEN, else undefined. */
+export function getBridgeTokenOverride(): string | undefined {
+ return (
+ (process.env.USER_TYPE === 'ant' &&
+ process.env.CLAUDE_BRIDGE_OAUTH_TOKEN) ||
+ undefined
+ )
+}
+
+/** Ant-only dev override: CLAUDE_BRIDGE_BASE_URL, else undefined. */
+export function getBridgeBaseUrlOverride(): string | undefined {
+ return (
+ (process.env.USER_TYPE === 'ant' && process.env.CLAUDE_BRIDGE_BASE_URL) ||
+ undefined
+ )
+}
+
+/**
+ * Access token for bridge API calls: dev override first, then the OAuth
+ * keychain. Undefined means "not logged in".
+ */
+export function getBridgeAccessToken(): string | undefined {
+ return getBridgeTokenOverride() ?? getClaudeAIOAuthTokens()?.accessToken
+}
+
+/**
+ * Base URL for bridge API calls: dev override first, then the production
+ * OAuth config. Always returns a URL.
+ */
+export function getBridgeBaseUrl(): string {
+ return getBridgeBaseUrlOverride() ?? getOauthConfig().BASE_API_URL
+}
diff --git a/original-source-code/src/bridge/bridgeDebug.ts b/original-source-code/src/bridge/bridgeDebug.ts
new file mode 100644
index 0000000..4d0f422
--- /dev/null
+++ b/original-source-code/src/bridge/bridgeDebug.ts
@@ -0,0 +1,135 @@
+import { logForDebugging } from '../utils/debug.js'
+import { BridgeFatalError } from './bridgeApi.js'
+import type { BridgeApiClient } from './types.js'
+
+/**
+ * Ant-only fault injection for manually testing bridge recovery paths.
+ *
+ * Real failure modes this targets (BQ 2026-03-12, 7-day window):
+ * poll 404 not_found_error โ 147K sessions/week, dead onEnvironmentLost gate
+ * ws_closed 1002/1006 โ 22K sessions/week, zombie poll after close
+ * register transient failure โ residual: network blips during doReconnect
+ *
+ * Usage: /bridge-kick from the REPL while Remote Control is
+ * connected, then tail debug.log to watch the recovery machinery react.
+ *
+ * Module-level state is intentional here: one bridge per REPL process, the
+ * /bridge-kick slash command has no other way to reach into initBridgeCore's
+ * closures, and teardown clears the slot.
+ */
+
+/** One-shot fault to inject on the next matching api call. */
+type BridgeFault = {
+ method:
+ | 'pollForWork'
+ | 'registerBridgeEnvironment'
+ | 'reconnectSession'
+ | 'heartbeatWork'
+ /** Fatal errors go through handleErrorStatus โ BridgeFatalError. Transient
+ * errors surface as plain axios rejections (5xx / network). Recovery code
+ * distinguishes the two: fatal โ teardown, transient โ retry/backoff. */
+ kind: 'fatal' | 'transient'
+ status: number
+ errorType?: string
+ /** Remaining injections. Decremented on consume; removed at 0. */
+ count: number
+}
+
+export type BridgeDebugHandle = {
+ /** Invoke the transport's permanent-close handler directly. Tests the
+ * ws_closed โ reconnectEnvironmentWithSession escalation (#22148). */
+ fireClose: (code: number) => void
+ /** Call reconnectEnvironmentWithSession() โ same as SIGUSR2 but
+ * reachable from the slash command. */
+ forceReconnect: () => void
+ /** Queue a fault for the next N calls to the named api method. */
+ injectFault: (fault: BridgeFault) => void
+ /** Abort the at-capacity sleep so an injected poll fault lands
+ * immediately instead of up to 10min later. */
+ wakePollLoop: () => void
+ /** env/session IDs for the debug.log grep. */
+ describe: () => string
+}
+
+let debugHandle: BridgeDebugHandle | null = null
+const faultQueue: BridgeFault[] = []
+
+export function registerBridgeDebugHandle(h: BridgeDebugHandle): void {
+ debugHandle = h
+}
+
+export function clearBridgeDebugHandle(): void {
+ debugHandle = null
+ faultQueue.length = 0
+}
+
+export function getBridgeDebugHandle(): BridgeDebugHandle | null {
+ return debugHandle
+}
+
+export function injectBridgeFault(fault: BridgeFault): void {
+ faultQueue.push(fault)
+ logForDebugging(
+ `[bridge:debug] Queued fault: ${fault.method} ${fault.kind}/${fault.status}${fault.errorType ? `/${fault.errorType}` : ''} ร${fault.count}`,
+ )
+}
+
+/**
+ * Wrap a BridgeApiClient so each call first checks the fault queue. If a
+ * matching fault is queued, throw the specified error instead of calling
+ * through. Delegates everything else to the real client.
+ *
+ * Only called when USER_TYPE === 'ant' โ zero overhead in external builds.
+ */
+export function wrapApiForFaultInjection(
+ api: BridgeApiClient,
+): BridgeApiClient {
+ function consume(method: BridgeFault['method']): BridgeFault | null {
+ const idx = faultQueue.findIndex(f => f.method === method)
+ if (idx === -1) return null
+ const fault = faultQueue[idx]!
+ fault.count--
+ if (fault.count <= 0) faultQueue.splice(idx, 1)
+ return fault
+ }
+
+ function throwFault(fault: BridgeFault, context: string): never {
+ logForDebugging(
+ `[bridge:debug] Injecting ${fault.kind} fault into ${context}: status=${fault.status} errorType=${fault.errorType ?? 'none'}`,
+ )
+ if (fault.kind === 'fatal') {
+ throw new BridgeFatalError(
+ `[injected] ${context} ${fault.status}`,
+ fault.status,
+ fault.errorType,
+ )
+ }
+ // Transient: mimic an axios rejection (5xx / network). No .status on
+ // the error itself โ that's how the catch blocks distinguish.
+ throw new Error(`[injected transient] ${context} ${fault.status}`)
+ }
+
+ return {
+ ...api,
+ async pollForWork(envId, secret, signal, reclaimMs) {
+ const f = consume('pollForWork')
+ if (f) throwFault(f, 'Poll')
+ return api.pollForWork(envId, secret, signal, reclaimMs)
+ },
+ async registerBridgeEnvironment(config) {
+ const f = consume('registerBridgeEnvironment')
+ if (f) throwFault(f, 'Registration')
+ return api.registerBridgeEnvironment(config)
+ },
+ async reconnectSession(envId, sessionId) {
+ const f = consume('reconnectSession')
+ if (f) throwFault(f, 'ReconnectSession')
+ return api.reconnectSession(envId, sessionId)
+ },
+ async heartbeatWork(envId, workId, token) {
+ const f = consume('heartbeatWork')
+ if (f) throwFault(f, 'Heartbeat')
+ return api.heartbeatWork(envId, workId, token)
+ },
+ }
+}
diff --git a/original-source-code/src/bridge/bridgeEnabled.ts b/original-source-code/src/bridge/bridgeEnabled.ts
new file mode 100644
index 0000000..b6eec41
--- /dev/null
+++ b/original-source-code/src/bridge/bridgeEnabled.ts
@@ -0,0 +1,202 @@
+import { feature } from 'bun:bundle'
+import {
+ checkGate_CACHED_OR_BLOCKING,
+ getDynamicConfig_CACHED_MAY_BE_STALE,
+ getFeatureValue_CACHED_MAY_BE_STALE,
+} from '../services/analytics/growthbook.js'
+// Namespace import breaks the bridgeEnabled โ auth โ config โ bridgeEnabled
+// cycle โ authModule.foo is a live binding, so by the time the helpers below
+// call it, auth.js is fully loaded. Previously used require() for the same
+// deferral, but require() hits a CJS cache that diverges from the ESM
+// namespace after mock.module() (daemon/auth.test.ts), breaking spyOn.
+import * as authModule from '../utils/auth.js'
+import { isEnvTruthy } from '../utils/envUtils.js'
+import { lt } from '../utils/semver.js'
+
+/**
+ * Runtime check for bridge mode entitlement.
+ *
+ * Remote Control requires a claude.ai subscription (the bridge auths to CCR
+ * with the claude.ai OAuth token). isClaudeAISubscriber() excludes
+ * Bedrock/Vertex/Foundry, apiKeyHelper/gateway deployments, env-var API keys,
+ * and Console API logins โ none of which have the OAuth token CCR needs.
+ * See github.com/deshaw/anthropic-issues/issues/24.
+ *
+ * The `feature('BRIDGE_MODE')` guard ensures the GrowthBook string literal
+ * is only referenced when bridge mode is enabled at build time.
+ */
+export function isBridgeEnabled(): boolean {
+ // Positive ternary pattern โ see docs/feature-gating.md.
+ // Negative pattern (if (!feature(...)) return) does not eliminate
+ // inline string literals from external builds.
+ return feature('BRIDGE_MODE')
+ ? isClaudeAISubscriber() &&
+ getFeatureValue_CACHED_MAY_BE_STALE('tengu_ccr_bridge', false)
+ : false
+}
+
+/**
+ * Blocking entitlement check for Remote Control.
+ *
+ * Returns cached `true` immediately (fast path). If the disk cache says
+ * `false` or is missing, awaits GrowthBook init and fetches the fresh
+ * server value (slow path, max ~5s), then writes it to disk.
+ *
+ * Use at entitlement gates where a stale `false` would unfairly block access.
+ * For user-facing error paths, prefer `getBridgeDisabledReason()` which gives
+ * a specific diagnostic. For render-body UI visibility checks, use
+ * `isBridgeEnabled()` instead.
+ */
+export async function isBridgeEnabledBlocking(): Promise {
+ return feature('BRIDGE_MODE')
+ ? isClaudeAISubscriber() &&
+ (await checkGate_CACHED_OR_BLOCKING('tengu_ccr_bridge'))
+ : false
+}
+
+/**
+ * Diagnostic message for why Remote Control is unavailable, or null if
+ * it's enabled. Call this instead of a bare `isBridgeEnabledBlocking()`
+ * check when you need to show the user an actionable error.
+ *
+ * The GrowthBook gate targets on organizationUUID, which comes from
+ * config.oauthAccount โ populated by /api/oauth/profile during login.
+ * That endpoint requires the user:profile scope. Tokens without it
+ * (setup-token, CLAUDE_CODE_OAUTH_TOKEN env var, or pre-scope-expansion
+ * logins) leave oauthAccount unpopulated, so the gate falls back to
+ * false and users see a dead-end "not enabled" message with no hint
+ * that re-login would fix it. See CC-1165 / gh-33105.
+ */
+export async function getBridgeDisabledReason(): Promise {
+ if (feature('BRIDGE_MODE')) {
+ if (!isClaudeAISubscriber()) {
+ return 'Remote Control requires a claude.ai subscription. Run `claude auth login` to sign in with your claude.ai account.'
+ }
+ if (!hasProfileScope()) {
+ return 'Remote Control requires a full-scope login token. Long-lived tokens (from `claude setup-token` or CLAUDE_CODE_OAUTH_TOKEN) are limited to inference-only for security reasons. Run `claude auth login` to use Remote Control.'
+ }
+ if (!getOauthAccountInfo()?.organizationUuid) {
+ return 'Unable to determine your organization for Remote Control eligibility. Run `claude auth login` to refresh your account information.'
+ }
+ if (!(await checkGate_CACHED_OR_BLOCKING('tengu_ccr_bridge'))) {
+ return 'Remote Control is not yet enabled for your account.'
+ }
+ return null
+ }
+ return 'Remote Control is not available in this build.'
+}
+
+// try/catch: main.tsx:5698 calls isBridgeEnabled() while defining the Commander
+// program, before enableConfigs() runs. isClaudeAISubscriber() โ getGlobalConfig()
+// throws "Config accessed before allowed" there. Pre-config, no OAuth token can
+// exist anyway โ false is correct. Same swallow getFeatureValue_CACHED_MAY_BE_STALE
+// already does at growthbook.ts:775-780.
+function isClaudeAISubscriber(): boolean {
+ try {
+ return authModule.isClaudeAISubscriber()
+ } catch {
+ return false
+ }
+}
+function hasProfileScope(): boolean {
+ try {
+ return authModule.hasProfileScope()
+ } catch {
+ return false
+ }
+}
+function getOauthAccountInfo(): ReturnType<
+ typeof authModule.getOauthAccountInfo
+> {
+ try {
+ return authModule.getOauthAccountInfo()
+ } catch {
+ return undefined
+ }
+}
+
+/**
+ * Runtime check for the env-less (v2) REPL bridge path.
+ * Returns true when the GrowthBook flag `tengu_bridge_repl_v2` is enabled.
+ *
+ * This gates which implementation initReplBridge uses โ NOT whether bridge
+ * is available at all (see isBridgeEnabled above). Daemon/print paths stay
+ * on the env-based implementation regardless of this gate.
+ */
+export function isEnvLessBridgeEnabled(): boolean {
+ return feature('BRIDGE_MODE')
+ ? getFeatureValue_CACHED_MAY_BE_STALE('tengu_bridge_repl_v2', false)
+ : false
+}
+
+/**
+ * Kill-switch for the `cse_*` โ `session_*` client-side retag shim.
+ *
+ * The shim exists because compat/convert.go:27 validates TagSession and the
+ * claude.ai frontend routes on `session_*`, while v2 worker endpoints hand out
+ * `cse_*`. Once the server tags by environment_kind and the frontend accepts
+ * `cse_*` directly, flip this to false to make toCompatSessionId a no-op.
+ * Defaults to true โ the shim stays active until explicitly disabled.
+ */
+export function isCseShimEnabled(): boolean {
+ return feature('BRIDGE_MODE')
+ ? getFeatureValue_CACHED_MAY_BE_STALE(
+ 'tengu_bridge_repl_v2_cse_shim_enabled',
+ true,
+ )
+ : true
+}
+
+/**
+ * Returns an error message if the current CLI version is below the
+ * minimum required for the v1 (env-based) Remote Control path, or null if the
+ * version is fine. The v2 (env-less) path uses checkEnvLessBridgeMinVersion()
+ * in envLessBridgeConfig.ts instead โ the two implementations have independent
+ * version floors.
+ *
+ * Uses cached (non-blocking) GrowthBook config. If GrowthBook hasn't
+ * loaded yet, the default '0.0.0' means the check passes โ a safe fallback.
+ */
+export function checkBridgeMinVersion(): string | null {
+ // Positive pattern โ see docs/feature-gating.md.
+ // Negative pattern (if (!feature(...)) return) does not eliminate
+ // inline string literals from external builds.
+ if (feature('BRIDGE_MODE')) {
+ const config = getDynamicConfig_CACHED_MAY_BE_STALE<{
+ minVersion: string
+ }>('tengu_bridge_min_version', { minVersion: '0.0.0' })
+ if (config.minVersion && lt(MACRO.VERSION, config.minVersion)) {
+ return `Your version of Claude Code (${MACRO.VERSION}) is too old for Remote Control.\nVersion ${config.minVersion} or higher is required. Run \`claude update\` to update.`
+ }
+ }
+ return null
+}
+
+/**
+ * Default for remoteControlAtStartup when the user hasn't explicitly set it.
+ * When the CCR_AUTO_CONNECT build flag is present (ant-only) and the
+ * tengu_cobalt_harbor GrowthBook gate is on, all sessions connect to CCR by
+ * default โ the user can still opt out by setting remoteControlAtStartup=false
+ * in config (explicit settings always win over this default).
+ *
+ * Defined here rather than in config.ts to avoid a direct
+ * config.ts โ growthbook.ts import cycle (growthbook.ts โ user.ts โ config.ts).
+ */
+export function getCcrAutoConnectDefault(): boolean {
+ return feature('CCR_AUTO_CONNECT')
+ ? getFeatureValue_CACHED_MAY_BE_STALE('tengu_cobalt_harbor', false)
+ : false
+}
+
+/**
+ * Opt-in CCR mirror mode โ every local session spawns an outbound-only
+ * Remote Control session that receives forwarded events. Separate from
+ * getCcrAutoConnectDefault (bidirectional Remote Control). Env var wins for
+ * local opt-in; GrowthBook controls rollout.
+ */
+export function isCcrMirrorEnabled(): boolean {
+ return feature('CCR_MIRROR')
+ ? isEnvTruthy(process.env.CLAUDE_CODE_CCR_MIRROR) ||
+ getFeatureValue_CACHED_MAY_BE_STALE('tengu_ccr_mirror', false)
+ : false
+}
diff --git a/original-source-code/src/bridge/bridgeMain.ts b/original-source-code/src/bridge/bridgeMain.ts
new file mode 100644
index 0000000..7aeacaf
--- /dev/null
+++ b/original-source-code/src/bridge/bridgeMain.ts
@@ -0,0 +1,2999 @@
+import { feature } from 'bun:bundle'
+import { randomUUID } from 'crypto'
+import { hostname, tmpdir } from 'os'
+import { basename, join, resolve } from 'path'
+import { getRemoteSessionUrl } from '../constants/product.js'
+import { shutdownDatadog } from '../services/analytics/datadog.js'
+import { shutdown1PEventLogging } from '../services/analytics/firstPartyEventLogger.js'
+import { checkGate_CACHED_OR_BLOCKING } from '../services/analytics/growthbook.js'
+import {
+ type AnalyticsMetadata_I_VERIFIED_THIS_IS_NOT_CODE_OR_FILEPATHS,
+ logEvent,
+ logEventAsync,
+} from '../services/analytics/index.js'
+import { isInBundledMode } from '../utils/bundledMode.js'
+import { logForDebugging } from '../utils/debug.js'
+import { logForDiagnosticsNoPII } from '../utils/diagLogs.js'
+import { isEnvTruthy, isInProtectedNamespace } from '../utils/envUtils.js'
+import { errorMessage } from '../utils/errors.js'
+import { truncateToWidth } from '../utils/format.js'
+import { logError } from '../utils/log.js'
+import { sleep } from '../utils/sleep.js'
+import { createAgentWorktree, removeAgentWorktree } from '../utils/worktree.js'
+import {
+ BridgeFatalError,
+ createBridgeApiClient,
+ isExpiredErrorType,
+ isSuppressible403,
+ validateBridgeId,
+} from './bridgeApi.js'
+import { formatDuration } from './bridgeStatusUtil.js'
+import { createBridgeLogger } from './bridgeUI.js'
+import { createCapacityWake } from './capacityWake.js'
+import { describeAxiosError } from './debugUtils.js'
+import { createTokenRefreshScheduler } from './jwtUtils.js'
+import { getPollIntervalConfig } from './pollConfig.js'
+import { toCompatSessionId, toInfraSessionId } from './sessionIdCompat.js'
+import { createSessionSpawner, safeFilenameId } from './sessionRunner.js'
+import { getTrustedDeviceToken } from './trustedDevice.js'
+import {
+ BRIDGE_LOGIN_ERROR,
+ type BridgeApiClient,
+ type BridgeConfig,
+ type BridgeLogger,
+ DEFAULT_SESSION_TIMEOUT_MS,
+ type SessionDoneStatus,
+ type SessionHandle,
+ type SessionSpawner,
+ type SessionSpawnOpts,
+ type SpawnMode,
+} from './types.js'
+import {
+ buildCCRv2SdkUrl,
+ buildSdkUrl,
+ decodeWorkSecret,
+ registerWorker,
+ sameSessionId,
+} from './workSecret.js'
+
+export type BackoffConfig = {
+ connInitialMs: number
+ connCapMs: number
+ connGiveUpMs: number
+ generalInitialMs: number
+ generalCapMs: number
+ generalGiveUpMs: number
+ /** SIGTERMโSIGKILL grace period on shutdown. Default 30s. */
+ shutdownGraceMs?: number
+ /** stopWorkWithRetry base delay (1s/2s/4s backoff). Default 1000ms. */
+ stopWorkBaseDelayMs?: number
+}
+
+const DEFAULT_BACKOFF: BackoffConfig = {
+ connInitialMs: 2_000,
+ connCapMs: 120_000, // 2 minutes
+ connGiveUpMs: 600_000, // 10 minutes
+ generalInitialMs: 500,
+ generalCapMs: 30_000,
+ generalGiveUpMs: 600_000, // 10 minutes
+}
+
+/** Status update interval for the live display (ms). */
+const STATUS_UPDATE_INTERVAL_MS = 1_000
+const SPAWN_SESSIONS_DEFAULT = 32
+
+/**
+ * GrowthBook gate for multi-session spawn modes (--spawn / --capacity / --create-session-in-dir).
+ * Sibling of tengu_ccr_bridge_multi_environment (multiple envs per host:dir) โ
+ * this one enables multiple sessions per environment.
+ * Rollout staged via targeting rules: ants first, then gradual external.
+ *
+ * Uses the blocking gate check so a stale disk-cache miss doesn't unfairly
+ * deny access. The fast path (cache has true) is still instant; only the
+ * cold-start path awaits the server fetch, and that fetch also seeds the
+ * disk cache for next time.
+ */
+async function isMultiSessionSpawnEnabled(): Promise {
+ return checkGate_CACHED_OR_BLOCKING('tengu_ccr_bridge_multi_session')
+}
+
+/**
+ * Returns the threshold for detecting system sleep/wake in the poll loop.
+ * Must exceed the max backoff cap โ otherwise normal backoff delays trigger
+ * false sleep detection (resetting the error budget indefinitely). Using
+ * 2ร the connection backoff cap, matching the pattern in WebSocketTransport
+ * and replBridge.
+ */
+function pollSleepDetectionThresholdMs(backoff: BackoffConfig): number {
+ return backoff.connCapMs * 2
+}
+
+/**
+ * Returns the args that must precede CLI flags when spawning a child claude
+ * process. In compiled binaries, process.execPath is the claude binary itself
+ * and args go directly to it. In npm installs (node running cli.js),
+ * process.execPath is the node runtime โ the child spawn must pass the script
+ * path as the first arg, otherwise node interprets --sdk-url as a node option
+ * and exits with "bad option: --sdk-url". See anthropics/claude-code#28334.
+ */
+function spawnScriptArgs(): string[] {
+ if (isInBundledMode() || !process.argv[1]) {
+ return []
+ }
+ return [process.argv[1]]
+}
+
+/** Attempt to spawn a session; returns error string if spawn throws. */
+function safeSpawn(
+ spawner: SessionSpawner,
+ opts: SessionSpawnOpts,
+ dir: string,
+): SessionHandle | string {
+ try {
+ return spawner.spawn(opts, dir)
+ } catch (err) {
+ const errMsg = errorMessage(err)
+ logError(new Error(`Session spawn failed: ${errMsg}`))
+ return errMsg
+ }
+}
+
+export async function runBridgeLoop(
+ config: BridgeConfig,
+ environmentId: string,
+ environmentSecret: string,
+ api: BridgeApiClient,
+ spawner: SessionSpawner,
+ logger: BridgeLogger,
+ signal: AbortSignal,
+ backoffConfig: BackoffConfig = DEFAULT_BACKOFF,
+ initialSessionId?: string,
+ getAccessToken?: () => string | undefined | Promise,
+): Promise {
+ // Local abort controller so that onSessionDone can stop the poll loop.
+ // Linked to the incoming signal so external aborts also work.
+ const controller = new AbortController()
+ if (signal.aborted) {
+ controller.abort()
+ } else {
+ signal.addEventListener('abort', () => controller.abort(), { once: true })
+ }
+ const loopSignal = controller.signal
+
+ const activeSessions = new Map()
+ const sessionStartTimes = new Map()
+ const sessionWorkIds = new Map()
+ // Compat-surface ID (session_*) computed once at spawn and cached so
+ // cleanup and status-update ticks use the same key regardless of whether
+ // the tengu_bridge_repl_v2_cse_shim_enabled gate flips mid-session.
+ const sessionCompatIds = new Map()
+ // Session ingress JWTs for heartbeat auth, keyed by sessionId.
+ // Stored separately from handle.accessToken because the token refresh
+ // scheduler overwrites that field with the OAuth token (~3h55m in).
+ const sessionIngressTokens = new Map()
+ const sessionTimers = new Map>()
+ const completedWorkIds = new Set()
+ const sessionWorktrees = new Map<
+ string,
+ {
+ worktreePath: string
+ worktreeBranch?: string
+ gitRoot?: string
+ hookBased?: boolean
+ }
+ >()
+ // Track sessions killed by the timeout watchdog so onSessionDone can
+ // distinguish them from server-initiated or shutdown interrupts.
+ const timedOutSessions = new Set()
+ // Sessions that already have a title (server-set or bridge-derived) so
+ // onFirstUserMessage doesn't clobber a user-assigned --name / web rename.
+ // Keyed by compatSessionId to match logger.setSessionTitle's key.
+ const titledSessions = new Set()
+ // Signal to wake the at-capacity sleep early when a session completes,
+ // so the bridge can immediately accept new work.
+ const capacityWake = createCapacityWake(loopSignal)
+
+ /**
+ * Heartbeat all active work items.
+ * Returns 'ok' if at least one heartbeat succeeded, 'auth_failed' if any
+ * got a 401/403 (JWT expired โ re-queued via reconnectSession so the next
+ * poll delivers fresh work), or 'failed' if all failed for other reasons.
+ */
+ async function heartbeatActiveWorkItems(): Promise<
+ 'ok' | 'auth_failed' | 'fatal' | 'failed'
+ > {
+ let anySuccess = false
+ let anyFatal = false
+ const authFailedSessions: string[] = []
+ for (const [sessionId] of activeSessions) {
+ const workId = sessionWorkIds.get(sessionId)
+ const ingressToken = sessionIngressTokens.get(sessionId)
+ if (!workId || !ingressToken) {
+ continue
+ }
+ try {
+ await api.heartbeatWork(environmentId, workId, ingressToken)
+ anySuccess = true
+ } catch (err) {
+ logForDebugging(
+ `[bridge:heartbeat] Failed for sessionId=${sessionId} workId=${workId}: ${errorMessage(err)}`,
+ )
+ if (err instanceof BridgeFatalError) {
+ logEvent('tengu_bridge_heartbeat_error', {
+ status:
+ err.status as unknown as AnalyticsMetadata_I_VERIFIED_THIS_IS_NOT_CODE_OR_FILEPATHS,
+ error_type: (err.status === 401 || err.status === 403
+ ? 'auth_failed'
+ : 'fatal') as AnalyticsMetadata_I_VERIFIED_THIS_IS_NOT_CODE_OR_FILEPATHS,
+ })
+ if (err.status === 401 || err.status === 403) {
+ authFailedSessions.push(sessionId)
+ } else {
+ // 404/410 = environment expired or deleted โ no point retrying
+ anyFatal = true
+ }
+ }
+ }
+ }
+ // JWT expired โ trigger server-side re-dispatch. Without this, work stays
+ // ACK'd out of the Redis PEL and poll returns empty forever (CC-1263).
+ // The existingHandle path below delivers the fresh token to the child.
+ // sessionId is already in the format /bridge/reconnect expects: it comes
+ // from work.data.id, which matches the server's EnvironmentInstance store
+ // (cse_* under the compat gate, session_* otherwise).
+ for (const sessionId of authFailedSessions) {
+ logger.logVerbose(
+ `Session ${sessionId} token expired โ re-queuing via bridge/reconnect`,
+ )
+ try {
+ await api.reconnectSession(environmentId, sessionId)
+ logForDebugging(
+ `[bridge:heartbeat] Re-queued sessionId=${sessionId} via bridge/reconnect`,
+ )
+ } catch (err) {
+ logger.logError(
+ `Failed to refresh session ${sessionId} token: ${errorMessage(err)}`,
+ )
+ logForDebugging(
+ `[bridge:heartbeat] reconnectSession(${sessionId}) failed: ${errorMessage(err)}`,
+ { level: 'error' },
+ )
+ }
+ }
+ if (anyFatal) {
+ return 'fatal'
+ }
+ if (authFailedSessions.length > 0) {
+ return 'auth_failed'
+ }
+ return anySuccess ? 'ok' : 'failed'
+ }
+
+ // Sessions spawned with CCR v2 env vars. v2 children cannot use OAuth
+ // tokens (CCR worker endpoints validate the JWT's session_id claim,
+ // register_worker.go:32), so onRefresh triggers server re-dispatch
+ // instead โ the next poll delivers fresh work with a new JWT via the
+ // existingHandle path below.
+ const v2Sessions = new Set()
+
+ // Proactive token refresh: schedules a timer 5min before the session
+ // ingress JWT expires. v1 delivers OAuth directly; v2 calls
+ // reconnectSession to trigger server re-dispatch (CC-1263: without
+ // this, v2 daemon sessions silently die at ~5h since the server does
+ // not auto-re-dispatch ACK'd work on lease expiry).
+ const tokenRefresh = getAccessToken
+ ? createTokenRefreshScheduler({
+ getAccessToken,
+ onRefresh: (sessionId, oauthToken) => {
+ const handle = activeSessions.get(sessionId)
+ if (!handle) {
+ return
+ }
+ if (v2Sessions.has(sessionId)) {
+ logger.logVerbose(
+ `Refreshing session ${sessionId} token via bridge/reconnect`,
+ )
+ void api
+ .reconnectSession(environmentId, sessionId)
+ .catch((err: unknown) => {
+ logger.logError(
+ `Failed to refresh session ${sessionId} token: ${errorMessage(err)}`,
+ )
+ logForDebugging(
+ `[bridge:token] reconnectSession(${sessionId}) failed: ${errorMessage(err)}`,
+ { level: 'error' },
+ )
+ })
+ } else {
+ handle.updateAccessToken(oauthToken)
+ }
+ },
+ label: 'bridge',
+ })
+ : null
+ const loopStartTime = Date.now()
+ // Track all in-flight cleanup promises (stopWork, worktree removal) so
+ // the shutdown sequence can await them before process.exit().
+ const pendingCleanups = new Set>()
+ function trackCleanup(p: Promise): void {
+ pendingCleanups.add(p)
+ void p.finally(() => pendingCleanups.delete(p))
+ }
+ let connBackoff = 0
+ let generalBackoff = 0
+ let connErrorStart: number | null = null
+ let generalErrorStart: number | null = null
+ let lastPollErrorTime: number | null = null
+ let statusUpdateTimer: ReturnType | null = null
+ // Set by BridgeFatalError and give-up paths so the shutdown block can
+ // skip the resume message (resume is impossible after env expiry/auth
+ // failure/sustained connection errors).
+ let fatalExit = false
+
+ logForDebugging(
+ `[bridge:work] Starting poll loop spawnMode=${config.spawnMode} maxSessions=${config.maxSessions} environmentId=${environmentId}`,
+ )
+ logForDiagnosticsNoPII('info', 'bridge_loop_started', {
+ max_sessions: config.maxSessions,
+ spawn_mode: config.spawnMode,
+ })
+
+ // For ant users, show where session debug logs will land so they can tail them.
+ // sessionRunner.ts uses the same base path. File appears once a session spawns.
+ if (process.env.USER_TYPE === 'ant') {
+ let debugGlob: string
+ if (config.debugFile) {
+ const ext = config.debugFile.lastIndexOf('.')
+ debugGlob =
+ ext > 0
+ ? `${config.debugFile.slice(0, ext)}-*${config.debugFile.slice(ext)}`
+ : `${config.debugFile}-*`
+ } else {
+ debugGlob = join(tmpdir(), 'claude', 'bridge-session-*.log')
+ }
+ logger.setDebugLogPath(debugGlob)
+ }
+
+ logger.printBanner(config, environmentId)
+
+ // Seed the logger's session count + spawn mode before any render. Without
+ // this, setAttached() below renders with the logger's default sessionMax=1,
+ // showing "Capacity: 0/1" until the status ticker kicks in (which is gated
+ // by !initialSessionId and only starts after the poll loop picks up work).
+ logger.updateSessionCount(0, config.maxSessions, config.spawnMode)
+
+ // If an initial session was pre-created, show its URL from the start so
+ // the user can click through immediately (matching /remote-control behavior).
+ if (initialSessionId) {
+ logger.setAttached(initialSessionId)
+ }
+
+ /** Refresh the inline status display. Shows idle or active depending on state. */
+ function updateStatusDisplay(): void {
+ // Push the session count (no-op when maxSessions === 1) so the
+ // next renderStatusLine tick shows the current count.
+ logger.updateSessionCount(
+ activeSessions.size,
+ config.maxSessions,
+ config.spawnMode,
+ )
+
+ // Push per-session activity into the multi-session display.
+ for (const [sid, handle] of activeSessions) {
+ const act = handle.currentActivity
+ if (act) {
+ logger.updateSessionActivity(sessionCompatIds.get(sid) ?? sid, act)
+ }
+ }
+
+ if (activeSessions.size === 0) {
+ logger.updateIdleStatus()
+ return
+ }
+
+ // Show the most recently started session that is still actively working.
+ // Sessions whose current activity is 'result' or 'error' are between
+ // turns โ the CLI emitted its result but the process stays alive waiting
+ // for the next user message. Skip updating so the status line keeps
+ // whatever state it had (Attached / session title).
+ const [sessionId, handle] = [...activeSessions.entries()].pop()!
+ const startTime = sessionStartTimes.get(sessionId)
+ if (!startTime) return
+
+ const activity = handle.currentActivity
+ if (!activity || activity.type === 'result' || activity.type === 'error') {
+ // Session is between turns โ keep current status (Attached/titled).
+ // In multi-session mode, still refresh so bullet-list activities stay current.
+ if (config.maxSessions > 1) logger.refreshDisplay()
+ return
+ }
+
+ const elapsed = formatDuration(Date.now() - startTime)
+
+ // Build trail from recent tool activities (last 5)
+ const trail = handle.activities
+ .filter(a => a.type === 'tool_start')
+ .slice(-5)
+ .map(a => a.summary)
+
+ logger.updateSessionStatus(sessionId, elapsed, activity, trail)
+ }
+
+ /** Start the status display update ticker. */
+ function startStatusUpdates(): void {
+ stopStatusUpdates()
+ // Call immediately so the first transition (e.g. Connecting โ Ready)
+ // happens without delay, avoiding concurrent timer races.
+ updateStatusDisplay()
+ statusUpdateTimer = setInterval(
+ updateStatusDisplay,
+ STATUS_UPDATE_INTERVAL_MS,
+ )
+ }
+
+ /** Stop the status display update ticker. */
+ function stopStatusUpdates(): void {
+ if (statusUpdateTimer) {
+ clearInterval(statusUpdateTimer)
+ statusUpdateTimer = null
+ }
+ }
+
+ function onSessionDone(
+ sessionId: string,
+ startTime: number,
+ handle: SessionHandle,
+ ): (status: SessionDoneStatus) => void {
+ return (rawStatus: SessionDoneStatus): void => {
+ const workId = sessionWorkIds.get(sessionId)
+ activeSessions.delete(sessionId)
+ sessionStartTimes.delete(sessionId)
+ sessionWorkIds.delete(sessionId)
+ sessionIngressTokens.delete(sessionId)
+ const compatId = sessionCompatIds.get(sessionId) ?? sessionId
+ sessionCompatIds.delete(sessionId)
+ logger.removeSession(compatId)
+ titledSessions.delete(compatId)
+ v2Sessions.delete(sessionId)
+ // Clear per-session timeout timer
+ const timer = sessionTimers.get(sessionId)
+ if (timer) {
+ clearTimeout(timer)
+ sessionTimers.delete(sessionId)
+ }
+ // Clear token refresh timer
+ tokenRefresh?.cancel(sessionId)
+ // Wake the at-capacity sleep so the bridge can accept new work immediately
+ capacityWake.wake()
+
+ // If the session was killed by the timeout watchdog, treat it as a
+ // failed session (not a server/shutdown interrupt) so we still call
+ // stopWork and archiveSession below.
+ const wasTimedOut = timedOutSessions.delete(sessionId)
+ const status: SessionDoneStatus =
+ wasTimedOut && rawStatus === 'interrupted' ? 'failed' : rawStatus
+ const durationMs = Date.now() - startTime
+
+ logForDebugging(
+ `[bridge:session] sessionId=${sessionId} workId=${workId ?? 'unknown'} exited status=${status} duration=${formatDuration(durationMs)}`,
+ )
+ logEvent('tengu_bridge_session_done', {
+ status:
+ status as AnalyticsMetadata_I_VERIFIED_THIS_IS_NOT_CODE_OR_FILEPATHS,
+ duration_ms: durationMs,
+ })
+ logForDiagnosticsNoPII('info', 'bridge_session_done', {
+ status,
+ duration_ms: durationMs,
+ })
+
+ // Clear the status display before printing final log
+ logger.clearStatus()
+ stopStatusUpdates()
+
+ // Build error message from stderr if available
+ const stderrSummary =
+ handle.lastStderr.length > 0 ? handle.lastStderr.join('\n') : undefined
+ let failureMessage: string | undefined
+
+ switch (status) {
+ case 'completed':
+ logger.logSessionComplete(sessionId, durationMs)
+ break
+ case 'failed':
+ // Skip failure log during shutdown โ the child exits non-zero when
+ // killed, which is expected and not a real failure.
+ // Also skip for timeout-killed sessions โ the timeout watchdog
+ // already logged a clear timeout message.
+ if (!wasTimedOut && !loopSignal.aborted) {
+ failureMessage = stderrSummary ?? 'Process exited with error'
+ logger.logSessionFailed(sessionId, failureMessage)
+ logError(new Error(`Bridge session failed: ${failureMessage}`))
+ }
+ break
+ case 'interrupted':
+ logger.logVerbose(`Session ${sessionId} interrupted`)
+ break
+ }
+
+ // Notify the server that this work item is done. Skip for interrupted
+ // sessions โ interrupts are either server-initiated (the server already
+ // knows) or caused by bridge shutdown (which calls stopWork() separately).
+ if (status !== 'interrupted' && workId) {
+ trackCleanup(
+ stopWorkWithRetry(
+ api,
+ environmentId,
+ workId,
+ logger,
+ backoffConfig.stopWorkBaseDelayMs,
+ ),
+ )
+ completedWorkIds.add(workId)
+ }
+
+ // Clean up worktree if one was created for this session
+ const wt = sessionWorktrees.get(sessionId)
+ if (wt) {
+ sessionWorktrees.delete(sessionId)
+ trackCleanup(
+ removeAgentWorktree(
+ wt.worktreePath,
+ wt.worktreeBranch,
+ wt.gitRoot,
+ wt.hookBased,
+ ).catch((err: unknown) =>
+ logger.logVerbose(
+ `Failed to remove worktree ${wt.worktreePath}: ${errorMessage(err)}`,
+ ),
+ ),
+ )
+ }
+
+ // Lifecycle decision: in multi-session mode, keep the bridge running
+ // after a session completes. In single-session mode, abort the poll
+ // loop so the bridge exits cleanly.
+ if (status !== 'interrupted' && !loopSignal.aborted) {
+ if (config.spawnMode !== 'single-session') {
+ // Multi-session: archive the completed session so it doesn't linger
+ // as stale in the web UI. archiveSession is idempotent (409 if already
+ // archived), so double-archiving at shutdown is safe.
+ // sessionId arrived as cse_* from the work poll (infrastructure-layer
+ // tag). archiveSession hits /v1/sessions/{id}/archive which is the
+ // compat surface and validates TagSession (session_*). Re-tag โ same
+ // UUID underneath.
+ trackCleanup(
+ api
+ .archiveSession(compatId)
+ .catch((err: unknown) =>
+ logger.logVerbose(
+ `Failed to archive session ${sessionId}: ${errorMessage(err)}`,
+ ),
+ ),
+ )
+ logForDebugging(
+ `[bridge:session] Session ${status}, returning to idle (multi-session mode)`,
+ )
+ } else {
+ // Single-session: coupled lifecycle โ tear down environment
+ logForDebugging(
+ `[bridge:session] Session ${status}, aborting poll loop to tear down environment`,
+ )
+ controller.abort()
+ return
+ }
+ }
+
+ if (!loopSignal.aborted) {
+ startStatusUpdates()
+ }
+ }
+ }
+
+ // Start the idle status display immediately โ unless we have a pre-created
+ // session, in which case setAttached() already set up the display and the
+ // poll loop will start status updates when it picks up the session.
+ if (!initialSessionId) {
+ startStatusUpdates()
+ }
+
+ while (!loopSignal.aborted) {
+ // Fetched once per iteration โ the GrowthBook cache refreshes every
+ // 5 min, so a loop running at the at-capacity rate picks up config
+ // changes within one sleep cycle.
+ const pollConfig = getPollIntervalConfig()
+
+ try {
+ const work = await api.pollForWork(
+ environmentId,
+ environmentSecret,
+ loopSignal,
+ pollConfig.reclaim_older_than_ms,
+ )
+
+ // Log reconnection if we were previously disconnected
+ const wasDisconnected =
+ connErrorStart !== null || generalErrorStart !== null
+ if (wasDisconnected) {
+ const disconnectedMs =
+ Date.now() - (connErrorStart ?? generalErrorStart ?? Date.now())
+ logger.logReconnected(disconnectedMs)
+ logForDebugging(
+ `[bridge:poll] Reconnected after ${formatDuration(disconnectedMs)}`,
+ )
+ logEvent('tengu_bridge_reconnected', {
+ disconnected_ms: disconnectedMs,
+ })
+ }
+
+ connBackoff = 0
+ generalBackoff = 0
+ connErrorStart = null
+ generalErrorStart = null
+ lastPollErrorTime = null
+
+ // Null response = no work available in the queue.
+ // Add a minimum delay to avoid hammering the server.
+ if (!work) {
+ // Use live check (not a snapshot) since sessions can end during poll.
+ const atCap = activeSessions.size >= config.maxSessions
+ if (atCap) {
+ const atCapMs = pollConfig.multisession_poll_interval_ms_at_capacity
+ // Heartbeat loops WITHOUT polling. When at-capacity polling is also
+ // enabled (atCapMs > 0), the loop tracks a deadline and breaks out
+ // to poll at that interval โ heartbeat and poll compose instead of
+ // one suppressing the other. We break out to poll when:
+ // - Poll deadline reached (atCapMs > 0 only)
+ // - Auth fails (JWT expired โ poll refreshes tokens)
+ // - Capacity wake fires (session ended โ poll for new work)
+ // - Loop aborted (shutdown)
+ if (pollConfig.non_exclusive_heartbeat_interval_ms > 0) {
+ logEvent('tengu_bridge_heartbeat_mode_entered', {
+ active_sessions: activeSessions.size,
+ heartbeat_interval_ms:
+ pollConfig.non_exclusive_heartbeat_interval_ms,
+ })
+ // Deadline computed once at entry โ GB updates to atCapMs don't
+ // shift an in-flight deadline (next entry picks up the new value).
+ const pollDeadline = atCapMs > 0 ? Date.now() + atCapMs : null
+ let hbResult: 'ok' | 'auth_failed' | 'fatal' | 'failed' = 'ok'
+ let hbCycles = 0
+ while (
+ !loopSignal.aborted &&
+ activeSessions.size >= config.maxSessions &&
+ (pollDeadline === null || Date.now() < pollDeadline)
+ ) {
+ // Re-read config each cycle so GrowthBook updates take effect
+ const hbConfig = getPollIntervalConfig()
+ if (hbConfig.non_exclusive_heartbeat_interval_ms <= 0) break
+
+ // Capture capacity signal BEFORE the async heartbeat call so
+ // a session ending during the HTTP request is caught by the
+ // subsequent sleep (instead of being lost to a replaced controller).
+ const cap = capacityWake.signal()
+
+ hbResult = await heartbeatActiveWorkItems()
+ if (hbResult === 'auth_failed' || hbResult === 'fatal') {
+ cap.cleanup()
+ break
+ }
+
+ hbCycles++
+ await sleep(
+ hbConfig.non_exclusive_heartbeat_interval_ms,
+ cap.signal,
+ )
+ cap.cleanup()
+ }
+
+ // Determine exit reason for telemetry
+ const exitReason =
+ hbResult === 'auth_failed' || hbResult === 'fatal'
+ ? hbResult
+ : loopSignal.aborted
+ ? 'shutdown'
+ : activeSessions.size < config.maxSessions
+ ? 'capacity_changed'
+ : pollDeadline !== null && Date.now() >= pollDeadline
+ ? 'poll_due'
+ : 'config_disabled'
+ logEvent('tengu_bridge_heartbeat_mode_exited', {
+ reason:
+ exitReason as AnalyticsMetadata_I_VERIFIED_THIS_IS_NOT_CODE_OR_FILEPATHS,
+ heartbeat_cycles: hbCycles,
+ active_sessions: activeSessions.size,
+ })
+ if (exitReason === 'poll_due') {
+ // bridgeApi throttles empty-poll logs (EMPTY_POLL_LOG_INTERVAL=100)
+ // so the once-per-10min poll_due poll is invisible at counter=2.
+ // Log it here so verification runs see both endpoints in the debug log.
+ logForDebugging(
+ `[bridge:poll] Heartbeat poll_due after ${hbCycles} cycles โ falling through to pollForWork`,
+ )
+ }
+
+ // On auth_failed or fatal, sleep before polling to avoid a tight
+ // poll+heartbeat loop. Auth_failed: heartbeatActiveWorkItems
+ // already called reconnectSession โ the sleep gives the server
+ // time to propagate the re-queue. Fatal (404/410): may be a
+ // single work item GCd while the environment is still valid.
+ // Use atCapMs if enabled, else the heartbeat interval as a floor
+ // (guaranteed > 0 here) so heartbeat-only configs don't tight-loop.
+ if (hbResult === 'auth_failed' || hbResult === 'fatal') {
+ const cap = capacityWake.signal()
+ await sleep(
+ atCapMs > 0
+ ? atCapMs
+ : pollConfig.non_exclusive_heartbeat_interval_ms,
+ cap.signal,
+ )
+ cap.cleanup()
+ }
+ } else if (atCapMs > 0) {
+ // Heartbeat disabled: slow poll as liveness signal.
+ const cap = capacityWake.signal()
+ await sleep(atCapMs, cap.signal)
+ cap.cleanup()
+ }
+ } else {
+ const interval =
+ activeSessions.size > 0
+ ? pollConfig.multisession_poll_interval_ms_partial_capacity
+ : pollConfig.multisession_poll_interval_ms_not_at_capacity
+ await sleep(interval, loopSignal)
+ }
+ continue
+ }
+
+ // At capacity โ we polled to keep the heartbeat alive, but cannot
+ // accept new work right now. We still enter the switch below so that
+ // token refreshes for existing sessions are processed (the case
+ // 'session' handler checks for existing sessions before the inner
+ // capacity guard).
+ const atCapacityBeforeSwitch = activeSessions.size >= config.maxSessions
+
+ // Skip work items that have already been completed and stopped.
+ // The server may re-deliver stale work before processing our stop
+ // request, which would otherwise cause a duplicate session spawn.
+ if (completedWorkIds.has(work.id)) {
+ logForDebugging(
+ `[bridge:work] Skipping already-completed workId=${work.id}`,
+ )
+ // Respect capacity throttle โ without a sleep here, persistent stale
+ // redeliveries would tight-loop at poll-request speed (the !work
+ // branch above is the only sleep, and work != null skips it).
+ if (atCapacityBeforeSwitch) {
+ const cap = capacityWake.signal()
+ if (pollConfig.non_exclusive_heartbeat_interval_ms > 0) {
+ await heartbeatActiveWorkItems()
+ await sleep(
+ pollConfig.non_exclusive_heartbeat_interval_ms,
+ cap.signal,
+ )
+ } else if (pollConfig.multisession_poll_interval_ms_at_capacity > 0) {
+ await sleep(
+ pollConfig.multisession_poll_interval_ms_at_capacity,
+ cap.signal,
+ )
+ }
+ cap.cleanup()
+ } else {
+ await sleep(1000, loopSignal)
+ }
+ continue
+ }
+
+ // Decode the work secret for session spawning and to extract the JWT
+ // used for the ack call below.
+ let secret
+ try {
+ secret = decodeWorkSecret(work.secret)
+ } catch (err) {
+ const errMsg = errorMessage(err)
+ logger.logError(
+ `Failed to decode work secret for workId=${work.id}: ${errMsg}`,
+ )
+ logEvent('tengu_bridge_work_secret_failed', {})
+ // Can't ack (needs the JWT we failed to decode). stopWork uses OAuth,
+ // so it's callable here โ prevents XAUTOCLAIM from re-delivering this
+ // poisoned item every reclaim_older_than_ms cycle.
+ completedWorkIds.add(work.id)
+ trackCleanup(
+ stopWorkWithRetry(
+ api,
+ environmentId,
+ work.id,
+ logger,
+ backoffConfig.stopWorkBaseDelayMs,
+ ),
+ )
+ // Respect capacity throttle before retrying โ without a sleep here,
+ // repeated decode failures at capacity would tight-loop at
+ // poll-request speed (work != null skips the !work sleep above).
+ if (atCapacityBeforeSwitch) {
+ const cap = capacityWake.signal()
+ if (pollConfig.non_exclusive_heartbeat_interval_ms > 0) {
+ await heartbeatActiveWorkItems()
+ await sleep(
+ pollConfig.non_exclusive_heartbeat_interval_ms,
+ cap.signal,
+ )
+ } else if (pollConfig.multisession_poll_interval_ms_at_capacity > 0) {
+ await sleep(
+ pollConfig.multisession_poll_interval_ms_at_capacity,
+ cap.signal,
+ )
+ }
+ cap.cleanup()
+ }
+ continue
+ }
+
+ // Explicitly acknowledge after committing to handle the work โ NOT
+ // before. The at-capacity guard inside case 'session' can break
+ // without spawning; acking there would permanently lose the work.
+ // Ack failures are non-fatal: server re-delivers, and existingHandle
+ // / completedWorkIds paths handle the dedup.
+ const ackWork = async (): Promise => {
+ logForDebugging(`[bridge:work] Acknowledging workId=${work.id}`)
+ try {
+ await api.acknowledgeWork(
+ environmentId,
+ work.id,
+ secret.session_ingress_token,
+ )
+ } catch (err) {
+ logForDebugging(
+ `[bridge:work] Acknowledge failed workId=${work.id}: ${errorMessage(err)}`,
+ )
+ }
+ }
+
+ const workType: string = work.data.type
+ switch (work.data.type) {
+ case 'healthcheck':
+ await ackWork()
+ logForDebugging('[bridge:work] Healthcheck received')
+ logger.logVerbose('Healthcheck received')
+ break
+ case 'session': {
+ const sessionId = work.data.id
+ try {
+ validateBridgeId(sessionId, 'session_id')
+ } catch {
+ await ackWork()
+ logger.logError(`Invalid session_id received: ${sessionId}`)
+ break
+ }
+
+ // If the session is already running, deliver the fresh token so
+ // the child process can reconnect its WebSocket with the new
+ // session ingress token. This handles the case where the server
+ // re-dispatches work for an existing session after the WS drops.
+ const existingHandle = activeSessions.get(sessionId)
+ if (existingHandle) {
+ existingHandle.updateAccessToken(secret.session_ingress_token)
+ sessionIngressTokens.set(sessionId, secret.session_ingress_token)
+ sessionWorkIds.set(sessionId, work.id)
+ // Re-schedule next refresh from the fresh JWT's expiry. onRefresh
+ // branches on v2Sessions so both v1 and v2 are safe here.
+ tokenRefresh?.schedule(sessionId, secret.session_ingress_token)
+ logForDebugging(
+ `[bridge:work] Updated access token for existing sessionId=${sessionId} workId=${work.id}`,
+ )
+ await ackWork()
+ break
+ }
+
+ // At capacity โ token refresh for existing sessions is handled
+ // above, but we cannot spawn new ones. The post-switch capacity
+ // sleep will throttle the loop; just break here.
+ if (activeSessions.size >= config.maxSessions) {
+ logForDebugging(
+ `[bridge:work] At capacity (${activeSessions.size}/${config.maxSessions}), cannot spawn new session for workId=${work.id}`,
+ )
+ break
+ }
+
+ await ackWork()
+ const spawnStartTime = Date.now()
+
+ // CCR v2 path: register this bridge as the session worker, get the
+ // epoch, and point the child at /v1/code/sessions/{id}. The child
+ // already has the full v2 client (SSETransport + CCRClient) โ same
+ // code path environment-manager launches in containers.
+ //
+ // v1 path: Session-Ingress WebSocket. Uses config.sessionIngressUrl
+ // (not secret.api_base_url, which may point to a remote proxy tunnel
+ // that doesn't know about locally-created sessions).
+ let sdkUrl: string
+ let useCcrV2 = false
+ let workerEpoch: number | undefined
+ // Server decides per-session via the work secret; env var is the
+ // ant-dev override (e.g. forcing v2 before the server flag is on).
+ if (
+ secret.use_code_sessions === true ||
+ isEnvTruthy(process.env.CLAUDE_BRIDGE_USE_CCR_V2)
+ ) {
+ sdkUrl = buildCCRv2SdkUrl(config.apiBaseUrl, sessionId)
+ // Retry once on transient failure (network blip, 500) before
+ // permanently giving up and killing the session.
+ for (let attempt = 1; attempt <= 2; attempt++) {
+ try {
+ workerEpoch = await registerWorker(
+ sdkUrl,
+ secret.session_ingress_token,
+ )
+ useCcrV2 = true
+ logForDebugging(
+ `[bridge:session] CCR v2: registered worker sessionId=${sessionId} epoch=${workerEpoch} attempt=${attempt}`,
+ )
+ break
+ } catch (err) {
+ const errMsg = errorMessage(err)
+ if (attempt < 2) {
+ logForDebugging(
+ `[bridge:session] CCR v2: registerWorker attempt ${attempt} failed, retrying: ${errMsg}`,
+ )
+ await sleep(2_000, loopSignal)
+ if (loopSignal.aborted) break
+ continue
+ }
+ logger.logError(
+ `CCR v2 worker registration failed for session ${sessionId}: ${errMsg}`,
+ )
+ logError(new Error(`registerWorker failed: ${errMsg}`))
+ completedWorkIds.add(work.id)
+ trackCleanup(
+ stopWorkWithRetry(
+ api,
+ environmentId,
+ work.id,
+ logger,
+ backoffConfig.stopWorkBaseDelayMs,
+ ),
+ )
+ }
+ }
+ if (!useCcrV2) break
+ } else {
+ sdkUrl = buildSdkUrl(config.sessionIngressUrl, sessionId)
+ }
+
+ // In worktree mode, on-demand sessions get an isolated git worktree
+ // so concurrent sessions don't interfere with each other's file
+ // changes. The pre-created initial session (if any) runs in
+ // config.dir so the user's first session lands in the directory they
+ // invoked `rc` from โ matching the old single-session UX.
+ // In same-dir and single-session modes, all sessions share config.dir.
+ // Capture spawnMode before the await below โ the `w` key handler
+ // mutates config.spawnMode directly, and createAgentWorktree can
+ // take 1-2s, so reading config.spawnMode after the await can
+ // produce contradictory analytics (spawn_mode:'same-dir', in_worktree:true).
+ const spawnModeAtDecision = config.spawnMode
+ let sessionDir = config.dir
+ let worktreeCreateMs = 0
+ if (
+ spawnModeAtDecision === 'worktree' &&
+ (initialSessionId === undefined ||
+ !sameSessionId(sessionId, initialSessionId))
+ ) {
+ const wtStart = Date.now()
+ try {
+ const wt = await createAgentWorktree(
+ `bridge-${safeFilenameId(sessionId)}`,
+ )
+ worktreeCreateMs = Date.now() - wtStart
+ sessionWorktrees.set(sessionId, {
+ worktreePath: wt.worktreePath,
+ worktreeBranch: wt.worktreeBranch,
+ gitRoot: wt.gitRoot,
+ hookBased: wt.hookBased,
+ })
+ sessionDir = wt.worktreePath
+ logForDebugging(
+ `[bridge:session] Created worktree for sessionId=${sessionId} at ${wt.worktreePath}`,
+ )
+ } catch (err) {
+ const errMsg = errorMessage(err)
+ logger.logError(
+ `Failed to create worktree for session ${sessionId}: ${errMsg}`,
+ )
+ logError(new Error(`Worktree creation failed: ${errMsg}`))
+ completedWorkIds.add(work.id)
+ trackCleanup(
+ stopWorkWithRetry(
+ api,
+ environmentId,
+ work.id,
+ logger,
+ backoffConfig.stopWorkBaseDelayMs,
+ ),
+ )
+ break
+ }
+ }
+
+ logForDebugging(
+ `[bridge:session] Spawning sessionId=${sessionId} sdkUrl=${sdkUrl}`,
+ )
+
+ // compat-surface session_* form for logger/Sessions-API calls.
+ // Work poll returns cse_* under v2 compat; convert before spawn so
+ // the onFirstUserMessage callback can close over it.
+ const compatSessionId = toCompatSessionId(sessionId)
+
+ const spawnResult = safeSpawn(
+ spawner,
+ {
+ sessionId,
+ sdkUrl,
+ accessToken: secret.session_ingress_token,
+ useCcrV2,
+ workerEpoch,
+ onFirstUserMessage: text => {
+ // Server-set titles (--name, web rename) win. fetchSessionTitle
+ // runs concurrently; if it already populated titledSessions,
+ // skip. If it hasn't resolved yet, the derived title sticks โ
+ // acceptable since the server had no title at spawn time.
+ if (titledSessions.has(compatSessionId)) return
+ titledSessions.add(compatSessionId)
+ const title = deriveSessionTitle(text)
+ logger.setSessionTitle(compatSessionId, title)
+ logForDebugging(
+ `[bridge:title] derived title for ${compatSessionId}: ${title}`,
+ )
+ void import('./createSession.js')
+ .then(({ updateBridgeSessionTitle }) =>
+ updateBridgeSessionTitle(compatSessionId, title, {
+ baseUrl: config.apiBaseUrl,
+ }),
+ )
+ .catch(err =>
+ logForDebugging(
+ `[bridge:title] failed to update title for ${compatSessionId}: ${err}`,
+ { level: 'error' },
+ ),
+ )
+ },
+ },
+ sessionDir,
+ )
+ if (typeof spawnResult === 'string') {
+ logger.logError(
+ `Failed to spawn session ${sessionId}: ${spawnResult}`,
+ )
+ // Clean up worktree if one was created for this session
+ const wt = sessionWorktrees.get(sessionId)
+ if (wt) {
+ sessionWorktrees.delete(sessionId)
+ trackCleanup(
+ removeAgentWorktree(
+ wt.worktreePath,
+ wt.worktreeBranch,
+ wt.gitRoot,
+ wt.hookBased,
+ ).catch((err: unknown) =>
+ logger.logVerbose(
+ `Failed to remove worktree ${wt.worktreePath}: ${errorMessage(err)}`,
+ ),
+ ),
+ )
+ }
+ completedWorkIds.add(work.id)
+ trackCleanup(
+ stopWorkWithRetry(
+ api,
+ environmentId,
+ work.id,
+ logger,
+ backoffConfig.stopWorkBaseDelayMs,
+ ),
+ )
+ break
+ }
+ const handle = spawnResult
+
+ const spawnDurationMs = Date.now() - spawnStartTime
+ logEvent('tengu_bridge_session_started', {
+ active_sessions: activeSessions.size,
+ spawn_mode:
+ spawnModeAtDecision as AnalyticsMetadata_I_VERIFIED_THIS_IS_NOT_CODE_OR_FILEPATHS,
+ in_worktree: sessionWorktrees.has(sessionId),
+ spawn_duration_ms: spawnDurationMs,
+ worktree_create_ms: worktreeCreateMs,
+ inProtectedNamespace: isInProtectedNamespace(),
+ })
+ logForDiagnosticsNoPII('info', 'bridge_session_started', {
+ spawn_mode: spawnModeAtDecision,
+ in_worktree: sessionWorktrees.has(sessionId),
+ spawn_duration_ms: spawnDurationMs,
+ worktree_create_ms: worktreeCreateMs,
+ })
+
+ activeSessions.set(sessionId, handle)
+ sessionWorkIds.set(sessionId, work.id)
+ sessionIngressTokens.set(sessionId, secret.session_ingress_token)
+ sessionCompatIds.set(sessionId, compatSessionId)
+
+ const startTime = Date.now()
+ sessionStartTimes.set(sessionId, startTime)
+
+ // Use a generic prompt description since we no longer get startup_context
+ logger.logSessionStart(sessionId, `Session ${sessionId}`)
+
+ // Compute the actual debug file path (mirrors sessionRunner.ts logic)
+ const safeId = safeFilenameId(sessionId)
+ let sessionDebugFile: string | undefined
+ if (config.debugFile) {
+ const ext = config.debugFile.lastIndexOf('.')
+ if (ext > 0) {
+ sessionDebugFile = `${config.debugFile.slice(0, ext)}-${safeId}${config.debugFile.slice(ext)}`
+ } else {
+ sessionDebugFile = `${config.debugFile}-${safeId}`
+ }
+ } else if (config.verbose || process.env.USER_TYPE === 'ant') {
+ sessionDebugFile = join(
+ tmpdir(),
+ 'claude',
+ `bridge-session-${safeId}.log`,
+ )
+ }
+
+ if (sessionDebugFile) {
+ logger.logVerbose(`Debug log: ${sessionDebugFile}`)
+ }
+
+ // Register in the sessions Map before starting status updates so the
+ // first render tick shows the correct count and bullet list in sync.
+ logger.addSession(
+ compatSessionId,
+ getRemoteSessionUrl(compatSessionId, config.sessionIngressUrl),
+ )
+
+ // Start live status updates and transition to "Attached" state.
+ startStatusUpdates()
+ logger.setAttached(compatSessionId)
+
+ // One-shot title fetch. If the session already has a title (set via
+ // --name, web rename, or /remote-control), display it and mark as
+ // titled so the first-user-message fallback doesn't overwrite it.
+ // Otherwise onFirstUserMessage derives one from the first prompt.
+ void fetchSessionTitle(compatSessionId, config.apiBaseUrl)
+ .then(title => {
+ if (title && activeSessions.has(sessionId)) {
+ titledSessions.add(compatSessionId)
+ logger.setSessionTitle(compatSessionId, title)
+ logForDebugging(
+ `[bridge:title] server title for ${compatSessionId}: ${title}`,
+ )
+ }
+ })
+ .catch(err =>
+ logForDebugging(
+ `[bridge:title] failed to fetch title for ${compatSessionId}: ${err}`,
+ { level: 'error' },
+ ),
+ )
+
+ // Start per-session timeout watchdog
+ const timeoutMs =
+ config.sessionTimeoutMs ?? DEFAULT_SESSION_TIMEOUT_MS
+ if (timeoutMs > 0) {
+ const timer = setTimeout(
+ onSessionTimeout,
+ timeoutMs,
+ sessionId,
+ timeoutMs,
+ logger,
+ timedOutSessions,
+ handle,
+ )
+ sessionTimers.set(sessionId, timer)
+ }
+
+ // Schedule proactive token refresh before the JWT expires.
+ // onRefresh branches on v2Sessions: v1 delivers OAuth to the
+ // child, v2 triggers server re-dispatch via reconnectSession.
+ if (useCcrV2) {
+ v2Sessions.add(sessionId)
+ }
+ tokenRefresh?.schedule(sessionId, secret.session_ingress_token)
+
+ void handle.done.then(onSessionDone(sessionId, startTime, handle))
+ break
+ }
+ default:
+ await ackWork()
+ // Gracefully ignore unknown work types. The backend may send new
+ // types before the bridge client is updated.
+ logForDebugging(
+ `[bridge:work] Unknown work type: ${workType}, skipping`,
+ )
+ break
+ }
+
+ // When at capacity, throttle the loop. The switch above still runs so
+ // existing-session token refreshes are processed, but we sleep here
+ // to avoid busy-looping. Include the capacity wake signal so the
+ // sleep is interrupted immediately when a session completes.
+ if (atCapacityBeforeSwitch) {
+ const cap = capacityWake.signal()
+ if (pollConfig.non_exclusive_heartbeat_interval_ms > 0) {
+ await heartbeatActiveWorkItems()
+ await sleep(
+ pollConfig.non_exclusive_heartbeat_interval_ms,
+ cap.signal,
+ )
+ } else if (pollConfig.multisession_poll_interval_ms_at_capacity > 0) {
+ await sleep(
+ pollConfig.multisession_poll_interval_ms_at_capacity,
+ cap.signal,
+ )
+ }
+ cap.cleanup()
+ }
+ } catch (err) {
+ if (loopSignal.aborted) {
+ break
+ }
+
+ // Fatal errors (401/403) โ no point retrying, auth won't fix itself
+ if (err instanceof BridgeFatalError) {
+ fatalExit = true
+ // Server-enforced expiry gets a clean status message, not an error
+ if (isExpiredErrorType(err.errorType)) {
+ logger.logStatus(err.message)
+ } else if (isSuppressible403(err)) {
+ // Cosmetic 403 errors (e.g., external_poll_sessions scope,
+ // environments:manage permission) โ don't show to user
+ logForDebugging(`[bridge:work] Suppressed 403 error: ${err.message}`)
+ } else {
+ logger.logError(err.message)
+ logError(err)
+ }
+ logEvent('tengu_bridge_fatal_error', {
+ status: err.status,
+ error_type:
+ err.errorType as AnalyticsMetadata_I_VERIFIED_THIS_IS_NOT_CODE_OR_FILEPATHS,
+ })
+ logForDiagnosticsNoPII(
+ isExpiredErrorType(err.errorType) ? 'info' : 'error',
+ 'bridge_fatal_error',
+ { status: err.status, error_type: err.errorType },
+ )
+ break
+ }
+
+ const errMsg = describeAxiosError(err)
+
+ if (isConnectionError(err) || isServerError(err)) {
+ const now = Date.now()
+
+ // Detect system sleep/wake: if the gap since the last poll error
+ // greatly exceeds the expected backoff, the machine likely slept.
+ // Reset error tracking so the bridge retries with a fresh budget.
+ if (
+ lastPollErrorTime !== null &&
+ now - lastPollErrorTime > pollSleepDetectionThresholdMs(backoffConfig)
+ ) {
+ logForDebugging(
+ `[bridge:work] Detected system sleep (${Math.round((now - lastPollErrorTime) / 1000)}s gap), resetting error budget`,
+ )
+ logForDiagnosticsNoPII('info', 'bridge_poll_sleep_detected', {
+ gapMs: now - lastPollErrorTime,
+ })
+ connErrorStart = null
+ connBackoff = 0
+ generalErrorStart = null
+ generalBackoff = 0
+ }
+ lastPollErrorTime = now
+
+ if (!connErrorStart) {
+ connErrorStart = now
+ }
+ const elapsed = now - connErrorStart
+ if (elapsed >= backoffConfig.connGiveUpMs) {
+ logger.logError(
+ `Server unreachable for ${Math.round(elapsed / 60_000)} minutes, giving up.`,
+ )
+ logEvent('tengu_bridge_poll_give_up', {
+ error_type:
+ 'connection' as AnalyticsMetadata_I_VERIFIED_THIS_IS_NOT_CODE_OR_FILEPATHS,
+ elapsed_ms: elapsed,
+ })
+ logForDiagnosticsNoPII('error', 'bridge_poll_give_up', {
+ error_type: 'connection',
+ elapsed_ms: elapsed,
+ })
+ fatalExit = true
+ break
+ }
+
+ // Reset the other track when switching error types
+ generalErrorStart = null
+ generalBackoff = 0
+
+ connBackoff = connBackoff
+ ? Math.min(connBackoff * 2, backoffConfig.connCapMs)
+ : backoffConfig.connInitialMs
+ const delay = addJitter(connBackoff)
+ logger.logVerbose(
+ `Connection error, retrying in ${formatDelay(delay)} (${Math.round(elapsed / 1000)}s elapsed): ${errMsg}`,
+ )
+ logger.updateReconnectingStatus(
+ formatDelay(delay),
+ formatDuration(elapsed),
+ )
+ // The poll_due heartbeat-loop exit leaves a healthy lease exposed to
+ // this backoff path. Heartbeat before each sleep so /poll outages
+ // (the VerifyEnvironmentSecretAuth DB path heartbeat was introduced
+ // to avoid) don't kill the 300s lease TTL. No-op when activeSessions
+ // is empty or heartbeat is disabled.
+ if (getPollIntervalConfig().non_exclusive_heartbeat_interval_ms > 0) {
+ await heartbeatActiveWorkItems()
+ }
+ await sleep(delay, loopSignal)
+ } else {
+ const now = Date.now()
+
+ // Sleep detection for general errors (same logic as connection errors)
+ if (
+ lastPollErrorTime !== null &&
+ now - lastPollErrorTime > pollSleepDetectionThresholdMs(backoffConfig)
+ ) {
+ logForDebugging(
+ `[bridge:work] Detected system sleep (${Math.round((now - lastPollErrorTime) / 1000)}s gap), resetting error budget`,
+ )
+ logForDiagnosticsNoPII('info', 'bridge_poll_sleep_detected', {
+ gapMs: now - lastPollErrorTime,
+ })
+ connErrorStart = null
+ connBackoff = 0
+ generalErrorStart = null
+ generalBackoff = 0
+ }
+ lastPollErrorTime = now
+
+ if (!generalErrorStart) {
+ generalErrorStart = now
+ }
+ const elapsed = now - generalErrorStart
+ if (elapsed >= backoffConfig.generalGiveUpMs) {
+ logger.logError(
+ `Persistent errors for ${Math.round(elapsed / 60_000)} minutes, giving up.`,
+ )
+ logEvent('tengu_bridge_poll_give_up', {
+ error_type:
+ 'general' as AnalyticsMetadata_I_VERIFIED_THIS_IS_NOT_CODE_OR_FILEPATHS,
+ elapsed_ms: elapsed,
+ })
+ logForDiagnosticsNoPII('error', 'bridge_poll_give_up', {
+ error_type: 'general',
+ elapsed_ms: elapsed,
+ })
+ fatalExit = true
+ break
+ }
+
+ // Reset the other track when switching error types
+ connErrorStart = null
+ connBackoff = 0
+
+ generalBackoff = generalBackoff
+ ? Math.min(generalBackoff * 2, backoffConfig.generalCapMs)
+ : backoffConfig.generalInitialMs
+ const delay = addJitter(generalBackoff)
+ logger.logVerbose(
+ `Poll failed, retrying in ${formatDelay(delay)} (${Math.round(elapsed / 1000)}s elapsed): ${errMsg}`,
+ )
+ logger.updateReconnectingStatus(
+ formatDelay(delay),
+ formatDuration(elapsed),
+ )
+ if (getPollIntervalConfig().non_exclusive_heartbeat_interval_ms > 0) {
+ await heartbeatActiveWorkItems()
+ }
+ await sleep(delay, loopSignal)
+ }
+ }
+ }
+
+ // Clean up
+ stopStatusUpdates()
+ logger.clearStatus()
+
+ const loopDurationMs = Date.now() - loopStartTime
+ logEvent('tengu_bridge_shutdown', {
+ active_sessions: activeSessions.size,
+ loop_duration_ms: loopDurationMs,
+ })
+ logForDiagnosticsNoPII('info', 'bridge_shutdown', {
+ active_sessions: activeSessions.size,
+ loop_duration_ms: loopDurationMs,
+ })
+
+ // Graceful shutdown: kill active sessions, report them as interrupted,
+ // archive sessions, then deregister the environment so the web UI shows
+ // the bridge as offline.
+
+ // Collect all session IDs to archive on exit. This includes:
+ // 1. Active sessions (snapshot before killing โ onSessionDone clears maps)
+ // 2. The initial auto-created session (may never have had work dispatched)
+ // api.archiveSession is idempotent (409 if already archived), so
+ // double-archiving is safe.
+ const sessionsToArchive = new Set(activeSessions.keys())
+ if (initialSessionId) {
+ sessionsToArchive.add(initialSessionId)
+ }
+ // Snapshot before killing โ onSessionDone clears sessionCompatIds.
+ const compatIdSnapshot = new Map(sessionCompatIds)
+
+ if (activeSessions.size > 0) {
+ logForDebugging(
+ `[bridge:shutdown] Shutting down ${activeSessions.size} active session(s)`,
+ )
+ logger.logStatus(
+ `Shutting down ${activeSessions.size} active session(s)\u2026`,
+ )
+
+ // Snapshot work IDs before killing โ onSessionDone clears the maps when
+ // each child exits, so we need a copy for the stopWork calls below.
+ const shutdownWorkIds = new Map(sessionWorkIds)
+
+ for (const [sessionId, handle] of activeSessions.entries()) {
+ logForDebugging(
+ `[bridge:shutdown] Sending SIGTERM to sessionId=${sessionId}`,
+ )
+ handle.kill()
+ }
+
+ const timeout = new AbortController()
+ await Promise.race([
+ Promise.allSettled([...activeSessions.values()].map(h => h.done)),
+ sleep(backoffConfig.shutdownGraceMs ?? 30_000, timeout.signal),
+ ])
+ timeout.abort()
+
+ // SIGKILL any processes that didn't respond to SIGTERM within the grace window
+ for (const [sid, handle] of activeSessions.entries()) {
+ logForDebugging(`[bridge:shutdown] Force-killing stuck sessionId=${sid}`)
+ handle.forceKill()
+ }
+
+ // Clear any remaining session timeout and refresh timers
+ for (const timer of sessionTimers.values()) {
+ clearTimeout(timer)
+ }
+ sessionTimers.clear()
+ tokenRefresh?.cancelAll()
+
+ // Clean up any remaining worktrees from active sessions.
+ // Snapshot and clear the map first so onSessionDone (which may fire
+ // during the await below when handle.done resolves) won't try to
+ // remove the same worktrees again.
+ if (sessionWorktrees.size > 0) {
+ const remainingWorktrees = [...sessionWorktrees.values()]
+ sessionWorktrees.clear()
+ logForDebugging(
+ `[bridge:shutdown] Cleaning up ${remainingWorktrees.length} worktree(s)`,
+ )
+ await Promise.allSettled(
+ remainingWorktrees.map(wt =>
+ removeAgentWorktree(
+ wt.worktreePath,
+ wt.worktreeBranch,
+ wt.gitRoot,
+ wt.hookBased,
+ ),
+ ),
+ )
+ }
+
+ // Stop all active work items so the server knows they're done
+ await Promise.allSettled(
+ [...shutdownWorkIds.entries()].map(([sessionId, workId]) => {
+ return api
+ .stopWork(environmentId, workId, true)
+ .catch(err =>
+ logger.logVerbose(
+ `Failed to stop work ${workId} for session ${sessionId}: ${errorMessage(err)}`,
+ ),
+ )
+ }),
+ )
+ }
+
+ // Ensure all in-flight cleanup (stopWork, worktree removal) from
+ // onSessionDone completes before deregistering โ otherwise
+ // process.exit() can kill them mid-flight.
+ if (pendingCleanups.size > 0) {
+ await Promise.allSettled([...pendingCleanups])
+ }
+
+ // In single-session mode with a known session, leave the session and
+ // environment alive so `claude remote-control --session-id=` can resume.
+ // The backend GCs stale environments via a 4h TTL (BRIDGE_LAST_POLL_TTL).
+ // Archiving the session or deregistering the environment would make the
+ // printed resume command a lie โ deregister deletes Firestore + Redis stream.
+ // Skip when the loop exited fatally (env expired, auth failed, give-up) โ
+ // resume is impossible in those cases and the message would contradict the
+ // error already printed.
+ // feature('KAIROS') gate: --session-id is ant-only; without the gate,
+ // revert to the pre-PR behavior (archive + deregister on every shutdown).
+ if (
+ feature('KAIROS') &&
+ config.spawnMode === 'single-session' &&
+ initialSessionId &&
+ !fatalExit
+ ) {
+ logger.logStatus(
+ `Resume this session by running \`claude remote-control --continue\``,
+ )
+ logForDebugging(
+ `[bridge:shutdown] Skipping archive+deregister to allow resume of session ${initialSessionId}`,
+ )
+ return
+ }
+
+ // Archive all known sessions so they don't linger as idle/running on the
+ // server after the bridge goes offline.
+ if (sessionsToArchive.size > 0) {
+ logForDebugging(
+ `[bridge:shutdown] Archiving ${sessionsToArchive.size} session(s)`,
+ )
+ await Promise.allSettled(
+ [...sessionsToArchive].map(sessionId =>
+ api
+ .archiveSession(
+ compatIdSnapshot.get(sessionId) ?? toCompatSessionId(sessionId),
+ )
+ .catch(err =>
+ logger.logVerbose(
+ `Failed to archive session ${sessionId}: ${errorMessage(err)}`,
+ ),
+ ),
+ ),
+ )
+ }
+
+ // Deregister the environment so the web UI shows the bridge as offline
+ // and the Redis stream is cleaned up.
+ try {
+ await api.deregisterEnvironment(environmentId)
+ logForDebugging(
+ `[bridge:shutdown] Environment deregistered, bridge offline`,
+ )
+ logger.logVerbose('Environment deregistered.')
+ } catch (err) {
+ logger.logVerbose(`Failed to deregister environment: ${errorMessage(err)}`)
+ }
+
+ // Clear the crash-recovery pointer โ the env is gone, pointer would be
+ // stale. The early return above (resumable SIGINT shutdown) skips this,
+ // leaving the pointer as a backup for the printed --session-id hint.
+ const { clearBridgePointer } = await import('./bridgePointer.js')
+ await clearBridgePointer(config.dir)
+
+ logger.logVerbose('Environment offline.')
+}
+
+const CONNECTION_ERROR_CODES = new Set([
+ 'ECONNREFUSED',
+ 'ECONNRESET',
+ 'ETIMEDOUT',
+ 'ENETUNREACH',
+ 'EHOSTUNREACH',
+])
+
+export function isConnectionError(err: unknown): boolean {
+ if (
+ err &&
+ typeof err === 'object' &&
+ 'code' in err &&
+ typeof err.code === 'string' &&
+ CONNECTION_ERROR_CODES.has(err.code)
+ ) {
+ return true
+ }
+ return false
+}
+
+/** Detect HTTP 5xx errors from axios (code: 'ERR_BAD_RESPONSE'). */
+export function isServerError(err: unknown): boolean {
+ return (
+ !!err &&
+ typeof err === 'object' &&
+ 'code' in err &&
+ typeof err.code === 'string' &&
+ err.code === 'ERR_BAD_RESPONSE'
+ )
+}
+
+/** Add ยฑ25% jitter to a delay value. */
+function addJitter(ms: number): number {
+ return Math.max(0, ms + ms * 0.25 * (2 * Math.random() - 1))
+}
+
+function formatDelay(ms: number): string {
+ return ms >= 1000 ? `${(ms / 1000).toFixed(1)}s` : `${Math.round(ms)}ms`
+}
+
+/**
+ * Retry stopWork with exponential backoff (3 attempts, 1s/2s/4s).
+ * Ensures the server learns the work item ended, preventing server-side zombies.
+ */
+async function stopWorkWithRetry(
+ api: BridgeApiClient,
+ environmentId: string,
+ workId: string,
+ logger: BridgeLogger,
+ baseDelayMs = 1000,
+): Promise {
+ const MAX_ATTEMPTS = 3
+
+ for (let attempt = 1; attempt <= MAX_ATTEMPTS; attempt++) {
+ try {
+ await api.stopWork(environmentId, workId, false)
+ logForDebugging(
+ `[bridge:work] stopWork succeeded for workId=${workId} on attempt ${attempt}/${MAX_ATTEMPTS}`,
+ )
+ return
+ } catch (err) {
+ // Auth/permission errors won't be fixed by retrying
+ if (err instanceof BridgeFatalError) {
+ if (isSuppressible403(err)) {
+ logForDebugging(
+ `[bridge:work] Suppressed stopWork 403 for ${workId}: ${err.message}`,
+ )
+ } else {
+ logger.logError(`Failed to stop work ${workId}: ${err.message}`)
+ }
+ logForDiagnosticsNoPII('error', 'bridge_stop_work_failed', {
+ attempts: attempt,
+ fatal: true,
+ })
+ return
+ }
+ const errMsg = errorMessage(err)
+ if (attempt < MAX_ATTEMPTS) {
+ const delay = addJitter(baseDelayMs * Math.pow(2, attempt - 1))
+ logger.logVerbose(
+ `Failed to stop work ${workId} (attempt ${attempt}/${MAX_ATTEMPTS}), retrying in ${formatDelay(delay)}: ${errMsg}`,
+ )
+ await sleep(delay)
+ } else {
+ logger.logError(
+ `Failed to stop work ${workId} after ${MAX_ATTEMPTS} attempts: ${errMsg}`,
+ )
+ logForDiagnosticsNoPII('error', 'bridge_stop_work_failed', {
+ attempts: MAX_ATTEMPTS,
+ })
+ }
+ }
+ }
+}
+
+function onSessionTimeout(
+ sessionId: string,
+ timeoutMs: number,
+ logger: BridgeLogger,
+ timedOutSessions: Set,
+ handle: SessionHandle,
+): void {
+ logForDebugging(
+ `[bridge:session] sessionId=${sessionId} timed out after ${formatDuration(timeoutMs)}`,
+ )
+ logEvent('tengu_bridge_session_timeout', {
+ timeout_ms: timeoutMs,
+ })
+ logger.logSessionFailed(
+ sessionId,
+ `Session timed out after ${formatDuration(timeoutMs)}`,
+ )
+ timedOutSessions.add(sessionId)
+ handle.kill()
+}
+
+export type ParsedArgs = {
+ verbose: boolean
+ sandbox: boolean
+ debugFile?: string
+ sessionTimeoutMs?: number
+ permissionMode?: string
+ name?: string
+ /** Value passed to --spawn (if any); undefined if no --spawn flag was given. */
+ spawnMode: SpawnMode | undefined
+ /** Value passed to --capacity (if any); undefined if no --capacity flag was given. */
+ capacity: number | undefined
+ /** --[no-]create-session-in-dir override; undefined = use default (on). */
+ createSessionInDir: boolean | undefined
+ /** Resume an existing session instead of creating a new one. */
+ sessionId?: string
+ /** Resume the last session in this directory (reads bridge-pointer.json). */
+ continueSession: boolean
+ help: boolean
+ error?: string
+}
+
+const SPAWN_FLAG_VALUES = ['session', 'same-dir', 'worktree'] as const
+
+function parseSpawnValue(raw: string | undefined): SpawnMode | string {
+ if (raw === 'session') return 'single-session'
+ if (raw === 'same-dir') return 'same-dir'
+ if (raw === 'worktree') return 'worktree'
+ return `--spawn requires one of: ${SPAWN_FLAG_VALUES.join(', ')} (got: ${raw ?? ''})`
+}
+
+function parseCapacityValue(raw: string | undefined): number | string {
+ const n = raw === undefined ? NaN : parseInt(raw, 10)
+ if (isNaN(n) || n < 1) {
+ return `--capacity requires a positive integer (got: ${raw ?? ''})`
+ }
+ return n
+}
+
+export function parseArgs(args: string[]): ParsedArgs {
+ let verbose = false
+ let sandbox = false
+ let debugFile: string | undefined
+ let sessionTimeoutMs: number | undefined
+ let permissionMode: string | undefined
+ let name: string | undefined
+ let help = false
+ let spawnMode: SpawnMode | undefined
+ let capacity: number | undefined
+ let createSessionInDir: boolean | undefined
+ let sessionId: string | undefined
+ let continueSession = false
+
+ for (let i = 0; i < args.length; i++) {
+ const arg = args[i]!
+ if (arg === '--help' || arg === '-h') {
+ help = true
+ } else if (arg === '--verbose' || arg === '-v') {
+ verbose = true
+ } else if (arg === '--sandbox') {
+ sandbox = true
+ } else if (arg === '--no-sandbox') {
+ sandbox = false
+ } else if (arg === '--debug-file' && i + 1 < args.length) {
+ debugFile = resolve(args[++i]!)
+ } else if (arg.startsWith('--debug-file=')) {
+ debugFile = resolve(arg.slice('--debug-file='.length))
+ } else if (arg === '--session-timeout' && i + 1 < args.length) {
+ sessionTimeoutMs = parseInt(args[++i]!, 10) * 1000
+ } else if (arg.startsWith('--session-timeout=')) {
+ sessionTimeoutMs =
+ parseInt(arg.slice('--session-timeout='.length), 10) * 1000
+ } else if (arg === '--permission-mode' && i + 1 < args.length) {
+ permissionMode = args[++i]!
+ } else if (arg.startsWith('--permission-mode=')) {
+ permissionMode = arg.slice('--permission-mode='.length)
+ } else if (arg === '--name' && i + 1 < args.length) {
+ name = args[++i]!
+ } else if (arg.startsWith('--name=')) {
+ name = arg.slice('--name='.length)
+ } else if (
+ feature('KAIROS') &&
+ arg === '--session-id' &&
+ i + 1 < args.length
+ ) {
+ sessionId = args[++i]!
+ if (!sessionId) {
+ return makeError('--session-id requires a value')
+ }
+ } else if (feature('KAIROS') && arg.startsWith('--session-id=')) {
+ sessionId = arg.slice('--session-id='.length)
+ if (!sessionId) {
+ return makeError('--session-id requires a value')
+ }
+ } else if (feature('KAIROS') && (arg === '--continue' || arg === '-c')) {
+ continueSession = true
+ } else if (arg === '--spawn' || arg.startsWith('--spawn=')) {
+ if (spawnMode !== undefined) {
+ return makeError('--spawn may only be specified once')
+ }
+ const raw = arg.startsWith('--spawn=')
+ ? arg.slice('--spawn='.length)
+ : args[++i]
+ const v = parseSpawnValue(raw)
+ if (v === 'single-session' || v === 'same-dir' || v === 'worktree') {
+ spawnMode = v
+ } else {
+ return makeError(v)
+ }
+ } else if (arg === '--capacity' || arg.startsWith('--capacity=')) {
+ if (capacity !== undefined) {
+ return makeError('--capacity may only be specified once')
+ }
+ const raw = arg.startsWith('--capacity=')
+ ? arg.slice('--capacity='.length)
+ : args[++i]
+ const v = parseCapacityValue(raw)
+ if (typeof v === 'number') capacity = v
+ else return makeError(v)
+ } else if (arg === '--create-session-in-dir') {
+ createSessionInDir = true
+ } else if (arg === '--no-create-session-in-dir') {
+ createSessionInDir = false
+ } else {
+ return makeError(
+ `Unknown argument: ${arg}\nRun 'claude remote-control --help' for usage.`,
+ )
+ }
+ }
+
+ // Note: gate check for --spawn/--capacity/--create-session-in-dir is in bridgeMain
+ // (gate-aware error). Flag cross-validation happens here.
+
+ // --capacity only makes sense for multi-session modes.
+ if (spawnMode === 'single-session' && capacity !== undefined) {
+ return makeError(
+ `--capacity cannot be used with --spawn=session (single-session mode has fixed capacity 1).`,
+ )
+ }
+
+ // --session-id / --continue resume a specific session on its original
+ // environment; incompatible with spawn-related flags (which configure
+ // fresh session creation), and mutually exclusive with each other.
+ if (
+ (sessionId || continueSession) &&
+ (spawnMode !== undefined ||
+ capacity !== undefined ||
+ createSessionInDir !== undefined)
+ ) {
+ return makeError(
+ `--session-id and --continue cannot be used with --spawn, --capacity, or --create-session-in-dir.`,
+ )
+ }
+ if (sessionId && continueSession) {
+ return makeError(`--session-id and --continue cannot be used together.`)
+ }
+
+ return {
+ verbose,
+ sandbox,
+ debugFile,
+ sessionTimeoutMs,
+ permissionMode,
+ name,
+ spawnMode,
+ capacity,
+ createSessionInDir,
+ sessionId,
+ continueSession,
+ help,
+ }
+
+ function makeError(error: string): ParsedArgs {
+ return {
+ verbose,
+ sandbox,
+ debugFile,
+ sessionTimeoutMs,
+ permissionMode,
+ name,
+ spawnMode,
+ capacity,
+ createSessionInDir,
+ sessionId,
+ continueSession,
+ help,
+ error,
+ }
+ }
+}
+
+async function printHelp(): Promise {
+ // Use EXTERNAL_PERMISSION_MODES for help text โ internal modes (bubble)
+ // are ant-only and auto is feature-gated; they're still accepted by validation.
+ const { EXTERNAL_PERMISSION_MODES } = await import('../types/permissions.js')
+ const modes = EXTERNAL_PERMISSION_MODES.join(', ')
+ const showServer = await isMultiSessionSpawnEnabled()
+ const serverOptions = showServer
+ ? ` --spawn Spawn mode: same-dir, worktree, session
+ (default: same-dir)
+ --capacity Max concurrent sessions in worktree or
+ same-dir mode (default: ${SPAWN_SESSIONS_DEFAULT})
+ --[no-]create-session-in-dir Pre-create a session in the current
+ directory; in worktree mode this session
+ stays in cwd while on-demand sessions get
+ isolated worktrees (default: on)
+`
+ : ''
+ const serverDescription = showServer
+ ? `
+ Remote Control runs as a persistent server that accepts multiple concurrent
+ sessions in the current directory. One session is pre-created on start so
+ you have somewhere to type immediately. Use --spawn=worktree to isolate
+ each on-demand session in its own git worktree, or --spawn=session for
+ the classic single-session mode (exits when that session ends). Press 'w'
+ during runtime to toggle between same-dir and worktree.
+`
+ : ''
+ const serverNote = showServer
+ ? ` - Worktree mode requires a git repository or WorktreeCreate/WorktreeRemove hooks
+`
+ : ''
+ const help = `
+Remote Control - Connect your local environment to claude.ai/code
+
+USAGE
+ claude remote-control [options]
+OPTIONS
+ --name Name for the session (shown in claude.ai/code)
+${
+ feature('KAIROS')
+ ? ` -c, --continue Resume the last session in this directory
+ --session-id Resume a specific session by ID (cannot be
+ used with spawn flags or --continue)
+`
+ : ''
+} --permission-mode Permission mode for spawned sessions
+ (${modes})
+ --debug-file Write debug logs to file
+ -v, --verbose Enable verbose output
+ -h, --help Show this help
+${serverOptions}
+DESCRIPTION
+ Remote Control allows you to control sessions on your local device from
+ claude.ai/code (https://claude.ai/code). Run this command in the
+ directory you want to work in, then connect from the Claude app or web.
+${serverDescription}
+NOTES
+ - You must be logged in with a Claude account that has a subscription
+ - Run \`claude\` first in the directory to accept the workspace trust dialog
+${serverNote}`
+ // biome-ignore lint/suspicious/noConsole: intentional help output
+ console.log(help)
+}
+
+const TITLE_MAX_LEN = 80
+
+/** Derive a session title from a user message: first line, truncated. */
+function deriveSessionTitle(text: string): string {
+ // Collapse whitespace โ newlines/tabs would break the single-line status display.
+ const flat = text.replace(/\s+/g, ' ').trim()
+ return truncateToWidth(flat, TITLE_MAX_LEN)
+}
+
+/**
+ * One-shot fetch of a session's title via GET /v1/sessions/{id}.
+ *
+ * Uses `getBridgeSession` from createSession.ts (ccr-byoc headers + org UUID)
+ * rather than the environments-level bridgeApi client, whose headers make the
+ * Sessions API return 404. Returns undefined if the session has no title yet
+ * or the fetch fails โ the caller falls back to deriving a title from the
+ * first user message.
+ */
+async function fetchSessionTitle(
+ compatSessionId: string,
+ baseUrl: string,
+): Promise {
+ const { getBridgeSession } = await import('./createSession.js')
+ const session = await getBridgeSession(compatSessionId, { baseUrl })
+ return session?.title || undefined
+}
+
+export async function bridgeMain(args: string[]): Promise {
+ const parsed = parseArgs(args)
+
+ if (parsed.help) {
+ await printHelp()
+ return
+ }
+ if (parsed.error) {
+ // biome-ignore lint/suspicious/noConsole: intentional error output
+ console.error(`Error: ${parsed.error}`)
+ // eslint-disable-next-line custom-rules/no-process-exit
+ process.exit(1)
+ }
+
+ const {
+ verbose,
+ sandbox,
+ debugFile,
+ sessionTimeoutMs,
+ permissionMode,
+ name,
+ spawnMode: parsedSpawnMode,
+ capacity: parsedCapacity,
+ createSessionInDir: parsedCreateSessionInDir,
+ sessionId: parsedSessionId,
+ continueSession,
+ } = parsed
+ // Mutable so --continue can set it from the pointer file. The #20460
+ // resume flow below then treats it the same as an explicit --session-id.
+ let resumeSessionId = parsedSessionId
+ // When --continue found a pointer, this is the directory it came from
+ // (may be a worktree sibling, not `dir`). On resume-flow deterministic
+ // failure, clear THIS file so --continue doesn't keep hitting the same
+ // dead session. Undefined for explicit --session-id (leaves pointer alone).
+ let resumePointerDir: string | undefined
+
+ const usedMultiSessionFeature =
+ parsedSpawnMode !== undefined ||
+ parsedCapacity !== undefined ||
+ parsedCreateSessionInDir !== undefined
+
+ // Validate permission mode early so the user gets an error before
+ // the bridge starts polling for work.
+ if (permissionMode !== undefined) {
+ const { PERMISSION_MODES } = await import('../types/permissions.js')
+ const valid: readonly string[] = PERMISSION_MODES
+ if (!valid.includes(permissionMode)) {
+ // biome-ignore lint/suspicious/noConsole: intentional error output
+ console.error(
+ `Error: Invalid permission mode '${permissionMode}'. Valid modes: ${valid.join(', ')}`,
+ )
+ // eslint-disable-next-line custom-rules/no-process-exit
+ process.exit(1)
+ }
+ }
+
+ const dir = resolve('.')
+
+ // The bridge fast-path bypasses init.ts, so we must enable config reading
+ // before any code that transitively calls getGlobalConfig()
+ const { enableConfigs, checkHasTrustDialogAccepted } = await import(
+ '../utils/config.js'
+ )
+ enableConfigs()
+
+ // Initialize analytics and error reporting sinks. The bridge bypasses the
+ // setup() init flow, so we call initSinks() directly to attach sinks here.
+ const { initSinks } = await import('../utils/sinks.js')
+ initSinks()
+
+ // Gate-aware validation: --spawn / --capacity / --create-session-in-dir require
+ // the multi-session gate. parseArgs has already validated flag combinations;
+ // here we only check the gate since that requires an async GrowthBook call.
+ // Runs after enableConfigs() (GrowthBook cache reads global config) and after
+ // initSinks() so the denial event can be enqueued.
+ const multiSessionEnabled = await isMultiSessionSpawnEnabled()
+ if (usedMultiSessionFeature && !multiSessionEnabled) {
+ await logEventAsync('tengu_bridge_multi_session_denied', {
+ used_spawn: parsedSpawnMode !== undefined,
+ used_capacity: parsedCapacity !== undefined,
+ used_create_session_in_dir: parsedCreateSessionInDir !== undefined,
+ })
+ // logEventAsync only enqueues โ process.exit() discards buffered events.
+ // Flush explicitly, capped at 500ms to match gracefulShutdown.ts.
+ // (sleep() doesn't unref its timer, but process.exit() follows immediately
+ // so the ref'd timer can't delay shutdown.)
+ await Promise.race([
+ Promise.all([shutdown1PEventLogging(), shutdownDatadog()]),
+ sleep(500, undefined, { unref: true }),
+ ]).catch(() => {})
+ // biome-ignore lint/suspicious/noConsole: intentional error output
+ console.error(
+ 'Error: Multi-session Remote Control is not enabled for your account yet.',
+ )
+ // eslint-disable-next-line custom-rules/no-process-exit
+ process.exit(1)
+ }
+
+ // Set the bootstrap CWD so that trust checks, project config lookups, and
+ // git utilities (getBranch, getRemoteUrl) resolve against the correct path.
+ const { setOriginalCwd, setCwdState } = await import('../bootstrap/state.js')
+ setOriginalCwd(dir)
+ setCwdState(dir)
+
+ // The bridge bypasses main.tsx (which renders the interactive TrustDialog via showSetupScreens),
+ // so we must verify trust was previously established by a normal `claude` session.
+ if (!checkHasTrustDialogAccepted()) {
+ // biome-ignore lint/suspicious/noConsole:: intentional console output
+ console.error(
+ `Error: Workspace not trusted. Please run \`claude\` in ${dir} first to review and accept the workspace trust dialog.`,
+ )
+ // eslint-disable-next-line custom-rules/no-process-exit
+ process.exit(1)
+ }
+
+ // Resolve auth
+ const { clearOAuthTokenCache, checkAndRefreshOAuthTokenIfNeeded } =
+ await import('../utils/auth.js')
+ const { getBridgeAccessToken, getBridgeBaseUrl } = await import(
+ './bridgeConfig.js'
+ )
+
+ const bridgeToken = getBridgeAccessToken()
+ if (!bridgeToken) {
+ // biome-ignore lint/suspicious/noConsole:: intentional console output
+ console.error(BRIDGE_LOGIN_ERROR)
+ // eslint-disable-next-line custom-rules/no-process-exit
+ process.exit(1)
+ }
+
+ // First-time remote dialog โ explain what bridge does and get consent
+ const {
+ getGlobalConfig,
+ saveGlobalConfig,
+ getCurrentProjectConfig,
+ saveCurrentProjectConfig,
+ } = await import('../utils/config.js')
+ if (!getGlobalConfig().remoteDialogSeen) {
+ const readline = await import('readline')
+ const rl = readline.createInterface({
+ input: process.stdin,
+ output: process.stdout,
+ })
+ // biome-ignore lint/suspicious/noConsole:: intentional console output
+ console.log(
+ '\nRemote Control lets you access this CLI session from the web (claude.ai/code)\nor the Claude app, so you can pick up where you left off on any device.\n\nYou can disconnect remote access anytime by running /remote-control again.\n',
+ )
+ const answer = await new Promise(resolve => {
+ rl.question('Enable Remote Control? (y/n) ', resolve)
+ })
+ rl.close()
+ saveGlobalConfig(current => {
+ if (current.remoteDialogSeen) return current
+ return { ...current, remoteDialogSeen: true }
+ })
+ if (answer.toLowerCase() !== 'y' && answer.toLowerCase() !== 'yes') {
+ // eslint-disable-next-line custom-rules/no-process-exit
+ process.exit(0)
+ }
+ }
+
+ // --continue: resolve the most recent session from the crash-recovery
+ // pointer and chain into the #20460 --session-id flow. Worktree-aware:
+ // checks current dir first (fast path, zero exec), then fans out to git
+ // worktree siblings if that misses โ the REPL bridge writes to
+ // getOriginalCwd() which EnterWorktreeTool/activeWorktreeSession can
+ // point at a worktree while the user's shell is at the repo root.
+ // KAIROS-gated at parseArgs โ continueSession is always false in external
+ // builds, so this block tree-shakes.
+ if (feature('KAIROS') && continueSession) {
+ const { readBridgePointerAcrossWorktrees } = await import(
+ './bridgePointer.js'
+ )
+ const found = await readBridgePointerAcrossWorktrees(dir)
+ if (!found) {
+ // biome-ignore lint/suspicious/noConsole: intentional error output
+ console.error(
+ `Error: No recent session found in this directory or its worktrees. Run \`claude remote-control\` to start a new one.`,
+ )
+ // eslint-disable-next-line custom-rules/no-process-exit
+ process.exit(1)
+ }
+ const { pointer, dir: pointerDir } = found
+ const ageMin = Math.round(pointer.ageMs / 60_000)
+ const ageStr = ageMin < 60 ? `${ageMin}m` : `${Math.round(ageMin / 60)}h`
+ const fromWt = pointerDir !== dir ? ` from worktree ${pointerDir}` : ''
+ // biome-ignore lint/suspicious/noConsole: intentional info output
+ console.error(
+ `Resuming session ${pointer.sessionId} (${ageStr} ago)${fromWt}\u2026`,
+ )
+ resumeSessionId = pointer.sessionId
+ // Track where the pointer came from so the #20460 exit(1) paths below
+ // clear the RIGHT file on deterministic failure โ otherwise --continue
+ // would keep hitting the same dead session. May be a worktree sibling.
+ resumePointerDir = pointerDir
+ }
+
+ // In production, baseUrl is the Anthropic API (from OAuth config).
+ // CLAUDE_BRIDGE_BASE_URL overrides this for ant local dev only.
+ const baseUrl = getBridgeBaseUrl()
+
+ // For non-localhost targets, require HTTPS to protect credentials.
+ if (
+ baseUrl.startsWith('http://') &&
+ !baseUrl.includes('localhost') &&
+ !baseUrl.includes('127.0.0.1')
+ ) {
+ // biome-ignore lint/suspicious/noConsole:: intentional console output
+ console.error(
+ 'Error: Remote Control base URL uses HTTP. Only HTTPS or localhost HTTP is allowed.',
+ )
+ // eslint-disable-next-line custom-rules/no-process-exit
+ process.exit(1)
+ }
+
+ // Session ingress URL for WebSocket connections. In production this is the
+ // same as baseUrl (Envoy routes /v1/session_ingress/* to session-ingress).
+ // Locally, session-ingress runs on a different port (9413) than the
+ // contain-provide-api (8211), so CLAUDE_BRIDGE_SESSION_INGRESS_URL must be
+ // set explicitly. Ant-only, matching CLAUDE_BRIDGE_BASE_URL.
+ const sessionIngressUrl =
+ process.env.USER_TYPE === 'ant' &&
+ process.env.CLAUDE_BRIDGE_SESSION_INGRESS_URL
+ ? process.env.CLAUDE_BRIDGE_SESSION_INGRESS_URL
+ : baseUrl
+
+ const { getBranch, getRemoteUrl, findGitRoot } = await import(
+ '../utils/git.js'
+ )
+
+ // Precheck worktree availability for the first-run dialog and the `w`
+ // toggle. Unconditional so we know upfront whether worktree is an option.
+ const { hasWorktreeCreateHook } = await import('../utils/hooks.js')
+ const worktreeAvailable = hasWorktreeCreateHook() || findGitRoot(dir) !== null
+
+ // Load saved per-project spawn-mode preference. Gated by multiSessionEnabled
+ // so a GrowthBook rollback cleanly reverts users to single-session โ
+ // otherwise a saved pref would silently re-enable multi-session behavior
+ // (worktree isolation, 32 max sessions, w toggle) despite the gate being off.
+ // Also guard against a stale worktree pref left over from when this dir WAS
+ // a git repo (or the user copied config) โ clear it on disk so the warning
+ // doesn't repeat on every launch.
+ let savedSpawnMode = multiSessionEnabled
+ ? getCurrentProjectConfig().remoteControlSpawnMode
+ : undefined
+ if (savedSpawnMode === 'worktree' && !worktreeAvailable) {
+ // biome-ignore lint/suspicious/noConsole: intentional warning output
+ console.error(
+ 'Warning: Saved spawn mode is worktree but this directory is not a git repository. Falling back to same-dir.',
+ )
+ savedSpawnMode = undefined
+ saveCurrentProjectConfig(current => {
+ if (current.remoteControlSpawnMode === undefined) return current
+ return { ...current, remoteControlSpawnMode: undefined }
+ })
+ }
+
+ // First-run spawn-mode choice: ask once per project when the choice is
+ // meaningful (gate on, both modes available, no explicit override, not
+ // resuming). Saves to ProjectConfig so subsequent runs skip this.
+ if (
+ multiSessionEnabled &&
+ !savedSpawnMode &&
+ worktreeAvailable &&
+ parsedSpawnMode === undefined &&
+ !resumeSessionId &&
+ process.stdin.isTTY
+ ) {
+ const readline = await import('readline')
+ const rl = readline.createInterface({
+ input: process.stdin,
+ output: process.stdout,
+ })
+ // biome-ignore lint/suspicious/noConsole: intentional dialog output
+ console.log(
+ `\nClaude Remote Control is launching in spawn mode which lets you create new sessions in this project from Claude Code on Web or your Mobile app. Learn more here: https://code.claude.com/docs/en/remote-control\n\n` +
+ `Spawn mode for this project:\n` +
+ ` [1] same-dir \u2014 sessions share the current directory (default)\n` +
+ ` [2] worktree \u2014 each session gets an isolated git worktree\n\n` +
+ `This can be changed later or explicitly set with --spawn=same-dir or --spawn=worktree.\n`,
+ )
+ const answer = await new Promise(resolve => {
+ rl.question('Choose [1/2] (default: 1): ', resolve)
+ })
+ rl.close()
+ const chosen: 'same-dir' | 'worktree' =
+ answer.trim() === '2' ? 'worktree' : 'same-dir'
+ savedSpawnMode = chosen
+ logEvent('tengu_bridge_spawn_mode_chosen', {
+ spawn_mode:
+ chosen as AnalyticsMetadata_I_VERIFIED_THIS_IS_NOT_CODE_OR_FILEPATHS,
+ })
+ saveCurrentProjectConfig(current => {
+ if (current.remoteControlSpawnMode === chosen) return current
+ return { ...current, remoteControlSpawnMode: chosen }
+ })
+ }
+
+ // Determine effective spawn mode.
+ // Precedence: resume > explicit --spawn > saved project pref > gate default
+ // - resuming via --continue / --session-id: always single-session (resume
+ // targets one specific session in its original directory)
+ // - explicit --spawn flag: use that value directly (does not persist)
+ // - saved ProjectConfig.remoteControlSpawnMode: set by first-run dialog or `w`
+ // - default with gate on: same-dir (persistent multi-session, shared cwd)
+ // - default with gate off: single-session (unchanged legacy behavior)
+ // Track how spawn mode was determined, for rollout analytics.
+ type SpawnModeSource = 'resume' | 'flag' | 'saved' | 'gate_default'
+ let spawnModeSource: SpawnModeSource
+ let spawnMode: SpawnMode
+ if (resumeSessionId) {
+ spawnMode = 'single-session'
+ spawnModeSource = 'resume'
+ } else if (parsedSpawnMode !== undefined) {
+ spawnMode = parsedSpawnMode
+ spawnModeSource = 'flag'
+ } else if (savedSpawnMode !== undefined) {
+ spawnMode = savedSpawnMode
+ spawnModeSource = 'saved'
+ } else {
+ spawnMode = multiSessionEnabled ? 'same-dir' : 'single-session'
+ spawnModeSource = 'gate_default'
+ }
+ const maxSessions =
+ spawnMode === 'single-session'
+ ? 1
+ : (parsedCapacity ?? SPAWN_SESSIONS_DEFAULT)
+ // Pre-create an empty session on start so the user has somewhere to type
+ // immediately, running in the current directory (exempted from worktree
+ // creation in the spawn loop). On by default; --no-create-session-in-dir
+ // opts out for a pure on-demand server where every session is isolated.
+ // The effectiveResumeSessionId guard at the creation site handles the
+ // resume case (skip creation when resume succeeded; fall through to
+ // fresh creation on env-mismatch fallback).
+ const preCreateSession = parsedCreateSessionInDir ?? true
+
+ // Without --continue: a leftover pointer means the previous run didn't
+ // shut down cleanly (crash, kill -9, terminal closed). Clear it so the
+ // stale env doesn't linger past its relevance. Runs in all modes
+ // (clearBridgePointer is a no-op when no file exists) โ covers the
+ // gate-transition case where a user crashed in single-session mode then
+ // starts fresh in worktree mode. Only single-session mode writes new
+ // pointers.
+ if (!resumeSessionId) {
+ const { clearBridgePointer } = await import('./bridgePointer.js')
+ await clearBridgePointer(dir)
+ }
+
+ // Worktree mode requires either git or WorktreeCreate/WorktreeRemove hooks.
+ // Only reachable via explicit --spawn=worktree (default is same-dir);
+ // saved worktree pref was already guarded above.
+ if (spawnMode === 'worktree' && !worktreeAvailable) {
+ // biome-ignore lint/suspicious/noConsole: intentional error output
+ console.error(
+ `Error: Worktree mode requires a git repository or WorktreeCreate hooks configured. Use --spawn=session for single-session mode.`,
+ )
+ // eslint-disable-next-line custom-rules/no-process-exit
+ process.exit(1)
+ }
+
+ const branch = await getBranch()
+ const gitRepoUrl = await getRemoteUrl()
+ const machineName = hostname()
+ const bridgeId = randomUUID()
+
+ const { handleOAuth401Error } = await import('../utils/auth.js')
+ const api = createBridgeApiClient({
+ baseUrl,
+ getAccessToken: getBridgeAccessToken,
+ runnerVersion: MACRO.VERSION,
+ onDebug: logForDebugging,
+ onAuth401: handleOAuth401Error,
+ getTrustedDeviceToken,
+ })
+
+ // When resuming a session via --session-id, fetch it to learn its
+ // environment_id and reuse that for registration (idempotent on the
+ // backend). Left undefined otherwise โ the backend rejects
+ // client-generated UUIDs and will allocate a fresh environment.
+ // feature('KAIROS') gate: --session-id is ant-only; parseArgs already
+ // rejects the flag when the gate is off, so resumeSessionId is always
+ // undefined here in external builds โ this guard is for tree-shaking.
+ let reuseEnvironmentId: string | undefined
+ if (feature('KAIROS') && resumeSessionId) {
+ try {
+ validateBridgeId(resumeSessionId, 'sessionId')
+ } catch {
+ // biome-ignore lint/suspicious/noConsole: intentional error output
+ console.error(
+ `Error: Invalid session ID "${resumeSessionId}". Session IDs must not contain unsafe characters.`,
+ )
+ // eslint-disable-next-line custom-rules/no-process-exit
+ process.exit(1)
+ }
+ // Proactively refresh the OAuth token โ getBridgeSession uses raw axios
+ // without the withOAuthRetry 401-refresh logic. An expired-but-present
+ // token would otherwise produce a misleading "not found" error.
+ await checkAndRefreshOAuthTokenIfNeeded()
+ clearOAuthTokenCache()
+ const { getBridgeSession } = await import('./createSession.js')
+ const session = await getBridgeSession(resumeSessionId, {
+ baseUrl,
+ getAccessToken: getBridgeAccessToken,
+ })
+ if (!session) {
+ // Session gone on server โ pointer is stale. Clear it so the user
+ // isn't re-prompted next launch. (Explicit --session-id leaves the
+ // pointer alone โ it's an independent file they may not even have.)
+ // resumePointerDir may be a worktree sibling โ clear THAT file.
+ if (resumePointerDir) {
+ const { clearBridgePointer } = await import('./bridgePointer.js')
+ await clearBridgePointer(resumePointerDir)
+ }
+ // biome-ignore lint/suspicious/noConsole: intentional error output
+ console.error(
+ `Error: Session ${resumeSessionId} not found. It may have been archived or expired, or your login may have lapsed (run \`claude /login\`).`,
+ )
+ // eslint-disable-next-line custom-rules/no-process-exit
+ process.exit(1)
+ }
+ if (!session.environment_id) {
+ if (resumePointerDir) {
+ const { clearBridgePointer } = await import('./bridgePointer.js')
+ await clearBridgePointer(resumePointerDir)
+ }
+ // biome-ignore lint/suspicious/noConsole: intentional error output
+ console.error(
+ `Error: Session ${resumeSessionId} has no environment_id. It may never have been attached to a bridge.`,
+ )
+ // eslint-disable-next-line custom-rules/no-process-exit
+ process.exit(1)
+ }
+ reuseEnvironmentId = session.environment_id
+ logForDebugging(
+ `[bridge:init] Resuming session ${resumeSessionId} on environment ${reuseEnvironmentId}`,
+ )
+ }
+
+ const config: BridgeConfig = {
+ dir,
+ machineName,
+ branch,
+ gitRepoUrl,
+ maxSessions,
+ spawnMode,
+ verbose,
+ sandbox,
+ bridgeId,
+ workerType: 'claude_code',
+ environmentId: randomUUID(),
+ reuseEnvironmentId,
+ apiBaseUrl: baseUrl,
+ sessionIngressUrl,
+ debugFile,
+ sessionTimeoutMs,
+ }
+
+ logForDebugging(
+ `[bridge:init] bridgeId=${bridgeId}${reuseEnvironmentId ? ` reuseEnvironmentId=${reuseEnvironmentId}` : ''} dir=${dir} branch=${branch} gitRepoUrl=${gitRepoUrl} machine=${machineName}`,
+ )
+ logForDebugging(
+ `[bridge:init] apiBaseUrl=${baseUrl} sessionIngressUrl=${sessionIngressUrl}`,
+ )
+ logForDebugging(
+ `[bridge:init] sandbox=${sandbox}${debugFile ? ` debugFile=${debugFile}` : ''}`,
+ )
+
+ // Register the bridge environment before entering the poll loop.
+ let environmentId: string
+ let environmentSecret: string
+ try {
+ const reg = await api.registerBridgeEnvironment(config)
+ environmentId = reg.environment_id
+ environmentSecret = reg.environment_secret
+ } catch (err) {
+ logEvent('tengu_bridge_registration_failed', {
+ status: err instanceof BridgeFatalError ? err.status : undefined,
+ })
+ // Registration failures are fatal โ print a clean message instead of a stack trace.
+ // biome-ignore lint/suspicious/noConsole:: intentional console output
+ console.error(
+ err instanceof BridgeFatalError && err.status === 404
+ ? 'Remote Control environments are not available for your account.'
+ : `Error: ${errorMessage(err)}`,
+ )
+ // eslint-disable-next-line custom-rules/no-process-exit
+ process.exit(1)
+ }
+
+ // Tracks whether the --session-id resume flow completed successfully.
+ // Used below to skip fresh session creation and seed initialSessionId.
+ // Cleared on env mismatch so we gracefully fall back to a new session.
+ let effectiveResumeSessionId: string | undefined
+ if (feature('KAIROS') && resumeSessionId) {
+ if (reuseEnvironmentId && environmentId !== reuseEnvironmentId) {
+ // Backend returned a different environment_id โ the original env
+ // expired or was reaped. Reconnect won't work against the new env
+ // (session is bound to the old one). Log to sentry for visibility
+ // and fall through to fresh session creation on the new env.
+ logError(
+ new Error(
+ `Bridge resume env mismatch: requested ${reuseEnvironmentId}, backend returned ${environmentId}. Falling back to fresh session.`,
+ ),
+ )
+ // biome-ignore lint/suspicious/noConsole: intentional warning output
+ console.warn(
+ `Warning: Could not resume session ${resumeSessionId} โ its environment has expired. Creating a fresh session instead.`,
+ )
+ // Don't deregister โ we're going to use this new environment.
+ // effectiveResumeSessionId stays undefined โ fresh session path below.
+ } else {
+ // Force-stop any stale worker instances for this session and re-queue
+ // it so our poll loop picks it up. Must happen after registration so
+ // the backend knows a live worker exists for the environment.
+ //
+ // The pointer stores a session_* ID but /bridge/reconnect looks
+ // sessions up by their infra tag (cse_*) when ccr_v2_compat_enabled
+ // is on. Try both; the conversion is a no-op if already cse_*.
+ const infraResumeId = toInfraSessionId(resumeSessionId)
+ const reconnectCandidates =
+ infraResumeId === resumeSessionId
+ ? [resumeSessionId]
+ : [resumeSessionId, infraResumeId]
+ let reconnected = false
+ let lastReconnectErr: unknown
+ for (const candidateId of reconnectCandidates) {
+ try {
+ await api.reconnectSession(environmentId, candidateId)
+ logForDebugging(
+ `[bridge:init] Session ${candidateId} re-queued via bridge/reconnect`,
+ )
+ effectiveResumeSessionId = resumeSessionId
+ reconnected = true
+ break
+ } catch (err) {
+ lastReconnectErr = err
+ logForDebugging(
+ `[bridge:init] reconnectSession(${candidateId}) failed: ${errorMessage(err)}`,
+ )
+ }
+ }
+ if (!reconnected) {
+ const err = lastReconnectErr
+
+ // Do NOT deregister on transient reconnect failure โ at this point
+ // environmentId IS the session's own environment. Deregistering
+ // would make retry impossible. The backend's 4h TTL cleans up.
+ const isFatal = err instanceof BridgeFatalError
+ // Clear pointer only on fatal reconnect failure. Transient failures
+ // ("try running the same command again") should keep the pointer so
+ // next launch re-prompts โ that IS the retry mechanism.
+ if (resumePointerDir && isFatal) {
+ const { clearBridgePointer } = await import('./bridgePointer.js')
+ await clearBridgePointer(resumePointerDir)
+ }
+ // biome-ignore lint/suspicious/noConsole: intentional error output
+ console.error(
+ isFatal
+ ? `Error: ${errorMessage(err)}`
+ : `Error: Failed to reconnect session ${resumeSessionId}: ${errorMessage(err)}\nThe session may still be resumable โ try running the same command again.`,
+ )
+ // eslint-disable-next-line custom-rules/no-process-exit
+ process.exit(1)
+ }
+ }
+ }
+
+ logForDebugging(
+ `[bridge:init] Registered, server environmentId=${environmentId}`,
+ )
+ const startupPollConfig = getPollIntervalConfig()
+ logEvent('tengu_bridge_started', {
+ max_sessions: config.maxSessions,
+ has_debug_file: !!config.debugFile,
+ sandbox: config.sandbox,
+ verbose: config.verbose,
+ heartbeat_interval_ms:
+ startupPollConfig.non_exclusive_heartbeat_interval_ms,
+ spawn_mode:
+ config.spawnMode as AnalyticsMetadata_I_VERIFIED_THIS_IS_NOT_CODE_OR_FILEPATHS,
+ spawn_mode_source:
+ spawnModeSource as AnalyticsMetadata_I_VERIFIED_THIS_IS_NOT_CODE_OR_FILEPATHS,
+ multi_session_gate: multiSessionEnabled,
+ pre_create_session: preCreateSession,
+ worktree_available: worktreeAvailable,
+ })
+ logForDiagnosticsNoPII('info', 'bridge_started', {
+ max_sessions: config.maxSessions,
+ sandbox: config.sandbox,
+ spawn_mode: config.spawnMode,
+ })
+
+ const spawner = createSessionSpawner({
+ execPath: process.execPath,
+ scriptArgs: spawnScriptArgs(),
+ env: process.env,
+ verbose,
+ sandbox,
+ debugFile,
+ permissionMode,
+ onDebug: logForDebugging,
+ onActivity: (sessionId, activity) => {
+ logForDebugging(
+ `[bridge:activity] sessionId=${sessionId} ${activity.type} ${activity.summary}`,
+ )
+ },
+ onPermissionRequest: (sessionId, request, _accessToken) => {
+ logForDebugging(
+ `[bridge:perm] sessionId=${sessionId} tool=${request.request.tool_name} request_id=${request.request_id} (not auto-approving)`,
+ )
+ },
+ })
+
+ const logger = createBridgeLogger({ verbose })
+ const { parseGitHubRepository } = await import('../utils/detectRepository.js')
+ const ownerRepo = gitRepoUrl ? parseGitHubRepository(gitRepoUrl) : null
+ // Use the repo name from the parsed owner/repo, or fall back to the dir basename
+ const repoName = ownerRepo ? ownerRepo.split('/').pop()! : basename(dir)
+ logger.setRepoInfo(repoName, branch)
+
+ // `w` toggle is available iff we're in a multi-session mode AND worktree
+ // is a valid option. When unavailable, the mode suffix and hint are hidden.
+ const toggleAvailable = spawnMode !== 'single-session' && worktreeAvailable
+ if (toggleAvailable) {
+ // Safe cast: spawnMode is not single-session (checked above), and the
+ // saved-worktree-in-non-git guard + exit check above ensure worktree
+ // is only reached when available.
+ logger.setSpawnModeDisplay(spawnMode as 'same-dir' | 'worktree')
+ }
+
+ // Listen for keys: space toggles QR code, w toggles spawn mode
+ const onStdinData = (data: Buffer): void => {
+ if (data[0] === 0x03 || data[0] === 0x04) {
+ // Ctrl+C / Ctrl+D โ trigger graceful shutdown
+ process.emit('SIGINT')
+ return
+ }
+ if (data[0] === 0x20 /* space */) {
+ logger.toggleQr()
+ return
+ }
+ if (data[0] === 0x77 /* 'w' */) {
+ if (!toggleAvailable) return
+ const newMode: 'same-dir' | 'worktree' =
+ config.spawnMode === 'same-dir' ? 'worktree' : 'same-dir'
+ config.spawnMode = newMode
+ logEvent('tengu_bridge_spawn_mode_toggled', {
+ spawn_mode:
+ newMode as AnalyticsMetadata_I_VERIFIED_THIS_IS_NOT_CODE_OR_FILEPATHS,
+ })
+ logger.logStatus(
+ newMode === 'worktree'
+ ? 'Spawn mode: worktree (new sessions get isolated git worktrees)'
+ : 'Spawn mode: same-dir (new sessions share the current directory)',
+ )
+ logger.setSpawnModeDisplay(newMode)
+ logger.refreshDisplay()
+ saveCurrentProjectConfig(current => {
+ if (current.remoteControlSpawnMode === newMode) return current
+ return { ...current, remoteControlSpawnMode: newMode }
+ })
+ return
+ }
+ }
+ if (process.stdin.isTTY) {
+ process.stdin.setRawMode(true)
+ process.stdin.resume()
+ process.stdin.on('data', onStdinData)
+ }
+
+ const controller = new AbortController()
+ const onSigint = (): void => {
+ logForDebugging('[bridge:shutdown] SIGINT received, shutting down')
+ controller.abort()
+ }
+ const onSigterm = (): void => {
+ logForDebugging('[bridge:shutdown] SIGTERM received, shutting down')
+ controller.abort()
+ }
+ process.on('SIGINT', onSigint)
+ process.on('SIGTERM', onSigterm)
+
+ // Auto-create an empty session so the user has somewhere to type
+ // immediately (matching /remote-control behavior). Controlled by
+ // preCreateSession: on by default; --no-create-session-in-dir opts out.
+ // When a --session-id resume succeeded, skip creation entirely โ the
+ // session already exists and bridge/reconnect has re-queued it.
+ // When resume was requested but failed on env mismatch, effectiveResumeSessionId
+ // is undefined, so we fall through to fresh session creation (honoring the
+ // "Creating a fresh session instead" warning printed above).
+ let initialSessionId: string | null =
+ feature('KAIROS') && effectiveResumeSessionId
+ ? effectiveResumeSessionId
+ : null
+ if (preCreateSession && !(feature('KAIROS') && effectiveResumeSessionId)) {
+ const { createBridgeSession } = await import('./createSession.js')
+ try {
+ initialSessionId = await createBridgeSession({
+ environmentId,
+ title: name,
+ events: [],
+ gitRepoUrl,
+ branch,
+ signal: controller.signal,
+ baseUrl,
+ getAccessToken: getBridgeAccessToken,
+ permissionMode,
+ })
+ if (initialSessionId) {
+ logForDebugging(
+ `[bridge:init] Created initial session ${initialSessionId}`,
+ )
+ }
+ } catch (err) {
+ logForDebugging(
+ `[bridge:init] Session creation failed (non-fatal): ${errorMessage(err)}`,
+ )
+ }
+ }
+
+ // Crash-recovery pointer: write immediately so kill -9 at any point
+ // after this leaves a recoverable trail. Covers both fresh sessions and
+ // resumed ones (so a second crash after resume is still recoverable).
+ // Cleared when runBridgeLoop falls through to archive+deregister; left in
+ // place on the SIGINT resumable-shutdown return (backup for when the user
+ // closes the terminal before copying the printed --session-id hint).
+ // Refreshed hourly so a 5h+ session that crashes still has a fresh
+ // pointer (staleness checks file mtime, backend TTL is rolling-from-poll).
+ let pointerRefreshTimer: ReturnType | null = null
+ // Single-session only: --continue forces single-session mode on resume,
+ // so a pointer written in multi-session mode would contradict the user's
+ // config when they try to resume. The resumable-shutdown path is also
+ // gated to single-session (line ~1254) so the pointer would be orphaned.
+ if (initialSessionId && spawnMode === 'single-session') {
+ const { writeBridgePointer } = await import('./bridgePointer.js')
+ const pointerPayload = {
+ sessionId: initialSessionId,
+ environmentId,
+ source: 'standalone' as const,
+ }
+ await writeBridgePointer(config.dir, pointerPayload)
+ pointerRefreshTimer = setInterval(
+ writeBridgePointer,
+ 60 * 60 * 1000,
+ config.dir,
+ pointerPayload,
+ )
+ // Don't let the interval keep the process alive on its own.
+ pointerRefreshTimer.unref?.()
+ }
+
+ try {
+ await runBridgeLoop(
+ config,
+ environmentId,
+ environmentSecret,
+ api,
+ spawner,
+ logger,
+ controller.signal,
+ undefined,
+ initialSessionId ?? undefined,
+ async () => {
+ // Clear the memoized OAuth token cache so we re-read from secure
+ // storage, picking up tokens refreshed by child processes.
+ clearOAuthTokenCache()
+ // Proactively refresh the token if it's expired on disk too.
+ await checkAndRefreshOAuthTokenIfNeeded()
+ return getBridgeAccessToken()
+ },
+ )
+ } finally {
+ if (pointerRefreshTimer !== null) {
+ clearInterval(pointerRefreshTimer)
+ }
+ process.off('SIGINT', onSigint)
+ process.off('SIGTERM', onSigterm)
+ process.stdin.off('data', onStdinData)
+ if (process.stdin.isTTY) {
+ process.stdin.setRawMode(false)
+ }
+ process.stdin.pause()
+ }
+
+ // The bridge bypasses init.ts (and its graceful shutdown handler), so we
+ // must exit explicitly.
+ // eslint-disable-next-line custom-rules/no-process-exit
+ process.exit(0)
+}
+
+// โโโ Headless bridge (daemon worker) โโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโ
+
+/**
+ * Thrown by runBridgeHeadless for configuration issues the supervisor should
+ * NOT retry (trust not accepted, worktree unavailable, http-not-https). The
+ * daemon worker catches this and exits with EXIT_CODE_PERMANENT so the
+ * supervisor parks the worker instead of respawning it on backoff.
+ */
+export class BridgeHeadlessPermanentError extends Error {
+ constructor(message: string) {
+ super(message)
+ this.name = 'BridgeHeadlessPermanentError'
+ }
+}
+
+export type HeadlessBridgeOpts = {
+ dir: string
+ name?: string
+ spawnMode: 'same-dir' | 'worktree'
+ capacity: number
+ permissionMode?: string
+ sandbox: boolean
+ sessionTimeoutMs?: number
+ createSessionOnStart: boolean
+ getAccessToken: () => string | undefined
+ onAuth401: (failedToken: string) => Promise
+ log: (s: string) => void
+}
+
+/**
+ * Non-interactive bridge entrypoint for the `remoteControl` daemon worker.
+ *
+ * Linear subset of bridgeMain(): no readline dialogs, no stdin key handlers,
+ * no TUI, no process.exit(). Config comes from the caller (daemon.json), auth
+ * comes via IPC (supervisor's AuthManager), logs go to the worker's stdout
+ * pipe. Throws on fatal errors โ the worker catches and maps permanent vs
+ * transient to the right exit code.
+ *
+ * Resolves cleanly when `signal` aborts and the poll loop tears down.
+ */
+export async function runBridgeHeadless(
+ opts: HeadlessBridgeOpts,
+ signal: AbortSignal,
+): Promise {
+ const { dir, log } = opts
+
+ // Worker inherits the supervisor's CWD. chdir first so git utilities
+ // (getBranch/getRemoteUrl) โ which read from bootstrap CWD state set
+ // below โ resolve against the right repo.
+ process.chdir(dir)
+ const { setOriginalCwd, setCwdState } = await import('../bootstrap/state.js')
+ setOriginalCwd(dir)
+ setCwdState(dir)
+
+ const { enableConfigs, checkHasTrustDialogAccepted } = await import(
+ '../utils/config.js'
+ )
+ enableConfigs()
+ const { initSinks } = await import('../utils/sinks.js')
+ initSinks()
+
+ if (!checkHasTrustDialogAccepted()) {
+ throw new BridgeHeadlessPermanentError(
+ `Workspace not trusted: ${dir}. Run \`claude\` in that directory first to accept the trust dialog.`,
+ )
+ }
+
+ if (!opts.getAccessToken()) {
+ // Transient โ supervisor's AuthManager may pick up a token on next cycle.
+ throw new Error(BRIDGE_LOGIN_ERROR)
+ }
+
+ const { getBridgeBaseUrl } = await import('./bridgeConfig.js')
+ const baseUrl = getBridgeBaseUrl()
+ if (
+ baseUrl.startsWith('http://') &&
+ !baseUrl.includes('localhost') &&
+ !baseUrl.includes('127.0.0.1')
+ ) {
+ throw new BridgeHeadlessPermanentError(
+ 'Remote Control base URL uses HTTP. Only HTTPS or localhost HTTP is allowed.',
+ )
+ }
+ const sessionIngressUrl =
+ process.env.USER_TYPE === 'ant' &&
+ process.env.CLAUDE_BRIDGE_SESSION_INGRESS_URL
+ ? process.env.CLAUDE_BRIDGE_SESSION_INGRESS_URL
+ : baseUrl
+
+ const { getBranch, getRemoteUrl, findGitRoot } = await import(
+ '../utils/git.js'
+ )
+ const { hasWorktreeCreateHook } = await import('../utils/hooks.js')
+
+ if (opts.spawnMode === 'worktree') {
+ const worktreeAvailable =
+ hasWorktreeCreateHook() || findGitRoot(dir) !== null
+ if (!worktreeAvailable) {
+ throw new BridgeHeadlessPermanentError(
+ `Worktree mode requires a git repository or WorktreeCreate hooks. Directory ${dir} has neither.`,
+ )
+ }
+ }
+
+ const branch = await getBranch()
+ const gitRepoUrl = await getRemoteUrl()
+ const machineName = hostname()
+ const bridgeId = randomUUID()
+
+ const config: BridgeConfig = {
+ dir,
+ machineName,
+ branch,
+ gitRepoUrl,
+ maxSessions: opts.capacity,
+ spawnMode: opts.spawnMode,
+ verbose: false,
+ sandbox: opts.sandbox,
+ bridgeId,
+ workerType: 'claude_code',
+ environmentId: randomUUID(),
+ apiBaseUrl: baseUrl,
+ sessionIngressUrl,
+ sessionTimeoutMs: opts.sessionTimeoutMs,
+ }
+
+ const api = createBridgeApiClient({
+ baseUrl,
+ getAccessToken: opts.getAccessToken,
+ runnerVersion: MACRO.VERSION,
+ onDebug: log,
+ onAuth401: opts.onAuth401,
+ getTrustedDeviceToken,
+ })
+
+ let environmentId: string
+ let environmentSecret: string
+ try {
+ const reg = await api.registerBridgeEnvironment(config)
+ environmentId = reg.environment_id
+ environmentSecret = reg.environment_secret
+ } catch (err) {
+ // Transient โ let supervisor backoff-retry.
+ throw new Error(`Bridge registration failed: ${errorMessage(err)}`)
+ }
+
+ const spawner = createSessionSpawner({
+ execPath: process.execPath,
+ scriptArgs: spawnScriptArgs(),
+ env: process.env,
+ verbose: false,
+ sandbox: opts.sandbox,
+ permissionMode: opts.permissionMode,
+ onDebug: log,
+ })
+
+ const logger = createHeadlessBridgeLogger(log)
+ logger.printBanner(config, environmentId)
+
+ let initialSessionId: string | undefined
+ if (opts.createSessionOnStart) {
+ const { createBridgeSession } = await import('./createSession.js')
+ try {
+ const sid = await createBridgeSession({
+ environmentId,
+ title: opts.name,
+ events: [],
+ gitRepoUrl,
+ branch,
+ signal,
+ baseUrl,
+ getAccessToken: opts.getAccessToken,
+ permissionMode: opts.permissionMode,
+ })
+ if (sid) {
+ initialSessionId = sid
+ log(`created initial session ${sid}`)
+ }
+ } catch (err) {
+ log(`session pre-creation failed (non-fatal): ${errorMessage(err)}`)
+ }
+ }
+
+ await runBridgeLoop(
+ config,
+ environmentId,
+ environmentSecret,
+ api,
+ spawner,
+ logger,
+ signal,
+ undefined,
+ initialSessionId,
+ async () => opts.getAccessToken(),
+ )
+}
+
+/** BridgeLogger adapter that routes everything to a single line-log fn. */
+function createHeadlessBridgeLogger(log: (s: string) => void): BridgeLogger {
+ const noop = (): void => {}
+ return {
+ printBanner: (cfg, envId) =>
+ log(
+ `registered environmentId=${envId} dir=${cfg.dir} spawnMode=${cfg.spawnMode} capacity=${cfg.maxSessions}`,
+ ),
+ logSessionStart: (id, _prompt) => log(`session start ${id}`),
+ logSessionComplete: (id, ms) => log(`session complete ${id} (${ms}ms)`),
+ logSessionFailed: (id, err) => log(`session failed ${id}: ${err}`),
+ logStatus: log,
+ logVerbose: log,
+ logError: s => log(`error: ${s}`),
+ logReconnected: ms => log(`reconnected after ${ms}ms`),
+ addSession: (id, _url) => log(`session attached ${id}`),
+ removeSession: id => log(`session detached ${id}`),
+ updateIdleStatus: noop,
+ updateReconnectingStatus: noop,
+ updateSessionStatus: noop,
+ updateSessionActivity: noop,
+ updateSessionCount: noop,
+ updateFailedStatus: noop,
+ setSpawnModeDisplay: noop,
+ setRepoInfo: noop,
+ setDebugLogPath: noop,
+ setAttached: noop,
+ setSessionTitle: noop,
+ clearStatus: noop,
+ toggleQr: noop,
+ refreshDisplay: noop,
+ }
+}
diff --git a/original-source-code/src/bridge/bridgeMessaging.ts b/original-source-code/src/bridge/bridgeMessaging.ts
new file mode 100644
index 0000000..98ece03
--- /dev/null
+++ b/original-source-code/src/bridge/bridgeMessaging.ts
@@ -0,0 +1,461 @@
+/**
+ * Shared transport-layer helpers for bridge message handling.
+ *
+ * Extracted from replBridge.ts so both the env-based core (initBridgeCore)
+ * and the env-less core (initEnvLessBridgeCore) can use the same ingress
+ * parsing, control-request handling, and echo-dedup machinery.
+ *
+ * Everything here is pure โ no closure over bridge-specific state. All
+ * collaborators (transport, sessionId, UUID sets, callbacks) are passed
+ * as params.
+ */
+
+import { randomUUID } from 'crypto'
+import type { SDKMessage } from '../entrypoints/agentSdkTypes.js'
+import type {
+ SDKControlRequest,
+ SDKControlResponse,
+} from '../entrypoints/sdk/controlTypes.js'
+import type { SDKResultSuccess } from '../entrypoints/sdk/coreTypes.js'
+import { logEvent } from '../services/analytics/index.js'
+import { EMPTY_USAGE } from '../services/api/emptyUsage.js'
+import type { Message } from '../types/message.js'
+import { normalizeControlMessageKeys } from '../utils/controlMessageCompat.js'
+import { logForDebugging } from '../utils/debug.js'
+import { stripDisplayTagsAllowEmpty } from '../utils/displayTags.js'
+import { errorMessage } from '../utils/errors.js'
+import type { PermissionMode } from '../utils/permissions/PermissionMode.js'
+import { jsonParse } from '../utils/slowOperations.js'
+import type { ReplBridgeTransport } from './replBridgeTransport.js'
+
+// โโโ Type guards โโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโ
+
+/** Type predicate for parsed WebSocket messages. SDKMessage is a
+ * discriminated union on `type` โ validating the discriminant is
+ * sufficient for the predicate; callers narrow further via the union. */
+export function isSDKMessage(value: unknown): value is SDKMessage {
+ return (
+ value !== null &&
+ typeof value === 'object' &&
+ 'type' in value &&
+ typeof value.type === 'string'
+ )
+}
+
+/** Type predicate for control_response messages from the server. */
+export function isSDKControlResponse(
+ value: unknown,
+): value is SDKControlResponse {
+ return (
+ value !== null &&
+ typeof value === 'object' &&
+ 'type' in value &&
+ value.type === 'control_response' &&
+ 'response' in value
+ )
+}
+
+/** Type predicate for control_request messages from the server. */
+export function isSDKControlRequest(
+ value: unknown,
+): value is SDKControlRequest {
+ return (
+ value !== null &&
+ typeof value === 'object' &&
+ 'type' in value &&
+ value.type === 'control_request' &&
+ 'request_id' in value &&
+ 'request' in value
+ )
+}
+
+/**
+ * True for message types that should be forwarded to the bridge transport.
+ * The server only wants user/assistant turns and slash-command system events;
+ * everything else (tool_result, progress, etc.) is internal REPL chatter.
+ */
+export function isEligibleBridgeMessage(m: Message): boolean {
+ // Virtual messages (REPL inner calls) are display-only โ bridge/SDK
+ // consumers see the REPL tool_use/result which summarizes the work.
+ if ((m.type === 'user' || m.type === 'assistant') && m.isVirtual) {
+ return false
+ }
+ return (
+ m.type === 'user' ||
+ m.type === 'assistant' ||
+ (m.type === 'system' && m.subtype === 'local_command')
+ )
+}
+
+/**
+ * Extract title-worthy text from a Message for onUserMessage. Returns
+ * undefined for messages that shouldn't title the session: non-user, meta
+ * (nudges), tool results, compact summaries, non-human origins (task
+ * notifications, channel messages), or pure display-tag content
+ * (, , etc.).
+ *
+ * Synthetic interrupts ([Request interrupted by user]) are NOT filtered here โ
+ * isSyntheticMessage lives in messages.ts (heavy import, pulls command
+ * registry). The initialMessages path in initReplBridge checks it; the
+ * writeMessages path reaching an interrupt as the *first* message is
+ * implausible (an interrupt implies a prior prompt already flowed through).
+ */
+export function extractTitleText(m: Message): string | undefined {
+ if (m.type !== 'user' || m.isMeta || m.toolUseResult || m.isCompactSummary)
+ return undefined
+ if (m.origin && m.origin.kind !== 'human') return undefined
+ const content = m.message.content
+ let raw: string | undefined
+ if (typeof content === 'string') {
+ raw = content
+ } else {
+ for (const block of content) {
+ if (block.type === 'text') {
+ raw = block.text
+ break
+ }
+ }
+ }
+ if (!raw) return undefined
+ const clean = stripDisplayTagsAllowEmpty(raw)
+ return clean || undefined
+}
+
+// โโโ Ingress routing โโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโ
+
+/**
+ * Parse an ingress WebSocket message and route it to the appropriate handler.
+ * Ignores messages whose UUID is in recentPostedUUIDs (echoes of what we sent)
+ * or in recentInboundUUIDs (re-deliveries we've already forwarded โ e.g.
+ * server replayed history after a transport swap lost the seq-num cursor).
+ */
+export function handleIngressMessage(
+ data: string,
+ recentPostedUUIDs: BoundedUUIDSet,
+ recentInboundUUIDs: BoundedUUIDSet,
+ onInboundMessage: ((msg: SDKMessage) => void | Promise) | undefined,
+ onPermissionResponse?: ((response: SDKControlResponse) => void) | undefined,
+ onControlRequest?: ((request: SDKControlRequest) => void) | undefined,
+): void {
+ try {
+ const parsed: unknown = normalizeControlMessageKeys(jsonParse(data))
+
+ // control_response is not an SDKMessage โ check before the type guard
+ if (isSDKControlResponse(parsed)) {
+ logForDebugging('[bridge:repl] Ingress message type=control_response')
+ onPermissionResponse?.(parsed)
+ return
+ }
+
+ // control_request from the server (initialize, set_model, can_use_tool).
+ // Must respond promptly or the server kills the WS (~10-14s timeout).
+ if (isSDKControlRequest(parsed)) {
+ logForDebugging(
+ `[bridge:repl] Inbound control_request subtype=${parsed.request.subtype}`,
+ )
+ onControlRequest?.(parsed)
+ return
+ }
+
+ if (!isSDKMessage(parsed)) return
+
+ // Check for UUID to detect echoes of our own messages
+ const uuid =
+ 'uuid' in parsed && typeof parsed.uuid === 'string'
+ ? parsed.uuid
+ : undefined
+
+ if (uuid && recentPostedUUIDs.has(uuid)) {
+ logForDebugging(
+ `[bridge:repl] Ignoring echo: type=${parsed.type} uuid=${uuid}`,
+ )
+ return
+ }
+
+ // Defensive dedup: drop inbound prompts we've already forwarded. The
+ // SSE seq-num carryover (lastTransportSequenceNum) is the primary fix
+ // for history-replay; this catches edge cases where that negotiation
+ // fails (server ignores from_sequence_num, transport died before
+ // receiving any frames, etc).
+ if (uuid && recentInboundUUIDs.has(uuid)) {
+ logForDebugging(
+ `[bridge:repl] Ignoring re-delivered inbound: type=${parsed.type} uuid=${uuid}`,
+ )
+ return
+ }
+
+ logForDebugging(
+ `[bridge:repl] Ingress message type=${parsed.type}${uuid ? ` uuid=${uuid}` : ''}`,
+ )
+
+ if (parsed.type === 'user') {
+ if (uuid) recentInboundUUIDs.add(uuid)
+ logEvent('tengu_bridge_message_received', {
+ is_repl: true,
+ })
+ // Fire-and-forget โ handler may be async (attachment resolution).
+ void onInboundMessage?.(parsed)
+ } else {
+ logForDebugging(
+ `[bridge:repl] Ignoring non-user inbound message: type=${parsed.type}`,
+ )
+ }
+ } catch (err) {
+ logForDebugging(
+ `[bridge:repl] Failed to parse ingress message: ${errorMessage(err)}`,
+ )
+ }
+}
+
+// โโโ Server-initiated control requests โโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโ
+
+export type ServerControlRequestHandlers = {
+ transport: ReplBridgeTransport | null
+ sessionId: string
+ /**
+ * When true, all mutable requests (interrupt, set_model, set_permission_mode,
+ * set_max_thinking_tokens) reply with an error instead of false-success.
+ * initialize still replies success โ the server kills the connection otherwise.
+ * Used by the outbound-only bridge mode and the SDK's /bridge subpath so claude.ai sees a
+ * proper error instead of "action succeeded but nothing happened locally".
+ */
+ outboundOnly?: boolean
+ onInterrupt?: () => void
+ onSetModel?: (model: string | undefined) => void
+ onSetMaxThinkingTokens?: (maxTokens: number | null) => void
+ onSetPermissionMode?: (
+ mode: PermissionMode,
+ ) => { ok: true } | { ok: false; error: string }
+}
+
+const OUTBOUND_ONLY_ERROR =
+ 'This session is outbound-only. Enable Remote Control locally to allow inbound control.'
+
+/**
+ * Respond to inbound control_request messages from the server. The server
+ * sends these for session lifecycle events (initialize, set_model) and
+ * for turn-level coordination (interrupt, set_max_thinking_tokens). If we
+ * don't respond, the server hangs and kills the WS after ~10-14s.
+ *
+ * Previously a closure inside initBridgeCore's onWorkReceived; now takes
+ * collaborators as params so both cores can use it.
+ */
+export function handleServerControlRequest(
+ request: SDKControlRequest,
+ handlers: ServerControlRequestHandlers,
+): void {
+ const {
+ transport,
+ sessionId,
+ outboundOnly,
+ onInterrupt,
+ onSetModel,
+ onSetMaxThinkingTokens,
+ onSetPermissionMode,
+ } = handlers
+ if (!transport) {
+ logForDebugging(
+ '[bridge:repl] Cannot respond to control_request: transport not configured',
+ )
+ return
+ }
+
+ let response: SDKControlResponse
+
+ // Outbound-only: reply error for mutable requests so claude.ai doesn't show
+ // false success. initialize must still succeed (server kills the connection
+ // if it doesn't โ see comment above).
+ if (outboundOnly && request.request.subtype !== 'initialize') {
+ response = {
+ type: 'control_response',
+ response: {
+ subtype: 'error',
+ request_id: request.request_id,
+ error: OUTBOUND_ONLY_ERROR,
+ },
+ }
+ const event = { ...response, session_id: sessionId }
+ void transport.write(event)
+ logForDebugging(
+ `[bridge:repl] Rejected ${request.request.subtype} (outbound-only) request_id=${request.request_id}`,
+ )
+ return
+ }
+
+ switch (request.request.subtype) {
+ case 'initialize':
+ // Respond with minimal capabilities โ the REPL handles
+ // commands, models, and account info itself.
+ response = {
+ type: 'control_response',
+ response: {
+ subtype: 'success',
+ request_id: request.request_id,
+ response: {
+ commands: [],
+ output_style: 'normal',
+ available_output_styles: ['normal'],
+ models: [],
+ account: {},
+ pid: process.pid,
+ },
+ },
+ }
+ break
+
+ case 'set_model':
+ onSetModel?.(request.request.model)
+ response = {
+ type: 'control_response',
+ response: {
+ subtype: 'success',
+ request_id: request.request_id,
+ },
+ }
+ break
+
+ case 'set_max_thinking_tokens':
+ onSetMaxThinkingTokens?.(request.request.max_thinking_tokens)
+ response = {
+ type: 'control_response',
+ response: {
+ subtype: 'success',
+ request_id: request.request_id,
+ },
+ }
+ break
+
+ case 'set_permission_mode': {
+ // The callback returns a policy verdict so we can send an error
+ // control_response without importing isAutoModeGateEnabled /
+ // isBypassPermissionsModeDisabled here (bootstrap-isolation). If no
+ // callback is registered (daemon context, which doesn't wire this โ
+ // see daemonBridge.ts), return an error verdict rather than a silent
+ // false-success: the mode is never actually applied in that context,
+ // so success would lie to the client.
+ const verdict = onSetPermissionMode?.(request.request.mode) ?? {
+ ok: false,
+ error:
+ 'set_permission_mode is not supported in this context (onSetPermissionMode callback not registered)',
+ }
+ if (verdict.ok) {
+ response = {
+ type: 'control_response',
+ response: {
+ subtype: 'success',
+ request_id: request.request_id,
+ },
+ }
+ } else {
+ response = {
+ type: 'control_response',
+ response: {
+ subtype: 'error',
+ request_id: request.request_id,
+ error: verdict.error,
+ },
+ }
+ }
+ break
+ }
+
+ case 'interrupt':
+ onInterrupt?.()
+ response = {
+ type: 'control_response',
+ response: {
+ subtype: 'success',
+ request_id: request.request_id,
+ },
+ }
+ break
+
+ default:
+ // Unknown subtype โ respond with error so the server doesn't
+ // hang waiting for a reply that never comes.
+ response = {
+ type: 'control_response',
+ response: {
+ subtype: 'error',
+ request_id: request.request_id,
+ error: `REPL bridge does not handle control_request subtype: ${request.request.subtype}`,
+ },
+ }
+ }
+
+ const event = { ...response, session_id: sessionId }
+ void transport.write(event)
+ logForDebugging(
+ `[bridge:repl] Sent control_response for ${request.request.subtype} request_id=${request.request_id} result=${response.response.subtype}`,
+ )
+}
+
+// โโโ Result message (for session archival on teardown) โโโโโโโโโโโโโโโโโโโโโโโ
+
+/**
+ * Build a minimal `SDKResultSuccess` message for session archival.
+ * The server needs this event before a WS close to trigger archival.
+ */
+export function makeResultMessage(sessionId: string): SDKResultSuccess {
+ return {
+ type: 'result',
+ subtype: 'success',
+ duration_ms: 0,
+ duration_api_ms: 0,
+ is_error: false,
+ num_turns: 0,
+ result: '',
+ stop_reason: null,
+ total_cost_usd: 0,
+ usage: { ...EMPTY_USAGE },
+ modelUsage: {},
+ permission_denials: [],
+ session_id: sessionId,
+ uuid: randomUUID(),
+ }
+}
+
+// โโโ BoundedUUIDSet (echo-dedup ring buffer) โโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโ
+
+/**
+ * FIFO-bounded set backed by a circular buffer. Evicts the oldest entry
+ * when capacity is reached, keeping memory usage constant at O(capacity).
+ *
+ * Messages are added in chronological order, so evicted entries are always
+ * the oldest. The caller relies on external ordering (the hook's
+ * lastWrittenIndexRef) as the primary dedup โ this set is a secondary
+ * safety net for echo filtering and race-condition dedup.
+ */
+export class BoundedUUIDSet {
+ private readonly capacity: number
+ private readonly ring: (string | undefined)[]
+ private readonly set = new Set()
+ private writeIdx = 0
+
+ constructor(capacity: number) {
+ this.capacity = capacity
+ this.ring = new Array(capacity)
+ }
+
+ add(uuid: string): void {
+ if (this.set.has(uuid)) return
+ // Evict the entry at the current write position (if occupied)
+ const evicted = this.ring[this.writeIdx]
+ if (evicted !== undefined) {
+ this.set.delete(evicted)
+ }
+ this.ring[this.writeIdx] = uuid
+ this.set.add(uuid)
+ this.writeIdx = (this.writeIdx + 1) % this.capacity
+ }
+
+ has(uuid: string): boolean {
+ return this.set.has(uuid)
+ }
+
+ clear(): void {
+ this.set.clear()
+ this.ring.fill(undefined)
+ this.writeIdx = 0
+ }
+}
diff --git a/original-source-code/src/bridge/bridgePermissionCallbacks.ts b/original-source-code/src/bridge/bridgePermissionCallbacks.ts
new file mode 100644
index 0000000..feaee66
--- /dev/null
+++ b/original-source-code/src/bridge/bridgePermissionCallbacks.ts
@@ -0,0 +1,43 @@
+import type { PermissionUpdate } from '../utils/permissions/PermissionUpdateSchema.js'
+
+type BridgePermissionResponse = {
+ behavior: 'allow' | 'deny'
+ updatedInput?: Record
+ updatedPermissions?: PermissionUpdate[]
+ message?: string
+}
+
+type BridgePermissionCallbacks = {
+ sendRequest(
+ requestId: string,
+ toolName: string,
+ input: Record,
+ toolUseId: string,
+ description: string,
+ permissionSuggestions?: PermissionUpdate[],
+ blockedPath?: string,
+ ): void
+ sendResponse(requestId: string, response: BridgePermissionResponse): void
+ /** Cancel a pending control_request so the web app can dismiss its prompt. */
+ cancelRequest(requestId: string): void
+ onResponse(
+ requestId: string,
+ handler: (response: BridgePermissionResponse) => void,
+ ): () => void // returns unsubscribe
+}
+
+/** Type predicate for validating a parsed control_response payload
+ * as a BridgePermissionResponse. Checks the required `behavior`
+ * discriminant rather than using an unsafe `as` cast. */
+function isBridgePermissionResponse(
+ value: unknown,
+): value is BridgePermissionResponse {
+ if (!value || typeof value !== 'object') return false
+ return (
+ 'behavior' in value &&
+ (value.behavior === 'allow' || value.behavior === 'deny')
+ )
+}
+
+export { isBridgePermissionResponse }
+export type { BridgePermissionCallbacks, BridgePermissionResponse }
diff --git a/original-source-code/src/bridge/bridgePointer.ts b/original-source-code/src/bridge/bridgePointer.ts
new file mode 100644
index 0000000..c32befc
--- /dev/null
+++ b/original-source-code/src/bridge/bridgePointer.ts
@@ -0,0 +1,210 @@
+import { mkdir, readFile, stat, unlink, writeFile } from 'fs/promises'
+import { dirname, join } from 'path'
+import { z } from 'zod/v4'
+import { logForDebugging } from '../utils/debug.js'
+import { isENOENT } from '../utils/errors.js'
+import { getWorktreePathsPortable } from '../utils/getWorktreePathsPortable.js'
+import { lazySchema } from '../utils/lazySchema.js'
+import {
+ getProjectsDir,
+ sanitizePath,
+} from '../utils/sessionStoragePortable.js'
+import { jsonParse, jsonStringify } from '../utils/slowOperations.js'
+
+/**
+ * Upper bound on worktree fanout. git worktree list is naturally bounded
+ * (50 is a LOT), but this caps the parallel stat() burst and guards against
+ * pathological setups. Above this, --continue falls back to current-dir-only.
+ */
+const MAX_WORKTREE_FANOUT = 50
+
+/**
+ * Crash-recovery pointer for Remote Control sessions.
+ *
+ * Written immediately after a bridge session is created, periodically
+ * refreshed during the session, and cleared on clean shutdown. If the
+ * process dies unclean (crash, kill -9, terminal closed), the pointer
+ * persists. On next startup, `claude remote-control` detects it and offers
+ * to resume via the --session-id flow from #20460.
+ *
+ * Staleness is checked against the file's mtime (not an embedded timestamp)
+ * so that a periodic re-write with the same content serves as a refresh โ
+ * matches the backend's rolling BRIDGE_LAST_POLL_TTL (4h) semantics. A
+ * bridge that's been polling for 5+ hours and then crashes still has a
+ * fresh pointer as long as the refresh ran within the window.
+ *
+ * Scoped per working directory (alongside transcript JSONL files) so two
+ * concurrent bridges in different repos don't clobber each other.
+ */
+
+export const BRIDGE_POINTER_TTL_MS = 4 * 60 * 60 * 1000
+
+const BridgePointerSchema = lazySchema(() =>
+ z.object({
+ sessionId: z.string(),
+ environmentId: z.string(),
+ source: z.enum(['standalone', 'repl']),
+ }),
+)
+
+export type BridgePointer = z.infer>
+
+export function getBridgePointerPath(dir: string): string {
+ return join(getProjectsDir(), sanitizePath(dir), 'bridge-pointer.json')
+}
+
+/**
+ * Write the pointer. Also used to refresh mtime during long sessions โ
+ * calling with the same IDs is a cheap no-content-change write that bumps
+ * the staleness clock. Best-effort โ a crash-recovery file must never
+ * itself cause a crash. Logs and swallows on error.
+ */
+export async function writeBridgePointer(
+ dir: string,
+ pointer: BridgePointer,
+): Promise {
+ const path = getBridgePointerPath(dir)
+ try {
+ await mkdir(dirname(path), { recursive: true })
+ await writeFile(path, jsonStringify(pointer), 'utf8')
+ logForDebugging(`[bridge:pointer] wrote ${path}`)
+ } catch (err: unknown) {
+ logForDebugging(`[bridge:pointer] write failed: ${err}`, { level: 'warn' })
+ }
+}
+
+/**
+ * Read the pointer and its age (ms since last write). Operates directly
+ * and handles errors โ no existence check (CLAUDE.md TOCTOU rule). Returns
+ * null on any failure: missing file, corrupted JSON, schema mismatch, or
+ * stale (mtime > 4h ago). Stale/invalid pointers are deleted so they don't
+ * keep re-prompting after the backend has already GC'd the env.
+ */
+export async function readBridgePointer(
+ dir: string,
+): Promise<(BridgePointer & { ageMs: number }) | null> {
+ const path = getBridgePointerPath(dir)
+ let raw: string
+ let mtimeMs: number
+ try {
+ // stat for mtime (staleness anchor), then read. Two syscalls, but both
+ // are needed โ mtime IS the data we return, not a TOCTOU guard.
+ mtimeMs = (await stat(path)).mtimeMs
+ raw = await readFile(path, 'utf8')
+ } catch {
+ return null
+ }
+
+ const parsed = BridgePointerSchema().safeParse(safeJsonParse(raw))
+ if (!parsed.success) {
+ logForDebugging(`[bridge:pointer] invalid schema, clearing: ${path}`)
+ await clearBridgePointer(dir)
+ return null
+ }
+
+ const ageMs = Math.max(0, Date.now() - mtimeMs)
+ if (ageMs > BRIDGE_POINTER_TTL_MS) {
+ logForDebugging(`[bridge:pointer] stale (>4h mtime), clearing: ${path}`)
+ await clearBridgePointer(dir)
+ return null
+ }
+
+ return { ...parsed.data, ageMs }
+}
+
+/**
+ * Worktree-aware read for `--continue`. The REPL bridge writes its pointer
+ * to `getOriginalCwd()` which EnterWorktreeTool/activeWorktreeSession can
+ * mutate to a worktree path โ but `claude remote-control --continue` runs
+ * with `resolve('.')` = shell CWD. This fans out across git worktree
+ * siblings to find the freshest pointer, matching /resume's semantics.
+ *
+ * Fast path: checks `dir` first. Only shells out to `git worktree list` if
+ * that misses โ the common case (pointer in launch dir) is one stat, zero
+ * exec. Fanout reads run in parallel; capped at MAX_WORKTREE_FANOUT.
+ *
+ * Returns the pointer AND the dir it was found in, so the caller can clear
+ * the right file on resume failure.
+ */
+export async function readBridgePointerAcrossWorktrees(
+ dir: string,
+): Promise<{ pointer: BridgePointer & { ageMs: number }; dir: string } | null> {
+ // Fast path: current dir. Covers standalone bridge (always matches) and
+ // REPL bridge when no worktree mutation happened.
+ const here = await readBridgePointer(dir)
+ if (here) {
+ return { pointer: here, dir }
+ }
+
+ // Fanout: scan worktree siblings. getWorktreePathsPortable has a 5s
+ // timeout and returns [] on any error (not a git repo, git not installed).
+ const worktrees = await getWorktreePathsPortable(dir)
+ if (worktrees.length <= 1) return null
+ if (worktrees.length > MAX_WORKTREE_FANOUT) {
+ logForDebugging(
+ `[bridge:pointer] ${worktrees.length} worktrees exceeds fanout cap ${MAX_WORKTREE_FANOUT}, skipping`,
+ )
+ return null
+ }
+
+ // Dedupe against `dir` so we don't re-stat it. sanitizePath normalizes
+ // case/separators so worktree-list output matches our fast-path key even
+ // on Windows where git may emit C:/ vs stored c:/.
+ const dirKey = sanitizePath(dir)
+ const candidates = worktrees.filter(wt => sanitizePath(wt) !== dirKey)
+
+ // Parallel stat+read. Each readBridgePointer is a stat() that ENOENTs
+ // for worktrees with no pointer (cheap) plus a ~100-byte read for the
+ // rare ones that have one. Promise.all โ latency โ slowest single stat.
+ const results = await Promise.all(
+ candidates.map(async wt => {
+ const p = await readBridgePointer(wt)
+ return p ? { pointer: p, dir: wt } : null
+ }),
+ )
+
+ // Pick freshest (lowest ageMs). The pointer stores environmentId so
+ // resume reconnects to the right env regardless of which worktree
+ // --continue was invoked from.
+ let freshest: {
+ pointer: BridgePointer & { ageMs: number }
+ dir: string
+ } | null = null
+ for (const r of results) {
+ if (r && (!freshest || r.pointer.ageMs < freshest.pointer.ageMs)) {
+ freshest = r
+ }
+ }
+ if (freshest) {
+ logForDebugging(
+ `[bridge:pointer] fanout found pointer in worktree ${freshest.dir} (ageMs=${freshest.pointer.ageMs})`,
+ )
+ }
+ return freshest
+}
+
+/**
+ * Delete the pointer. Idempotent โ ENOENT is expected when the process
+ * shut down clean previously.
+ */
+export async function clearBridgePointer(dir: string): Promise {
+ const path = getBridgePointerPath(dir)
+ try {
+ await unlink(path)
+ logForDebugging(`[bridge:pointer] cleared ${path}`)
+ } catch (err: unknown) {
+ if (!isENOENT(err)) {
+ logForDebugging(`[bridge:pointer] clear failed: ${err}`, {
+ level: 'warn',
+ })
+ }
+ }
+}
+
+function safeJsonParse(raw: string): unknown {
+ try {
+ return jsonParse(raw)
+ } catch {
+ return null
+ }
+}
diff --git a/original-source-code/src/bridge/bridgeStatusUtil.ts b/original-source-code/src/bridge/bridgeStatusUtil.ts
new file mode 100644
index 0000000..90de462
--- /dev/null
+++ b/original-source-code/src/bridge/bridgeStatusUtil.ts
@@ -0,0 +1,163 @@
+import {
+ getClaudeAiBaseUrl,
+ getRemoteSessionUrl,
+} from '../constants/product.js'
+import { stringWidth } from '../ink/stringWidth.js'
+import { formatDuration, truncateToWidth } from '../utils/format.js'
+import { getGraphemeSegmenter } from '../utils/intl.js'
+
+/** Bridge status state machine states. */
+export type StatusState =
+ | 'idle'
+ | 'attached'
+ | 'titled'
+ | 'reconnecting'
+ | 'failed'
+
+/** How long a tool activity line stays visible after last tool_start (ms). */
+export const TOOL_DISPLAY_EXPIRY_MS = 30_000
+
+/** Interval for the shimmer animation tick (ms). */
+export const SHIMMER_INTERVAL_MS = 150
+
+export function timestamp(): string {
+ const now = new Date()
+ const h = String(now.getHours()).padStart(2, '0')
+ const m = String(now.getMinutes()).padStart(2, '0')
+ const s = String(now.getSeconds()).padStart(2, '0')
+ return `${h}:${m}:${s}`
+}
+
+export { formatDuration, truncateToWidth as truncatePrompt }
+
+/** Abbreviate a tool activity summary for the trail display. */
+export function abbreviateActivity(summary: string): string {
+ return truncateToWidth(summary, 30)
+}
+
+/** Build the connect URL shown when the bridge is idle. */
+export function buildBridgeConnectUrl(
+ environmentId: string,
+ ingressUrl?: string,
+): string {
+ const baseUrl = getClaudeAiBaseUrl(undefined, ingressUrl)
+ return `${baseUrl}/code?bridge=${environmentId}`
+}
+
+/**
+ * Build the session URL shown when a session is attached. Delegates to
+ * getRemoteSessionUrl for the cse_โsession_ prefix translation, then appends
+ * the v1-specific ?bridge={environmentId} query.
+ */
+export function buildBridgeSessionUrl(
+ sessionId: string,
+ environmentId: string,
+ ingressUrl?: string,
+): string {
+ return `${getRemoteSessionUrl(sessionId, ingressUrl)}?bridge=${environmentId}`
+}
+
+/** Compute the glimmer index for a reverse-sweep shimmer animation. */
+export function computeGlimmerIndex(
+ tick: number,
+ messageWidth: number,
+): number {
+ const cycleLength = messageWidth + 20
+ return messageWidth + 10 - (tick % cycleLength)
+}
+
+/**
+ * Split text into three segments by visual column position for shimmer rendering.
+ *
+ * Uses grapheme segmentation and `stringWidth` so the split is correct for
+ * multi-byte characters, emoji, and CJK glyphs.
+ *
+ * Returns `{ before, shimmer, after }` strings. Both renderers (chalk in
+ * bridgeUI.ts and React/Ink in bridge.tsx) apply their own coloring to
+ * these segments.
+ */
+export function computeShimmerSegments(
+ text: string,
+ glimmerIndex: number,
+): { before: string; shimmer: string; after: string } {
+ const messageWidth = stringWidth(text)
+ const shimmerStart = glimmerIndex - 1
+ const shimmerEnd = glimmerIndex + 1
+
+ // When shimmer is offscreen, return all text as "before"
+ if (shimmerStart >= messageWidth || shimmerEnd < 0) {
+ return { before: text, shimmer: '', after: '' }
+ }
+
+ // Split into at most 3 segments by visual column position
+ const clampedStart = Math.max(0, shimmerStart)
+ let colPos = 0
+ let before = ''
+ let shimmer = ''
+ let after = ''
+ for (const { segment } of getGraphemeSegmenter().segment(text)) {
+ const segWidth = stringWidth(segment)
+ if (colPos + segWidth <= clampedStart) {
+ before += segment
+ } else if (colPos > shimmerEnd) {
+ after += segment
+ } else {
+ shimmer += segment
+ }
+ colPos += segWidth
+ }
+
+ return { before, shimmer, after }
+}
+
+/** Computed bridge status label and color from connection state. */
+export type BridgeStatusInfo = {
+ label:
+ | 'Remote Control failed'
+ | 'Remote Control reconnecting'
+ | 'Remote Control active'
+ | 'Remote Control connecting\u2026'
+ color: 'error' | 'warning' | 'success'
+}
+
+/** Derive a status label and color from the bridge connection state. */
+export function getBridgeStatus({
+ error,
+ connected,
+ sessionActive,
+ reconnecting,
+}: {
+ error: string | undefined
+ connected: boolean
+ sessionActive: boolean
+ reconnecting: boolean
+}): BridgeStatusInfo {
+ if (error) return { label: 'Remote Control failed', color: 'error' }
+ if (reconnecting)
+ return { label: 'Remote Control reconnecting', color: 'warning' }
+ if (sessionActive || connected)
+ return { label: 'Remote Control active', color: 'success' }
+ return { label: 'Remote Control connecting\u2026', color: 'warning' }
+}
+
+/** Footer text shown when bridge is idle (Ready state). */
+export function buildIdleFooterText(url: string): string {
+ return `Code everywhere with the Claude app or ${url}`
+}
+
+/** Footer text shown when a session is active (Connected state). */
+export function buildActiveFooterText(url: string): string {
+ return `Continue coding in the Claude app or ${url}`
+}
+
+/** Footer text shown when the bridge has failed. */
+export const FAILED_FOOTER_TEXT = 'Something went wrong, please try again'
+
+/**
+ * Wrap text in an OSC 8 terminal hyperlink. Zero visual width for layout purposes.
+ * strip-ansi (used by stringWidth) correctly strips these sequences, so
+ * countVisualLines in bridgeUI.ts remains accurate.
+ */
+export function wrapWithOsc8Link(text: string, url: string): string {
+ return `\x1b]8;;${url}\x07${text}\x1b]8;;\x07`
+}
diff --git a/original-source-code/src/bridge/bridgeUI.ts b/original-source-code/src/bridge/bridgeUI.ts
new file mode 100644
index 0000000..5149839
--- /dev/null
+++ b/original-source-code/src/bridge/bridgeUI.ts
@@ -0,0 +1,530 @@
+import chalk from 'chalk'
+import { toString as qrToString } from 'qrcode'
+import {
+ BRIDGE_FAILED_INDICATOR,
+ BRIDGE_READY_INDICATOR,
+ BRIDGE_SPINNER_FRAMES,
+} from '../constants/figures.js'
+import { stringWidth } from '../ink/stringWidth.js'
+import { logForDebugging } from '../utils/debug.js'
+import {
+ buildActiveFooterText,
+ buildBridgeConnectUrl,
+ buildBridgeSessionUrl,
+ buildIdleFooterText,
+ FAILED_FOOTER_TEXT,
+ formatDuration,
+ type StatusState,
+ TOOL_DISPLAY_EXPIRY_MS,
+ timestamp,
+ truncatePrompt,
+ wrapWithOsc8Link,
+} from './bridgeStatusUtil.js'
+import type {
+ BridgeConfig,
+ BridgeLogger,
+ SessionActivity,
+ SpawnMode,
+} from './types.js'
+
+const QR_OPTIONS = {
+ type: 'utf8' as const,
+ errorCorrectionLevel: 'L' as const,
+ small: true,
+}
+
+/** Generate a QR code and return its lines. */
+async function generateQr(url: string): Promise {
+ const qr = await qrToString(url, QR_OPTIONS)
+ return qr.split('\n').filter((line: string) => line.length > 0)
+}
+
+export function createBridgeLogger(options: {
+ verbose: boolean
+ write?: (s: string) => void
+}): BridgeLogger {
+ const write = options.write ?? ((s: string) => process.stdout.write(s))
+ const verbose = options.verbose
+
+ // Track how many status lines are currently displayed at the bottom
+ let statusLineCount = 0
+
+ // Status state machine
+ let currentState: StatusState = 'idle'
+ let currentStateText = 'Ready'
+ let repoName = ''
+ let branch = ''
+ let debugLogPath = ''
+
+ // Connect URL (built in printBanner with correct base for staging/prod)
+ let connectUrl = ''
+ let cachedIngressUrl = ''
+ let cachedEnvironmentId = ''
+ let activeSessionUrl: string | null = null
+
+ // QR code lines for the current URL
+ let qrLines: string[] = []
+ let qrVisible = false
+
+ // Tool activity for the second status line
+ let lastToolSummary: string | null = null
+ let lastToolTime = 0
+
+ // Session count indicator (shown when multi-session mode is enabled)
+ let sessionActive = 0
+ let sessionMax = 1
+ // Spawn mode shown in the session-count line + gates the `w` hint
+ let spawnModeDisplay: 'same-dir' | 'worktree' | null = null
+ let spawnMode: SpawnMode = 'single-session'
+
+ // Per-session display info for the multi-session bullet list (keyed by compat sessionId)
+ const sessionDisplayInfo = new Map<
+ string,
+ { title?: string; url: string; activity?: SessionActivity }
+ >()
+
+ // Connecting spinner state
+ let connectingTimer: ReturnType | null = null
+ let connectingTick = 0
+
+ /**
+ * Count how many visual terminal rows a string occupies, accounting for
+ * line wrapping. Each `\n` is one row, and content wider than the terminal
+ * wraps to additional rows.
+ */
+ function countVisualLines(text: string): number {
+ // eslint-disable-next-line custom-rules/prefer-use-terminal-size
+ const cols = process.stdout.columns || 80 // non-React CLI context
+ let count = 0
+ // Split on newlines to get logical lines
+ for (const logical of text.split('\n')) {
+ if (logical.length === 0) {
+ // Empty segment between consecutive \n โ counts as 1 row
+ count++
+ continue
+ }
+ const width = stringWidth(logical)
+ count += Math.max(1, Math.ceil(width / cols))
+ }
+ // The trailing \n in "line\n" produces an empty last element โ don't count it
+ // because the cursor sits at the start of the next line, not a new visual row.
+ if (text.endsWith('\n')) {
+ count--
+ }
+ return count
+ }
+
+ /** Write a status line and track its visual line count. */
+ function writeStatus(text: string): void {
+ write(text)
+ statusLineCount += countVisualLines(text)
+ }
+
+ /** Clear any currently displayed status lines. */
+ function clearStatusLines(): void {
+ if (statusLineCount <= 0) return
+ logForDebugging(`[bridge:ui] clearStatusLines count=${statusLineCount}`)
+ // Move cursor up to the start of the status block, then erase everything below
+ write(`\x1b[${statusLineCount}A`) // cursor up N lines
+ write('\x1b[J') // erase from cursor to end of screen
+ statusLineCount = 0
+ }
+
+ /** Print a permanent log line, clearing status first and restoring after. */
+ function printLog(line: string): void {
+ clearStatusLines()
+ write(line)
+ }
+
+ /** Regenerate the QR code with the given URL. */
+ function regenerateQr(url: string): void {
+ generateQr(url)
+ .then(lines => {
+ qrLines = lines
+ renderStatusLine()
+ })
+ .catch(e => {
+ logForDebugging(`QR code generation failed: ${e}`, { level: 'error' })
+ })
+ }
+
+ /** Render the connecting spinner line (shown before first updateIdleStatus). */
+ function renderConnectingLine(): void {
+ clearStatusLines()
+
+ const frame =
+ BRIDGE_SPINNER_FRAMES[connectingTick % BRIDGE_SPINNER_FRAMES.length]!
+ let suffix = ''
+ if (repoName) {
+ suffix += chalk.dim(' \u00b7 ') + chalk.dim(repoName)
+ }
+ if (branch) {
+ suffix += chalk.dim(' \u00b7 ') + chalk.dim(branch)
+ }
+ writeStatus(
+ `${chalk.yellow(frame)} ${chalk.yellow('Connecting')}${suffix}\n`,
+ )
+ }
+
+ /** Start the connecting spinner. Stopped by first updateIdleStatus(). */
+ function startConnecting(): void {
+ stopConnecting()
+ renderConnectingLine()
+ connectingTimer = setInterval(() => {
+ connectingTick++
+ renderConnectingLine()
+ }, 150)
+ }
+
+ /** Stop the connecting spinner. */
+ function stopConnecting(): void {
+ if (connectingTimer) {
+ clearInterval(connectingTimer)
+ connectingTimer = null
+ }
+ }
+
+ /** Render and write the current status lines based on state. */
+ function renderStatusLine(): void {
+ if (currentState === 'reconnecting' || currentState === 'failed') {
+ // These states are handled separately (updateReconnectingStatus /
+ // updateFailedStatus). Return before clearing so callers like toggleQr
+ // and setSpawnModeDisplay don't blank the display during these states.
+ return
+ }
+
+ clearStatusLines()
+
+ const isIdle = currentState === 'idle'
+
+ // QR code above the status line
+ if (qrVisible) {
+ for (const line of qrLines) {
+ writeStatus(`${chalk.dim(line)}\n`)
+ }
+ }
+
+ // Determine indicator and colors based on state
+ const indicator = BRIDGE_READY_INDICATOR
+ const indicatorColor = isIdle ? chalk.green : chalk.cyan
+ const baseColor = isIdle ? chalk.green : chalk.cyan
+ const stateText = baseColor(currentStateText)
+
+ // Build the suffix with repo and branch
+ let suffix = ''
+ if (repoName) {
+ suffix += chalk.dim(' \u00b7 ') + chalk.dim(repoName)
+ }
+ // In worktree mode each session gets its own branch, so showing the
+ // bridge's branch would be misleading.
+ if (branch && spawnMode !== 'worktree') {
+ suffix += chalk.dim(' \u00b7 ') + chalk.dim(branch)
+ }
+
+ if (process.env.USER_TYPE === 'ant' && debugLogPath) {
+ writeStatus(
+ `${chalk.yellow('[ANT-ONLY] Logs:')} ${chalk.dim(debugLogPath)}\n`,
+ )
+ }
+ writeStatus(`${indicatorColor(indicator)} ${stateText}${suffix}\n`)
+
+ // Session count and per-session list (multi-session mode only)
+ if (sessionMax > 1) {
+ const modeHint =
+ spawnMode === 'worktree'
+ ? 'New sessions will be created in an isolated worktree'
+ : 'New sessions will be created in the current directory'
+ writeStatus(
+ ` ${chalk.dim(`Capacity: ${sessionActive}/${sessionMax} \u00b7 ${modeHint}`)}\n`,
+ )
+ for (const [, info] of sessionDisplayInfo) {
+ const titleText = info.title
+ ? truncatePrompt(info.title, 35)
+ : chalk.dim('Attached')
+ const titleLinked = wrapWithOsc8Link(titleText, info.url)
+ const act = info.activity
+ const showAct = act && act.type !== 'result' && act.type !== 'error'
+ const actText = showAct
+ ? chalk.dim(` ${truncatePrompt(act.summary, 40)}`)
+ : ''
+ writeStatus(` ${titleLinked}${actText}
+`)
+ }
+ }
+
+ // Mode line for spawn modes with a single slot (or true single-session mode)
+ if (sessionMax === 1) {
+ const modeText =
+ spawnMode === 'single-session'
+ ? 'Single session \u00b7 exits when complete'
+ : spawnMode === 'worktree'
+ ? `Capacity: ${sessionActive}/1 \u00b7 New sessions will be created in an isolated worktree`
+ : `Capacity: ${sessionActive}/1 \u00b7 New sessions will be created in the current directory`
+ writeStatus(` ${chalk.dim(modeText)}\n`)
+ }
+
+ // Tool activity line for single-session mode
+ if (
+ sessionMax === 1 &&
+ !isIdle &&
+ lastToolSummary &&
+ Date.now() - lastToolTime < TOOL_DISPLAY_EXPIRY_MS
+ ) {
+ writeStatus(` ${chalk.dim(truncatePrompt(lastToolSummary, 60))}\n`)
+ }
+
+ // Blank line separator before footer
+ const url = activeSessionUrl ?? connectUrl
+ if (url) {
+ writeStatus('\n')
+ const footerText = isIdle
+ ? buildIdleFooterText(url)
+ : buildActiveFooterText(url)
+ const qrHint = qrVisible
+ ? chalk.dim.italic('space to hide QR code')
+ : chalk.dim.italic('space to show QR code')
+ const toggleHint = spawnModeDisplay
+ ? chalk.dim.italic(' \u00b7 w to toggle spawn mode')
+ : ''
+ writeStatus(`${chalk.dim(footerText)}\n`)
+ writeStatus(`${qrHint}${toggleHint}\n`)
+ }
+ }
+
+ return {
+ printBanner(config: BridgeConfig, environmentId: string): void {
+ cachedIngressUrl = config.sessionIngressUrl
+ cachedEnvironmentId = environmentId
+ connectUrl = buildBridgeConnectUrl(environmentId, cachedIngressUrl)
+ regenerateQr(connectUrl)
+
+ if (verbose) {
+ write(chalk.dim(`Remote Control`) + ` v${MACRO.VERSION}\n`)
+ }
+ if (verbose) {
+ if (config.spawnMode !== 'single-session') {
+ write(chalk.dim(`Spawn mode: `) + `${config.spawnMode}\n`)
+ write(
+ chalk.dim(`Max concurrent sessions: `) + `${config.maxSessions}\n`,
+ )
+ }
+ write(chalk.dim(`Environment ID: `) + `${environmentId}\n`)
+ }
+ if (config.sandbox) {
+ write(chalk.dim(`Sandbox: `) + `${chalk.green('Enabled')}\n`)
+ }
+ write('\n')
+
+ // Start connecting spinner โ first updateIdleStatus() will stop it
+ startConnecting()
+ },
+
+ logSessionStart(sessionId: string, prompt: string): void {
+ if (verbose) {
+ const short = truncatePrompt(prompt, 80)
+ printLog(
+ chalk.dim(`[${timestamp()}]`) +
+ ` Session started: ${chalk.white(`"${short}"`)} (${chalk.dim(sessionId)})\n`,
+ )
+ }
+ },
+
+ logSessionComplete(sessionId: string, durationMs: number): void {
+ printLog(
+ chalk.dim(`[${timestamp()}]`) +
+ ` Session ${chalk.green('completed')} (${formatDuration(durationMs)}) ${chalk.dim(sessionId)}\n`,
+ )
+ },
+
+ logSessionFailed(sessionId: string, error: string): void {
+ printLog(
+ chalk.dim(`[${timestamp()}]`) +
+ ` Session ${chalk.red('failed')}: ${error} ${chalk.dim(sessionId)}\n`,
+ )
+ },
+
+ logStatus(message: string): void {
+ printLog(chalk.dim(`[${timestamp()}]`) + ` ${message}\n`)
+ },
+
+ logVerbose(message: string): void {
+ if (verbose) {
+ printLog(chalk.dim(`[${timestamp()}] ${message}`) + '\n')
+ }
+ },
+
+ logError(message: string): void {
+ printLog(chalk.red(`[${timestamp()}] Error: ${message}`) + '\n')
+ },
+
+ logReconnected(disconnectedMs: number): void {
+ printLog(
+ chalk.dim(`[${timestamp()}]`) +
+ ` ${chalk.green('Reconnected')} after ${formatDuration(disconnectedMs)}\n`,
+ )
+ },
+
+ setRepoInfo(repo: string, branchName: string): void {
+ repoName = repo
+ branch = branchName
+ },
+
+ setDebugLogPath(path: string): void {
+ debugLogPath = path
+ },
+
+ updateIdleStatus(): void {
+ stopConnecting()
+
+ currentState = 'idle'
+ currentStateText = 'Ready'
+ lastToolSummary = null
+ lastToolTime = 0
+ activeSessionUrl = null
+ regenerateQr(connectUrl)
+ renderStatusLine()
+ },
+
+ setAttached(sessionId: string): void {
+ stopConnecting()
+ currentState = 'attached'
+ currentStateText = 'Connected'
+ lastToolSummary = null
+ lastToolTime = 0
+ // Multi-session: keep footer/QR on the environment connect URL so users
+ // can spawn more sessions. Per-session links are in the bullet list.
+ if (sessionMax <= 1) {
+ activeSessionUrl = buildBridgeSessionUrl(
+ sessionId,
+ cachedEnvironmentId,
+ cachedIngressUrl,
+ )
+ regenerateQr(activeSessionUrl)
+ }
+ renderStatusLine()
+ },
+
+ updateReconnectingStatus(delayStr: string, elapsedStr: string): void {
+ stopConnecting()
+ clearStatusLines()
+ currentState = 'reconnecting'
+
+ // QR code above the status line
+ if (qrVisible) {
+ for (const line of qrLines) {
+ writeStatus(`${chalk.dim(line)}\n`)
+ }
+ }
+
+ const frame =
+ BRIDGE_SPINNER_FRAMES[connectingTick % BRIDGE_SPINNER_FRAMES.length]!
+ connectingTick++
+ writeStatus(
+ `${chalk.yellow(frame)} ${chalk.yellow('Reconnecting')} ${chalk.dim('\u00b7')} ${chalk.dim(`retrying in ${delayStr}`)} ${chalk.dim('\u00b7')} ${chalk.dim(`disconnected ${elapsedStr}`)}\n`,
+ )
+ },
+
+ updateFailedStatus(error: string): void {
+ stopConnecting()
+ clearStatusLines()
+ currentState = 'failed'
+
+ let suffix = ''
+ if (repoName) {
+ suffix += chalk.dim(' \u00b7 ') + chalk.dim(repoName)
+ }
+ if (branch) {
+ suffix += chalk.dim(' \u00b7 ') + chalk.dim(branch)
+ }
+
+ writeStatus(
+ `${chalk.red(BRIDGE_FAILED_INDICATOR)} ${chalk.red('Remote Control Failed')}${suffix}\n`,
+ )
+ writeStatus(`${chalk.dim(FAILED_FOOTER_TEXT)}\n`)
+
+ if (error) {
+ writeStatus(`${chalk.red(error)}\n`)
+ }
+ },
+
+ updateSessionStatus(
+ _sessionId: string,
+ _elapsed: string,
+ activity: SessionActivity,
+ _trail: string[],
+ ): void {
+ // Cache tool activity for the second status line
+ if (activity.type === 'tool_start') {
+ lastToolSummary = activity.summary
+ lastToolTime = Date.now()
+ }
+ renderStatusLine()
+ },
+
+ clearStatus(): void {
+ stopConnecting()
+ clearStatusLines()
+ },
+
+ toggleQr(): void {
+ qrVisible = !qrVisible
+ renderStatusLine()
+ },
+
+ updateSessionCount(active: number, max: number, mode: SpawnMode): void {
+ if (sessionActive === active && sessionMax === max && spawnMode === mode)
+ return
+ sessionActive = active
+ sessionMax = max
+ spawnMode = mode
+ // Don't re-render here โ the status ticker calls renderStatusLine
+ // on its own cadence, and the next tick will pick up the new values.
+ },
+
+ setSpawnModeDisplay(mode: 'same-dir' | 'worktree' | null): void {
+ if (spawnModeDisplay === mode) return
+ spawnModeDisplay = mode
+ // Also sync the #21118-added spawnMode so the next render shows correct
+ // mode hint + branch visibility. Don't render here โ matches
+ // updateSessionCount: called before printBanner (initial setup) and
+ // again from the `w` handler (which follows with refreshDisplay).
+ if (mode) spawnMode = mode
+ },
+
+ addSession(sessionId: string, url: string): void {
+ sessionDisplayInfo.set(sessionId, { url })
+ },
+
+ updateSessionActivity(sessionId: string, activity: SessionActivity): void {
+ const info = sessionDisplayInfo.get(sessionId)
+ if (!info) return
+ info.activity = activity
+ },
+
+ setSessionTitle(sessionId: string, title: string): void {
+ const info = sessionDisplayInfo.get(sessionId)
+ if (!info) return
+ info.title = title
+ // Guard against reconnecting/failed โ renderStatusLine clears then returns
+ // early for those states, which would erase the spinner/error.
+ if (currentState === 'reconnecting' || currentState === 'failed') return
+ if (sessionMax === 1) {
+ // Single-session: show title in the main status line too.
+ currentState = 'titled'
+ currentStateText = truncatePrompt(title, 40)
+ }
+ renderStatusLine()
+ },
+
+ removeSession(sessionId: string): void {
+ sessionDisplayInfo.delete(sessionId)
+ },
+
+ refreshDisplay(): void {
+ // Skip during reconnecting/failed โ renderStatusLine clears then returns
+ // early for those states, which would erase the spinner/error.
+ if (currentState === 'reconnecting' || currentState === 'failed') return
+ renderStatusLine()
+ },
+ }
+}
diff --git a/original-source-code/src/bridge/capacityWake.ts b/original-source-code/src/bridge/capacityWake.ts
new file mode 100644
index 0000000..e58c50d
--- /dev/null
+++ b/original-source-code/src/bridge/capacityWake.ts
@@ -0,0 +1,56 @@
+/**
+ * Shared capacity-wake primitive for bridge poll loops.
+ *
+ * Both replBridge.ts and bridgeMain.ts need to sleep while "at capacity"
+ * but wake early when either (a) the outer loop signal aborts (shutdown),
+ * or (b) capacity frees up (session done / transport lost). This module
+ * encapsulates the mutable wake-controller + two-signal merger that both
+ * poll loops previously duplicated byte-for-byte.
+ */
+
+export type CapacitySignal = { signal: AbortSignal; cleanup: () => void }
+
+export type CapacityWake = {
+ /**
+ * Create a signal that aborts when either the outer loop signal or the
+ * capacity-wake controller fires. Returns the merged signal and a cleanup
+ * function that removes listeners when the sleep resolves normally
+ * (without abort).
+ */
+ signal(): CapacitySignal
+ /**
+ * Abort the current at-capacity sleep and arm a fresh controller so the
+ * poll loop immediately re-checks for new work.
+ */
+ wake(): void
+}
+
+export function createCapacityWake(outerSignal: AbortSignal): CapacityWake {
+ let wakeController = new AbortController()
+
+ function wake(): void {
+ wakeController.abort()
+ wakeController = new AbortController()
+ }
+
+ function signal(): CapacitySignal {
+ const merged = new AbortController()
+ const abort = (): void => merged.abort()
+ if (outerSignal.aborted || wakeController.signal.aborted) {
+ merged.abort()
+ return { signal: merged.signal, cleanup: () => {} }
+ }
+ outerSignal.addEventListener('abort', abort, { once: true })
+ const capSig = wakeController.signal
+ capSig.addEventListener('abort', abort, { once: true })
+ return {
+ signal: merged.signal,
+ cleanup: () => {
+ outerSignal.removeEventListener('abort', abort)
+ capSig.removeEventListener('abort', abort)
+ },
+ }
+ }
+
+ return { signal, wake }
+}
diff --git a/original-source-code/src/bridge/codeSessionApi.ts b/original-source-code/src/bridge/codeSessionApi.ts
new file mode 100644
index 0000000..65b46a3
--- /dev/null
+++ b/original-source-code/src/bridge/codeSessionApi.ts
@@ -0,0 +1,168 @@
+/**
+ * Thin HTTP wrappers for the CCR v2 code-session API.
+ *
+ * Separate file from remoteBridgeCore.ts so the SDK /bridge subpath can
+ * export createCodeSession + fetchRemoteCredentials without bundling the
+ * heavy CLI tree (analytics, transport, etc.). Callers supply explicit
+ * accessToken + baseUrl โ no implicit auth or config reads.
+ */
+
+import axios from 'axios'
+import { logForDebugging } from '../utils/debug.js'
+import { errorMessage } from '../utils/errors.js'
+import { jsonStringify } from '../utils/slowOperations.js'
+import { extractErrorDetail } from './debugUtils.js'
+
+const ANTHROPIC_VERSION = '2023-06-01'
+
+function oauthHeaders(accessToken: string): Record {
+ return {
+ Authorization: `Bearer ${accessToken}`,
+ 'Content-Type': 'application/json',
+ 'anthropic-version': ANTHROPIC_VERSION,
+ }
+}
+
+export async function createCodeSession(
+ baseUrl: string,
+ accessToken: string,
+ title: string,
+ timeoutMs: number,
+ tags?: string[],
+): Promise {
+ const url = `${baseUrl}/v1/code/sessions`
+ let response
+ try {
+ response = await axios.post(
+ url,
+ // bridge: {} is the positive signal for the oneof runner โ omitting it
+ // (or sending environment_id: "") now 400s. BridgeRunner is an empty
+ // message today; it's a placeholder for future bridge-specific options.
+ { title, bridge: {}, ...(tags?.length ? { tags } : {}) },
+ {
+ headers: oauthHeaders(accessToken),
+ timeout: timeoutMs,
+ validateStatus: s => s < 500,
+ },
+ )
+ } catch (err: unknown) {
+ logForDebugging(
+ `[code-session] Session create request failed: ${errorMessage(err)}`,
+ )
+ return null
+ }
+
+ if (response.status !== 200 && response.status !== 201) {
+ const detail = extractErrorDetail(response.data)
+ logForDebugging(
+ `[code-session] Session create failed ${response.status}${detail ? `: ${detail}` : ''}`,
+ )
+ return null
+ }
+
+ const data: unknown = response.data
+ if (
+ !data ||
+ typeof data !== 'object' ||
+ !('session' in data) ||
+ !data.session ||
+ typeof data.session !== 'object' ||
+ !('id' in data.session) ||
+ typeof data.session.id !== 'string' ||
+ !data.session.id.startsWith('cse_')
+ ) {
+ logForDebugging(
+ `[code-session] No session.id (cse_*) in response: ${jsonStringify(data).slice(0, 200)}`,
+ )
+ return null
+ }
+ return data.session.id
+}
+
+/**
+ * Credentials from POST /bridge. JWT is opaque โ do not decode.
+ * Each /bridge call bumps worker_epoch server-side (it IS the register).
+ */
+export type RemoteCredentials = {
+ worker_jwt: string
+ api_base_url: string
+ expires_in: number
+ worker_epoch: number
+}
+
+export async function fetchRemoteCredentials(
+ sessionId: string,
+ baseUrl: string,
+ accessToken: string,
+ timeoutMs: number,
+ trustedDeviceToken?: string,
+): Promise {
+ const url = `${baseUrl}/v1/code/sessions/${sessionId}/bridge`
+ const headers = oauthHeaders(accessToken)
+ if (trustedDeviceToken) {
+ headers['X-Trusted-Device-Token'] = trustedDeviceToken
+ }
+ let response
+ try {
+ response = await axios.post(
+ url,
+ {},
+ {
+ headers,
+ timeout: timeoutMs,
+ validateStatus: s => s < 500,
+ },
+ )
+ } catch (err: unknown) {
+ logForDebugging(
+ `[code-session] /bridge request failed: ${errorMessage(err)}`,
+ )
+ return null
+ }
+
+ if (response.status !== 200) {
+ const detail = extractErrorDetail(response.data)
+ logForDebugging(
+ `[code-session] /bridge failed ${response.status}${detail ? `: ${detail}` : ''}`,
+ )
+ return null
+ }
+
+ const data: unknown = response.data
+ if (
+ data === null ||
+ typeof data !== 'object' ||
+ !('worker_jwt' in data) ||
+ typeof data.worker_jwt !== 'string' ||
+ !('expires_in' in data) ||
+ typeof data.expires_in !== 'number' ||
+ !('api_base_url' in data) ||
+ typeof data.api_base_url !== 'string' ||
+ !('worker_epoch' in data)
+ ) {
+ logForDebugging(
+ `[code-session] /bridge response malformed (need worker_jwt, expires_in, api_base_url, worker_epoch): ${jsonStringify(data).slice(0, 200)}`,
+ )
+ return null
+ }
+ // protojson serializes int64 as a string to avoid JS precision loss;
+ // Go may also return a number depending on encoder settings.
+ const rawEpoch = data.worker_epoch
+ const epoch = typeof rawEpoch === 'string' ? Number(rawEpoch) : rawEpoch
+ if (
+ typeof epoch !== 'number' ||
+ !Number.isFinite(epoch) ||
+ !Number.isSafeInteger(epoch)
+ ) {
+ logForDebugging(
+ `[code-session] /bridge worker_epoch invalid: ${jsonStringify(rawEpoch)}`,
+ )
+ return null
+ }
+ return {
+ worker_jwt: data.worker_jwt,
+ api_base_url: data.api_base_url,
+ expires_in: data.expires_in,
+ worker_epoch: epoch,
+ }
+}
diff --git a/original-source-code/src/bridge/createSession.ts b/original-source-code/src/bridge/createSession.ts
new file mode 100644
index 0000000..d5bc83a
--- /dev/null
+++ b/original-source-code/src/bridge/createSession.ts
@@ -0,0 +1,384 @@
+import type { SDKMessage } from '../entrypoints/agentSdkTypes.js'
+import { logForDebugging } from '../utils/debug.js'
+import { errorMessage } from '../utils/errors.js'
+import { extractErrorDetail } from './debugUtils.js'
+import { toCompatSessionId } from './sessionIdCompat.js'
+
+type GitSource = {
+ type: 'git_repository'
+ url: string
+ revision?: string
+}
+
+type GitOutcome = {
+ type: 'git_repository'
+ git_info: { type: 'github'; repo: string; branches: string[] }
+}
+
+// Events must be wrapped in { type: 'event', data: } for the
+// POST /v1/sessions endpoint (discriminated union format).
+type SessionEvent = {
+ type: 'event'
+ data: SDKMessage
+}
+
+/**
+ * Create a session on a bridge environment via POST /v1/sessions.
+ *
+ * Used by both `claude remote-control` (empty session so the user has somewhere to
+ * type immediately) and `/remote-control` (session pre-populated with conversation
+ * history).
+ *
+ * Returns the session ID on success, or null if creation fails (non-fatal).
+ */
+export async function createBridgeSession({
+ environmentId,
+ title,
+ events,
+ gitRepoUrl,
+ branch,
+ signal,
+ baseUrl: baseUrlOverride,
+ getAccessToken,
+ permissionMode,
+}: {
+ environmentId: string
+ title?: string
+ events: SessionEvent[]
+ gitRepoUrl: string | null
+ branch: string
+ signal: AbortSignal
+ baseUrl?: string
+ getAccessToken?: () => string | undefined
+ permissionMode?: string
+}): Promise {
+ const { getClaudeAIOAuthTokens } = await import('../utils/auth.js')
+ const { getOrganizationUUID } = await import('../services/oauth/client.js')
+ const { getOauthConfig } = await import('../constants/oauth.js')
+ const { getOAuthHeaders } = await import('../utils/teleport/api.js')
+ const { parseGitHubRepository } = await import('../utils/detectRepository.js')
+ const { getDefaultBranch } = await import('../utils/git.js')
+ const { getMainLoopModel } = await import('../utils/model/model.js')
+ const { default: axios } = await import('axios')
+
+ const accessToken =
+ getAccessToken?.() ?? getClaudeAIOAuthTokens()?.accessToken
+ if (!accessToken) {
+ logForDebugging('[bridge] No access token for session creation')
+ return null
+ }
+
+ const orgUUID = await getOrganizationUUID()
+ if (!orgUUID) {
+ logForDebugging('[bridge] No org UUID for session creation')
+ return null
+ }
+
+ // Build git source and outcome context
+ let gitSource: GitSource | null = null
+ let gitOutcome: GitOutcome | null = null
+
+ if (gitRepoUrl) {
+ const { parseGitRemote } = await import('../utils/detectRepository.js')
+ const parsed = parseGitRemote(gitRepoUrl)
+ if (parsed) {
+ const { host, owner, name } = parsed
+ const revision = branch || (await getDefaultBranch()) || undefined
+ gitSource = {
+ type: 'git_repository',
+ url: `https://${host}/${owner}/${name}`,
+ revision,
+ }
+ gitOutcome = {
+ type: 'git_repository',
+ git_info: {
+ type: 'github',
+ repo: `${owner}/${name}`,
+ branches: [`claude/${branch || 'task'}`],
+ },
+ }
+ } else {
+ // Fallback: try parseGitHubRepository for owner/repo format
+ const ownerRepo = parseGitHubRepository(gitRepoUrl)
+ if (ownerRepo) {
+ const [owner, name] = ownerRepo.split('/')
+ if (owner && name) {
+ const revision = branch || (await getDefaultBranch()) || undefined
+ gitSource = {
+ type: 'git_repository',
+ url: `https://github.com/${owner}/${name}`,
+ revision,
+ }
+ gitOutcome = {
+ type: 'git_repository',
+ git_info: {
+ type: 'github',
+ repo: `${owner}/${name}`,
+ branches: [`claude/${branch || 'task'}`],
+ },
+ }
+ }
+ }
+ }
+ }
+
+ const requestBody = {
+ ...(title !== undefined && { title }),
+ events,
+ session_context: {
+ sources: gitSource ? [gitSource] : [],
+ outcomes: gitOutcome ? [gitOutcome] : [],
+ model: getMainLoopModel(),
+ },
+ environment_id: environmentId,
+ source: 'remote-control',
+ ...(permissionMode && { permission_mode: permissionMode }),
+ }
+
+ const headers = {
+ ...getOAuthHeaders(accessToken),
+ 'anthropic-beta': 'ccr-byoc-2025-07-29',
+ 'x-organization-uuid': orgUUID,
+ }
+
+ const url = `${baseUrlOverride ?? getOauthConfig().BASE_API_URL}/v1/sessions`
+ let response
+ try {
+ response = await axios.post(url, requestBody, {
+ headers,
+ signal,
+ validateStatus: s => s < 500,
+ })
+ } catch (err: unknown) {
+ logForDebugging(
+ `[bridge] Session creation request failed: ${errorMessage(err)}`,
+ )
+ return null
+ }
+ const isSuccess = response.status === 200 || response.status === 201
+
+ if (!isSuccess) {
+ const detail = extractErrorDetail(response.data)
+ logForDebugging(
+ `[bridge] Session creation failed with status ${response.status}${detail ? `: ${detail}` : ''}`,
+ )
+ return null
+ }
+
+ const sessionData: unknown = response.data
+ if (
+ !sessionData ||
+ typeof sessionData !== 'object' ||
+ !('id' in sessionData) ||
+ typeof sessionData.id !== 'string'
+ ) {
+ logForDebugging('[bridge] No session ID in response')
+ return null
+ }
+
+ return sessionData.id
+}
+
+/**
+ * Fetch a bridge session via GET /v1/sessions/{id}.
+ *
+ * Returns the session's environment_id (for `--session-id` resume) and title.
+ * Uses the same org-scoped headers as create/archive โ the environments-level
+ * client in bridgeApi.ts uses a different beta header and no org UUID, which
+ * makes the Sessions API return 404.
+ */
+export async function getBridgeSession(
+ sessionId: string,
+ opts?: { baseUrl?: string; getAccessToken?: () => string | undefined },
+): Promise<{ environment_id?: string; title?: string } | null> {
+ const { getClaudeAIOAuthTokens } = await import('../utils/auth.js')
+ const { getOrganizationUUID } = await import('../services/oauth/client.js')
+ const { getOauthConfig } = await import('../constants/oauth.js')
+ const { getOAuthHeaders } = await import('../utils/teleport/api.js')
+ const { default: axios } = await import('axios')
+
+ const accessToken =
+ opts?.getAccessToken?.() ?? getClaudeAIOAuthTokens()?.accessToken
+ if (!accessToken) {
+ logForDebugging('[bridge] No access token for session fetch')
+ return null
+ }
+
+ const orgUUID = await getOrganizationUUID()
+ if (!orgUUID) {
+ logForDebugging('[bridge] No org UUID for session fetch')
+ return null
+ }
+
+ const headers = {
+ ...getOAuthHeaders(accessToken),
+ 'anthropic-beta': 'ccr-byoc-2025-07-29',
+ 'x-organization-uuid': orgUUID,
+ }
+
+ const url = `${opts?.baseUrl ?? getOauthConfig().BASE_API_URL}/v1/sessions/${sessionId}`
+ logForDebugging(`[bridge] Fetching session ${sessionId}`)
+
+ let response
+ try {
+ response = await axios.get<{ environment_id?: string; title?: string }>(
+ url,
+ { headers, timeout: 10_000, validateStatus: s => s < 500 },
+ )
+ } catch (err: unknown) {
+ logForDebugging(
+ `[bridge] Session fetch request failed: ${errorMessage(err)}`,
+ )
+ return null
+ }
+
+ if (response.status !== 200) {
+ const detail = extractErrorDetail(response.data)
+ logForDebugging(
+ `[bridge] Session fetch failed with status ${response.status}${detail ? `: ${detail}` : ''}`,
+ )
+ return null
+ }
+
+ return response.data
+}
+
+/**
+ * Archive a bridge session via POST /v1/sessions/{id}/archive.
+ *
+ * The CCR server never auto-archives sessions โ archival is always an
+ * explicit client action. Both `claude remote-control` (standalone bridge) and the
+ * always-on `/remote-control` REPL bridge call this during shutdown to archive any
+ * sessions that are still alive.
+ *
+ * The archive endpoint accepts sessions in any status (running, idle,
+ * requires_action, pending) and returns 409 if already archived, making
+ * it safe to call even if the server-side runner already archived the
+ * session.
+ *
+ * Callers must handle errors โ this function has no try/catch; 5xx,
+ * timeouts, and network errors throw. Archival is best-effort during
+ * cleanup; call sites wrap with .catch().
+ */
+export async function archiveBridgeSession(
+ sessionId: string,
+ opts?: {
+ baseUrl?: string
+ getAccessToken?: () => string | undefined
+ timeoutMs?: number
+ },
+): Promise {
+ const { getClaudeAIOAuthTokens } = await import('../utils/auth.js')
+ const { getOrganizationUUID } = await import('../services/oauth/client.js')
+ const { getOauthConfig } = await import('../constants/oauth.js')
+ const { getOAuthHeaders } = await import('../utils/teleport/api.js')
+ const { default: axios } = await import('axios')
+
+ const accessToken =
+ opts?.getAccessToken?.() ?? getClaudeAIOAuthTokens()?.accessToken
+ if (!accessToken) {
+ logForDebugging('[bridge] No access token for session archive')
+ return
+ }
+
+ const orgUUID = await getOrganizationUUID()
+ if (!orgUUID) {
+ logForDebugging('[bridge] No org UUID for session archive')
+ return
+ }
+
+ const headers = {
+ ...getOAuthHeaders(accessToken),
+ 'anthropic-beta': 'ccr-byoc-2025-07-29',
+ 'x-organization-uuid': orgUUID,
+ }
+
+ const url = `${opts?.baseUrl ?? getOauthConfig().BASE_API_URL}/v1/sessions/${sessionId}/archive`
+ logForDebugging(`[bridge] Archiving session ${sessionId}`)
+
+ const response = await axios.post(
+ url,
+ {},
+ {
+ headers,
+ timeout: opts?.timeoutMs ?? 10_000,
+ validateStatus: s => s < 500,
+ },
+ )
+
+ if (response.status === 200) {
+ logForDebugging(`[bridge] Session ${sessionId} archived successfully`)
+ } else {
+ const detail = extractErrorDetail(response.data)
+ logForDebugging(
+ `[bridge] Session archive failed with status ${response.status}${detail ? `: ${detail}` : ''}`,
+ )
+ }
+}
+
+/**
+ * Update the title of a bridge session via PATCH /v1/sessions/{id}.
+ *
+ * Called when the user renames a session via /rename while a bridge
+ * connection is active, so the title stays in sync on claude.ai/code.
+ *
+ * Errors are swallowed โ title sync is best-effort.
+ */
+export async function updateBridgeSessionTitle(
+ sessionId: string,
+ title: string,
+ opts?: { baseUrl?: string; getAccessToken?: () => string | undefined },
+): Promise {
+ const { getClaudeAIOAuthTokens } = await import('../utils/auth.js')
+ const { getOrganizationUUID } = await import('../services/oauth/client.js')
+ const { getOauthConfig } = await import('../constants/oauth.js')
+ const { getOAuthHeaders } = await import('../utils/teleport/api.js')
+ const { default: axios } = await import('axios')
+
+ const accessToken =
+ opts?.getAccessToken?.() ?? getClaudeAIOAuthTokens()?.accessToken
+ if (!accessToken) {
+ logForDebugging('[bridge] No access token for session title update')
+ return
+ }
+
+ const orgUUID = await getOrganizationUUID()
+ if (!orgUUID) {
+ logForDebugging('[bridge] No org UUID for session title update')
+ return
+ }
+
+ const headers = {
+ ...getOAuthHeaders(accessToken),
+ 'anthropic-beta': 'ccr-byoc-2025-07-29',
+ 'x-organization-uuid': orgUUID,
+ }
+
+ // Compat gateway only accepts session_* (compat/convert.go:27). v2 callers
+ // pass raw cse_*; retag here so all callers can pass whatever they hold.
+ // Idempotent for v1's session_* and bridgeMain's pre-converted compatSessionId.
+ const compatId = toCompatSessionId(sessionId)
+ const url = `${opts?.baseUrl ?? getOauthConfig().BASE_API_URL}/v1/sessions/${compatId}`
+ logForDebugging(`[bridge] Updating session title: ${compatId} โ ${title}`)
+
+ try {
+ const response = await axios.patch(
+ url,
+ { title },
+ { headers, timeout: 10_000, validateStatus: s => s < 500 },
+ )
+
+ if (response.status === 200) {
+ logForDebugging(`[bridge] Session title updated successfully`)
+ } else {
+ const detail = extractErrorDetail(response.data)
+ logForDebugging(
+ `[bridge] Session title update failed with status ${response.status}${detail ? `: ${detail}` : ''}`,
+ )
+ }
+ } catch (err: unknown) {
+ logForDebugging(
+ `[bridge] Session title update request failed: ${errorMessage(err)}`,
+ )
+ }
+}
diff --git a/original-source-code/src/bridge/debugUtils.ts b/original-source-code/src/bridge/debugUtils.ts
new file mode 100644
index 0000000..e9f7293
--- /dev/null
+++ b/original-source-code/src/bridge/debugUtils.ts
@@ -0,0 +1,141 @@
+import {
+ type AnalyticsMetadata_I_VERIFIED_THIS_IS_NOT_CODE_OR_FILEPATHS,
+ logEvent,
+} from '../services/analytics/index.js'
+import { logForDebugging } from '../utils/debug.js'
+import { errorMessage } from '../utils/errors.js'
+import { jsonStringify } from '../utils/slowOperations.js'
+
+const DEBUG_MSG_LIMIT = 2000
+
+const SECRET_FIELD_NAMES = [
+ 'session_ingress_token',
+ 'environment_secret',
+ 'access_token',
+ 'secret',
+ 'token',
+]
+
+const SECRET_PATTERN = new RegExp(
+ `"(${SECRET_FIELD_NAMES.join('|')})"\\s*:\\s*"([^"]*)"`,
+ 'g',
+)
+
+const REDACT_MIN_LENGTH = 16
+
+export function redactSecrets(s: string): string {
+ return s.replace(SECRET_PATTERN, (_match, field: string, value: string) => {
+ if (value.length < REDACT_MIN_LENGTH) {
+ return `"${field}":"[REDACTED]"`
+ }
+ const redacted = `${value.slice(0, 8)}...${value.slice(-4)}`
+ return `"${field}":"${redacted}"`
+ })
+}
+
+/** Truncate a string for debug logging, collapsing newlines. */
+export function debugTruncate(s: string): string {
+ const flat = s.replace(/\n/g, '\\n')
+ if (flat.length <= DEBUG_MSG_LIMIT) {
+ return flat
+ }
+ return flat.slice(0, DEBUG_MSG_LIMIT) + `... (${flat.length} chars)`
+}
+
+/** Truncate a JSON-serializable value for debug logging. */
+export function debugBody(data: unknown): string {
+ const raw = typeof data === 'string' ? data : jsonStringify(data)
+ const s = redactSecrets(raw)
+ if (s.length <= DEBUG_MSG_LIMIT) {
+ return s
+ }
+ return s.slice(0, DEBUG_MSG_LIMIT) + `... (${s.length} chars)`
+}
+
+/**
+ * Extract a descriptive error message from an axios error (or any error).
+ * For HTTP errors, appends the server's response body message if available,
+ * since axios's default message only includes the status code.
+ */
+export function describeAxiosError(err: unknown): string {
+ const msg = errorMessage(err)
+ if (err && typeof err === 'object' && 'response' in err) {
+ const response = (err as { response?: { data?: unknown } }).response
+ if (response?.data && typeof response.data === 'object') {
+ const data = response.data as Record
+ const detail =
+ typeof data.message === 'string'
+ ? data.message
+ : typeof data.error === 'object' &&
+ data.error &&
+ 'message' in data.error &&
+ typeof (data.error as Record).message ===
+ 'string'
+ ? (data.error as Record).message
+ : undefined
+ if (detail) {
+ return `${msg}: ${detail}`
+ }
+ }
+ }
+ return msg
+}
+
+/**
+ * Extract the HTTP status code from an axios error, if present.
+ * Returns undefined for non-HTTP errors (e.g. network failures).
+ */
+export function extractHttpStatus(err: unknown): number | undefined {
+ if (
+ err &&
+ typeof err === 'object' &&
+ 'response' in err &&
+ (err as { response?: { status?: unknown } }).response &&
+ typeof (err as { response: { status?: unknown } }).response.status ===
+ 'number'
+ ) {
+ return (err as { response: { status: number } }).response.status
+ }
+ return undefined
+}
+
+/**
+ * Pull a human-readable message out of an API error response body.
+ * Checks `data.message` first, then `data.error.message`.
+ */
+export function extractErrorDetail(data: unknown): string | undefined {
+ if (!data || typeof data !== 'object') return undefined
+ if ('message' in data && typeof data.message === 'string') {
+ return data.message
+ }
+ if (
+ 'error' in data &&
+ data.error !== null &&
+ typeof data.error === 'object' &&
+ 'message' in data.error &&
+ typeof data.error.message === 'string'
+ ) {
+ return data.error.message
+ }
+ return undefined
+}
+
+/**
+ * Log a bridge init skip โ debug message + `tengu_bridge_repl_skipped`
+ * analytics event. Centralizes the event name and the AnalyticsMetadata
+ * cast so call sites don't each repeat the 5-line boilerplate.
+ */
+export function logBridgeSkip(
+ reason: string,
+ debugMsg?: string,
+ v2?: boolean,
+): void {
+ if (debugMsg) {
+ logForDebugging(debugMsg)
+ }
+ logEvent('tengu_bridge_repl_skipped', {
+ reason:
+ reason as AnalyticsMetadata_I_VERIFIED_THIS_IS_NOT_CODE_OR_FILEPATHS,
+ ...(v2 !== undefined && { v2 }),
+ })
+}
diff --git a/original-source-code/src/bridge/envLessBridgeConfig.ts b/original-source-code/src/bridge/envLessBridgeConfig.ts
new file mode 100644
index 0000000..de0cb5e
--- /dev/null
+++ b/original-source-code/src/bridge/envLessBridgeConfig.ts
@@ -0,0 +1,165 @@
+import { z } from 'zod/v4'
+import { getFeatureValue_DEPRECATED } from '../services/analytics/growthbook.js'
+import { lazySchema } from '../utils/lazySchema.js'
+import { lt } from '../utils/semver.js'
+import { isEnvLessBridgeEnabled } from './bridgeEnabled.js'
+
+export type EnvLessBridgeConfig = {
+ // withRetry โ init-phase backoff (createSession, POST /bridge, recovery /bridge)
+ init_retry_max_attempts: number
+ init_retry_base_delay_ms: number
+ init_retry_jitter_fraction: number
+ init_retry_max_delay_ms: number
+ // axios timeout for POST /sessions, POST /bridge, POST /archive
+ http_timeout_ms: number
+ // BoundedUUIDSet ring size (echo + re-delivery dedup)
+ uuid_dedup_buffer_size: number
+ // CCRClient worker heartbeat cadence. Server TTL is 60s โ 20s gives 3ร margin.
+ heartbeat_interval_ms: number
+ // ยฑfraction of interval โ per-beat jitter to spread fleet load.
+ heartbeat_jitter_fraction: number
+ // Fire proactive JWT refresh this long before expires_in. Larger buffer =
+ // more frequent refresh (refresh cadence โ expires_in - buffer).
+ token_refresh_buffer_ms: number
+ // Archive POST timeout in teardown(). Distinct from http_timeout_ms because
+ // gracefulShutdown races runCleanupFunctions() against a 2s cap โ a 10s
+ // axios timeout on a slow/stalled archive burns the whole budget on a
+ // request that forceExit will kill anyway.
+ teardown_archive_timeout_ms: number
+ // Deadline for onConnect after transport.connect(). If neither onConnect
+ // nor onClose fires before this, emit tengu_bridge_repl_connect_timeout
+ // โ the only telemetry for the ~1% of sessions that emit `started` then
+ // go silent (no error, no event, just nothing).
+ connect_timeout_ms: number
+ // Semver floor for the env-less bridge path. Separate from the v1
+ // tengu_bridge_min_version config so a v2-specific bug can force upgrades
+ // without blocking v1 (env-based) clients, and vice versa.
+ min_version: string
+ // When true, tell users their claude.ai app may be too old to see v2
+ // sessions โ lets us roll the v2 bridge before the app ships the new
+ // session-list query.
+ should_show_app_upgrade_message: boolean
+}
+
+export const DEFAULT_ENV_LESS_BRIDGE_CONFIG: EnvLessBridgeConfig = {
+ init_retry_max_attempts: 3,
+ init_retry_base_delay_ms: 500,
+ init_retry_jitter_fraction: 0.25,
+ init_retry_max_delay_ms: 4000,
+ http_timeout_ms: 10_000,
+ uuid_dedup_buffer_size: 2000,
+ heartbeat_interval_ms: 20_000,
+ heartbeat_jitter_fraction: 0.1,
+ token_refresh_buffer_ms: 300_000,
+ teardown_archive_timeout_ms: 1500,
+ connect_timeout_ms: 15_000,
+ min_version: '0.0.0',
+ should_show_app_upgrade_message: false,
+}
+
+// Floors reject the whole object on violation (fall back to DEFAULT) rather
+// than partially trusting โ same defense-in-depth as pollConfig.ts.
+const envLessBridgeConfigSchema = lazySchema(() =>
+ z.object({
+ init_retry_max_attempts: z.number().int().min(1).max(10).default(3),
+ init_retry_base_delay_ms: z.number().int().min(100).default(500),
+ init_retry_jitter_fraction: z.number().min(0).max(1).default(0.25),
+ init_retry_max_delay_ms: z.number().int().min(500).default(4000),
+ http_timeout_ms: z.number().int().min(2000).default(10_000),
+ uuid_dedup_buffer_size: z.number().int().min(100).max(50_000).default(2000),
+ // Server TTL is 60s. Floor 5s prevents thrash; cap 30s keeps โฅ2ร margin.
+ heartbeat_interval_ms: z
+ .number()
+ .int()
+ .min(5000)
+ .max(30_000)
+ .default(20_000),
+ // ยฑfraction per beat. Cap 0.5: at max interval (30s) ร 1.5 = 45s worst case,
+ // still under the 60s TTL.
+ heartbeat_jitter_fraction: z.number().min(0).max(0.5).default(0.1),
+ // Floor 30s prevents tight-looping. Cap 30min rejects buffer-vs-delay
+ // semantic inversion: ops entering expires_in-5min (the *delay until
+ // refresh*) instead of 5min (the *buffer before expiry*) yields
+ // delayMs = expires_in - buffer โ 5min instead of โ4h. Both are positive
+ // durations so .min() alone can't distinguish; .max() catches the
+ // inverted value since buffer โฅ 30min is nonsensical for a multi-hour JWT.
+ token_refresh_buffer_ms: z
+ .number()
+ .int()
+ .min(30_000)
+ .max(1_800_000)
+ .default(300_000),
+ // Cap 2000 keeps this under gracefulShutdown's 2s cleanup race โ a higher
+ // timeout just lies to axios since forceExit kills the socket regardless.
+ teardown_archive_timeout_ms: z
+ .number()
+ .int()
+ .min(500)
+ .max(2000)
+ .default(1500),
+ // Observed p99 connect is ~2-3s; 15s is ~5ร headroom. Floor 5s bounds
+ // false-positive rate under transient slowness; cap 60s bounds how long
+ // a truly-stalled session stays dark.
+ connect_timeout_ms: z.number().int().min(5_000).max(60_000).default(15_000),
+ min_version: z
+ .string()
+ .refine(v => {
+ try {
+ lt(v, '0.0.0')
+ return true
+ } catch {
+ return false
+ }
+ })
+ .default('0.0.0'),
+ should_show_app_upgrade_message: z.boolean().default(false),
+ }),
+)
+
+/**
+ * Fetch the env-less bridge timing config from GrowthBook. Read once per
+ * initEnvLessBridgeCore call โ config is fixed for the lifetime of a bridge
+ * session.
+ *
+ * Uses the blocking getter (not _CACHED_MAY_BE_STALE) because /remote-control
+ * runs well after GrowthBook init โ initializeGrowthBook() resolves instantly,
+ * so there's no startup penalty, and we get the fresh in-memory remoteEval
+ * value instead of the stale-on-first-read disk cache. The _DEPRECATED suffix
+ * warns against startup-path usage, which this isn't.
+ */
+export async function getEnvLessBridgeConfig(): Promise {
+ const raw = await getFeatureValue_DEPRECATED(
+ 'tengu_bridge_repl_v2_config',
+ DEFAULT_ENV_LESS_BRIDGE_CONFIG,
+ )
+ const parsed = envLessBridgeConfigSchema().safeParse(raw)
+ return parsed.success ? parsed.data : DEFAULT_ENV_LESS_BRIDGE_CONFIG
+}
+
+/**
+ * Returns an error message if the current CLI version is below the minimum
+ * required for the env-less (v2) bridge path, or null if the version is fine.
+ *
+ * v2 analogue of checkBridgeMinVersion() โ reads from tengu_bridge_repl_v2_config
+ * instead of tengu_bridge_min_version so the two implementations can enforce
+ * independent floors.
+ */
+export async function checkEnvLessBridgeMinVersion(): Promise {
+ const cfg = await getEnvLessBridgeConfig()
+ if (cfg.min_version && lt(MACRO.VERSION, cfg.min_version)) {
+ return `Your version of Claude Code (${MACRO.VERSION}) is too old for Remote Control.\nVersion ${cfg.min_version} or higher is required. Run \`claude update\` to update.`
+ }
+ return null
+}
+
+/**
+ * Whether to nudge users toward upgrading their claude.ai app when a
+ * Remote Control session starts. True only when the v2 bridge is active
+ * AND the should_show_app_upgrade_message config bit is set โ lets us
+ * roll the v2 bridge before the app ships the new session-list query.
+ */
+export async function shouldShowAppUpgradeMessage(): Promise {
+ if (!isEnvLessBridgeEnabled()) return false
+ const cfg = await getEnvLessBridgeConfig()
+ return cfg.should_show_app_upgrade_message
+}
diff --git a/original-source-code/src/bridge/flushGate.ts b/original-source-code/src/bridge/flushGate.ts
new file mode 100644
index 0000000..6216334
--- /dev/null
+++ b/original-source-code/src/bridge/flushGate.ts
@@ -0,0 +1,71 @@
+/**
+ * State machine for gating message writes during an initial flush.
+ *
+ * When a bridge session starts, historical messages are flushed to the
+ * server via a single HTTP POST. During that flush, new messages must
+ * be queued to prevent them from arriving at the server interleaved
+ * with the historical messages.
+ *
+ * Lifecycle:
+ * start() โ enqueue() returns true, items are queued
+ * end() โ returns queued items for draining, enqueue() returns false
+ * drop() โ discards queued items (permanent transport close)
+ * deactivate() โ clears active flag without dropping items
+ * (transport replacement โ new transport will drain)
+ */
+export class FlushGate {
+ private _active = false
+ private _pending: T[] = []
+
+ get active(): boolean {
+ return this._active
+ }
+
+ get pendingCount(): number {
+ return this._pending.length
+ }
+
+ /** Mark flush as in-progress. enqueue() will start queuing items. */
+ start(): void {
+ this._active = true
+ }
+
+ /**
+ * End the flush and return any queued items for draining.
+ * Caller is responsible for sending the returned items.
+ */
+ end(): T[] {
+ this._active = false
+ return this._pending.splice(0)
+ }
+
+ /**
+ * If flush is active, queue the items and return true.
+ * If flush is not active, return false (caller should send directly).
+ */
+ enqueue(...items: T[]): boolean {
+ if (!this._active) return false
+ this._pending.push(...items)
+ return true
+ }
+
+ /**
+ * Discard all queued items (permanent transport close).
+ * Returns the number of items dropped.
+ */
+ drop(): number {
+ this._active = false
+ const count = this._pending.length
+ this._pending.length = 0
+ return count
+ }
+
+ /**
+ * Clear the active flag without dropping queued items.
+ * Used when the transport is replaced (onWorkReceived) โ the new
+ * transport's flush will drain the pending items.
+ */
+ deactivate(): void {
+ this._active = false
+ }
+}
diff --git a/original-source-code/src/bridge/inboundAttachments.ts b/original-source-code/src/bridge/inboundAttachments.ts
new file mode 100644
index 0000000..f7c13c8
--- /dev/null
+++ b/original-source-code/src/bridge/inboundAttachments.ts
@@ -0,0 +1,175 @@
+/**
+ * Resolve file_uuid attachments on inbound bridge user messages.
+ *
+ * Web composer uploads via cookie-authed /api/{org}/upload, sends file_uuid
+ * alongside the message. Here we fetch each via GET /api/oauth/files/{uuid}/content
+ * (oauth-authed, same store), write to ~/.claude/uploads/{sessionId}/, and
+ * return @path refs to prepend. Claude's Read tool takes it from there.
+ *
+ * Best-effort: any failure (no token, network, non-2xx, disk) logs debug and
+ * skips that attachment. The message still reaches Claude, just without @path.
+ */
+
+import type { ContentBlockParam } from '@anthropic-ai/sdk/resources/messages.mjs'
+import axios from 'axios'
+import { randomUUID } from 'crypto'
+import { mkdir, writeFile } from 'fs/promises'
+import { basename, join } from 'path'
+import { z } from 'zod/v4'
+import { getSessionId } from '../bootstrap/state.js'
+import { logForDebugging } from '../utils/debug.js'
+import { getClaudeConfigHomeDir } from '../utils/envUtils.js'
+import { lazySchema } from '../utils/lazySchema.js'
+import { getBridgeAccessToken, getBridgeBaseUrl } from './bridgeConfig.js'
+
+const DOWNLOAD_TIMEOUT_MS = 30_000
+
+function debug(msg: string): void {
+ logForDebugging(`[bridge:inbound-attach] ${msg}`)
+}
+
+const attachmentSchema = lazySchema(() =>
+ z.object({
+ file_uuid: z.string(),
+ file_name: z.string(),
+ }),
+)
+const attachmentsArraySchema = lazySchema(() => z.array(attachmentSchema()))
+
+export type InboundAttachment = z.infer>
+
+/** Pull file_attachments off a loosely-typed inbound message. */
+export function extractInboundAttachments(msg: unknown): InboundAttachment[] {
+ if (typeof msg !== 'object' || msg === null || !('file_attachments' in msg)) {
+ return []
+ }
+ const parsed = attachmentsArraySchema().safeParse(msg.file_attachments)
+ return parsed.success ? parsed.data : []
+}
+
+/**
+ * Strip path components and keep only filename-safe chars. file_name comes
+ * from the network (web composer), so treat it as untrusted even though the
+ * composer controls it.
+ */
+function sanitizeFileName(name: string): string {
+ const base = basename(name).replace(/[^a-zA-Z0-9._-]/g, '_')
+ return base || 'attachment'
+}
+
+function uploadsDir(): string {
+ return join(getClaudeConfigHomeDir(), 'uploads', getSessionId())
+}
+
+/**
+ * Fetch + write one attachment. Returns the absolute path on success,
+ * undefined on any failure.
+ */
+async function resolveOne(att: InboundAttachment): Promise {
+ const token = getBridgeAccessToken()
+ if (!token) {
+ debug('skip: no oauth token')
+ return undefined
+ }
+
+ let data: Buffer
+ try {
+ // getOauthConfig() (via getBridgeBaseUrl) throws on a non-allowlisted
+ // CLAUDE_CODE_CUSTOM_OAUTH_URL โ keep it inside the try so a bad
+ // FedStart URL degrades to "no @path" instead of crashing print.ts's
+ // reader loop (which has no catch around the await).
+ const url = `${getBridgeBaseUrl()}/api/oauth/files/${encodeURIComponent(att.file_uuid)}/content`
+ const response = await axios.get(url, {
+ headers: { Authorization: `Bearer ${token}` },
+ responseType: 'arraybuffer',
+ timeout: DOWNLOAD_TIMEOUT_MS,
+ validateStatus: () => true,
+ })
+ if (response.status !== 200) {
+ debug(`fetch ${att.file_uuid} failed: status=${response.status}`)
+ return undefined
+ }
+ data = Buffer.from(response.data)
+ } catch (e) {
+ debug(`fetch ${att.file_uuid} threw: ${e}`)
+ return undefined
+ }
+
+ // uuid-prefix makes collisions impossible across messages and within one
+ // (same filename, different files). 8 chars is enough โ this isn't security.
+ const safeName = sanitizeFileName(att.file_name)
+ const prefix = (
+ att.file_uuid.slice(0, 8) || randomUUID().slice(0, 8)
+ ).replace(/[^a-zA-Z0-9_-]/g, '_')
+ const dir = uploadsDir()
+ const outPath = join(dir, `${prefix}-${safeName}`)
+
+ try {
+ await mkdir(dir, { recursive: true })
+ await writeFile(outPath, data)
+ } catch (e) {
+ debug(`write ${outPath} failed: ${e}`)
+ return undefined
+ }
+
+ debug(`resolved ${att.file_uuid} โ ${outPath} (${data.length} bytes)`)
+ return outPath
+}
+
+/**
+ * Resolve all attachments on an inbound message to a prefix string of
+ * @path refs. Empty string if none resolved.
+ */
+export async function resolveInboundAttachments(
+ attachments: InboundAttachment[],
+): Promise {
+ if (attachments.length === 0) return ''
+ debug(`resolving ${attachments.length} attachment(s)`)
+ const paths = await Promise.all(attachments.map(resolveOne))
+ const ok = paths.filter((p): p is string => p !== undefined)
+ if (ok.length === 0) return ''
+ // Quoted form โ extractAtMentionedFiles truncates unquoted @refs at the
+ // first space, which breaks any home dir with spaces (/Users/John Smith/).
+ return ok.map(p => `@"${p}"`).join(' ') + ' '
+}
+
+/**
+ * Prepend @path refs to content, whichever form it's in.
+ * Targets the LAST text block โ processUserInputBase reads inputString
+ * from processedBlocks[processedBlocks.length - 1], so putting refs in
+ * block[0] means they're silently ignored for [text, image] content.
+ */
+export function prependPathRefs(
+ content: string | Array,
+ prefix: string,
+): string | Array {
+ if (!prefix) return content
+ if (typeof content === 'string') return prefix + content
+ const i = content.findLastIndex(b => b.type === 'text')
+ if (i !== -1) {
+ const b = content[i]!
+ if (b.type === 'text') {
+ return [
+ ...content.slice(0, i),
+ { ...b, text: prefix + b.text },
+ ...content.slice(i + 1),
+ ]
+ }
+ }
+ // No text block โ append one at the end so it's last.
+ return [...content, { type: 'text', text: prefix.trimEnd() }]
+}
+
+/**
+ * Convenience: extract + resolve + prepend. No-op when the message has no
+ * file_attachments field (fast path โ no network, returns same reference).
+ */
+export async function resolveAndPrepend(
+ msg: unknown,
+ content: string | Array,
+): Promise> {
+ const attachments = extractInboundAttachments(msg)
+ if (attachments.length === 0) return content
+ const prefix = await resolveInboundAttachments(attachments)
+ return prependPathRefs(content, prefix)
+}
diff --git a/original-source-code/src/bridge/inboundMessages.ts b/original-source-code/src/bridge/inboundMessages.ts
new file mode 100644
index 0000000..2c02f50
--- /dev/null
+++ b/original-source-code/src/bridge/inboundMessages.ts
@@ -0,0 +1,80 @@
+import type {
+ Base64ImageSource,
+ ContentBlockParam,
+ ImageBlockParam,
+} from '@anthropic-ai/sdk/resources/messages.mjs'
+import type { UUID } from 'crypto'
+import type { SDKMessage } from '../entrypoints/agentSdkTypes.js'
+import { detectImageFormatFromBase64 } from '../utils/imageResizer.js'
+
+/**
+ * Process an inbound user message from the bridge, extracting content
+ * and UUID for enqueueing. Supports both string content and
+ * ContentBlockParam[] (e.g. messages containing images).
+ *
+ * Normalizes image blocks from bridge clients that may use camelCase
+ * `mediaType` instead of snake_case `media_type` (mobile-apps#5825).
+ *
+ * Returns the extracted fields, or undefined if the message should be
+ * skipped (non-user type, missing/empty content).
+ */
+export function extractInboundMessageFields(
+ msg: SDKMessage,
+):
+ | { content: string | Array; uuid: UUID | undefined }
+ | undefined {
+ if (msg.type !== 'user') return undefined
+ const content = msg.message?.content
+ if (!content) return undefined
+ if (Array.isArray(content) && content.length === 0) return undefined
+
+ const uuid =
+ 'uuid' in msg && typeof msg.uuid === 'string'
+ ? (msg.uuid as UUID)
+ : undefined
+
+ return {
+ content: Array.isArray(content) ? normalizeImageBlocks(content) : content,
+ uuid,
+ }
+}
+
+/**
+ * Normalize image content blocks from bridge clients. iOS/web clients may
+ * send `mediaType` (camelCase) instead of `media_type` (snake_case), or
+ * omit the field entirely. Without normalization, the bad block poisons
+ * the session โ every subsequent API call fails with
+ * "media_type: Field required".
+ *
+ * Fast-path scan returns the original array reference when no
+ * normalization is needed (zero allocation on the happy path).
+ */
+export function normalizeImageBlocks(
+ blocks: Array,
+): Array {
+ if (!blocks.some(isMalformedBase64Image)) return blocks
+
+ return blocks.map(block => {
+ if (!isMalformedBase64Image(block)) return block
+ const src = block.source as unknown as Record
+ const mediaType =
+ typeof src.mediaType === 'string' && src.mediaType
+ ? src.mediaType
+ : detectImageFormatFromBase64(block.source.data)
+ return {
+ ...block,
+ source: {
+ type: 'base64' as const,
+ media_type: mediaType as Base64ImageSource['media_type'],
+ data: block.source.data,
+ },
+ }
+ })
+}
+
+function isMalformedBase64Image(
+ block: ContentBlockParam,
+): block is ImageBlockParam & { source: Base64ImageSource } {
+ if (block.type !== 'image' || block.source?.type !== 'base64') return false
+ return !(block.source as unknown as Record).media_type
+}
diff --git a/original-source-code/src/bridge/initReplBridge.ts b/original-source-code/src/bridge/initReplBridge.ts
new file mode 100644
index 0000000..85e403d
--- /dev/null
+++ b/original-source-code/src/bridge/initReplBridge.ts
@@ -0,0 +1,569 @@
+/**
+ * REPL-specific wrapper around initBridgeCore. Owns the parts that read
+ * bootstrap state โ gates, cwd, session ID, git context, OAuth, title
+ * derivation โ then delegates to the bootstrap-free core.
+ *
+ * Split out of replBridge.ts because the sessionStorage import
+ * (getCurrentSessionTitle) transitively pulls in src/commands.ts โ the
+ * entire slash command + React component tree (~1300 modules). Keeping
+ * initBridgeCore in a file that doesn't touch sessionStorage lets
+ * daemonBridge.ts import the core without bloating the Agent SDK bundle.
+ *
+ * Called via dynamic import by useReplBridge (auto-start) and print.ts
+ * (SDK -p mode via query.enableRemoteControl).
+ */
+
+import { feature } from 'bun:bundle'
+import { hostname } from 'os'
+import { getOriginalCwd, getSessionId } from '../bootstrap/state.js'
+import type { SDKMessage } from '../entrypoints/agentSdkTypes.js'
+import type { SDKControlResponse } from '../entrypoints/sdk/controlTypes.js'
+import { getFeatureValue_CACHED_WITH_REFRESH } from '../services/analytics/growthbook.js'
+import { getOrganizationUUID } from '../services/oauth/client.js'
+import {
+ isPolicyAllowed,
+ waitForPolicyLimitsToLoad,
+} from '../services/policyLimits/index.js'
+import type { Message } from '../types/message.js'
+import {
+ checkAndRefreshOAuthTokenIfNeeded,
+ getClaudeAIOAuthTokens,
+ handleOAuth401Error,
+} from '../utils/auth.js'
+import { getGlobalConfig, saveGlobalConfig } from '../utils/config.js'
+import { logForDebugging } from '../utils/debug.js'
+import { stripDisplayTagsAllowEmpty } from '../utils/displayTags.js'
+import { errorMessage } from '../utils/errors.js'
+import { getBranch, getRemoteUrl } from '../utils/git.js'
+import { toSDKMessages } from '../utils/messages/mappers.js'
+import {
+ getContentText,
+ getMessagesAfterCompactBoundary,
+ isSyntheticMessage,
+} from '../utils/messages.js'
+import type { PermissionMode } from '../utils/permissions/PermissionMode.js'
+import { getCurrentSessionTitle } from '../utils/sessionStorage.js'
+import {
+ extractConversationText,
+ generateSessionTitle,
+} from '../utils/sessionTitle.js'
+import { generateShortWordSlug } from '../utils/words.js'
+import {
+ getBridgeAccessToken,
+ getBridgeBaseUrl,
+ getBridgeTokenOverride,
+} from './bridgeConfig.js'
+import {
+ checkBridgeMinVersion,
+ isBridgeEnabledBlocking,
+ isCseShimEnabled,
+ isEnvLessBridgeEnabled,
+} from './bridgeEnabled.js'
+import {
+ archiveBridgeSession,
+ createBridgeSession,
+ updateBridgeSessionTitle,
+} from './createSession.js'
+import { logBridgeSkip } from './debugUtils.js'
+import { checkEnvLessBridgeMinVersion } from './envLessBridgeConfig.js'
+import { getPollIntervalConfig } from './pollConfig.js'
+import type { BridgeState, ReplBridgeHandle } from './replBridge.js'
+import { initBridgeCore } from './replBridge.js'
+import { setCseShimGate } from './sessionIdCompat.js'
+import type { BridgeWorkerType } from './types.js'
+
+export type InitBridgeOptions = {
+ onInboundMessage?: (msg: SDKMessage) => void | Promise
+ onPermissionResponse?: (response: SDKControlResponse) => void
+ onInterrupt?: () => void
+ onSetModel?: (model: string | undefined) => void
+ onSetMaxThinkingTokens?: (maxTokens: number | null) => void
+ onSetPermissionMode?: (
+ mode: PermissionMode,
+ ) => { ok: true } | { ok: false; error: string }
+ onStateChange?: (state: BridgeState, detail?: string) => void
+ initialMessages?: Message[]
+ // Explicit session name from `/remote-control `. When set, overrides
+ // the title derived from the conversation or /rename.
+ initialName?: string
+ // Fresh view of the full conversation at call time. Used by onUserMessage's
+ // count-3 derivation to call generateSessionTitle over the full conversation.
+ // Optional โ print.ts's SDK enableRemoteControl path has no REPL message
+ // array; count-3 falls back to the single message text when absent.
+ getMessages?: () => Message[]
+ // UUIDs already flushed in a prior bridge session. Messages with these
+ // UUIDs are excluded from the initial flush to avoid poisoning the
+ // server (duplicate UUIDs across sessions cause the WS to be killed).
+ // Mutated in place โ newly flushed UUIDs are added after each flush.
+ previouslyFlushedUUIDs?: Set
+ /** See BridgeCoreParams.perpetual. */
+ perpetual?: boolean
+ /**
+ * When true, the bridge only forwards events outbound (no SSE inbound
+ * stream). Used by CCR mirror mode โ local sessions visible on claude.ai
+ * without enabling inbound control.
+ */
+ outboundOnly?: boolean
+ tags?: string[]
+}
+
+export async function initReplBridge(
+ options?: InitBridgeOptions,
+): Promise {
+ const {
+ onInboundMessage,
+ onPermissionResponse,
+ onInterrupt,
+ onSetModel,
+ onSetMaxThinkingTokens,
+ onSetPermissionMode,
+ onStateChange,
+ initialMessages,
+ getMessages,
+ previouslyFlushedUUIDs,
+ initialName,
+ perpetual,
+ outboundOnly,
+ tags,
+ } = options ?? {}
+
+ // Wire the cse_ shim kill switch so toCompatSessionId respects the
+ // GrowthBook gate. Daemon/SDK paths skip this โ shim defaults to active.
+ setCseShimGate(isCseShimEnabled)
+
+ // 1. Runtime gate
+ if (!(await isBridgeEnabledBlocking())) {
+ logBridgeSkip('not_enabled', '[bridge:repl] Skipping: bridge not enabled')
+ return null
+ }
+
+ // 1b. Minimum version check โ deferred to after the v1/v2 branch below,
+ // since each implementation has its own floor (tengu_bridge_min_version
+ // for v1, tengu_bridge_repl_v2_config.min_version for v2).
+
+ // 2. Check OAuth โ must be signed in with claude.ai. Runs before the
+ // policy check so console-auth users get the actionable "/login" hint
+ // instead of a misleading policy error from a stale/wrong-org cache.
+ if (!getBridgeAccessToken()) {
+ logBridgeSkip('no_oauth', '[bridge:repl] Skipping: no OAuth tokens')
+ onStateChange?.('failed', '/login')
+ return null
+ }
+
+ // 3. Check organization policy โ remote control may be disabled
+ await waitForPolicyLimitsToLoad()
+ if (!isPolicyAllowed('allow_remote_control')) {
+ logBridgeSkip(
+ 'policy_denied',
+ '[bridge:repl] Skipping: allow_remote_control policy not allowed',
+ )
+ onStateChange?.('failed', "disabled by your organization's policy")
+ return null
+ }
+
+ // When CLAUDE_BRIDGE_OAUTH_TOKEN is set (ant-only local dev), the bridge
+ // uses that token directly via getBridgeAccessToken() โ keychain state is
+ // irrelevant. Skip 2b/2c to preserve that decoupling: an expired keychain
+ // token shouldn't block a bridge connection that doesn't use it.
+ if (!getBridgeTokenOverride()) {
+ // 2a. Cross-process backoff. If N prior processes already saw this exact
+ // dead token (matched by expiresAt), skip silently โ no event, no refresh
+ // attempt. The count threshold tolerates transient refresh failures (auth
+ // server 5xx, lockfile errors per auth.ts:1437/1444/1485): each process
+ // independently retries until 3 consecutive failures prove the token dead.
+ // Mirrors useReplBridge's MAX_CONSECUTIVE_INIT_FAILURES for in-process.
+ // The expiresAt key is content-addressed: /login โ new token โ new expiresAt
+ // โ this stops matching without any explicit clear.
+ const cfg = getGlobalConfig()
+ if (
+ cfg.bridgeOauthDeadExpiresAt != null &&
+ (cfg.bridgeOauthDeadFailCount ?? 0) >= 3 &&
+ getClaudeAIOAuthTokens()?.expiresAt === cfg.bridgeOauthDeadExpiresAt
+ ) {
+ logForDebugging(
+ `[bridge:repl] Skipping: cross-process backoff (dead token seen ${cfg.bridgeOauthDeadFailCount} times)`,
+ )
+ return null
+ }
+
+ // 2b. Proactively refresh if expired. Mirrors bridgeMain.ts:2096 โ the REPL
+ // bridge fires at useEffect mount BEFORE any v1/messages call, making this
+ // usually the first OAuth request of the session. Without this, ~9% of
+ // registrations hit the server with a >8h-expired token โ 401 โ withOAuthRetry
+ // recovers, but the server logs a 401 we can avoid. VPN egress IPs observed
+ // at 30:1 401:200 when many unrelated users cluster at the 8h TTL boundary.
+ //
+ // Fresh-token cost: one memoized read + one Date.now() comparison (~ยตs).
+ // checkAndRefreshOAuthTokenIfNeeded clears its own cache in every path that
+ // touches the keychain (refresh success, lockfile race, throw), so no
+ // explicit clearOAuthTokenCache() here โ that would force a blocking
+ // keychain spawn on the 91%+ fresh-token path.
+ await checkAndRefreshOAuthTokenIfNeeded()
+
+ // 2c. Skip if token is still expired post-refresh-attempt. Env-var / FD
+ // tokens (auth.ts:894-917) have expiresAt=null โ never trip this. But a
+ // keychain token whose refresh token is dead (password change, org left,
+ // token GC'd) has expiresAt ({
+ ...c,
+ bridgeOauthDeadExpiresAt: deadExpiresAt,
+ bridgeOauthDeadFailCount:
+ c.bridgeOauthDeadExpiresAt === deadExpiresAt
+ ? (c.bridgeOauthDeadFailCount ?? 0) + 1
+ : 1,
+ }))
+ return null
+ }
+ }
+
+ // 4. Compute baseUrl โ needed by both v1 (env-based) and v2 (env-less)
+ // paths. Hoisted above the v2 gate so both can use it.
+ const baseUrl = getBridgeBaseUrl()
+
+ // 5. Derive session title. Precedence: explicit initialName โ /rename
+ // (session storage) โ last meaningful user message โ generated slug.
+ // Cosmetic only (claude.ai session list); the model never sees it.
+ // Two flags: `hasExplicitTitle` (initialName or /rename โ never auto-
+ // overwrite) vs. `hasTitle` (any title, including auto-derived โ blocks
+ // the count-1 re-derivation but not count-3). The onUserMessage callback
+ // (wired to both v1 and v2 below) derives from the 1st prompt and again
+ // from the 3rd so mobile/web show a title that reflects more context.
+ // The slug fallback (e.g. "remote-control-graceful-unicorn") makes
+ // auto-started sessions distinguishable in the claude.ai list before the
+ // first prompt.
+ let title = `remote-control-${generateShortWordSlug()}`
+ let hasTitle = false
+ let hasExplicitTitle = false
+ if (initialName) {
+ title = initialName
+ hasTitle = true
+ hasExplicitTitle = true
+ } else {
+ const sessionId = getSessionId()
+ const customTitle = sessionId
+ ? getCurrentSessionTitle(sessionId)
+ : undefined
+ if (customTitle) {
+ title = customTitle
+ hasTitle = true
+ hasExplicitTitle = true
+ } else if (initialMessages && initialMessages.length > 0) {
+ // Find the last user message that has meaningful content. Skip meta
+ // (nudges), tool results, compact summaries ("This session is being
+ // continuedโฆ"), non-human origins (task notifications, channel pushes),
+ // and synthetic interrupts ([Request interrupted by user]) โ none are
+ // human-authored. Same filter as extractTitleText + isSyntheticMessage.
+ for (let i = initialMessages.length - 1; i >= 0; i--) {
+ const msg = initialMessages[i]!
+ if (
+ msg.type !== 'user' ||
+ msg.isMeta ||
+ msg.toolUseResult ||
+ msg.isCompactSummary ||
+ (msg.origin && msg.origin.kind !== 'human') ||
+ isSyntheticMessage(msg)
+ )
+ continue
+ const rawContent = getContentText(msg.message.content)
+ if (!rawContent) continue
+ const derived = deriveTitle(rawContent)
+ if (!derived) continue
+ title = derived
+ hasTitle = true
+ break
+ }
+ }
+ }
+
+ // Shared by both v1 and v2 โ fires on every title-worthy user message until
+ // it returns true. At count 1: deriveTitle placeholder immediately, then
+ // generateSessionTitle (Haiku, sentence-case) fire-and-forget upgrade. At
+ // count 3: re-generate over the full conversation. Skips entirely if the
+ // title is explicit (/remote-control or /rename) โ re-checks
+ // sessionStorage at call time so /rename between messages isn't clobbered.
+ // Skips count 1 if initialMessages already derived (that title is fresh);
+ // still refreshes at count 3. v2 passes cse_*; updateBridgeSessionTitle
+ // retags internally.
+ let userMessageCount = 0
+ let lastBridgeSessionId: string | undefined
+ let genSeq = 0
+ const patch = (
+ derived: string,
+ bridgeSessionId: string,
+ atCount: number,
+ ): void => {
+ hasTitle = true
+ title = derived
+ logForDebugging(
+ `[bridge:repl] derived title from message ${atCount}: ${derived}`,
+ )
+ void updateBridgeSessionTitle(bridgeSessionId, derived, {
+ baseUrl,
+ getAccessToken: getBridgeAccessToken,
+ }).catch(() => {})
+ }
+ // Fire-and-forget Haiku generation with post-await guards. Re-checks /rename
+ // (sessionStorage), v1 env-lost (lastBridgeSessionId), and same-session
+ // out-of-order resolution (genSeq โ count-1's Haiku resolving after count-3
+ // would clobber the richer title). generateSessionTitle never rejects.
+ const generateAndPatch = (input: string, bridgeSessionId: string): void => {
+ const gen = ++genSeq
+ const atCount = userMessageCount
+ void generateSessionTitle(input, AbortSignal.timeout(15_000)).then(
+ generated => {
+ if (
+ generated &&
+ gen === genSeq &&
+ lastBridgeSessionId === bridgeSessionId &&
+ !getCurrentSessionTitle(getSessionId())
+ ) {
+ patch(generated, bridgeSessionId, atCount)
+ }
+ },
+ )
+ }
+ const onUserMessage = (text: string, bridgeSessionId: string): boolean => {
+ if (hasExplicitTitle || getCurrentSessionTitle(getSessionId())) {
+ return true
+ }
+ // v1 env-lost re-creates the session with a new ID. Reset the count so
+ // the new session gets its own count-3 derivation; hasTitle stays true
+ // (new session was created via getCurrentTitle(), which reads the count-1
+ // title from this closure), so count-1 of the fresh cycle correctly skips.
+ if (
+ lastBridgeSessionId !== undefined &&
+ lastBridgeSessionId !== bridgeSessionId
+ ) {
+ userMessageCount = 0
+ }
+ lastBridgeSessionId = bridgeSessionId
+ userMessageCount++
+ if (userMessageCount === 1 && !hasTitle) {
+ const placeholder = deriveTitle(text)
+ if (placeholder) patch(placeholder, bridgeSessionId, userMessageCount)
+ generateAndPatch(text, bridgeSessionId)
+ } else if (userMessageCount === 3) {
+ const msgs = getMessages?.()
+ const input = msgs
+ ? extractConversationText(getMessagesAfterCompactBoundary(msgs))
+ : text
+ generateAndPatch(input, bridgeSessionId)
+ }
+ // Also re-latches if v1 env-lost resets the transport's done flag past 3.
+ return userMessageCount >= 3
+ }
+
+ const initialHistoryCap = getFeatureValue_CACHED_WITH_REFRESH(
+ 'tengu_bridge_initial_history_cap',
+ 200,
+ 5 * 60 * 1000,
+ )
+
+ // Fetch orgUUID before the v1/v2 branch โ both paths need it. v1 for
+ // environment registration; v2 for archive (which lives at the compat
+ // /v1/sessions/{id}/archive, not /v1/code/sessions). Without it, v2
+ // archive 404s and sessions stay alive in CCR after /exit.
+ const orgUUID = await getOrganizationUUID()
+ if (!orgUUID) {
+ logBridgeSkip('no_org_uuid', '[bridge:repl] Skipping: no org UUID')
+ onStateChange?.('failed', '/login')
+ return null
+ }
+
+ // โโ GrowthBook gate: env-less bridge โโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโ
+ // When enabled, skips the Environments API layer entirely (no register/
+ // poll/ack/heartbeat) and connects directly via POST /bridge โ worker_jwt.
+ // See server PR #292605 (renamed in #293280). REPL-only โ daemon/print stay
+ // on env-based.
+ //
+ // NAMING: "env-less" is distinct from "CCR v2" (the /worker/* transport).
+ // The env-based path below can ALSO use CCR v2 via CLAUDE_CODE_USE_CCR_V2.
+ // tengu_bridge_repl_v2 gates env-less (no poll loop), not transport version.
+ //
+ // perpetual (assistant-mode session continuity via bridge-pointer.json) is
+ // env-coupled and not yet implemented here โ fall back to env-based when set
+ // so KAIROS users don't silently lose cross-restart continuity.
+ if (isEnvLessBridgeEnabled() && !perpetual) {
+ const versionError = await checkEnvLessBridgeMinVersion()
+ if (versionError) {
+ logBridgeSkip(
+ 'version_too_old',
+ `[bridge:repl] Skipping: ${versionError}`,
+ true,
+ )
+ onStateChange?.('failed', 'run `claude update` to upgrade')
+ return null
+ }
+ logForDebugging(
+ '[bridge:repl] Using env-less bridge path (tengu_bridge_repl_v2)',
+ )
+ const { initEnvLessBridgeCore } = await import('./remoteBridgeCore.js')
+ return initEnvLessBridgeCore({
+ baseUrl,
+ orgUUID,
+ title,
+ getAccessToken: getBridgeAccessToken,
+ onAuth401: handleOAuth401Error,
+ toSDKMessages,
+ initialHistoryCap,
+ initialMessages,
+ // v2 always creates a fresh server session (new cse_* id), so
+ // previouslyFlushedUUIDs is not passed โ there's no cross-session
+ // UUID collision risk, and the ref persists across enableโdisableโ
+ // re-enable cycles which would cause the new session to receive zero
+ // history (all UUIDs already in the set from the prior enable).
+ // v1 handles this by calling previouslyFlushedUUIDs.clear() on fresh
+ // session creation (replBridge.ts:768); v2 skips the param entirely.
+ onInboundMessage,
+ onUserMessage,
+ onPermissionResponse,
+ onInterrupt,
+ onSetModel,
+ onSetMaxThinkingTokens,
+ onSetPermissionMode,
+ onStateChange,
+ outboundOnly,
+ tags,
+ })
+ }
+
+ // โโ v1 path: env-based (register/poll/ack/heartbeat) โโโโโโโโโโโโโโโโโโ
+
+ const versionError = checkBridgeMinVersion()
+ if (versionError) {
+ logBridgeSkip('version_too_old', `[bridge:repl] Skipping: ${versionError}`)
+ onStateChange?.('failed', 'run `claude update` to upgrade')
+ return null
+ }
+
+ // Gather git context โ this is the bootstrap-read boundary.
+ // Everything from here down is passed explicitly to bridgeCore.
+ const branch = await getBranch()
+ const gitRepoUrl = await getRemoteUrl()
+ const sessionIngressUrl =
+ process.env.USER_TYPE === 'ant' &&
+ process.env.CLAUDE_BRIDGE_SESSION_INGRESS_URL
+ ? process.env.CLAUDE_BRIDGE_SESSION_INGRESS_URL
+ : baseUrl
+
+ // Assistant-mode sessions advertise a distinct worker_type so the web UI
+ // can filter them into a dedicated picker. KAIROS guard keeps the
+ // assistant module out of external builds entirely.
+ let workerType: BridgeWorkerType = 'claude_code'
+ if (feature('KAIROS')) {
+ /* eslint-disable @typescript-eslint/no-require-imports */
+ const { isAssistantMode } =
+ require('../assistant/index.js') as typeof import('../assistant/index.js')
+ /* eslint-enable @typescript-eslint/no-require-imports */
+ if (isAssistantMode()) {
+ workerType = 'claude_code_assistant'
+ }
+ }
+
+ // 6. Delegate. BridgeCoreHandle is a structural superset of
+ // ReplBridgeHandle (adds writeSdkMessages which REPL callers don't use),
+ // so no adapter needed โ just the narrower type on the way out.
+ return initBridgeCore({
+ dir: getOriginalCwd(),
+ machineName: hostname(),
+ branch,
+ gitRepoUrl,
+ title,
+ baseUrl,
+ sessionIngressUrl,
+ workerType,
+ getAccessToken: getBridgeAccessToken,
+ createSession: opts =>
+ createBridgeSession({
+ ...opts,
+ events: [],
+ baseUrl,
+ getAccessToken: getBridgeAccessToken,
+ }),
+ archiveSession: sessionId =>
+ archiveBridgeSession(sessionId, {
+ baseUrl,
+ getAccessToken: getBridgeAccessToken,
+ // gracefulShutdown.ts:407 races runCleanupFunctions against 2s.
+ // Teardown also does stopWork (parallel) + deregister (sequential),
+ // so archive can't have the full budget. 1.5s matches v2's
+ // teardown_archive_timeout_ms default.
+ timeoutMs: 1500,
+ }).catch((err: unknown) => {
+ // archiveBridgeSession has no try/catch โ 5xx/timeout/network throw
+ // straight through. Previously swallowed silently, making archive
+ // failures BQ-invisible and undiagnosable from debug logs.
+ logForDebugging(
+ `[bridge:repl] archiveBridgeSession threw: ${errorMessage(err)}`,
+ { level: 'error' },
+ )
+ }),
+ // getCurrentTitle is read on reconnect-after-env-lost to re-title the new
+ // session. /rename writes to session storage; onUserMessage mutates
+ // `title` directly โ both paths are picked up here.
+ getCurrentTitle: () => getCurrentSessionTitle(getSessionId()) ?? title,
+ onUserMessage,
+ toSDKMessages,
+ onAuth401: handleOAuth401Error,
+ getPollIntervalConfig,
+ initialHistoryCap,
+ initialMessages,
+ previouslyFlushedUUIDs,
+ onInboundMessage,
+ onPermissionResponse,
+ onInterrupt,
+ onSetModel,
+ onSetMaxThinkingTokens,
+ onSetPermissionMode,
+ onStateChange,
+ perpetual,
+ })
+}
+
+const TITLE_MAX_LEN = 50
+
+/**
+ * Quick placeholder title: strip display tags, take the first sentence,
+ * collapse whitespace, truncate to 50 chars. Returns undefined if the result
+ * is empty (e.g. message was only ). Replaced by
+ * generateSessionTitle once Haiku resolves (~1-15s).
+ */
+function deriveTitle(raw: string): string | undefined {
+ // Strip , , etc. โ these appear in
+ // user messages when IDE/hooks inject context. stripDisplayTagsAllowEmpty
+ // returns '' (not the original) so pure-tag messages are skipped.
+ const clean = stripDisplayTagsAllowEmpty(raw)
+ // First sentence is usually the intent; rest is often context/detail.
+ // Capture group instead of lookbehind โ keeps YARR JIT happy.
+ const firstSentence = /^(.*?[.!?])\s/.exec(clean)?.[1] ?? clean
+ // Collapse newlines/tabs โ titles are single-line in the claude.ai list.
+ const flat = firstSentence.replace(/\s+/g, ' ').trim()
+ if (!flat) return undefined
+ return flat.length > TITLE_MAX_LEN
+ ? flat.slice(0, TITLE_MAX_LEN - 1) + '\u2026'
+ : flat
+}
diff --git a/original-source-code/src/bridge/jwtUtils.ts b/original-source-code/src/bridge/jwtUtils.ts
new file mode 100644
index 0000000..030c001
--- /dev/null
+++ b/original-source-code/src/bridge/jwtUtils.ts
@@ -0,0 +1,256 @@
+import { logEvent } from '../services/analytics/index.js'
+import { logForDebugging } from '../utils/debug.js'
+import { logForDiagnosticsNoPII } from '../utils/diagLogs.js'
+import { errorMessage } from '../utils/errors.js'
+import { jsonParse } from '../utils/slowOperations.js'
+
+/** Format a millisecond duration as a human-readable string (e.g. "5m 30s"). */
+function formatDuration(ms: number): string {
+ if (ms < 60_000) return `${Math.round(ms / 1000)}s`
+ const m = Math.floor(ms / 60_000)
+ const s = Math.round((ms % 60_000) / 1000)
+ return s > 0 ? `${m}m ${s}s` : `${m}m`
+}
+
+/**
+ * Decode a JWT's payload segment without verifying the signature.
+ * Strips the `sk-ant-si-` session-ingress prefix if present.
+ * Returns the parsed JSON payload as `unknown`, or `null` if the
+ * token is malformed or the payload is not valid JSON.
+ */
+export function decodeJwtPayload(token: string): unknown | null {
+ const jwt = token.startsWith('sk-ant-si-')
+ ? token.slice('sk-ant-si-'.length)
+ : token
+ const parts = jwt.split('.')
+ if (parts.length !== 3 || !parts[1]) return null
+ try {
+ return jsonParse(Buffer.from(parts[1], 'base64url').toString('utf8'))
+ } catch {
+ return null
+ }
+}
+
+/**
+ * Decode the `exp` (expiry) claim from a JWT without verifying the signature.
+ * @returns The `exp` value in Unix seconds, or `null` if unparseable
+ */
+export function decodeJwtExpiry(token: string): number | null {
+ const payload = decodeJwtPayload(token)
+ if (
+ payload !== null &&
+ typeof payload === 'object' &&
+ 'exp' in payload &&
+ typeof payload.exp === 'number'
+ ) {
+ return payload.exp
+ }
+ return null
+}
+
+/** Refresh buffer: request a new token before expiry. */
+const TOKEN_REFRESH_BUFFER_MS = 5 * 60 * 1000
+
+/** Fallback refresh interval when the new token's expiry is unknown. */
+const FALLBACK_REFRESH_INTERVAL_MS = 30 * 60 * 1000 // 30 minutes
+
+/** Max consecutive failures before giving up on the refresh chain. */
+const MAX_REFRESH_FAILURES = 3
+
+/** Retry delay when getAccessToken returns undefined. */
+const REFRESH_RETRY_DELAY_MS = 60_000
+
+/**
+ * Creates a token refresh scheduler that proactively refreshes session tokens
+ * before they expire. Used by both the standalone bridge and the REPL bridge.
+ *
+ * When a token is about to expire, the scheduler calls `onRefresh` with the
+ * session ID and the bridge's OAuth access token. The caller is responsible
+ * for delivering the token to the appropriate transport (child process stdin
+ * for standalone bridge, WebSocket reconnect for REPL bridge).
+ */
+export function createTokenRefreshScheduler({
+ getAccessToken,
+ onRefresh,
+ label,
+ refreshBufferMs = TOKEN_REFRESH_BUFFER_MS,
+}: {
+ getAccessToken: () => string | undefined | Promise
+ onRefresh: (sessionId: string, oauthToken: string) => void
+ label: string
+ /** How long before expiry to fire refresh. Defaults to 5 min. */
+ refreshBufferMs?: number
+}): {
+ schedule: (sessionId: string, token: string) => void
+ scheduleFromExpiresIn: (sessionId: string, expiresInSeconds: number) => void
+ cancel: (sessionId: string) => void
+ cancelAll: () => void
+} {
+ const timers = new Map>()
+ const failureCounts = new Map()
+ // Generation counter per session โ incremented by schedule() and cancel()
+ // so that in-flight async doRefresh() calls can detect when they've been
+ // superseded and should skip setting follow-up timers.
+ const generations = new Map()
+
+ function nextGeneration(sessionId: string): number {
+ const gen = (generations.get(sessionId) ?? 0) + 1
+ generations.set(sessionId, gen)
+ return gen
+ }
+
+ function schedule(sessionId: string, token: string): void {
+ const expiry = decodeJwtExpiry(token)
+ if (!expiry) {
+ // Token is not a decodable JWT (e.g. an OAuth token passed from the
+ // REPL bridge WebSocket open handler). Preserve any existing timer
+ // (such as the follow-up refresh set by doRefresh) so the refresh
+ // chain is not broken.
+ logForDebugging(
+ `[${label}:token] Could not decode JWT expiry for sessionId=${sessionId}, token prefix=${token.slice(0, 15)}โฆ, keeping existing timer`,
+ )
+ return
+ }
+
+ // Clear any existing refresh timer โ we have a concrete expiry to replace it.
+ const existing = timers.get(sessionId)
+ if (existing) {
+ clearTimeout(existing)
+ }
+
+ // Bump generation to invalidate any in-flight async doRefresh.
+ const gen = nextGeneration(sessionId)
+
+ const expiryDate = new Date(expiry * 1000).toISOString()
+ const delayMs = expiry * 1000 - Date.now() - refreshBufferMs
+ if (delayMs <= 0) {
+ logForDebugging(
+ `[${label}:token] Token for sessionId=${sessionId} expires=${expiryDate} (past or within buffer), refreshing immediately`,
+ )
+ void doRefresh(sessionId, gen)
+ return
+ }
+
+ logForDebugging(
+ `[${label}:token] Scheduled token refresh for sessionId=${sessionId} in ${formatDuration(delayMs)} (expires=${expiryDate}, buffer=${refreshBufferMs / 1000}s)`,
+ )
+
+ const timer = setTimeout(doRefresh, delayMs, sessionId, gen)
+ timers.set(sessionId, timer)
+ }
+
+ /**
+ * Schedule refresh using an explicit TTL (seconds until expiry) rather
+ * than decoding a JWT's exp claim. Used by callers whose JWT is opaque
+ * (e.g. POST /v1/code/sessions/{id}/bridge returns expires_in directly).
+ */
+ function scheduleFromExpiresIn(
+ sessionId: string,
+ expiresInSeconds: number,
+ ): void {
+ const existing = timers.get(sessionId)
+ if (existing) clearTimeout(existing)
+ const gen = nextGeneration(sessionId)
+ // Clamp to 30s floor โ if refreshBufferMs exceeds the server's expires_in
+ // (e.g. very large buffer for frequent-refresh testing, or server shortens
+ // expires_in unexpectedly), unclamped delayMs โค 0 would tight-loop.
+ const delayMs = Math.max(expiresInSeconds * 1000 - refreshBufferMs, 30_000)
+ logForDebugging(
+ `[${label}:token] Scheduled token refresh for sessionId=${sessionId} in ${formatDuration(delayMs)} (expires_in=${expiresInSeconds}s, buffer=${refreshBufferMs / 1000}s)`,
+ )
+ const timer = setTimeout(doRefresh, delayMs, sessionId, gen)
+ timers.set(sessionId, timer)
+ }
+
+ async function doRefresh(sessionId: string, gen: number): Promise {
+ let oauthToken: string | undefined
+ try {
+ oauthToken = await getAccessToken()
+ } catch (err) {
+ logForDebugging(
+ `[${label}:token] getAccessToken threw for sessionId=${sessionId}: ${errorMessage(err)}`,
+ { level: 'error' },
+ )
+ }
+
+ // If the session was cancelled or rescheduled while we were awaiting,
+ // the generation will have changed โ bail out to avoid orphaned timers.
+ if (generations.get(sessionId) !== gen) {
+ logForDebugging(
+ `[${label}:token] doRefresh for sessionId=${sessionId} stale (gen ${gen} vs ${generations.get(sessionId)}), skipping`,
+ )
+ return
+ }
+
+ if (!oauthToken) {
+ const failures = (failureCounts.get(sessionId) ?? 0) + 1
+ failureCounts.set(sessionId, failures)
+ logForDebugging(
+ `[${label}:token] No OAuth token available for refresh, sessionId=${sessionId} (failure ${failures}/${MAX_REFRESH_FAILURES})`,
+ { level: 'error' },
+ )
+ logForDiagnosticsNoPII('error', 'bridge_token_refresh_no_oauth')
+ // Schedule a retry so the refresh chain can recover if the token
+ // becomes available again (e.g. transient cache clear during refresh).
+ // Cap retries to avoid spamming on genuine failures.
+ if (failures < MAX_REFRESH_FAILURES) {
+ const retryTimer = setTimeout(
+ doRefresh,
+ REFRESH_RETRY_DELAY_MS,
+ sessionId,
+ gen,
+ )
+ timers.set(sessionId, retryTimer)
+ }
+ return
+ }
+
+ // Reset failure counter on successful token retrieval
+ failureCounts.delete(sessionId)
+
+ logForDebugging(
+ `[${label}:token] Refreshing token for sessionId=${sessionId}: new token prefix=${oauthToken.slice(0, 15)}โฆ`,
+ )
+ logEvent('tengu_bridge_token_refreshed', {})
+ onRefresh(sessionId, oauthToken)
+
+ // Schedule a follow-up refresh so long-running sessions stay authenticated.
+ // Without this, the initial one-shot timer leaves the session vulnerable
+ // to token expiry if it runs past the first refresh window.
+ const timer = setTimeout(
+ doRefresh,
+ FALLBACK_REFRESH_INTERVAL_MS,
+ sessionId,
+ gen,
+ )
+ timers.set(sessionId, timer)
+ logForDebugging(
+ `[${label}:token] Scheduled follow-up refresh for sessionId=${sessionId} in ${formatDuration(FALLBACK_REFRESH_INTERVAL_MS)}`,
+ )
+ }
+
+ function cancel(sessionId: string): void {
+ // Bump generation to invalidate any in-flight async doRefresh.
+ nextGeneration(sessionId)
+ const timer = timers.get(sessionId)
+ if (timer) {
+ clearTimeout(timer)
+ timers.delete(sessionId)
+ }
+ failureCounts.delete(sessionId)
+ }
+
+ function cancelAll(): void {
+ // Bump all generations so in-flight doRefresh calls are invalidated.
+ for (const sessionId of generations.keys()) {
+ nextGeneration(sessionId)
+ }
+ for (const timer of timers.values()) {
+ clearTimeout(timer)
+ }
+ timers.clear()
+ failureCounts.clear()
+ }
+
+ return { schedule, scheduleFromExpiresIn, cancel, cancelAll }
+}
diff --git a/original-source-code/src/bridge/pollConfig.ts b/original-source-code/src/bridge/pollConfig.ts
new file mode 100644
index 0000000..024b476
--- /dev/null
+++ b/original-source-code/src/bridge/pollConfig.ts
@@ -0,0 +1,110 @@
+import { z } from 'zod/v4'
+import { getFeatureValue_CACHED_WITH_REFRESH } from '../services/analytics/growthbook.js'
+import { lazySchema } from '../utils/lazySchema.js'
+import {
+ DEFAULT_POLL_CONFIG,
+ type PollIntervalConfig,
+} from './pollConfigDefaults.js'
+
+// .min(100) on the seek-work intervals restores the old Math.max(..., 100)
+// defense-in-depth floor against fat-fingered GrowthBook values. Unlike a
+// clamp, Zod rejects the whole object on violation โ a config with one bad
+// field falls back to DEFAULT_POLL_CONFIG entirely rather than being
+// partially trusted.
+//
+// The at_capacity intervals use a 0-or-โฅ100 refinement: 0 means "disabled"
+// (heartbeat-only mode), โฅ100 is the fat-finger floor. Values 1โ99 are
+// rejected so unit confusion (ops thinks seconds, enters 10) doesn't poll
+// every 10ms against the VerifyEnvironmentSecretAuth DB path.
+//
+// The object-level refines require at least one at-capacity liveness
+// mechanism enabled: heartbeat OR the relevant poll interval. Without this,
+// the hb=0, atCapMs=0 drift config (ops disables heartbeat without
+// restoring at_capacity) falls through every throttle site with no sleep โ
+// tight-looping /poll at HTTP-round-trip speed.
+const zeroOrAtLeast100 = {
+ message: 'must be 0 (disabled) or โฅ100ms',
+}
+const pollIntervalConfigSchema = lazySchema(() =>
+ z
+ .object({
+ poll_interval_ms_not_at_capacity: z.number().int().min(100),
+ // 0 = no at-capacity polling. Independent of heartbeat โ both can be
+ // enabled (heartbeat runs, periodically breaks out to poll).
+ poll_interval_ms_at_capacity: z
+ .number()
+ .int()
+ .refine(v => v === 0 || v >= 100, zeroOrAtLeast100),
+ // 0 = disabled; positive value = heartbeat at this interval while at
+ // capacity. Runs alongside at-capacity polling, not instead of it.
+ // Named non_exclusive to distinguish from the old heartbeat_interval_ms
+ // (either-or semantics in pre-#22145 clients). .default(0) so existing
+ // GrowthBook configs without this field parse successfully.
+ non_exclusive_heartbeat_interval_ms: z.number().int().min(0).default(0),
+ // Multisession (bridgeMain.ts) intervals. Defaults match the
+ // single-session values so existing configs without these fields
+ // preserve current behavior.
+ multisession_poll_interval_ms_not_at_capacity: z
+ .number()
+ .int()
+ .min(100)
+ .default(
+ DEFAULT_POLL_CONFIG.multisession_poll_interval_ms_not_at_capacity,
+ ),
+ multisession_poll_interval_ms_partial_capacity: z
+ .number()
+ .int()
+ .min(100)
+ .default(
+ DEFAULT_POLL_CONFIG.multisession_poll_interval_ms_partial_capacity,
+ ),
+ multisession_poll_interval_ms_at_capacity: z
+ .number()
+ .int()
+ .refine(v => v === 0 || v >= 100, zeroOrAtLeast100)
+ .default(DEFAULT_POLL_CONFIG.multisession_poll_interval_ms_at_capacity),
+ // .min(1) matches the server's ge=1 constraint (work_v1.py:230).
+ reclaim_older_than_ms: z.number().int().min(1).default(5000),
+ session_keepalive_interval_v2_ms: z
+ .number()
+ .int()
+ .min(0)
+ .default(120_000),
+ })
+ .refine(
+ cfg =>
+ cfg.non_exclusive_heartbeat_interval_ms > 0 ||
+ cfg.poll_interval_ms_at_capacity > 0,
+ {
+ message:
+ 'at-capacity liveness requires non_exclusive_heartbeat_interval_ms > 0 or poll_interval_ms_at_capacity > 0',
+ },
+ )
+ .refine(
+ cfg =>
+ cfg.non_exclusive_heartbeat_interval_ms > 0 ||
+ cfg.multisession_poll_interval_ms_at_capacity > 0,
+ {
+ message:
+ 'at-capacity liveness requires non_exclusive_heartbeat_interval_ms > 0 or multisession_poll_interval_ms_at_capacity > 0',
+ },
+ ),
+)
+
+/**
+ * Fetch the bridge poll interval config from GrowthBook with a 5-minute
+ * refresh window. Validates the served JSON against the schema; falls back
+ * to defaults if the flag is absent, malformed, or partially-specified.
+ *
+ * Shared by bridgeMain.ts (standalone) and replBridge.ts (REPL) so ops
+ * can tune both poll rates fleet-wide with a single config push.
+ */
+export function getPollIntervalConfig(): PollIntervalConfig {
+ const raw = getFeatureValue_CACHED_WITH_REFRESH(
+ 'tengu_bridge_poll_interval_config',
+ DEFAULT_POLL_CONFIG,
+ 5 * 60 * 1000,
+ )
+ const parsed = pollIntervalConfigSchema().safeParse(raw)
+ return parsed.success ? parsed.data : DEFAULT_POLL_CONFIG
+}
diff --git a/original-source-code/src/bridge/pollConfigDefaults.ts b/original-source-code/src/bridge/pollConfigDefaults.ts
new file mode 100644
index 0000000..7a4e6d8
--- /dev/null
+++ b/original-source-code/src/bridge/pollConfigDefaults.ts
@@ -0,0 +1,82 @@
+/**
+ * Bridge poll interval defaults. Extracted from pollConfig.ts so callers
+ * that don't need live GrowthBook tuning (daemon via Agent SDK) can avoid
+ * the growthbook.ts โ config.ts โ file.ts โ sessionStorage.ts โ commands.ts
+ * transitive dependency chain.
+ */
+
+/**
+ * Poll interval when actively seeking work (no transport / below maxSessions).
+ * Governs user-visible "connectingโฆ" latency on initial work pickup and
+ * recovery speed after the server re-dispatches a work item.
+ */
+const POLL_INTERVAL_MS_NOT_AT_CAPACITY = 2000
+
+/**
+ * Poll interval when the transport is connected. Runs independently of
+ * heartbeat โ when both are enabled, the heartbeat loop breaks out to poll
+ * at this interval. Set to 0 to disable at-capacity polling entirely.
+ *
+ * Server-side constraints that bound this value:
+ * - BRIDGE_LAST_POLL_TTL = 4h (Redis key expiry โ environment auto-archived)
+ * - max_poll_stale_seconds = 24h (session-creation health gate, currently disabled)
+ *
+ * 10 minutes gives 24ร headroom on the Redis TTL while still picking up
+ * server-initiated token-rotation redispatches within one poll cycle.
+ * The transport auto-reconnects internally for 10 minutes on transient WS
+ * failures, so poll is not the recovery path โ it's strictly a liveness
+ * signal plus a backstop for permanent close.
+ */
+const POLL_INTERVAL_MS_AT_CAPACITY = 600_000
+
+/**
+ * Multisession bridge (bridgeMain.ts) poll intervals. Defaults match the
+ * single-session values so existing GrowthBook configs without these fields
+ * preserve current behavior. Ops can tune these independently via the
+ * tengu_bridge_poll_interval_config GB flag.
+ */
+const MULTISESSION_POLL_INTERVAL_MS_NOT_AT_CAPACITY =
+ POLL_INTERVAL_MS_NOT_AT_CAPACITY
+const MULTISESSION_POLL_INTERVAL_MS_PARTIAL_CAPACITY =
+ POLL_INTERVAL_MS_NOT_AT_CAPACITY
+const MULTISESSION_POLL_INTERVAL_MS_AT_CAPACITY = POLL_INTERVAL_MS_AT_CAPACITY
+
+export type PollIntervalConfig = {
+ poll_interval_ms_not_at_capacity: number
+ poll_interval_ms_at_capacity: number
+ non_exclusive_heartbeat_interval_ms: number
+ multisession_poll_interval_ms_not_at_capacity: number
+ multisession_poll_interval_ms_partial_capacity: number
+ multisession_poll_interval_ms_at_capacity: number
+ reclaim_older_than_ms: number
+ session_keepalive_interval_v2_ms: number
+}
+
+export const DEFAULT_POLL_CONFIG: PollIntervalConfig = {
+ poll_interval_ms_not_at_capacity: POLL_INTERVAL_MS_NOT_AT_CAPACITY,
+ poll_interval_ms_at_capacity: POLL_INTERVAL_MS_AT_CAPACITY,
+ // 0 = disabled. When > 0, at-capacity loops send per-work-item heartbeats
+ // at this interval. Independent of poll_interval_ms_at_capacity โ both may
+ // run (heartbeat periodically yields to poll). 60s gives 5ร headroom under
+ // the server's 300s heartbeat TTL. Named non_exclusive to distinguish from
+ // the old heartbeat_interval_ms field (either-or semantics in pre-#22145
+ // clients โ heartbeat suppressed poll). Old clients ignore this key; ops
+ // can set both fields during rollout.
+ non_exclusive_heartbeat_interval_ms: 0,
+ multisession_poll_interval_ms_not_at_capacity:
+ MULTISESSION_POLL_INTERVAL_MS_NOT_AT_CAPACITY,
+ multisession_poll_interval_ms_partial_capacity:
+ MULTISESSION_POLL_INTERVAL_MS_PARTIAL_CAPACITY,
+ multisession_poll_interval_ms_at_capacity:
+ MULTISESSION_POLL_INTERVAL_MS_AT_CAPACITY,
+ // Poll query param: reclaim unacknowledged work items older than this.
+ // Matches the server's DEFAULT_RECLAIM_OLDER_THAN_MS (work_service.py:24).
+ // Enables picking up stale-pending work after JWT expiry, when the prior
+ // ack failed because the session_ingress_token was already stale.
+ reclaim_older_than_ms: 5000,
+ // 0 = disabled. When > 0, push a silent {type:'keep_alive'} frame to
+ // session-ingress at this interval so upstream proxies don't GC an idle
+ // remote-control session. 2 min is the default. _v2: bridge-only gate
+ // (pre-v2 clients read the old key, new clients ignore it).
+ session_keepalive_interval_v2_ms: 120_000,
+}
diff --git a/original-source-code/src/bridge/remoteBridgeCore.ts b/original-source-code/src/bridge/remoteBridgeCore.ts
new file mode 100644
index 0000000..76545f6
--- /dev/null
+++ b/original-source-code/src/bridge/remoteBridgeCore.ts
@@ -0,0 +1,1008 @@
+// biome-ignore-all assist/source/organizeImports: ANT-ONLY import markers must not be reordered
+/**
+ * Env-less Remote Control bridge core.
+ *
+ * "Env-less" = no Environments API layer. Distinct from "CCR v2" (the
+ * /worker/* transport protocol) โ the env-based path (replBridge.ts) can also
+ * use CCR v2 transport via CLAUDE_CODE_USE_CCR_V2. This file is about removing
+ * the poll/dispatch layer, not about which transport protocol is underneath.
+ *
+ * Unlike initBridgeCore (env-based, ~2400 lines), this connects directly
+ * to the session-ingress layer without the Environments API work-dispatch
+ * layer:
+ *
+ * 1. POST /v1/code/sessions (OAuth, no env_id) โ session.id
+ * 2. POST /v1/code/sessions/{id}/bridge (OAuth) โ {worker_jwt, expires_in, api_base_url, worker_epoch}
+ * Each /bridge call bumps epoch โ it IS the register. No separate /worker/register.
+ * 3. createV2ReplTransport(worker_jwt, worker_epoch) โ SSE + CCRClient
+ * 4. createTokenRefreshScheduler โ proactive /bridge re-call (new JWT + new epoch)
+ * 5. 401 on SSE โ rebuild transport with fresh /bridge credentials (same seq-num)
+ *
+ * No register/poll/ack/stop/heartbeat/deregister environment lifecycle.
+ * The Environments API historically existed because CCR's /worker/*
+ * endpoints required a session_id+role=worker JWT that only the work-dispatch
+ * layer could mint. Server PR #292605 (renamed in #293280) adds the /bridge endpoint as a direct
+ * OAuthโworker_jwt exchange, making the env layer optional for REPL sessions.
+ *
+ * Gated by `tengu_bridge_repl_v2` GrowthBook flag in initReplBridge.ts.
+ * REPL-only โ daemon/print stay on env-based.
+ */
+
+import { feature } from 'bun:bundle'
+import axios from 'axios'
+import {
+ createV2ReplTransport,
+ type ReplBridgeTransport,
+} from './replBridgeTransport.js'
+import { buildCCRv2SdkUrl } from './workSecret.js'
+import { toCompatSessionId } from './sessionIdCompat.js'
+import { FlushGate } from './flushGate.js'
+import { createTokenRefreshScheduler } from './jwtUtils.js'
+import { getTrustedDeviceToken } from './trustedDevice.js'
+import {
+ getEnvLessBridgeConfig,
+ type EnvLessBridgeConfig,
+} from './envLessBridgeConfig.js'
+import {
+ handleIngressMessage,
+ handleServerControlRequest,
+ makeResultMessage,
+ isEligibleBridgeMessage,
+ extractTitleText,
+ BoundedUUIDSet,
+} from './bridgeMessaging.js'
+import { logBridgeSkip } from './debugUtils.js'
+import { logForDebugging } from '../utils/debug.js'
+import { logForDiagnosticsNoPII } from '../utils/diagLogs.js'
+import { isInProtectedNamespace } from '../utils/envUtils.js'
+import { errorMessage } from '../utils/errors.js'
+import { sleep } from '../utils/sleep.js'
+import { registerCleanup } from '../utils/cleanupRegistry.js'
+import {
+ type AnalyticsMetadata_I_VERIFIED_THIS_IS_NOT_CODE_OR_FILEPATHS,
+ logEvent,
+} from '../services/analytics/index.js'
+import type { ReplBridgeHandle, BridgeState } from './replBridge.js'
+import type { Message } from '../types/message.js'
+import type { SDKMessage } from '../entrypoints/agentSdkTypes.js'
+import type {
+ SDKControlRequest,
+ SDKControlResponse,
+} from '../entrypoints/sdk/controlTypes.js'
+import type { PermissionMode } from '../utils/permissions/PermissionMode.js'
+
+const ANTHROPIC_VERSION = '2023-06-01'
+
+// Telemetry discriminator for ws_connected. 'initial' is the default and
+// never passed to rebuildTransport (which can only be called post-init);
+// Exclude<> makes that constraint explicit at both signatures.
+type ConnectCause = 'initial' | 'proactive_refresh' | 'auth_401_recovery'
+
+function oauthHeaders(accessToken: string): Record {
+ return {
+ Authorization: `Bearer ${accessToken}`,
+ 'Content-Type': 'application/json',
+ 'anthropic-version': ANTHROPIC_VERSION,
+ }
+}
+
+export type EnvLessBridgeParams = {
+ baseUrl: string
+ orgUUID: string
+ title: string
+ getAccessToken: () => string | undefined
+ onAuth401?: (staleAccessToken: string) => Promise
+ /**
+ * Converts internal Message[] โ SDKMessage[] for writeMessages() and the
+ * initial-flush/drain paths. Injected rather than imported โ mappers.ts
+ * transitively pulls in src/commands.ts (entire command registry + React
+ * tree) which would bloat bundles that don't already have it.
+ */
+ toSDKMessages: (messages: Message[]) => SDKMessage[]
+ initialHistoryCap: number
+ initialMessages?: Message[]
+ onInboundMessage?: (msg: SDKMessage) => void | Promise
+ /**
+ * Fired on each title-worthy user message seen in writeMessages() until
+ * the callback returns true (done). Mirrors replBridge.ts's onUserMessage โ
+ * caller derives a title and PATCHes /v1/sessions/{id} so auto-started
+ * sessions don't stay at the generic fallback. The caller owns the
+ * derive-at-count-1-and-3 policy; the transport just keeps calling until
+ * told to stop. sessionId is the raw cse_* โ updateBridgeSessionTitle
+ * retags internally.
+ */
+ onUserMessage?: (text: string, sessionId: string) => boolean
+ onPermissionResponse?: (response: SDKControlResponse) => void
+ onInterrupt?: () => void
+ onSetModel?: (model: string | undefined) => void
+ onSetMaxThinkingTokens?: (maxTokens: number | null) => void
+ onSetPermissionMode?: (
+ mode: PermissionMode,
+ ) => { ok: true } | { ok: false; error: string }
+ onStateChange?: (state: BridgeState, detail?: string) => void
+ /**
+ * When true, skip opening the SSE read stream โ only the CCRClient write
+ * path is activated. Threaded to createV2ReplTransport and
+ * handleServerControlRequest.
+ */
+ outboundOnly?: boolean
+ /** Free-form tags for session categorization (e.g. ['ccr-mirror']). */
+ tags?: string[]
+}
+
+/**
+ * Create a session, fetch a worker JWT, connect the v2 transport.
+ *
+ * Returns null on any pre-flight failure (session create failed, /bridge
+ * failed, transport setup failed). Caller (initReplBridge) surfaces this
+ * as a generic "initialization failed" state.
+ */
+export async function initEnvLessBridgeCore(
+ params: EnvLessBridgeParams,
+): Promise {
+ const {
+ baseUrl,
+ orgUUID,
+ title,
+ getAccessToken,
+ onAuth401,
+ toSDKMessages,
+ initialHistoryCap,
+ initialMessages,
+ onInboundMessage,
+ onUserMessage,
+ onPermissionResponse,
+ onInterrupt,
+ onSetModel,
+ onSetMaxThinkingTokens,
+ onSetPermissionMode,
+ onStateChange,
+ outboundOnly,
+ tags,
+ } = params
+
+ const cfg = await getEnvLessBridgeConfig()
+
+ // โโ 1. Create session (POST /v1/code/sessions, no env_id) โโโโโโโโโโโโโโโ
+ const accessToken = getAccessToken()
+ if (!accessToken) {
+ logForDebugging('[remote-bridge] No OAuth token')
+ return null
+ }
+
+ const createdSessionId = await withRetry(
+ () =>
+ createCodeSession(baseUrl, accessToken, title, cfg.http_timeout_ms, tags),
+ 'createCodeSession',
+ cfg,
+ )
+ if (!createdSessionId) {
+ onStateChange?.('failed', 'Session creation failed โ see debug log')
+ logBridgeSkip('v2_session_create_failed', undefined, true)
+ return null
+ }
+ const sessionId: string = createdSessionId
+ logForDebugging(`[remote-bridge] Created session ${sessionId}`)
+ logForDiagnosticsNoPII('info', 'bridge_repl_v2_session_created')
+
+ // โโ 2. Fetch bridge credentials (POST /bridge โ worker_jwt, expires_in, api_base_url) โโ
+ const credentials = await withRetry(
+ () =>
+ fetchRemoteCredentials(
+ sessionId,
+ baseUrl,
+ accessToken,
+ cfg.http_timeout_ms,
+ ),
+ 'fetchRemoteCredentials',
+ cfg,
+ )
+ if (!credentials) {
+ onStateChange?.('failed', 'Remote credentials fetch failed โ see debug log')
+ logBridgeSkip('v2_remote_creds_failed', undefined, true)
+ void archiveSession(
+ sessionId,
+ baseUrl,
+ accessToken,
+ orgUUID,
+ cfg.http_timeout_ms,
+ )
+ return null
+ }
+ logForDebugging(
+ `[remote-bridge] Fetched bridge credentials (expires_in=${credentials.expires_in}s)`,
+ )
+
+ // โโ 3. Build v2 transport (SSETransport + CCRClient) โโโโโโโโโโโโโโโโโโโโ
+ const sessionUrl = buildCCRv2SdkUrl(credentials.api_base_url, sessionId)
+ logForDebugging(`[remote-bridge] v2 session URL: ${sessionUrl}`)
+
+ let transport: ReplBridgeTransport
+ try {
+ transport = await createV2ReplTransport({
+ sessionUrl,
+ ingressToken: credentials.worker_jwt,
+ sessionId,
+ epoch: credentials.worker_epoch,
+ heartbeatIntervalMs: cfg.heartbeat_interval_ms,
+ heartbeatJitterFraction: cfg.heartbeat_jitter_fraction,
+ // Per-instance closure โ keeps the worker JWT out of
+ // process.env.CLAUDE_CODE_SESSION_ACCESS_TOKEN, which mcp/client.ts
+ // reads ungatedly and would otherwise send to user-configured ws/http
+ // MCP servers. Frozen-at-construction is correct: transport is fully
+ // rebuilt on refresh (rebuildTransport below).
+ getAuthToken: () => credentials.worker_jwt,
+ outboundOnly,
+ })
+ } catch (err) {
+ logForDebugging(
+ `[remote-bridge] v2 transport setup failed: ${errorMessage(err)}`,
+ { level: 'error' },
+ )
+ onStateChange?.('failed', `Transport setup failed: ${errorMessage(err)}`)
+ logBridgeSkip('v2_transport_setup_failed', undefined, true)
+ void archiveSession(
+ sessionId,
+ baseUrl,
+ accessToken,
+ orgUUID,
+ cfg.http_timeout_ms,
+ )
+ return null
+ }
+ logForDebugging(
+ `[remote-bridge] v2 transport created (epoch=${credentials.worker_epoch})`,
+ )
+ onStateChange?.('ready')
+
+ // โโ 4. State โโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโ
+
+ // Echo dedup: messages we POST come back on the read stream. Seeded with
+ // initial message UUIDs so server echoes of flushed history are recognized.
+ // Both sets cover initial UUIDs โ recentPostedUUIDs is a 2000-cap ring buffer
+ // and could evict them after enough live writes; initialMessageUUIDs is the
+ // unbounded fallback. Defense-in-depth; mirrors replBridge.ts.
+ const recentPostedUUIDs = new BoundedUUIDSet(cfg.uuid_dedup_buffer_size)
+ const initialMessageUUIDs = new Set()
+ if (initialMessages) {
+ for (const msg of initialMessages) {
+ initialMessageUUIDs.add(msg.uuid)
+ recentPostedUUIDs.add(msg.uuid)
+ }
+ }
+
+ // Defensive dedup for re-delivered inbound prompts (seq-num negotiation
+ // edge cases, server history replay after transport swap).
+ const recentInboundUUIDs = new BoundedUUIDSet(cfg.uuid_dedup_buffer_size)
+
+ // FlushGate: queue live writes while the history flush POST is in flight,
+ // so the server receives [history..., live...] in order.
+ const flushGate = new FlushGate()
+
+ let initialFlushDone = false
+ let tornDown = false
+ let authRecoveryInFlight = false
+ // Latch for onUserMessage โ flips true when the callback returns true
+ // (policy says "done deriving"). sessionId is const (no re-create path โ
+ // rebuildTransport swaps JWT/epoch, same session), so no reset needed.
+ let userMessageCallbackDone = !onUserMessage
+
+ // Telemetry: why did onConnect fire? Set by rebuildTransport before
+ // wireTransportCallbacks; read asynchronously by onConnect. Race-safe
+ // because authRecoveryInFlight serializes rebuild callers, and a fresh
+ // initEnvLessBridgeCore() call gets a fresh closure defaulting to 'initial'.
+ let connectCause: ConnectCause = 'initial'
+
+ // Deadline for onConnect after transport.connect(). Cleared by onConnect
+ // (connected) and onClose (got a close โ not silent). If neither fires
+ // before cfg.connect_timeout_ms, onConnectTimeout emits โ the only
+ // signal for the `started โ (silence)` gap.
+ let connectDeadline: ReturnType | undefined
+ function onConnectTimeout(cause: ConnectCause): void {
+ if (tornDown) return
+ logEvent('tengu_bridge_repl_connect_timeout', {
+ v2: true,
+ elapsed_ms: cfg.connect_timeout_ms,
+ cause:
+ cause as AnalyticsMetadata_I_VERIFIED_THIS_IS_NOT_CODE_OR_FILEPATHS,
+ })
+ }
+
+ // โโ 5. JWT refresh scheduler โโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโ
+ // Schedule a callback 5min before expiry (per response.expires_in). On fire,
+ // re-fetch /bridge with OAuth โ rebuild transport with fresh credentials.
+ // Each /bridge call bumps epoch server-side, so a JWT-only swap would leave
+ // the old CCRClient heartbeating with a stale epoch โ 409 within 20s.
+ // JWT is opaque โ do not decode.
+ const refresh = createTokenRefreshScheduler({
+ refreshBufferMs: cfg.token_refresh_buffer_ms,
+ getAccessToken: async () => {
+ // Unconditionally refresh OAuth before calling /bridge โ getAccessToken()
+ // returns expired tokens as non-null strings (doesn't check expiresAt),
+ // so truthiness doesn't mean valid. Pass the stale token to onAuth401
+ // so handleOAuth401Error's keychain-comparison can detect parallel refresh.
+ const stale = getAccessToken()
+ if (onAuth401) await onAuth401(stale ?? '')
+ return getAccessToken() ?? stale
+ },
+ onRefresh: (sid, oauthToken) => {
+ void (async () => {
+ // Laptop wake: overdue proactive timer + SSE 401 fire ~simultaneously.
+ // Claim the flag BEFORE the /bridge fetch so the other path skips
+ // entirely โ prevents double epoch bump (each /bridge call bumps; if
+ // both fetch, the first rebuild gets a stale epoch and 409s).
+ if (authRecoveryInFlight || tornDown) {
+ logForDebugging(
+ '[remote-bridge] Recovery already in flight, skipping proactive refresh',
+ )
+ return
+ }
+ authRecoveryInFlight = true
+ try {
+ const fresh = await withRetry(
+ () =>
+ fetchRemoteCredentials(
+ sid,
+ baseUrl,
+ oauthToken,
+ cfg.http_timeout_ms,
+ ),
+ 'fetchRemoteCredentials (proactive)',
+ cfg,
+ )
+ if (!fresh || tornDown) return
+ await rebuildTransport(fresh, 'proactive_refresh')
+ logForDebugging(
+ '[remote-bridge] Transport rebuilt (proactive refresh)',
+ )
+ } catch (err) {
+ logForDebugging(
+ `[remote-bridge] Proactive refresh rebuild failed: ${errorMessage(err)}`,
+ { level: 'error' },
+ )
+ logForDiagnosticsNoPII(
+ 'error',
+ 'bridge_repl_v2_proactive_refresh_failed',
+ )
+ if (!tornDown) {
+ onStateChange?.('failed', `Refresh failed: ${errorMessage(err)}`)
+ }
+ } finally {
+ authRecoveryInFlight = false
+ }
+ })()
+ },
+ label: 'remote',
+ })
+ refresh.scheduleFromExpiresIn(sessionId, credentials.expires_in)
+
+ // โโ 6. Wire callbacks (extracted so transport-rebuild can re-wire) โโโโโโ
+ function wireTransportCallbacks(): void {
+ transport.setOnConnect(() => {
+ clearTimeout(connectDeadline)
+ logForDebugging('[remote-bridge] v2 transport connected')
+ logForDiagnosticsNoPII('info', 'bridge_repl_v2_transport_connected')
+ logEvent('tengu_bridge_repl_ws_connected', {
+ v2: true,
+ cause:
+ connectCause as AnalyticsMetadata_I_VERIFIED_THIS_IS_NOT_CODE_OR_FILEPATHS,
+ })
+
+ if (!initialFlushDone && initialMessages && initialMessages.length > 0) {
+ initialFlushDone = true
+ // Capture current transport โ if 401/teardown happens mid-flush,
+ // the stale .finally() must not drain the gate or signal connected.
+ // (Same guard pattern as replBridge.ts:1119.)
+ const flushTransport = transport
+ void flushHistory(initialMessages)
+ .catch(e =>
+ logForDebugging(`[remote-bridge] flushHistory failed: ${e}`),
+ )
+ .finally(() => {
+ // authRecoveryInFlight catches the v1-vs-v2 asymmetry: v1 nulls
+ // transport synchronously in setOnClose (replBridge.ts:1175), so
+ // transport !== flushTransport trips immediately. v2 doesn't null โ
+ // transport reassigned only at rebuildTransport:346, 3 awaits deep.
+ // authRecoveryInFlight is set synchronously at rebuildTransport entry.
+ if (
+ transport !== flushTransport ||
+ tornDown ||
+ authRecoveryInFlight
+ ) {
+ return
+ }
+ drainFlushGate()
+ onStateChange?.('connected')
+ })
+ } else if (!flushGate.active) {
+ onStateChange?.('connected')
+ }
+ })
+
+ transport.setOnData((data: string) => {
+ handleIngressMessage(
+ data,
+ recentPostedUUIDs,
+ recentInboundUUIDs,
+ onInboundMessage,
+ // Remote client answered the permission prompt โ the turn resumes.
+ // Without this the server stays on requires_action until the next
+ // user message or turn-end result.
+ onPermissionResponse
+ ? res => {
+ transport.reportState('running')
+ onPermissionResponse(res)
+ }
+ : undefined,
+ req =>
+ handleServerControlRequest(req, {
+ transport,
+ sessionId,
+ onInterrupt,
+ onSetModel,
+ onSetMaxThinkingTokens,
+ onSetPermissionMode,
+ outboundOnly,
+ }),
+ )
+ })
+
+ transport.setOnClose((code?: number) => {
+ clearTimeout(connectDeadline)
+ if (tornDown) return
+ logForDebugging(`[remote-bridge] v2 transport closed (code=${code})`)
+ logEvent('tengu_bridge_repl_ws_closed', { code, v2: true })
+ // onClose fires only for TERMINAL failures: 401 (JWT invalid),
+ // 4090 (CCR epoch mismatch), 4091 (CCR init failed), or SSE 10-min
+ // reconnect budget exhausted. Transient disconnects are handled
+ // transparently inside SSETransport. 401 we can recover from (fetch
+ // fresh JWT, rebuild transport); all other codes are dead-ends.
+ if (code === 401 && !authRecoveryInFlight) {
+ void recoverFromAuthFailure()
+ return
+ }
+ onStateChange?.('failed', `Transport closed (code ${code})`)
+ })
+ }
+
+ // โโ 7. Transport rebuild (shared by proactive refresh + 401 recovery) โโ
+ // Every /bridge call bumps epoch server-side. Both refresh paths must
+ // rebuild the transport with the new epoch โ a JWT-only swap leaves the
+ // old CCRClient heartbeating stale epoch โ 409. SSE resumes from the old
+ // transport's high-water-mark seq-num so no server-side replay.
+ // Caller MUST set authRecoveryInFlight = true before calling (synchronously,
+ // before any await) and clear it in a finally. This function doesn't manage
+ // the flag โ moving it here would be too late to prevent a double /bridge
+ // fetch, and each fetch bumps epoch.
+ async function rebuildTransport(
+ fresh: RemoteCredentials,
+ cause: Exclude,
+ ): Promise {
+ connectCause = cause
+ // Queue writes during rebuild โ once /bridge returns, the old transport's
+ // epoch is stale and its next write/heartbeat 409s. Without this gate,
+ // writeMessages adds UUIDs to recentPostedUUIDs then writeBatch silently
+ // no-ops (closed uploader after 409) โ permanent silent message loss.
+ flushGate.start()
+ try {
+ const seq = transport.getLastSequenceNum()
+ transport.close()
+ transport = await createV2ReplTransport({
+ sessionUrl: buildCCRv2SdkUrl(fresh.api_base_url, sessionId),
+ ingressToken: fresh.worker_jwt,
+ sessionId,
+ epoch: fresh.worker_epoch,
+ heartbeatIntervalMs: cfg.heartbeat_interval_ms,
+ heartbeatJitterFraction: cfg.heartbeat_jitter_fraction,
+ initialSequenceNum: seq,
+ getAuthToken: () => fresh.worker_jwt,
+ outboundOnly,
+ })
+ if (tornDown) {
+ // Teardown fired during the async createV2ReplTransport window.
+ // Don't wire/connect/schedule โ we'd re-arm timers after cancelAll()
+ // and fire onInboundMessage into a torn-down bridge.
+ transport.close()
+ return
+ }
+ wireTransportCallbacks()
+ transport.connect()
+ connectDeadline = setTimeout(
+ onConnectTimeout,
+ cfg.connect_timeout_ms,
+ connectCause,
+ )
+ refresh.scheduleFromExpiresIn(sessionId, fresh.expires_in)
+ // Drain queued writes into the new uploader. Runs before
+ // ccr.initialize() resolves (transport.connect() is fire-and-forget),
+ // but the uploader serializes behind the initial PUT /worker. If
+ // init fails (4091), events drop โ but only recentPostedUUIDs
+ // (per-instance) is populated, so re-enabling the bridge re-flushes.
+ drainFlushGate()
+ } finally {
+ // End the gate on failure paths too โ drainFlushGate already ended
+ // it on success. Queued messages are dropped (transport still dead).
+ flushGate.drop()
+ }
+ }
+
+ // โโ 8. 401 recovery (OAuth refresh + rebuild) โโโโโโโโโโโโโโโโโโโโโโโโโโโ
+ async function recoverFromAuthFailure(): Promise {
+ // setOnClose already guards `!authRecoveryInFlight` but that check and
+ // this set must be atomic against onRefresh โ claim synchronously before
+ // any await. Laptop wake fires both paths ~simultaneously.
+ if (authRecoveryInFlight) return
+ authRecoveryInFlight = true
+ onStateChange?.('reconnecting', 'JWT expired โ refreshing')
+ logForDebugging('[remote-bridge] 401 on SSE โ attempting JWT refresh')
+ try {
+ // Unconditionally try OAuth refresh โ getAccessToken() returns expired
+ // tokens as non-null strings, so !oauthToken doesn't catch expiry.
+ // Pass the stale token so handleOAuth401Error's keychain-comparison
+ // can detect if another tab already refreshed.
+ const stale = getAccessToken()
+ if (onAuth401) await onAuth401(stale ?? '')
+ const oauthToken = getAccessToken() ?? stale
+ if (!oauthToken || tornDown) {
+ if (!tornDown) {
+ onStateChange?.('failed', 'JWT refresh failed: no OAuth token')
+ }
+ return
+ }
+
+ const fresh = await withRetry(
+ () =>
+ fetchRemoteCredentials(
+ sessionId,
+ baseUrl,
+ oauthToken,
+ cfg.http_timeout_ms,
+ ),
+ 'fetchRemoteCredentials (recovery)',
+ cfg,
+ )
+ if (!fresh || tornDown) {
+ if (!tornDown) {
+ onStateChange?.('failed', 'JWT refresh failed after 401')
+ }
+ return
+ }
+ // If 401 interrupted the initial flush, writeBatch may have silently
+ // no-op'd on the closed uploader (ccr.close() ran in the SSE wrapper
+ // before our setOnClose callback). Reset so the new onConnect re-flushes.
+ // (v1 scopes initialFlushDone inside the per-transport closure at
+ // replBridge.ts:1027 so it resets naturally; v2 has it at outer scope.)
+ initialFlushDone = false
+ await rebuildTransport(fresh, 'auth_401_recovery')
+ logForDebugging('[remote-bridge] Transport rebuilt after 401')
+ } catch (err) {
+ logForDebugging(
+ `[remote-bridge] 401 recovery failed: ${errorMessage(err)}`,
+ { level: 'error' },
+ )
+ logForDiagnosticsNoPII('error', 'bridge_repl_v2_jwt_refresh_failed')
+ if (!tornDown) {
+ onStateChange?.('failed', `JWT refresh failed: ${errorMessage(err)}`)
+ }
+ } finally {
+ authRecoveryInFlight = false
+ }
+ }
+
+ wireTransportCallbacks()
+
+ // Start flushGate BEFORE connect so writeMessages() during handshake
+ // queues instead of racing the history POST.
+ if (initialMessages && initialMessages.length > 0) {
+ flushGate.start()
+ }
+ transport.connect()
+ connectDeadline = setTimeout(
+ onConnectTimeout,
+ cfg.connect_timeout_ms,
+ connectCause,
+ )
+
+ // โโ 8. History flush + drain helpers โโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโ
+ function drainFlushGate(): void {
+ const msgs = flushGate.end()
+ if (msgs.length === 0) return
+ for (const msg of msgs) recentPostedUUIDs.add(msg.uuid)
+ const events = toSDKMessages(msgs).map(m => ({
+ ...m,
+ session_id: sessionId,
+ }))
+ if (msgs.some(m => m.type === 'user')) {
+ transport.reportState('running')
+ }
+ logForDebugging(
+ `[remote-bridge] Drained ${msgs.length} queued message(s) after flush`,
+ )
+ void transport.writeBatch(events)
+ }
+
+ async function flushHistory(msgs: Message[]): Promise {
+ // v2 always creates a fresh server session (unconditional createCodeSession
+ // above) โ no session reuse, no double-post risk. Unlike v1, we do NOT
+ // filter by previouslyFlushedUUIDs: that set persists across REPL enable/
+ // disable cycles (useRef), so it would wrongly suppress history on re-enable.
+ const eligible = msgs.filter(isEligibleBridgeMessage)
+ const capped =
+ initialHistoryCap > 0 && eligible.length > initialHistoryCap
+ ? eligible.slice(-initialHistoryCap)
+ : eligible
+ if (capped.length < eligible.length) {
+ logForDebugging(
+ `[remote-bridge] Capped initial flush: ${eligible.length} -> ${capped.length} (cap=${initialHistoryCap})`,
+ )
+ }
+ const events = toSDKMessages(capped).map(m => ({
+ ...m,
+ session_id: sessionId,
+ }))
+ if (events.length === 0) return
+ // Mid-turn init: if Remote Control is enabled while a query is running,
+ // the last eligible message is a user prompt or tool_result (both 'user'
+ // type). Without this the init PUT's 'idle' sticks until the next user-
+ // type message forwards via writeMessages โ which for a pure-text turn
+ // is never (only assistant chunks stream post-init). Check eligible (pre-
+ // cap), not capped: the cap may truncate to a user message even when the
+ // actual trailing message is assistant.
+ if (eligible.at(-1)?.type === 'user') {
+ transport.reportState('running')
+ }
+ logForDebugging(`[remote-bridge] Flushing ${events.length} history events`)
+ await transport.writeBatch(events)
+ }
+
+ // โโ 9. Teardown โโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโ
+ // On SIGINT/SIGTERM/โ /exit, gracefulShutdown races runCleanupFunctions()
+ // against a 2s cap before forceExit kills the process. Budget accordingly:
+ // - archive: teardown_archive_timeout_ms (default 1500, cap 2000)
+ // - result write: fire-and-forget, archive latency covers the drain
+ // - 401 retry: only if first archive 401s, shares the same budget
+ async function teardown(): Promise {
+ if (tornDown) return
+ tornDown = true
+ refresh.cancelAll()
+ clearTimeout(connectDeadline)
+ flushGate.drop()
+
+ // Fire the result message before archive โ transport.write() only awaits
+ // enqueue (SerialBatchEventUploader resolves once buffered, drain is
+ // async). Archiving before close() gives the uploader's drain loop a
+ // window (typical archive โ 100-500ms) to POST the result without an
+ // explicit sleep. close() sets closed=true which interrupts drain at the
+ // next while-check, so close-before-archive drops the result.
+ transport.reportState('idle')
+ void transport.write(makeResultMessage(sessionId))
+
+ let token = getAccessToken()
+ let status = await archiveSession(
+ sessionId,
+ baseUrl,
+ token,
+ orgUUID,
+ cfg.teardown_archive_timeout_ms,
+ )
+
+ // Token is usually fresh (refresh scheduler runs 5min before expiry) but
+ // laptop-wake past the refresh window leaves getAccessToken() returning a
+ // stale string. Retry once on 401 โ onAuth401 (= handleOAuth401Error)
+ // clears keychain cache + force-refreshes. No proactive refresh on the
+ // happy path: handleOAuth401Error force-refreshes even valid tokens,
+ // which would waste budget 99% of the time. try/catch mirrors
+ // recoverFromAuthFailure: keychain reads can throw (macOS locked after
+ // wake); an uncaught throw here would skip transport.close + telemetry.
+ if (status === 401 && onAuth401) {
+ try {
+ await onAuth401(token ?? '')
+ token = getAccessToken()
+ status = await archiveSession(
+ sessionId,
+ baseUrl,
+ token,
+ orgUUID,
+ cfg.teardown_archive_timeout_ms,
+ )
+ } catch (err) {
+ logForDebugging(
+ `[remote-bridge] Teardown 401 retry threw: ${errorMessage(err)}`,
+ { level: 'error' },
+ )
+ }
+ }
+
+ transport.close()
+
+ const archiveStatus: ArchiveTelemetryStatus =
+ status === 'no_token'
+ ? 'skipped_no_token'
+ : status === 'timeout' || status === 'error'
+ ? 'network_error'
+ : status >= 500
+ ? 'server_5xx'
+ : status >= 400
+ ? 'server_4xx'
+ : 'ok'
+
+ logForDebugging(`[remote-bridge] Torn down (archive=${status})`)
+ logForDiagnosticsNoPII('info', 'bridge_repl_v2_teardown')
+ logEvent(
+ feature('CCR_MIRROR') && outboundOnly
+ ? 'tengu_ccr_mirror_teardown'
+ : 'tengu_bridge_repl_teardown',
+ {
+ v2: true,
+ archive_status:
+ archiveStatus as AnalyticsMetadata_I_VERIFIED_THIS_IS_NOT_CODE_OR_FILEPATHS,
+ archive_ok: typeof status === 'number' && status < 400,
+ archive_http_status: typeof status === 'number' ? status : undefined,
+ archive_timeout: status === 'timeout',
+ archive_no_token: status === 'no_token',
+ },
+ )
+ }
+ const unregister = registerCleanup(teardown)
+
+ if (feature('CCR_MIRROR') && outboundOnly) {
+ logEvent('tengu_ccr_mirror_started', {
+ v2: true,
+ expires_in_s: credentials.expires_in,
+ })
+ } else {
+ logEvent('tengu_bridge_repl_started', {
+ has_initial_messages: !!(initialMessages && initialMessages.length > 0),
+ v2: true,
+ expires_in_s: credentials.expires_in,
+ inProtectedNamespace: isInProtectedNamespace(),
+ })
+ }
+
+ // โโ 10. Handle โโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโ
+ return {
+ bridgeSessionId: sessionId,
+ environmentId: '',
+ sessionIngressUrl: credentials.api_base_url,
+ writeMessages(messages) {
+ const filtered = messages.filter(
+ m =>
+ isEligibleBridgeMessage(m) &&
+ !initialMessageUUIDs.has(m.uuid) &&
+ !recentPostedUUIDs.has(m.uuid),
+ )
+ if (filtered.length === 0) return
+
+ // Fire onUserMessage for title derivation. Scan before the flushGate
+ // check โ prompts are title-worthy even if they queue. Keeps calling
+ // on every title-worthy message until the callback returns true; the
+ // caller owns the policy (derive at 1st and 3rd, skip if explicit).
+ if (!userMessageCallbackDone) {
+ for (const m of filtered) {
+ const text = extractTitleText(m)
+ if (text !== undefined && onUserMessage?.(text, sessionId)) {
+ userMessageCallbackDone = true
+ break
+ }
+ }
+ }
+
+ if (flushGate.enqueue(...filtered)) {
+ logForDebugging(
+ `[remote-bridge] Queued ${filtered.length} message(s) during flush`,
+ )
+ return
+ }
+
+ for (const msg of filtered) recentPostedUUIDs.add(msg.uuid)
+ const events = toSDKMessages(filtered).map(m => ({
+ ...m,
+ session_id: sessionId,
+ }))
+ // v2 does not derive worker_status from events server-side (unlike v1
+ // session-ingress session_status_updater.go). Push it from here so the
+ // CCR web session list shows Running instead of stuck on Idle. A user
+ // message in the batch marks turn start. CCRClient.reportState dedupes
+ // consecutive same-state pushes.
+ if (filtered.some(m => m.type === 'user')) {
+ transport.reportState('running')
+ }
+ logForDebugging(`[remote-bridge] Sending ${filtered.length} message(s)`)
+ void transport.writeBatch(events)
+ },
+ writeSdkMessages(messages: SDKMessage[]) {
+ const filtered = messages.filter(
+ m => !m.uuid || !recentPostedUUIDs.has(m.uuid),
+ )
+ if (filtered.length === 0) return
+ for (const msg of filtered) {
+ if (msg.uuid) recentPostedUUIDs.add(msg.uuid)
+ }
+ const events = filtered.map(m => ({ ...m, session_id: sessionId }))
+ void transport.writeBatch(events)
+ },
+ sendControlRequest(request: SDKControlRequest) {
+ if (authRecoveryInFlight) {
+ logForDebugging(
+ `[remote-bridge] Dropping control_request during 401 recovery: ${request.request_id}`,
+ )
+ return
+ }
+ const event = { ...request, session_id: sessionId }
+ if (request.request.subtype === 'can_use_tool') {
+ transport.reportState('requires_action')
+ }
+ void transport.write(event)
+ logForDebugging(
+ `[remote-bridge] Sent control_request request_id=${request.request_id}`,
+ )
+ },
+ sendControlResponse(response: SDKControlResponse) {
+ if (authRecoveryInFlight) {
+ logForDebugging(
+ '[remote-bridge] Dropping control_response during 401 recovery',
+ )
+ return
+ }
+ const event = { ...response, session_id: sessionId }
+ transport.reportState('running')
+ void transport.write(event)
+ logForDebugging('[remote-bridge] Sent control_response')
+ },
+ sendControlCancelRequest(requestId: string) {
+ if (authRecoveryInFlight) {
+ logForDebugging(
+ `[remote-bridge] Dropping control_cancel_request during 401 recovery: ${requestId}`,
+ )
+ return
+ }
+ const event = {
+ type: 'control_cancel_request' as const,
+ request_id: requestId,
+ session_id: sessionId,
+ }
+ // Hook/classifier/channel/recheck resolved the permission locally โ
+ // interactiveHandler calls only cancelRequest (no sendResponse) on
+ // those paths, so without this the server stays on requires_action.
+ transport.reportState('running')
+ void transport.write(event)
+ logForDebugging(
+ `[remote-bridge] Sent control_cancel_request request_id=${requestId}`,
+ )
+ },
+ sendResult() {
+ if (authRecoveryInFlight) {
+ logForDebugging('[remote-bridge] Dropping result during 401 recovery')
+ return
+ }
+ transport.reportState('idle')
+ void transport.write(makeResultMessage(sessionId))
+ logForDebugging(`[remote-bridge] Sent result`)
+ },
+ async teardown() {
+ unregister()
+ await teardown()
+ },
+ }
+}
+
+// โโโ Session API (v2 /code/sessions, no env) โโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโ
+
+/** Retry an async init call with exponential backoff + jitter. */
+async function withRetry(
+ fn: () => Promise,
+ label: string,
+ cfg: EnvLessBridgeConfig,
+): Promise {
+ const max = cfg.init_retry_max_attempts
+ for (let attempt = 1; attempt <= max; attempt++) {
+ const result = await fn()
+ if (result !== null) return result
+ if (attempt < max) {
+ const base = cfg.init_retry_base_delay_ms * 2 ** (attempt - 1)
+ const jitter =
+ base * cfg.init_retry_jitter_fraction * (2 * Math.random() - 1)
+ const delay = Math.min(base + jitter, cfg.init_retry_max_delay_ms)
+ logForDebugging(
+ `[remote-bridge] ${label} failed (attempt ${attempt}/${max}), retrying in ${Math.round(delay)}ms`,
+ )
+ await sleep(delay)
+ }
+ }
+ return null
+}
+
+// Moved to codeSessionApi.ts so the SDK /bridge subpath can bundle them
+// without pulling in this file's heavy CLI tree (analytics, transport).
+export {
+ createCodeSession,
+ type RemoteCredentials,
+} from './codeSessionApi.js'
+import {
+ createCodeSession,
+ fetchRemoteCredentials as fetchRemoteCredentialsRaw,
+ type RemoteCredentials,
+} from './codeSessionApi.js'
+import { getBridgeBaseUrlOverride } from './bridgeConfig.js'
+
+// CLI-side wrapper that applies the CLAUDE_BRIDGE_BASE_URL dev override and
+// injects the trusted-device token (both are env/GrowthBook reads that the
+// SDK-facing codeSessionApi.ts export must stay free of).
+export async function fetchRemoteCredentials(
+ sessionId: string,
+ baseUrl: string,
+ accessToken: string,
+ timeoutMs: number,
+): Promise {
+ const creds = await fetchRemoteCredentialsRaw(
+ sessionId,
+ baseUrl,
+ accessToken,
+ timeoutMs,
+ getTrustedDeviceToken(),
+ )
+ if (!creds) return null
+ return getBridgeBaseUrlOverride()
+ ? { ...creds, api_base_url: baseUrl }
+ : creds
+}
+
+type ArchiveStatus = number | 'timeout' | 'error' | 'no_token'
+
+// Single categorical for BQ `GROUP BY archive_status`. The booleans on
+// _teardown predate this and are redundant with it (except archive_timeout,
+// which distinguishes ECONNABORTED from other network errors โ both map to
+// 'network_error' here since the dominant cause in a 1.5s window is timeout).
+type ArchiveTelemetryStatus =
+ | 'ok'
+ | 'skipped_no_token'
+ | 'network_error'
+ | 'server_4xx'
+ | 'server_5xx'
+
+async function archiveSession(
+ sessionId: string,
+ baseUrl: string,
+ accessToken: string | undefined,
+ orgUUID: string,
+ timeoutMs: number,
+): Promise {
+ if (!accessToken) return 'no_token'
+ // Archive lives at the compat layer (/v1/sessions/*, not /v1/code/sessions).
+ // compat.parseSessionID only accepts TagSession (session_*), so retag cse_*.
+ // anthropic-beta + x-organization-uuid are required โ without them the
+ // compat gateway 404s before reaching the handler.
+ //
+ // Unlike bridgeMain.ts (which caches compatId in sessionCompatIds to keep
+ // in-memory titledSessions/logger keys consistent across a mid-session
+ // gate flip), this compatId is only a server URL path segment โ no
+ // in-memory state. Fresh compute matches whatever the server currently
+ // validates: if the gate is OFF, the server has been updated to accept
+ // cse_* and we correctly send it.
+ const compatId = toCompatSessionId(sessionId)
+ try {
+ const response = await axios.post(
+ `${baseUrl}/v1/sessions/${compatId}/archive`,
+ {},
+ {
+ headers: {
+ ...oauthHeaders(accessToken),
+ 'anthropic-beta': 'ccr-byoc-2025-07-29',
+ 'x-organization-uuid': orgUUID,
+ },
+ timeout: timeoutMs,
+ validateStatus: () => true,
+ },
+ )
+ logForDebugging(
+ `[remote-bridge] Archive ${compatId} status=${response.status}`,
+ )
+ return response.status
+ } catch (err) {
+ const msg = errorMessage(err)
+ logForDebugging(`[remote-bridge] Archive failed: ${msg}`)
+ return axios.isAxiosError(err) && err.code === 'ECONNABORTED'
+ ? 'timeout'
+ : 'error'
+ }
+}
diff --git a/original-source-code/src/bridge/replBridge.ts b/original-source-code/src/bridge/replBridge.ts
new file mode 100644
index 0000000..7d7ac6a
--- /dev/null
+++ b/original-source-code/src/bridge/replBridge.ts
@@ -0,0 +1,2406 @@
+// biome-ignore-all assist/source/organizeImports: ANT-ONLY import markers must not be reordered
+import { randomUUID } from 'crypto'
+import {
+ createBridgeApiClient,
+ BridgeFatalError,
+ isExpiredErrorType,
+ isSuppressible403,
+} from './bridgeApi.js'
+import type { BridgeConfig, BridgeApiClient } from './types.js'
+import { logForDebugging } from '../utils/debug.js'
+import { logForDiagnosticsNoPII } from '../utils/diagLogs.js'
+import {
+ type AnalyticsMetadata_I_VERIFIED_THIS_IS_NOT_CODE_OR_FILEPATHS,
+ logEvent,
+} from '../services/analytics/index.js'
+import { registerCleanup } from '../utils/cleanupRegistry.js'
+import {
+ handleIngressMessage,
+ handleServerControlRequest,
+ makeResultMessage,
+ isEligibleBridgeMessage,
+ extractTitleText,
+ BoundedUUIDSet,
+} from './bridgeMessaging.js'
+import {
+ decodeWorkSecret,
+ buildSdkUrl,
+ buildCCRv2SdkUrl,
+ sameSessionId,
+} from './workSecret.js'
+import { toCompatSessionId, toInfraSessionId } from './sessionIdCompat.js'
+import { updateSessionBridgeId } from '../utils/concurrentSessions.js'
+import { getTrustedDeviceToken } from './trustedDevice.js'
+import { HybridTransport } from '../cli/transports/HybridTransport.js'
+import {
+ type ReplBridgeTransport,
+ createV1ReplTransport,
+ createV2ReplTransport,
+} from './replBridgeTransport.js'
+import { updateSessionIngressAuthToken } from '../utils/sessionIngressAuth.js'
+import { isEnvTruthy, isInProtectedNamespace } from '../utils/envUtils.js'
+import { validateBridgeId } from './bridgeApi.js'
+import {
+ describeAxiosError,
+ extractHttpStatus,
+ logBridgeSkip,
+} from './debugUtils.js'
+import type { Message } from '../types/message.js'
+import type { SDKMessage } from '../entrypoints/agentSdkTypes.js'
+import type { PermissionMode } from '../utils/permissions/PermissionMode.js'
+import type {
+ SDKControlRequest,
+ SDKControlResponse,
+} from '../entrypoints/sdk/controlTypes.js'
+import { createCapacityWake, type CapacitySignal } from './capacityWake.js'
+import { FlushGate } from './flushGate.js'
+import {
+ DEFAULT_POLL_CONFIG,
+ type PollIntervalConfig,
+} from './pollConfigDefaults.js'
+import { errorMessage } from '../utils/errors.js'
+import { sleep } from '../utils/sleep.js'
+import {
+ wrapApiForFaultInjection,
+ registerBridgeDebugHandle,
+ clearBridgeDebugHandle,
+ injectBridgeFault,
+} from './bridgeDebug.js'
+
+export type ReplBridgeHandle = {
+ bridgeSessionId: string
+ environmentId: string
+ sessionIngressUrl: string
+ writeMessages(messages: Message[]): void
+ writeSdkMessages(messages: SDKMessage[]): void
+ sendControlRequest(request: SDKControlRequest): void
+ sendControlResponse(response: SDKControlResponse): void
+ sendControlCancelRequest(requestId: string): void
+ sendResult(): void
+ teardown(): Promise
+}
+
+export type BridgeState = 'ready' | 'connected' | 'reconnecting' | 'failed'
+
+/**
+ * Explicit-param input to initBridgeCore. Everything initReplBridge reads
+ * from bootstrap state (cwd, session ID, git, OAuth) becomes a field here.
+ * A daemon caller (Agent SDK, PR 4) that never runs main.tsx fills these
+ * in itself.
+ */
+export type BridgeCoreParams = {
+ dir: string
+ machineName: string
+ branch: string
+ gitRepoUrl: string | null
+ title: string
+ baseUrl: string
+ sessionIngressUrl: string
+ /**
+ * Opaque string sent as metadata.worker_type. Use BridgeWorkerType for
+ * the two CLI-originated values; daemon callers may send any string the
+ * backend recognizes (it's just a filter key on the web side).
+ */
+ workerType: string
+ getAccessToken: () => string | undefined
+ /**
+ * POST /v1/sessions. Injected because `createSession.ts` lazy-loads
+ * `auth.ts`/`model.ts`/`oauth/client.ts` and `bun --outfile` inlines
+ * dynamic imports โ the lazy-load doesn't help, the whole REPL tree ends
+ * up in the Agent SDK bundle.
+ *
+ * REPL wrapper passes `createBridgeSession` from `createSession.ts`.
+ * Daemon wrapper passes `createBridgeSessionLean` from `sessionApi.ts`
+ * (HTTP-only, orgUUID+model supplied by the daemon caller).
+ *
+ * Receives `gitRepoUrl`+`branch` so the REPL wrapper can build the git
+ * source/outcome for claude.ai's session card. Daemon ignores them.
+ */
+ createSession: (opts: {
+ environmentId: string
+ title: string
+ gitRepoUrl: string | null
+ branch: string
+ signal: AbortSignal
+ }) => Promise
+ /**
+ * POST /v1/sessions/{id}/archive. Same injection rationale. Best-effort;
+ * the callback MUST NOT throw.
+ */
+ archiveSession: (sessionId: string) => Promise
+ /**
+ * Invoked on reconnect-after-env-lost to refresh the title. REPL wrapper
+ * reads session storage (picks up /rename); daemon returns the static
+ * title. Defaults to () => title.
+ */
+ getCurrentTitle?: () => string
+ /**
+ * Converts internal Message[] โ SDKMessage[] for writeMessages() and the
+ * initial-flush/drain paths. REPL wrapper passes the real toSDKMessages
+ * from utils/messages/mappers.ts. Daemon callers that only use
+ * writeSdkMessages() and pass no initialMessages can omit this โ those
+ * code paths are unreachable.
+ *
+ * Injected rather than imported because mappers.ts transitively pulls in
+ * src/commands.ts via messages.ts โ api.ts โ prompts.ts, dragging the
+ * entire command registry + React tree into the Agent SDK bundle.
+ */
+ toSDKMessages?: (messages: Message[]) => SDKMessage[]
+ /**
+ * OAuth 401 refresh handler passed to createBridgeApiClient. REPL wrapper
+ * passes handleOAuth401Error; daemon passes its AuthManager's handler.
+ * Injected because utils/auth.ts transitively pulls in the command
+ * registry via config.ts โ file.ts โ permissions/filesystem.ts โ
+ * sessionStorage.ts โ commands.ts.
+ */
+ onAuth401?: (staleAccessToken: string) => Promise
+ /**
+ * Poll interval config getter for the work-poll heartbeat loop. REPL
+ * wrapper passes the GrowthBook-backed getPollIntervalConfig (allows ops
+ * to live-tune poll rates fleet-wide). Daemon passes a static config
+ * with a 60s heartbeat (5ร headroom under the 300s work-lease TTL).
+ * Injected because growthbook.ts transitively pulls in the command
+ * registry via the same config.ts chain.
+ */
+ getPollIntervalConfig?: () => PollIntervalConfig
+ /**
+ * Max initial messages to replay on connect. REPL wrapper reads from the
+ * tengu_bridge_initial_history_cap GrowthBook flag. Daemon passes no
+ * initialMessages so this is never read. Default 200 matches the flag
+ * default.
+ */
+ initialHistoryCap?: number
+ // Same REPL-flush machinery as InitBridgeOptions โ daemon omits these.
+ initialMessages?: Message[]
+ previouslyFlushedUUIDs?: Set
+ onInboundMessage?: (msg: SDKMessage) => void
+ onPermissionResponse?: (response: SDKControlResponse) => void
+ onInterrupt?: () => void
+ onSetModel?: (model: string | undefined) => void
+ onSetMaxThinkingTokens?: (maxTokens: number | null) => void
+ /**
+ * Returns a policy verdict so this module can emit an error control_response
+ * without importing the policy checks itself (bootstrap-isolation constraint).
+ * The callback must guard `auto` (isAutoModeGateEnabled) and
+ * `bypassPermissions` (isBypassPermissionsModeDisabled AND
+ * isBypassPermissionsModeAvailable) BEFORE calling transitionPermissionMode โ
+ * that function's internal auto-gate check is a defensive throw, not a
+ * graceful guard, and its side-effect order is setAutoModeActive(true) then
+ * throw, which corrupts the 3-way invariant documented in src/CLAUDE.md if
+ * the callback lets the throw escape here.
+ */
+ onSetPermissionMode?: (
+ mode: PermissionMode,
+ ) => { ok: true } | { ok: false; error: string }
+ onStateChange?: (state: BridgeState, detail?: string) => void
+ /**
+ * Fires on each real user message to flow through writeMessages() until
+ * the callback returns true (done). Mirrors remoteBridgeCore.ts's
+ * onUserMessage so the REPL bridge can derive a session title from early
+ * prompts when none was set at init time (e.g. user runs /remote-control
+ * on an empty conversation, then types). Tool-result wrappers, meta
+ * messages, and display-tag-only messages are skipped. Receives
+ * currentSessionId so the wrapper can PATCH the title without a closure
+ * dance to reach the not-yet-returned handle. The caller owns the
+ * derive-at-count-1-and-3 policy; the transport just keeps calling until
+ * told to stop. Not fired for the writeSdkMessages daemon path (daemon
+ * sets its own title at init). Distinct from SessionSpawnOpts's
+ * onFirstUserMessage (spawn-bridge, PR #21250), which stays fire-once.
+ */
+ onUserMessage?: (text: string, sessionId: string) => boolean
+ /** See InitBridgeOptions.perpetual. */
+ perpetual?: boolean
+ /**
+ * Seeds lastTransportSequenceNum โ the SSE event-stream high-water mark
+ * that's carried across transport swaps within one process. Daemon callers
+ * pass the value they persisted at shutdown so the FIRST SSE connect of a
+ * fresh process sends from_sequence_num and the server doesn't replay full
+ * history. REPL callers omit (fresh session each run โ 0 is correct).
+ */
+ initialSSESequenceNum?: number
+}
+
+/**
+ * Superset of ReplBridgeHandle. Adds getSSESequenceNum for daemon callers
+ * that persist the SSE seq-num across process restarts and pass it back as
+ * initialSSESequenceNum on the next start.
+ */
+export type BridgeCoreHandle = ReplBridgeHandle & {
+ /**
+ * Current SSE sequence-number high-water mark. Updates as transports
+ * swap. Daemon callers persist this on shutdown and pass it back as
+ * initialSSESequenceNum on next start.
+ */
+ getSSESequenceNum(): number
+}
+
+/**
+ * Poll error recovery constants. When the work poll starts failing (e.g.
+ * server 500s), we use exponential backoff and give up after this timeout.
+ * This is deliberately long โ the server is the authority on when a session
+ * is truly dead. As long as the server accepts our poll, we keep waiting
+ * for it to re-dispatch the work item.
+ */
+const POLL_ERROR_INITIAL_DELAY_MS = 2_000
+const POLL_ERROR_MAX_DELAY_MS = 60_000
+const POLL_ERROR_GIVE_UP_MS = 15 * 60 * 1000
+
+// Monotonically increasing counter for distinguishing init calls in logs
+let initSequence = 0
+
+/**
+ * Bootstrap-free core: env registration โ session creation โ poll loop โ
+ * ingress WS โ teardown. Reads nothing from bootstrap/state or
+ * sessionStorage โ all context comes from params. Caller (initReplBridge
+ * below, or a daemon in PR 4) has already passed entitlement gates and
+ * gathered git/auth/title.
+ *
+ * Returns null on registration or session-creation failure.
+ */
+export async function initBridgeCore(
+ params: BridgeCoreParams,
+): Promise {
+ const {
+ dir,
+ machineName,
+ branch,
+ gitRepoUrl,
+ title,
+ baseUrl,
+ sessionIngressUrl,
+ workerType,
+ getAccessToken,
+ createSession,
+ archiveSession,
+ getCurrentTitle = () => title,
+ toSDKMessages = () => {
+ throw new Error(
+ 'BridgeCoreParams.toSDKMessages not provided. Pass it if you use writeMessages() or initialMessages โ daemon callers that only use writeSdkMessages() never hit this path.',
+ )
+ },
+ onAuth401,
+ getPollIntervalConfig = () => DEFAULT_POLL_CONFIG,
+ initialHistoryCap = 200,
+ initialMessages,
+ previouslyFlushedUUIDs,
+ onInboundMessage,
+ onPermissionResponse,
+ onInterrupt,
+ onSetModel,
+ onSetMaxThinkingTokens,
+ onSetPermissionMode,
+ onStateChange,
+ onUserMessage,
+ perpetual,
+ initialSSESequenceNum = 0,
+ } = params
+
+ const seq = ++initSequence
+
+ // bridgePointer import hoisted: perpetual mode reads it before register;
+ // non-perpetual writes it after session create; both use clear at teardown.
+ const { writeBridgePointer, clearBridgePointer, readBridgePointer } =
+ await import('./bridgePointer.js')
+
+ // Perpetual mode: read the crash-recovery pointer and treat it as prior
+ // state. The pointer is written unconditionally after session create
+ // (crash-recovery for all sessions); perpetual mode just skips the
+ // teardown clear so it survives clean exits too. Only reuse 'repl'
+ // pointers โ a crashed standalone bridge (`claude remote-control`)
+ // writes source:'standalone' with a different workerType.
+ const rawPrior = perpetual ? await readBridgePointer(dir) : null
+ const prior = rawPrior?.source === 'repl' ? rawPrior : null
+
+ logForDebugging(
+ `[bridge:repl] initBridgeCore #${seq} starting (initialMessages=${initialMessages?.length ?? 0}${prior ? ` perpetual prior=env:${prior.environmentId}` : ''})`,
+ )
+
+ // 5. Register bridge environment
+ const rawApi = createBridgeApiClient({
+ baseUrl,
+ getAccessToken,
+ runnerVersion: MACRO.VERSION,
+ onDebug: logForDebugging,
+ onAuth401,
+ getTrustedDeviceToken,
+ })
+ // Ant-only: interpose so /bridge-kick can inject poll/register/heartbeat
+ // failures. Zero cost in external builds (rawApi passes through unchanged).
+ const api =
+ process.env.USER_TYPE === 'ant' ? wrapApiForFaultInjection(rawApi) : rawApi
+
+ const bridgeConfig: BridgeConfig = {
+ dir,
+ machineName,
+ branch,
+ gitRepoUrl,
+ maxSessions: 1,
+ spawnMode: 'single-session',
+ verbose: false,
+ sandbox: false,
+ bridgeId: randomUUID(),
+ workerType,
+ environmentId: randomUUID(),
+ reuseEnvironmentId: prior?.environmentId,
+ apiBaseUrl: baseUrl,
+ sessionIngressUrl,
+ }
+
+ let environmentId: string
+ let environmentSecret: string
+ try {
+ const reg = await api.registerBridgeEnvironment(bridgeConfig)
+ environmentId = reg.environment_id
+ environmentSecret = reg.environment_secret
+ } catch (err) {
+ logBridgeSkip(
+ 'registration_failed',
+ `[bridge:repl] Environment registration failed: ${errorMessage(err)}`,
+ )
+ // Stale pointer may be the cause (expired/deleted env) โ clear it so
+ // the next start doesn't retry the same dead ID.
+ if (prior) {
+ await clearBridgePointer(dir)
+ }
+ onStateChange?.('failed', errorMessage(err))
+ return null
+ }
+
+ logForDebugging(`[bridge:repl] Environment registered: ${environmentId}`)
+ logForDiagnosticsNoPII('info', 'bridge_repl_env_registered')
+ logEvent('tengu_bridge_repl_env_registered', {})
+
+ /**
+ * Reconnect-in-place: if the just-registered environmentId matches what
+ * was requested, call reconnectSession to force-stop stale workers and
+ * re-queue the session. Used at init (perpetual mode โ env is alive but
+ * idle after clean teardown) and in doReconnect() Strategy 1 (env lost
+ * then resurrected). Returns true on success; caller falls back to
+ * fresh session creation on false.
+ */
+ async function tryReconnectInPlace(
+ requestedEnvId: string,
+ sessionId: string,
+ ): Promise {
+ if (environmentId !== requestedEnvId) {
+ logForDebugging(
+ `[bridge:repl] Env mismatch (requested ${requestedEnvId}, got ${environmentId}) โ cannot reconnect in place`,
+ )
+ return false
+ }
+ // The pointer stores what createBridgeSession returned (session_*,
+ // compat/convert.go:41). /bridge/reconnect is an environments-layer
+ // endpoint โ once the server's ccr_v2_compat_enabled gate is on it
+ // looks sessions up by their infra tag (cse_*) and returns "Session
+ // not found" for the session_* costume. We don't know the gate state
+ // pre-poll, so try both; the re-tag is a no-op if the ID is already
+ // cse_* (doReconnect Strategy 1 path โ currentSessionId never mutates
+ // to cse_* but future-proof the check).
+ const infraId = toInfraSessionId(sessionId)
+ const candidates =
+ infraId === sessionId ? [sessionId] : [sessionId, infraId]
+ for (const id of candidates) {
+ try {
+ await api.reconnectSession(environmentId, id)
+ logForDebugging(
+ `[bridge:repl] Reconnected session ${id} in place on env ${environmentId}`,
+ )
+ return true
+ } catch (err) {
+ logForDebugging(
+ `[bridge:repl] reconnectSession(${id}) failed: ${errorMessage(err)}`,
+ )
+ }
+ }
+ logForDebugging(
+ '[bridge:repl] reconnectSession exhausted โ falling through to fresh session',
+ )
+ return false
+ }
+
+ // Perpetual init: env is alive but has no queued work after clean
+ // teardown. reconnectSession re-queues it. doReconnect() has the same
+ // call but only fires on poll 404 (env dead);
+ // here the env is alive but idle.
+ const reusedPriorSession = prior
+ ? await tryReconnectInPlace(prior.environmentId, prior.sessionId)
+ : false
+ if (prior && !reusedPriorSession) {
+ await clearBridgePointer(dir)
+ }
+
+ // 6. Create session on the bridge. Initial messages are NOT included as
+ // session creation events because those use STREAM_ONLY persistence and
+ // are published before the CCR UI subscribes, so they get lost. Instead,
+ // initial messages are flushed via the ingress WebSocket once it connects.
+
+ // Mutable session ID โ updated when the environment+session pair is
+ // re-created after a connection loss.
+ let currentSessionId: string
+
+
+ if (reusedPriorSession && prior) {
+ currentSessionId = prior.sessionId
+ logForDebugging(
+ `[bridge:repl] Perpetual session reused: ${currentSessionId}`,
+ )
+ // Server already has all initialMessages from the prior CLI run. Mark
+ // them as previously-flushed so the initial flush filter excludes them
+ // (previouslyFlushedUUIDs is a fresh Set on every CLI start). Duplicate
+ // UUIDs cause the server to kill the WebSocket.
+ if (initialMessages && previouslyFlushedUUIDs) {
+ for (const msg of initialMessages) {
+ previouslyFlushedUUIDs.add(msg.uuid)
+ }
+ }
+ } else {
+ const createdSessionId = await createSession({
+ environmentId,
+ title,
+ gitRepoUrl,
+ branch,
+ signal: AbortSignal.timeout(15_000),
+ })
+
+ if (!createdSessionId) {
+ logForDebugging(
+ '[bridge:repl] Session creation failed, deregistering environment',
+ )
+ logEvent('tengu_bridge_repl_session_failed', {})
+ await api.deregisterEnvironment(environmentId).catch(() => {})
+ onStateChange?.('failed', 'Session creation failed')
+ return null
+ }
+
+ currentSessionId = createdSessionId
+ logForDebugging(`[bridge:repl] Session created: ${currentSessionId}`)
+ }
+
+ // Crash-recovery pointer: written now so a kill -9 at any point after
+ // this leaves a recoverable trail. Cleared in teardown (non-perpetual)
+ // or left alone (perpetual mode โ pointer survives clean exit too).
+ // `claude remote-control --continue` from the same directory will detect
+ // it and offer to resume.
+ await writeBridgePointer(dir, {
+ sessionId: currentSessionId,
+ environmentId,
+ source: 'repl',
+ })
+ logForDiagnosticsNoPII('info', 'bridge_repl_session_created')
+ logEvent('tengu_bridge_repl_started', {
+ has_initial_messages: !!(initialMessages && initialMessages.length > 0),
+ inProtectedNamespace: isInProtectedNamespace(),
+ })
+
+ // UUIDs of initial messages. Used for dedup in writeMessages to avoid
+ // re-sending messages that were already flushed on WebSocket open.
+ const initialMessageUUIDs = new Set()
+ if (initialMessages) {
+ for (const msg of initialMessages) {
+ initialMessageUUIDs.add(msg.uuid)
+ }
+ }
+
+ // Bounded ring buffer of UUIDs for messages we've already sent to the
+ // server via the ingress WebSocket. Serves two purposes:
+ // 1. Echo filtering โ ignore our own messages bouncing back on the WS.
+ // 2. Secondary dedup in writeMessages โ catch race conditions where
+ // the hook's index-based tracking isn't sufficient.
+ //
+ // Seeded with initialMessageUUIDs so that when the server echoes back
+ // the initial conversation context over the ingress WebSocket, those
+ // messages are recognized as echoes and not re-injected into the REPL.
+ //
+ // Capacity of 2000 covers well over any realistic echo window (echoes
+ // arrive within milliseconds) and any messages that might be re-encountered
+ // after compaction. The hook's lastWrittenIndexRef is the primary dedup;
+ // this is a safety net.
+ const recentPostedUUIDs = new BoundedUUIDSet(2000)
+ for (const uuid of initialMessageUUIDs) {
+ recentPostedUUIDs.add(uuid)
+ }
+
+ // Bounded set of INBOUND prompt UUIDs we've already forwarded to the REPL.
+ // Defensive dedup for when the server re-delivers prompts (seq-num
+ // negotiation failure, server edge cases, transport swap races). The
+ // seq-num carryover below is the primary fix; this is the safety net.
+ const recentInboundUUIDs = new BoundedUUIDSet(2000)
+
+ // 7. Start poll loop for work items โ this is what makes the session
+ // "live" on claude.ai. When a user types there, the backend dispatches
+ // a work item to our environment. We poll for it, get the ingress token,
+ // and connect the ingress WebSocket.
+ //
+ // The poll loop keeps running: when work arrives it connects the ingress
+ // WebSocket, and if the WebSocket drops unexpectedly (code != 1000) it
+ // resumes polling to get a fresh ingress token and reconnect.
+ const pollController = new AbortController()
+ // Adapter over either HybridTransport (v1: WS reads + POST writes to
+ // Session-Ingress) or SSETransport+CCRClient (v2: SSE reads + POST
+ // writes to CCR /worker/*). The v1/v2 choice is made in onWorkReceived:
+ // server-driven via secret.use_code_sessions, with CLAUDE_BRIDGE_USE_CCR_V2
+ // as an ant-dev override.
+ let transport: ReplBridgeTransport | null = null
+ // Bumped on every onWorkReceived. Captured in createV2ReplTransport's .then()
+ // closure to detect stale resolutions: if two calls race while transport is
+ // null, both registerWorker() (bumping server epoch), and whichever resolves
+ // SECOND is the correct one โ but the transport !== null check gets this
+ // backwards (first-to-resolve installs, second discards). The generation
+ // counter catches it independent of transport state.
+ let v2Generation = 0
+ // SSE sequence-number high-water mark carried across transport swaps.
+ // Without this, each new SSETransport starts at 0, sends no
+ // from_sequence_num / Last-Event-ID on its first connect, and the server
+ // replays the entire session event history โ every prompt ever sent
+ // re-delivered as fresh inbound messages on every onWorkReceived.
+ //
+ // Seed only when we actually reconnected the prior session. If
+ // `reusedPriorSession` is false we fell through to `createSession()` โ
+ // the caller's persisted seq-num belongs to a dead session and applying
+ // it to the fresh stream (starting at 1) silently drops events. Same
+ // hazard as doReconnect Strategy 2; same fix as the reset there.
+ let lastTransportSequenceNum = reusedPriorSession ? initialSSESequenceNum : 0
+ // Track the current work ID so teardown can call stopWork
+ let currentWorkId: string | null = null
+ // Session ingress JWT for the current work item โ used for heartbeat auth.
+ let currentIngressToken: string | null = null
+ // Signal to wake the at-capacity sleep early when the transport is lost,
+ // so the poll loop immediately switches back to fast polling for new work.
+ const capacityWake = createCapacityWake(pollController.signal)
+ const wakePollLoop = capacityWake.wake
+ const capacitySignal = capacityWake.signal
+ // Gates message writes during the initial flush to prevent ordering
+ // races where new messages arrive at the server interleaved with history.
+ const flushGate = new FlushGate()
+
+ // Latch for onUserMessage โ flips true when the callback returns true
+ // (policy says "done deriving"). If no callback, skip scanning entirely
+ // (daemon path โ no title derivation needed).
+ let userMessageCallbackDone = !onUserMessage
+
+ // Shared counter for environment re-creations, used by both
+ // onEnvironmentLost and the abnormal-close handler.
+ const MAX_ENVIRONMENT_RECREATIONS = 3
+ let environmentRecreations = 0
+ let reconnectPromise: Promise | null = null
+
+ /**
+ * Recover from onEnvironmentLost (poll returned 404 โ env was reaped
+ * server-side). Tries two strategies in order:
+ *
+ * 1. Reconnect-in-place: idempotent re-register with reuseEnvironmentId
+ * โ if the backend returns the same env ID, call reconnectSession()
+ * to re-queue the existing session. currentSessionId stays the same;
+ * the URL on the user's phone stays valid; previouslyFlushedUUIDs is
+ * preserved so history isn't re-sent.
+ *
+ * 2. Fresh session fallback: if the backend returns a different env ID
+ * (original TTL-expired, e.g. laptop slept >4h) or reconnectSession()
+ * throws, archive the old session and create a new one on the
+ * now-registered env. Old behavior before #20460 primitives landed.
+ *
+ * Uses a promise-based reentrancy guard so concurrent callers share the
+ * same reconnection attempt.
+ */
+ async function reconnectEnvironmentWithSession(): Promise {
+ if (reconnectPromise) {
+ return reconnectPromise
+ }
+ reconnectPromise = doReconnect()
+ try {
+ return await reconnectPromise
+ } finally {
+ reconnectPromise = null
+ }
+ }
+
+ async function doReconnect(): Promise {
+ environmentRecreations++
+ // Invalidate any in-flight v2 handshake โ the environment is being
+ // recreated, so a stale transport arriving post-reconnect would be
+ // pointed at a dead session.
+ v2Generation++
+ logForDebugging(
+ `[bridge:repl] Reconnecting after env lost (attempt ${environmentRecreations}/${MAX_ENVIRONMENT_RECREATIONS})`,
+ )
+
+ if (environmentRecreations > MAX_ENVIRONMENT_RECREATIONS) {
+ logForDebugging(
+ `[bridge:repl] Environment reconnect limit reached (${MAX_ENVIRONMENT_RECREATIONS}), giving up`,
+ )
+ return false
+ }
+
+ // Close the stale transport. Capture seq BEFORE close โ if Strategy 1
+ // (tryReconnectInPlace) succeeds we keep the SAME session, and the
+ // next transport must resume where this one left off, not replay from
+ // the last transport-swap checkpoint.
+ if (transport) {
+ const seq = transport.getLastSequenceNum()
+ if (seq > lastTransportSequenceNum) {
+ lastTransportSequenceNum = seq
+ }
+ transport.close()
+ transport = null
+ }
+ // Transport is gone โ wake the poll loop out of its at-capacity
+ // heartbeat sleep so it can fast-poll for re-dispatched work.
+ wakePollLoop()
+ // Reset flush gate so writeMessages() hits the !transport guard
+ // instead of silently queuing into a dead buffer.
+ flushGate.drop()
+
+ // Release the current work item (force=false โ we may want the session
+ // back). Best-effort: the env is probably gone, so this likely 404s.
+ if (currentWorkId) {
+ const workIdBeingCleared = currentWorkId
+ await api
+ .stopWork(environmentId, workIdBeingCleared, false)
+ .catch(() => {})
+ // When doReconnect runs concurrently with the poll loop (ws_closed
+ // handler case โ void-called, unlike the awaited onEnvironmentLost
+ // path), onWorkReceived can fire during the stopWork await and set
+ // a fresh currentWorkId. If it did, the poll loop has already
+ // recovered on its own โ defer to it rather than proceeding to
+ // archiveSession, which would destroy the session its new
+ // transport is connected to.
+ if (currentWorkId !== workIdBeingCleared) {
+ logForDebugging(
+ '[bridge:repl] Poll loop recovered during stopWork await โ deferring to it',
+ )
+ environmentRecreations = 0
+ return true
+ }
+ currentWorkId = null
+ currentIngressToken = null
+ }
+
+ // Bail out if teardown started while we were awaiting
+ if (pollController.signal.aborted) {
+ logForDebugging('[bridge:repl] Reconnect aborted by teardown')
+ return false
+ }
+
+ // Strategy 1: idempotent re-register with the server-issued env ID.
+ // If the backend resurrects the same env (fresh secret), we can
+ // reconnect the existing session. If it hands back a different ID, the
+ // original env is truly gone and we fall through to a fresh session.
+ const requestedEnvId = environmentId
+ bridgeConfig.reuseEnvironmentId = requestedEnvId
+ try {
+ const reg = await api.registerBridgeEnvironment(bridgeConfig)
+ environmentId = reg.environment_id
+ environmentSecret = reg.environment_secret
+ } catch (err) {
+ bridgeConfig.reuseEnvironmentId = undefined
+ logForDebugging(
+ `[bridge:repl] Environment re-registration failed: ${errorMessage(err)}`,
+ )
+ return false
+ }
+ // Clear before any await โ a stale value would poison the next fresh
+ // registration if doReconnect runs again.
+ bridgeConfig.reuseEnvironmentId = undefined
+
+ logForDebugging(
+ `[bridge:repl] Re-registered: requested=${requestedEnvId} got=${environmentId}`,
+ )
+
+ // Bail out if teardown started while we were registering
+ if (pollController.signal.aborted) {
+ logForDebugging(
+ '[bridge:repl] Reconnect aborted after env registration, cleaning up',
+ )
+ await api.deregisterEnvironment(environmentId).catch(() => {})
+ return false
+ }
+
+ // Same race as above, narrower window: poll loop may have set up a
+ // transport during the registerBridgeEnvironment await. Bail before
+ // tryReconnectInPlace/archiveSession kill it server-side.
+ if (transport !== null) {
+ logForDebugging(
+ '[bridge:repl] Poll loop recovered during registerBridgeEnvironment await โ deferring to it',
+ )
+ environmentRecreations = 0
+ return true
+ }
+
+ // Strategy 1: same helper as perpetual init. currentSessionId stays
+ // the same on success; URL on mobile/web stays valid;
+ // previouslyFlushedUUIDs preserved (no re-flush).
+ if (await tryReconnectInPlace(requestedEnvId, currentSessionId)) {
+ logEvent('tengu_bridge_repl_reconnected_in_place', {})
+ environmentRecreations = 0
+ return true
+ }
+ // Env differs โ TTL-expired/reaped; or reconnect failed.
+ // Don't deregister โ we have a fresh secret for this env either way.
+ if (environmentId !== requestedEnvId) {
+ logEvent('tengu_bridge_repl_env_expired_fresh_session', {})
+ }
+
+ // Strategy 2: fresh session on the now-registered environment.
+ // Archive the old session first โ it's orphaned (bound to a dead env,
+ // or reconnectSession rejected it). Don't deregister the env โ we just
+ // got a fresh secret for it and are about to use it.
+ await archiveSession(currentSessionId)
+
+ // Bail out if teardown started while we were archiving
+ if (pollController.signal.aborted) {
+ logForDebugging(
+ '[bridge:repl] Reconnect aborted after archive, cleaning up',
+ )
+ await api.deregisterEnvironment(environmentId).catch(() => {})
+ return false
+ }
+
+ // Re-read the current title in case the user renamed the session.
+ // REPL wrapper reads session storage; daemon wrapper returns the
+ // original title (nothing to refresh).
+ const currentTitle = getCurrentTitle()
+
+ // Create a new session on the now-registered environment
+ const newSessionId = await createSession({
+ environmentId,
+ title: currentTitle,
+ gitRepoUrl,
+ branch,
+ signal: AbortSignal.timeout(15_000),
+ })
+
+ if (!newSessionId) {
+ logForDebugging(
+ '[bridge:repl] Session creation failed during reconnection',
+ )
+ return false
+ }
+
+ // Bail out if teardown started during session creation (up to 15s)
+ if (pollController.signal.aborted) {
+ logForDebugging(
+ '[bridge:repl] Reconnect aborted after session creation, cleaning up',
+ )
+ await archiveSession(newSessionId)
+ return false
+ }
+
+ currentSessionId = newSessionId
+ // Re-publish to the PID file so peer dedup (peerRegistry.ts) picks up the
+ // new ID โ setReplBridgeHandle only fires at init/teardown, not reconnect.
+ void updateSessionBridgeId(toCompatSessionId(newSessionId)).catch(() => {})
+ // Reset per-session transport state IMMEDIATELY after the session swap,
+ // before any await. If this runs after `await writeBridgePointer` below,
+ // there's a window where handle.bridgeSessionId already returns session B
+ // but getSSESequenceNum() still returns session A's seq โ a daemon
+ // persistState() in that window writes {bridgeSessionId: B, seq: OLD_A},
+ // which PASSES the session-ID validation check and defeats it entirely.
+ //
+ // The SSE seq-num is scoped to the session's event stream โ carrying it
+ // over leaves the transport's lastSequenceNum stuck high (seq only
+ // advances when received > last), and its next internal reconnect would
+ // send from_sequence_num=OLD_SEQ against a stream starting at 1 โ all
+ // events in the gap silently dropped. Inbound UUID dedup is also
+ // session-scoped.
+ lastTransportSequenceNum = 0
+ recentInboundUUIDs.clear()
+ // Title derivation is session-scoped too: if the user typed during the
+ // createSession await above, the callback fired against the OLD archived
+ // session ID (PATCH lost) and the new session got `currentTitle` captured
+ // BEFORE they typed. Reset so the next prompt can re-derive. Self-
+ // correcting: if the caller's policy is already done (explicit title or
+ // count โฅ 3), it returns true on the first post-reset call and re-latches.
+ userMessageCallbackDone = !onUserMessage
+ logForDebugging(`[bridge:repl] Re-created session: ${currentSessionId}`)
+
+ // Rewrite the crash-recovery pointer with the new IDs so a crash after
+ // this point resumes the right session. (The reconnect-in-place path
+ // above doesn't touch the pointer โ same session, same env.)
+ await writeBridgePointer(dir, {
+ sessionId: currentSessionId,
+ environmentId,
+ source: 'repl',
+ })
+
+ // Clear flushed UUIDs so initial messages are re-sent to the new session.
+ // UUIDs are scoped per-session on the server, so re-flushing is safe.
+ previouslyFlushedUUIDs?.clear()
+
+
+ // Reset the counter so independent reconnections hours apart don't
+ // exhaust the limit โ it guards against rapid consecutive failures,
+ // not lifetime total.
+ environmentRecreations = 0
+
+ return true
+ }
+
+ // Helper: get the current OAuth access token for session ingress auth.
+ // Unlike the JWT path, OAuth tokens are refreshed by the standard OAuth
+ // flow โ no proactive scheduler needed.
+ function getOAuthToken(): string | undefined {
+ return getAccessToken()
+ }
+
+ // Drain any messages that were queued during the initial flush.
+ // Called after writeBatch completes (or fails) so queued messages
+ // are sent in order after the historical messages.
+ function drainFlushGate(): void {
+ const msgs = flushGate.end()
+ if (msgs.length === 0) return
+ if (!transport) {
+ logForDebugging(
+ `[bridge:repl] Cannot drain ${msgs.length} pending message(s): no transport`,
+ )
+ return
+ }
+ for (const msg of msgs) {
+ recentPostedUUIDs.add(msg.uuid)
+ }
+ const sdkMessages = toSDKMessages(msgs)
+ const events = sdkMessages.map(sdkMsg => ({
+ ...sdkMsg,
+ session_id: currentSessionId,
+ }))
+ logForDebugging(
+ `[bridge:repl] Drained ${msgs.length} pending message(s) after flush`,
+ )
+ void transport.writeBatch(events)
+ }
+
+ // Teardown reference โ set after definition below. All callers are async
+ // callbacks that run after assignment, so the reference is always valid.
+ let doTeardownImpl: (() => Promise) | null = null
+ function triggerTeardown(): void {
+ void doTeardownImpl?.()
+ }
+
+ /**
+ * Body of the transport's setOnClose callback, hoisted to initBridgeCore
+ * scope so /bridge-kick can fire it directly. setOnClose wraps this with
+ * a stale-transport guard; debugFireClose calls it bare.
+ *
+ * With autoReconnect:true, this only fires on: clean close (1000),
+ * permanent server rejection (4001/1002/4003), or 10-min budget
+ * exhaustion. Transient drops are retried internally by the transport.
+ */
+ function handleTransportPermanentClose(closeCode: number | undefined): void {
+ logForDebugging(
+ `[bridge:repl] Transport permanently closed: code=${closeCode}`,
+ )
+ logEvent('tengu_bridge_repl_ws_closed', {
+ code: closeCode,
+ })
+ // Capture SSE seq high-water mark before nulling. When called from
+ // setOnClose the guard guarantees transport !== null; when fired from
+ // /bridge-kick it may already be null (e.g. fired twice) โ skip.
+ if (transport) {
+ const closedSeq = transport.getLastSequenceNum()
+ if (closedSeq > lastTransportSequenceNum) {
+ lastTransportSequenceNum = closedSeq
+ }
+ transport = null
+ }
+ // Transport is gone โ wake the poll loop out of its at-capacity
+ // heartbeat sleep so it's fast-polling by the time the reconnect
+ // below completes and the server re-queues work.
+ wakePollLoop()
+ // Reset flush state so writeMessages() hits the !transport guard
+ // (with a warning log) instead of silently queuing into a buffer
+ // that will never be drained. Unlike onWorkReceived (which
+ // preserves pending messages for the new transport), onClose is
+ // a permanent close โ no new transport will drain these.
+ const dropped = flushGate.drop()
+ if (dropped > 0) {
+ logForDebugging(
+ `[bridge:repl] Dropping ${dropped} pending message(s) on transport close (code=${closeCode})`,
+ { level: 'warn' },
+ )
+ }
+
+ if (closeCode === 1000) {
+ // Clean close โ session ended normally. Tear down the bridge.
+ onStateChange?.('failed', 'session ended')
+ pollController.abort()
+ triggerTeardown()
+ return
+ }
+
+ // Transport reconnect budget exhausted or permanent server
+ // rejection. By this point the env has usually been reaped
+ // server-side (BQ 2026-03-12: ~98% of ws_closed never recover
+ // via poll alone). stopWork(force=false) can't re-dispatch work
+ // from an archived env; reconnectEnvironmentWithSession can
+ // re-activate it via POST /bridge/reconnect, or fall through
+ // to a fresh session if the env is truly gone. The poll loop
+ // (already woken above) picks up the re-queued work once
+ // doReconnect completes.
+ onStateChange?.(
+ 'reconnecting',
+ `Remote Control connection lost (code ${closeCode})`,
+ )
+ logForDebugging(
+ `[bridge:repl] Transport reconnect budget exhausted (code=${closeCode}), attempting env reconnect`,
+ )
+ void reconnectEnvironmentWithSession().then(success => {
+ if (success) return
+ // doReconnect has four abort-check return-false sites for
+ // teardown-in-progress. Don't pollute the BQ failure signal
+ // or double-teardown when the user just quit.
+ if (pollController.signal.aborted) return
+ // doReconnect returns false (never throws) on genuine failure.
+ // The dangerous case: registerBridgeEnvironment succeeded (so
+ // environmentId now points at a fresh valid env) but
+ // createSession failed โ poll loop would poll a sessionless
+ // env getting null work with no errors, never hitting any
+ // give-up path. Tear down explicitly.
+ logForDebugging(
+ '[bridge:repl] reconnectEnvironmentWithSession resolved false โ tearing down',
+ )
+ logEvent('tengu_bridge_repl_reconnect_failed', {
+ close_code: closeCode,
+ })
+ onStateChange?.('failed', 'reconnection failed')
+ triggerTeardown()
+ })
+ }
+
+ // Ant-only: SIGUSR2 โ force doReconnect() for manual testing. Skips the
+ // ~30s poll wait โ fire-and-observe in the debug log immediately.
+ // Windows has no USR signals; `process.on` would throw there.
+ let sigusr2Handler: (() => void) | undefined
+ if (process.env.USER_TYPE === 'ant' && process.platform !== 'win32') {
+ sigusr2Handler = () => {
+ logForDebugging(
+ '[bridge:repl] SIGUSR2 received โ forcing doReconnect() for testing',
+ )
+ void reconnectEnvironmentWithSession()
+ }
+ process.on('SIGUSR2', sigusr2Handler)
+ }
+
+ // Ant-only: /bridge-kick fault injection. handleTransportPermanentClose
+ // is defined below and assigned into this slot so the slash command can
+ // invoke it directly โ the real setOnClose callback is buried inside
+ // wireTransport which is itself inside onWorkReceived.
+ let debugFireClose: ((code: number) => void) | null = null
+ if (process.env.USER_TYPE === 'ant') {
+ registerBridgeDebugHandle({
+ fireClose: code => {
+ if (!debugFireClose) {
+ logForDebugging('[bridge:debug] fireClose: no transport wired yet')
+ return
+ }
+ logForDebugging(`[bridge:debug] fireClose(${code}) โ injecting`)
+ debugFireClose(code)
+ },
+ forceReconnect: () => {
+ logForDebugging('[bridge:debug] forceReconnect โ injecting')
+ void reconnectEnvironmentWithSession()
+ },
+ injectFault: injectBridgeFault,
+ wakePollLoop,
+ describe: () =>
+ `env=${environmentId} session=${currentSessionId} transport=${transport?.getStateLabel() ?? 'null'} workId=${currentWorkId ?? 'null'}`,
+ })
+ }
+
+ const pollOpts = {
+ api,
+ getCredentials: () => ({ environmentId, environmentSecret }),
+ signal: pollController.signal,
+ getPollIntervalConfig,
+ onStateChange,
+ getWsState: () => transport?.getStateLabel() ?? 'null',
+ // REPL bridge is single-session: having any transport == at capacity.
+ // No need to check isConnectedStatus() โ even while the transport is
+ // auto-reconnecting internally (up to 10 min), poll is heartbeat-only.
+ isAtCapacity: () => transport !== null,
+ capacitySignal,
+ onFatalError: triggerTeardown,
+ getHeartbeatInfo: () => {
+ if (!currentWorkId || !currentIngressToken) {
+ return null
+ }
+ return {
+ environmentId,
+ workId: currentWorkId,
+ sessionToken: currentIngressToken,
+ }
+ },
+ // Work-item JWT expired (or work gone). The transport is useless โ
+ // SSE reconnects and CCR writes use the same stale token. Without
+ // this callback the poll loop would do a 10-min at-capacity backoff,
+ // during which the work lease (300s TTL) expires and the server stops
+ // forwarding prompts โ ~25-min dead window observed in daemon logs.
+ // Kill the transport + work state so isAtCapacity()=false; the loop
+ // fast-polls and picks up the server's re-dispatched work in seconds.
+ onHeartbeatFatal: (err: BridgeFatalError) => {
+ logForDebugging(
+ `[bridge:repl] heartbeatWork fatal (status=${err.status}) โ tearing down work item for fast re-dispatch`,
+ )
+ if (transport) {
+ const seq = transport.getLastSequenceNum()
+ if (seq > lastTransportSequenceNum) {
+ lastTransportSequenceNum = seq
+ }
+ transport.close()
+ transport = null
+ }
+ flushGate.drop()
+ // force=false โ server re-queues. Likely already expired, but
+ // idempotent and makes re-dispatch immediate if not.
+ if (currentWorkId) {
+ void api
+ .stopWork(environmentId, currentWorkId, false)
+ .catch((e: unknown) => {
+ logForDebugging(
+ `[bridge:repl] stopWork after heartbeat fatal: ${errorMessage(e)}`,
+ )
+ })
+ }
+ currentWorkId = null
+ currentIngressToken = null
+ wakePollLoop()
+ onStateChange?.(
+ 'reconnecting',
+ 'Work item lease expired, fetching fresh token',
+ )
+ },
+ async onEnvironmentLost() {
+ const success = await reconnectEnvironmentWithSession()
+ if (!success) {
+ return null
+ }
+ return { environmentId, environmentSecret }
+ },
+ onWorkReceived: (
+ workSessionId: string,
+ ingressToken: string,
+ workId: string,
+ serverUseCcrV2: boolean,
+ ) => {
+ // When new work arrives while a transport is already open, the
+ // server has decided to re-dispatch (e.g. token rotation, server
+ // restart). Close the existing transport and reconnect โ discarding
+ // the work causes a stuck 'reconnecting' state if the old WS dies
+ // shortly after (the server won't re-dispatch a work item it
+ // already delivered).
+ // ingressToken (JWT) is stored for heartbeat auth (both v1 and v2).
+ // Transport auth diverges โ see the v1/v2 split below.
+ if (transport?.isConnectedStatus()) {
+ logForDebugging(
+ `[bridge:repl] Work received while transport connected, replacing with fresh token (workId=${workId})`,
+ )
+ }
+
+ logForDebugging(
+ `[bridge:repl] Work received: workId=${workId} workSessionId=${workSessionId} currentSessionId=${currentSessionId} match=${sameSessionId(workSessionId, currentSessionId)}`,
+ )
+
+ // Refresh the crash-recovery pointer's mtime. Staleness checks file
+ // mtime (not embedded timestamp) so this re-write bumps the clock โ
+ // a 5h+ session that crashes still has a fresh pointer. Fires once
+ // per work dispatch (infrequent โ bounded by user message rate).
+ void writeBridgePointer(dir, {
+ sessionId: currentSessionId,
+ environmentId,
+ source: 'repl',
+ })
+
+ // Reject foreign session IDs โ the server shouldn't assign sessions
+ // from other environments. Since we create env+session as a pair,
+ // a mismatch indicates an unexpected server-side reassignment.
+ //
+ // Compare by underlying UUID, not by tagged-ID prefix. When CCR
+ // v2's compat layer serves the session, createBridgeSession gets
+ // session_* from the v1-facing API (compat/convert.go:41) but the
+ // infrastructure layer delivers cse_* in the work queue
+ // (container_manager.go:129). Same UUID, different tag.
+ if (!sameSessionId(workSessionId, currentSessionId)) {
+ logForDebugging(
+ `[bridge:repl] Rejecting foreign session: expected=${currentSessionId} got=${workSessionId}`,
+ )
+ return
+ }
+
+ currentWorkId = workId
+ currentIngressToken = ingressToken
+
+ // Server decides per-session (secret.use_code_sessions from the work
+ // secret, threaded through runWorkPollLoop). The env var is an ant-dev
+ // override for forcing v2 before the server flag is on for your user โ
+ // requires ccr_v2_compat_enabled server-side or registerWorker 404s.
+ //
+ // Kept separate from CLAUDE_CODE_USE_CCR_V2 (the child-SDK transport
+ // selector set by sessionRunner/environment-manager) to avoid the
+ // inheritance hazard in spawn mode where the parent's orchestrator
+ // var would leak into a v1 child.
+ const useCcrV2 =
+ serverUseCcrV2 || isEnvTruthy(process.env.CLAUDE_BRIDGE_USE_CCR_V2)
+
+ // Auth is the one place v1 and v2 diverge hard:
+ //
+ // - v1 (Session-Ingress): accepts OAuth OR JWT. We prefer OAuth
+ // because the standard OAuth refresh flow handles expiry โ no
+ // separate JWT refresh scheduler needed.
+ //
+ // - v2 (CCR /worker/*): REQUIRES the JWT. register_worker.go:32
+ // validates the session_id claim, which OAuth tokens don't carry.
+ // The JWT from the work secret has both that claim and the worker
+ // role (environment_auth.py:856). JWT refresh: when it expires the
+ // server re-dispatches work with a fresh one, and onWorkReceived
+ // fires again. createV2ReplTransport stores it via
+ // updateSessionIngressAuthToken() before touching the network.
+ let v1OauthToken: string | undefined
+ if (!useCcrV2) {
+ v1OauthToken = getOAuthToken()
+ if (!v1OauthToken) {
+ logForDebugging(
+ '[bridge:repl] No OAuth token available for session ingress, skipping work',
+ )
+ return
+ }
+ updateSessionIngressAuthToken(v1OauthToken)
+ }
+ logEvent('tengu_bridge_repl_work_received', {})
+
+ // Close the previous transport. Nullify BEFORE calling close() so
+ // the close callback doesn't treat the programmatic close as
+ // "session ended normally" and trigger a full teardown.
+ if (transport) {
+ const oldTransport = transport
+ transport = null
+ // Capture the SSE sequence high-water mark so the next transport
+ // resumes the stream instead of replaying from seq 0. Use max() โ
+ // a transport that died early (never received any frames) would
+ // otherwise reset a non-zero mark back to 0.
+ const oldSeq = oldTransport.getLastSequenceNum()
+ if (oldSeq > lastTransportSequenceNum) {
+ lastTransportSequenceNum = oldSeq
+ }
+ oldTransport.close()
+ }
+ // Reset flush state โ the old flush (if any) is no longer relevant.
+ // Preserve pending messages so they're drained after the new
+ // transport's flush completes (the hook has already advanced its
+ // lastWrittenIndex and won't re-send them).
+ flushGate.deactivate()
+
+ // Closure adapter over the shared handleServerControlRequest โ
+ // captures transport/currentSessionId so the transport.setOnData
+ // callback below doesn't need to thread them through.
+ const onServerControlRequest = (request: SDKControlRequest): void =>
+ handleServerControlRequest(request, {
+ transport,
+ sessionId: currentSessionId,
+ onInterrupt,
+ onSetModel,
+ onSetMaxThinkingTokens,
+ onSetPermissionMode,
+ })
+
+ let initialFlushDone = false
+
+ // Wire callbacks onto a freshly constructed transport and connect.
+ // Extracted so the (sync) v1 and (async) v2 construction paths can
+ // share the identical callback + flush machinery.
+ const wireTransport = (newTransport: ReplBridgeTransport): void => {
+ transport = newTransport
+
+ newTransport.setOnConnect(() => {
+ // Guard: if transport was replaced by a newer onWorkReceived call
+ // while the WS was connecting, ignore this stale callback.
+ if (transport !== newTransport) return
+
+ logForDebugging('[bridge:repl] Ingress transport connected')
+ logEvent('tengu_bridge_repl_ws_connected', {})
+
+ // Update the env var with the latest OAuth token so POST writes
+ // (which read via getSessionIngressAuthToken()) use a fresh token.
+ // v2 skips this โ createV2ReplTransport already stored the JWT,
+ // and overwriting it with OAuth would break subsequent /worker/*
+ // requests (session_id claim check).
+ if (!useCcrV2) {
+ const freshToken = getOAuthToken()
+ if (freshToken) {
+ updateSessionIngressAuthToken(freshToken)
+ }
+ }
+
+ // Reset teardownStarted so future teardowns are not blocked.
+ teardownStarted = false
+
+ // Flush initial messages only on first connect, not on every
+ // WS reconnection. Re-flushing would cause duplicate messages.
+ // IMPORTANT: onStateChange('connected') is deferred until the
+ // flush completes. This prevents writeMessages() from sending
+ // new messages that could arrive at the server interleaved with
+ // the historical messages, and delays the web UI from showing
+ // the session as active until history is persisted.
+ if (
+ !initialFlushDone &&
+ initialMessages &&
+ initialMessages.length > 0
+ ) {
+ initialFlushDone = true
+
+ // Cap the initial flush to the most recent N messages. The full
+ // history is UI-only (model doesn't see it) and large replays cause
+ // slow session-ingress persistence (each event is a threadstore write)
+ // plus elevated Firestore pressure. A 0 or negative cap disables it.
+ const historyCap = initialHistoryCap
+ const eligibleMessages = initialMessages.filter(
+ m =>
+ isEligibleBridgeMessage(m) &&
+ !previouslyFlushedUUIDs?.has(m.uuid),
+ )
+ const cappedMessages =
+ historyCap > 0 && eligibleMessages.length > historyCap
+ ? eligibleMessages.slice(-historyCap)
+ : eligibleMessages
+ if (cappedMessages.length < eligibleMessages.length) {
+ logForDebugging(
+ `[bridge:repl] Capped initial flush: ${eligibleMessages.length} -> ${cappedMessages.length} (cap=${historyCap})`,
+ )
+ logEvent('tengu_bridge_repl_history_capped', {
+ eligible_count: eligibleMessages.length,
+ capped_count: cappedMessages.length,
+ })
+ }
+ const sdkMessages = toSDKMessages(cappedMessages)
+ if (sdkMessages.length > 0) {
+ logForDebugging(
+ `[bridge:repl] Flushing ${sdkMessages.length} initial message(s) via transport`,
+ )
+ const events = sdkMessages.map(sdkMsg => ({
+ ...sdkMsg,
+ session_id: currentSessionId,
+ }))
+ const dropsBefore = newTransport.droppedBatchCount
+ void newTransport
+ .writeBatch(events)
+ .then(() => {
+ // If any batch was dropped during this flush (SI down for
+ // maxConsecutiveFailures attempts), flush() still resolved
+ // normally but the events were NOT delivered. Don't mark
+ // UUIDs as flushed โ keep them eligible for re-send on the
+ // next onWorkReceived (JWT refresh re-dispatch, line ~1144).
+ if (newTransport.droppedBatchCount > dropsBefore) {
+ logForDebugging(
+ `[bridge:repl] Initial flush dropped ${newTransport.droppedBatchCount - dropsBefore} batch(es) โ not marking ${sdkMessages.length} UUID(s) as flushed`,
+ )
+ return
+ }
+ if (previouslyFlushedUUIDs) {
+ for (const sdkMsg of sdkMessages) {
+ if (sdkMsg.uuid) {
+ previouslyFlushedUUIDs.add(sdkMsg.uuid)
+ }
+ }
+ }
+ })
+ .catch(e =>
+ logForDebugging(`[bridge:repl] Initial flush failed: ${e}`),
+ )
+ .finally(() => {
+ // Guard: if transport was replaced during the flush,
+ // don't signal connected or drain โ the new transport
+ // owns the lifecycle now.
+ if (transport !== newTransport) return
+ drainFlushGate()
+ onStateChange?.('connected')
+ })
+ } else {
+ // All initial messages were already flushed (filtered by
+ // previouslyFlushedUUIDs). No flush POST needed โ clear
+ // the flag and signal connected immediately. This is the
+ // first connect for this transport (inside !initialFlushDone),
+ // so no flush POST is in-flight โ the flag was set before
+ // connect() and must be cleared here.
+ drainFlushGate()
+ onStateChange?.('connected')
+ }
+ } else if (!flushGate.active) {
+ // No initial messages or already flushed on first connect.
+ // WS auto-reconnect path โ only signal connected if no flush
+ // POST is in-flight. If one is, .finally() owns the lifecycle.
+ onStateChange?.('connected')
+ }
+ })
+
+ newTransport.setOnData(data => {
+ handleIngressMessage(
+ data,
+ recentPostedUUIDs,
+ recentInboundUUIDs,
+ onInboundMessage,
+ onPermissionResponse,
+ onServerControlRequest,
+ )
+ })
+
+ // Body lives at initBridgeCore scope so /bridge-kick can call it
+ // directly via debugFireClose. All referenced closures (transport,
+ // wakePollLoop, flushGate, reconnectEnvironmentWithSession, etc.)
+ // are already at that scope. The only lexical dependency on
+ // wireTransport was `newTransport.getLastSequenceNum()` โ but after
+ // the guard below passes we know transport === newTransport.
+ debugFireClose = handleTransportPermanentClose
+ newTransport.setOnClose(closeCode => {
+ // Guard: if transport was replaced, ignore stale close.
+ if (transport !== newTransport) return
+ handleTransportPermanentClose(closeCode)
+ })
+
+ // Start the flush gate before connect() to cover the WS handshake
+ // window. Between transport assignment and setOnConnect firing,
+ // writeMessages() could send messages via HTTP POST before the
+ // initial flush starts. Starting the gate here ensures those
+ // calls are queued. If there are no initial messages, the gate
+ // stays inactive.
+ if (
+ !initialFlushDone &&
+ initialMessages &&
+ initialMessages.length > 0
+ ) {
+ flushGate.start()
+ }
+
+ newTransport.connect()
+ } // end wireTransport
+
+ // Bump unconditionally โ ANY new transport (v1 or v2) invalidates an
+ // in-flight v2 handshake. Also bumped in doReconnect().
+ v2Generation++
+
+ if (useCcrV2) {
+ // workSessionId is the cse_* form (infrastructure-layer ID from the
+ // work queue), which is what /v1/code/sessions/{id}/worker/* wants.
+ // The session_* form (currentSessionId) is NOT usable here โ
+ // handler/convert.go:30 validates TagCodeSession.
+ const sessionUrl = buildCCRv2SdkUrl(baseUrl, workSessionId)
+ const thisGen = v2Generation
+ logForDebugging(
+ `[bridge:repl] CCR v2: sessionUrl=${sessionUrl} session=${workSessionId} gen=${thisGen}`,
+ )
+ void createV2ReplTransport({
+ sessionUrl,
+ ingressToken,
+ sessionId: workSessionId,
+ initialSequenceNum: lastTransportSequenceNum,
+ }).then(
+ t => {
+ // Teardown started while registerWorker was in flight. Teardown
+ // saw transport === null and skipped close(); installing now
+ // would leak CCRClient heartbeat timers and reset
+ // teardownStarted via wireTransport's side effects.
+ if (pollController.signal.aborted) {
+ t.close()
+ return
+ }
+ // onWorkReceived may have fired again while registerWorker()
+ // was in flight (server re-dispatch with a fresh JWT). The
+ // transport !== null check alone gets the race wrong when BOTH
+ // attempts saw transport === null โ it keeps the first resolver
+ // (stale epoch) and discards the second (correct epoch). The
+ // generation check catches it regardless of transport state.
+ if (thisGen !== v2Generation) {
+ logForDebugging(
+ `[bridge:repl] CCR v2: discarding stale handshake gen=${thisGen} current=${v2Generation}`,
+ )
+ t.close()
+ return
+ }
+ wireTransport(t)
+ },
+ (err: unknown) => {
+ logForDebugging(
+ `[bridge:repl] CCR v2: createV2ReplTransport failed: ${errorMessage(err)}`,
+ { level: 'error' },
+ )
+ logEvent('tengu_bridge_repl_ccr_v2_init_failed', {})
+ // If a newer attempt is in flight or already succeeded, don't
+ // touch its work item โ our failure is irrelevant.
+ if (thisGen !== v2Generation) return
+ // Release the work item so the server re-dispatches immediately
+ // instead of waiting for its own timeout. currentWorkId was set
+ // above; without this, the session looks stuck to the user.
+ if (currentWorkId) {
+ void api
+ .stopWork(environmentId, currentWorkId, false)
+ .catch((e: unknown) => {
+ logForDebugging(
+ `[bridge:repl] stopWork after v2 init failure: ${errorMessage(e)}`,
+ )
+ })
+ currentWorkId = null
+ currentIngressToken = null
+ }
+ wakePollLoop()
+ },
+ )
+ } else {
+ // v1: HybridTransport (WS reads + POST writes to Session-Ingress).
+ // autoReconnect is true (default) โ when the WS dies, the transport
+ // reconnects automatically with exponential backoff. POST writes
+ // continue during reconnection (they use getSessionIngressAuthToken()
+ // independently of WS state). The poll loop remains as a secondary
+ // fallback if the reconnect budget is exhausted (10 min).
+ //
+ // Auth: uses OAuth tokens directly instead of the JWT from the work
+ // secret. refreshHeaders picks up the latest OAuth token on each
+ // WS reconnect attempt.
+ const wsUrl = buildSdkUrl(sessionIngressUrl, workSessionId)
+ logForDebugging(`[bridge:repl] Ingress URL: ${wsUrl}`)
+ logForDebugging(
+ `[bridge:repl] Creating HybridTransport: session=${workSessionId}`,
+ )
+ // v1OauthToken was validated non-null above (we'd have returned early).
+ const oauthToken = v1OauthToken ?? ''
+ wireTransport(
+ createV1ReplTransport(
+ new HybridTransport(
+ new URL(wsUrl),
+ {
+ Authorization: `Bearer ${oauthToken}`,
+ 'anthropic-version': '2023-06-01',
+ },
+ workSessionId,
+ () => ({
+ Authorization: `Bearer ${getOAuthToken() ?? oauthToken}`,
+ 'anthropic-version': '2023-06-01',
+ }),
+ // Cap retries so a persistently-failing session-ingress can't
+ // pin the uploader drain loop for the lifetime of the bridge.
+ // 50 attempts โ 20 min (15s POST timeout + 8s backoff + jitter
+ // per cycle at steady state). Bridge-only โ 1P keeps indefinite.
+ {
+ maxConsecutiveFailures: 50,
+ isBridge: true,
+ onBatchDropped: () => {
+ onStateChange?.(
+ 'reconnecting',
+ 'Lost sync with Remote Control โ events could not be delivered',
+ )
+ // SI has been down ~20 min. Wake the poll loop so that when
+ // SI recovers, next poll โ onWorkReceived โ fresh transport
+ // โ initial flush succeeds โ onStateChange('connected') at
+ // ~line 1420. Without this, state stays 'reconnecting' even
+ // after SI recovers โ daemon.ts:437 denies all permissions,
+ // useReplBridge.ts:311 keeps replBridgeSessionActive=false.
+ // If the env was archived during the outage, poll 404 โ
+ // onEnvironmentLost recovery path handles it.
+ wakePollLoop()
+ },
+ },
+ ),
+ ),
+ )
+ }
+ },
+ }
+ void startWorkPollLoop(pollOpts)
+
+ // Perpetual mode: hourly mtime refresh of the crash-recovery pointer.
+ // The onWorkReceived refresh only fires per user prompt โ a
+ // daemon idle for >4h would have a stale pointer, and the next restart
+ // would clear it (readBridgePointer TTL check) โ fresh session. The
+ // standalone bridge (bridgeMain.ts) has an identical hourly timer.
+ const pointerRefreshTimer = perpetual
+ ? setInterval(() => {
+ // doReconnect() reassigns currentSessionId/environmentId non-
+ // atomically (env at ~:634, session at ~:719, awaits in between).
+ // If this timer fires in that window, its fire-and-forget write can
+ // race with (and overwrite) doReconnect's own pointer write at ~:740,
+ // leaving the pointer at the now-archived old session. doReconnect
+ // writes the pointer itself, so skipping here is free.
+ if (reconnectPromise) return
+ void writeBridgePointer(dir, {
+ sessionId: currentSessionId,
+ environmentId,
+ source: 'repl',
+ })
+ }, 60 * 60_000)
+ : null
+ pointerRefreshTimer?.unref?.()
+
+ // Push a silent keep_alive frame on a fixed interval so upstream proxies
+ // and the session-ingress layer don't GC an otherwise-idle remote control
+ // session. The keep_alive type is filtered before reaching any client UI
+ // (Query.ts drops it; web/iOS/Android never see it in their message loop).
+ // Interval comes from GrowthBook (tengu_bridge_poll_interval_config
+ // session_keepalive_interval_v2_ms, default 120s); 0 = disabled.
+ const keepAliveIntervalMs =
+ getPollIntervalConfig().session_keepalive_interval_v2_ms
+ const keepAliveTimer =
+ keepAliveIntervalMs > 0
+ ? setInterval(() => {
+ if (!transport) return
+ logForDebugging('[bridge:repl] keep_alive sent')
+ void transport.write({ type: 'keep_alive' }).catch((err: unknown) => {
+ logForDebugging(
+ `[bridge:repl] keep_alive write failed: ${errorMessage(err)}`,
+ )
+ })
+ }, keepAliveIntervalMs)
+ : null
+ keepAliveTimer?.unref?.()
+
+ // Shared teardown sequence used by both cleanup registration and
+ // the explicit teardown() method on the returned handle.
+ let teardownStarted = false
+ doTeardownImpl = async (): Promise => {
+ if (teardownStarted) {
+ logForDebugging(
+ `[bridge:repl] Teardown already in progress, skipping duplicate call env=${environmentId} session=${currentSessionId}`,
+ )
+ return
+ }
+ teardownStarted = true
+ const teardownStart = Date.now()
+ logForDebugging(
+ `[bridge:repl] Teardown starting: env=${environmentId} session=${currentSessionId} workId=${currentWorkId ?? 'none'} transportState=${transport?.getStateLabel() ?? 'null'}`,
+ )
+
+ if (pointerRefreshTimer !== null) {
+ clearInterval(pointerRefreshTimer)
+ }
+ if (keepAliveTimer !== null) {
+ clearInterval(keepAliveTimer)
+ }
+ if (sigusr2Handler) {
+ process.off('SIGUSR2', sigusr2Handler)
+ }
+ if (process.env.USER_TYPE === 'ant') {
+ clearBridgeDebugHandle()
+ debugFireClose = null
+ }
+ pollController.abort()
+ logForDebugging('[bridge:repl] Teardown: poll loop aborted')
+
+ // Capture the live transport's seq BEFORE close() โ close() is sync
+ // (just aborts the SSE fetch) and does NOT invoke onClose, so the
+ // setOnClose capture path never runs for explicit teardown.
+ // Without this, getSSESequenceNum() after teardown returns the stale
+ // lastTransportSequenceNum (captured at the last transport swap), and
+ // daemon callers persisting that value lose all events since then.
+ if (transport) {
+ const finalSeq = transport.getLastSequenceNum()
+ if (finalSeq > lastTransportSequenceNum) {
+ lastTransportSequenceNum = finalSeq
+ }
+ }
+
+ if (perpetual) {
+ // Perpetual teardown is LOCAL-ONLY โ do not send result, do not call
+ // stopWork, do not close the transport. All of those signal the
+ // server (and any mobile/attach subscribers) that the session is
+ // ending. Instead: stop polling, let the socket die with the
+ // process; the backend times the work-item lease back to pending on
+ // its own (TTL 300s). Next daemon start reads the pointer and
+ // reconnectSession re-queues work.
+ transport = null
+ flushGate.drop()
+ // Refresh the pointer mtime so that sessions lasting longer than
+ // BRIDGE_POINTER_TTL_MS (4h) don't appear stale on next start.
+ await writeBridgePointer(dir, {
+ sessionId: currentSessionId,
+ environmentId,
+ source: 'repl',
+ })
+ logForDebugging(
+ `[bridge:repl] Teardown (perpetual): leaving env=${environmentId} session=${currentSessionId} alive on server, duration=${Date.now() - teardownStart}ms`,
+ )
+ return
+ }
+
+ // Fire the result message, then archive, THEN close. transport.write()
+ // only enqueues (SerialBatchEventUploader resolves on buffer-add); the
+ // stopWork/archive latency (~200-500ms) is the drain window for the
+ // result POST. Closing BEFORE archive meant relying on HybridTransport's
+ // void-ed 3s grace period, which nothing awaits โ forceExit can kill the
+ // socket mid-POST. Same reorder as remoteBridgeCore.ts teardown (#22803).
+ const teardownTransport = transport
+ transport = null
+ flushGate.drop()
+ if (teardownTransport) {
+ void teardownTransport.write(makeResultMessage(currentSessionId))
+ }
+
+ const stopWorkP = currentWorkId
+ ? api
+ .stopWork(environmentId, currentWorkId, true)
+ .then(() => {
+ logForDebugging('[bridge:repl] Teardown: stopWork completed')
+ })
+ .catch((err: unknown) => {
+ logForDebugging(
+ `[bridge:repl] Teardown stopWork failed: ${errorMessage(err)}`,
+ )
+ })
+ : Promise.resolve()
+
+ // Run stopWork and archiveSession in parallel. gracefulShutdown.ts:407
+ // races runCleanupFunctions() against 2s (NOT the 5s outer failsafe),
+ // so archive is capped at 1.5s at the injection site to stay under budget.
+ // archiveSession is contractually no-throw; the injected implementations
+ // log their own success/failure internally.
+ await Promise.all([stopWorkP, archiveSession(currentSessionId)])
+
+ teardownTransport?.close()
+ logForDebugging('[bridge:repl] Teardown: transport closed')
+
+ await api.deregisterEnvironment(environmentId).catch((err: unknown) => {
+ logForDebugging(
+ `[bridge:repl] Teardown deregister failed: ${errorMessage(err)}`,
+ )
+ })
+
+ // Clear the crash-recovery pointer โ explicit disconnect or clean REPL
+ // exit means the user is done with this session. Crash/kill-9 never
+ // reaches this line, leaving the pointer for next-launch recovery.
+ await clearBridgePointer(dir)
+
+ logForDebugging(
+ `[bridge:repl] Teardown complete: env=${environmentId} duration=${Date.now() - teardownStart}ms`,
+ )
+ }
+
+ // 8. Register cleanup for graceful shutdown
+ const unregister = registerCleanup(() => doTeardownImpl?.())
+
+ logForDebugging(
+ `[bridge:repl] Ready: env=${environmentId} session=${currentSessionId}`,
+ )
+ onStateChange?.('ready')
+
+ return {
+ get bridgeSessionId() {
+ return currentSessionId
+ },
+ get environmentId() {
+ return environmentId
+ },
+ getSSESequenceNum() {
+ // lastTransportSequenceNum only updates when a transport is CLOSED
+ // (captured at swap/onClose). During normal operation the CURRENT
+ // transport's live seq isn't reflected there. Merge both so callers
+ // (e.g. daemon persistState()) get the actual high-water mark.
+ const live = transport?.getLastSequenceNum() ?? 0
+ return Math.max(lastTransportSequenceNum, live)
+ },
+ sessionIngressUrl,
+ writeMessages(messages) {
+ // Filter to user/assistant messages that haven't already been sent.
+ // Two layers of dedup:
+ // - initialMessageUUIDs: messages sent as session creation events
+ // - recentPostedUUIDs: messages recently sent via POST
+ const filtered = messages.filter(
+ m =>
+ isEligibleBridgeMessage(m) &&
+ !initialMessageUUIDs.has(m.uuid) &&
+ !recentPostedUUIDs.has(m.uuid),
+ )
+ if (filtered.length === 0) return
+
+ // Fire onUserMessage for title derivation. Scan before the flushGate
+ // check โ prompts are title-worthy even if they queue behind the
+ // initial history flush. Keeps calling on every title-worthy message
+ // until the callback returns true; the caller owns the policy.
+ if (!userMessageCallbackDone) {
+ for (const m of filtered) {
+ const text = extractTitleText(m)
+ if (text !== undefined && onUserMessage?.(text, currentSessionId)) {
+ userMessageCallbackDone = true
+ break
+ }
+ }
+ }
+
+ // Queue messages while the initial flush is in progress to prevent
+ // them from arriving at the server interleaved with history.
+ if (flushGate.enqueue(...filtered)) {
+ logForDebugging(
+ `[bridge:repl] Queued ${filtered.length} message(s) during initial flush`,
+ )
+ return
+ }
+
+ if (!transport) {
+ const types = filtered.map(m => m.type).join(',')
+ logForDebugging(
+ `[bridge:repl] Transport not configured, dropping ${filtered.length} message(s) [${types}] for session=${currentSessionId}`,
+ { level: 'warn' },
+ )
+ return
+ }
+
+ // Track in the bounded ring buffer for echo filtering and dedup.
+ for (const msg of filtered) {
+ recentPostedUUIDs.add(msg.uuid)
+ }
+
+ logForDebugging(
+ `[bridge:repl] Sending ${filtered.length} message(s) via transport`,
+ )
+
+ // Convert to SDK format and send via HTTP POST (HybridTransport).
+ // The web UI receives them via the subscribe WebSocket.
+ const sdkMessages = toSDKMessages(filtered)
+ const events = sdkMessages.map(sdkMsg => ({
+ ...sdkMsg,
+ session_id: currentSessionId,
+ }))
+ void transport.writeBatch(events)
+ },
+ writeSdkMessages(messages) {
+ // Daemon path: query() already yields SDKMessage, skip conversion.
+ // Still run echo dedup (server bounces writes back on the WS).
+ // No initialMessageUUIDs filter โ daemon has no initial messages.
+ // No flushGate โ daemon never starts it (no initial flush).
+ const filtered = messages.filter(
+ m => !m.uuid || !recentPostedUUIDs.has(m.uuid),
+ )
+ if (filtered.length === 0) return
+ if (!transport) {
+ logForDebugging(
+ `[bridge:repl] Transport not configured, dropping ${filtered.length} SDK message(s) for session=${currentSessionId}`,
+ { level: 'warn' },
+ )
+ return
+ }
+ for (const msg of filtered) {
+ if (msg.uuid) recentPostedUUIDs.add(msg.uuid)
+ }
+ const events = filtered.map(m => ({ ...m, session_id: currentSessionId }))
+ void transport.writeBatch(events)
+ },
+ sendControlRequest(request: SDKControlRequest) {
+ if (!transport) {
+ logForDebugging(
+ '[bridge:repl] Transport not configured, skipping control_request',
+ )
+ return
+ }
+ const event = { ...request, session_id: currentSessionId }
+ void transport.write(event)
+ logForDebugging(
+ `[bridge:repl] Sent control_request request_id=${request.request_id}`,
+ )
+ },
+ sendControlResponse(response: SDKControlResponse) {
+ if (!transport) {
+ logForDebugging(
+ '[bridge:repl] Transport not configured, skipping control_response',
+ )
+ return
+ }
+ const event = { ...response, session_id: currentSessionId }
+ void transport.write(event)
+ logForDebugging('[bridge:repl] Sent control_response')
+ },
+ sendControlCancelRequest(requestId: string) {
+ if (!transport) {
+ logForDebugging(
+ '[bridge:repl] Transport not configured, skipping control_cancel_request',
+ )
+ return
+ }
+ const event = {
+ type: 'control_cancel_request' as const,
+ request_id: requestId,
+ session_id: currentSessionId,
+ }
+ void transport.write(event)
+ logForDebugging(
+ `[bridge:repl] Sent control_cancel_request request_id=${requestId}`,
+ )
+ },
+ sendResult() {
+ if (!transport) {
+ logForDebugging(
+ `[bridge:repl] sendResult: skipping, transport not configured session=${currentSessionId}`,
+ )
+ return
+ }
+ void transport.write(makeResultMessage(currentSessionId))
+ logForDebugging(
+ `[bridge:repl] Sent result for session=${currentSessionId}`,
+ )
+ },
+ async teardown() {
+ unregister()
+ await doTeardownImpl?.()
+ logForDebugging('[bridge:repl] Torn down')
+ logEvent('tengu_bridge_repl_teardown', {})
+ },
+ }
+}
+
+/**
+ * Persistent poll loop for work items. Runs in the background for the
+ * lifetime of the bridge connection.
+ *
+ * When a work item arrives, acknowledges it and calls onWorkReceived
+ * with the session ID and ingress token (which connects the ingress
+ * WebSocket). Then continues polling โ the server will dispatch a new
+ * work item if the ingress WebSocket drops, allowing automatic
+ * reconnection without tearing down the bridge.
+ */
+async function startWorkPollLoop({
+ api,
+ getCredentials,
+ signal,
+ onStateChange,
+ onWorkReceived,
+ onEnvironmentLost,
+ getWsState,
+ isAtCapacity,
+ capacitySignal,
+ onFatalError,
+ getPollIntervalConfig = () => DEFAULT_POLL_CONFIG,
+ getHeartbeatInfo,
+ onHeartbeatFatal,
+}: {
+ api: BridgeApiClient
+ getCredentials: () => { environmentId: string; environmentSecret: string }
+ signal: AbortSignal
+ onStateChange?: (state: BridgeState, detail?: string) => void
+ onWorkReceived: (
+ sessionId: string,
+ ingressToken: string,
+ workId: string,
+ useCodeSessions: boolean,
+ ) => void
+ /** Called when the environment has been deleted. Returns new credentials or null. */
+ onEnvironmentLost?: () => Promise<{
+ environmentId: string
+ environmentSecret: string
+ } | null>
+ /** Returns the current WebSocket readyState label for diagnostic logging. */
+ getWsState?: () => string
+ /**
+ * Returns true when the caller cannot accept new work (transport already
+ * connected). When true, the loop polls at the configured at-capacity
+ * interval as a heartbeat only. Server-side BRIDGE_LAST_POLL_TTL is
+ * 4 hours โ anything shorter than that is sufficient for liveness.
+ */
+ isAtCapacity?: () => boolean
+ /**
+ * Produces a signal that aborts when capacity frees up (transport lost),
+ * merged with the loop signal. Used to interrupt the at-capacity sleep
+ * so recovery polling starts immediately.
+ */
+ capacitySignal?: () => CapacitySignal
+ /** Called on unrecoverable errors (e.g. server-side expiry) to trigger full teardown. */
+ onFatalError?: () => void
+ /** Poll interval config getter โ defaults to DEFAULT_POLL_CONFIG. */
+ getPollIntervalConfig?: () => PollIntervalConfig
+ /**
+ * Returns the current work ID and session ingress token for heartbeat.
+ * When null, heartbeat is not possible (no active work item).
+ */
+ getHeartbeatInfo?: () => {
+ environmentId: string
+ workId: string
+ sessionToken: string
+ } | null
+ /**
+ * Called when heartbeatWork throws BridgeFatalError (401/403/404/410 โ
+ * JWT expired or work item gone). Caller should tear down the transport
+ * + work state so isAtCapacity() flips to false and the loop fast-polls
+ * for the server's re-dispatched work item. When provided, the loop
+ * SKIPS the at-capacity backoff sleep (which would otherwise cause a
+ * ~10-minute dead window before recovery). When omitted, falls back to
+ * the backoff sleep to avoid a tight poll+heartbeat loop.
+ */
+ onHeartbeatFatal?: (err: BridgeFatalError) => void
+}): Promise {
+ const MAX_ENVIRONMENT_RECREATIONS = 3
+
+ logForDebugging(
+ `[bridge:repl] Starting work poll loop for env=${getCredentials().environmentId}`,
+ )
+
+ let consecutiveErrors = 0
+ let firstErrorTime: number | null = null
+ let lastPollErrorTime: number | null = null
+ let environmentRecreations = 0
+ // Set when the at-capacity sleep overruns its deadline by a large margin
+ // (process suspension). Consumed at the top of the next iteration to
+ // force one fast-poll cycle โ isAtCapacity() is `transport !== null`,
+ // which stays true while the transport auto-reconnects, so the poll
+ // loop would otherwise go straight back to a 10-minute sleep on a
+ // transport that may be pointed at a dead socket.
+ let suspensionDetected = false
+
+ while (!signal.aborted) {
+ // Capture credentials outside try so the catch block can detect
+ // whether a concurrent reconnection replaced the environment.
+ const { environmentId: envId, environmentSecret: envSecret } =
+ getCredentials()
+ const pollConfig = getPollIntervalConfig()
+ try {
+ const work = await api.pollForWork(
+ envId,
+ envSecret,
+ signal,
+ pollConfig.reclaim_older_than_ms,
+ )
+
+ // A successful poll proves the env is genuinely healthy โ reset the
+ // env-loss counter so events hours apart each start fresh. Outside
+ // the state-change guard below because onEnvLost's success path
+ // already emits 'ready'; emitting again here would be a duplicate.
+ // (onEnvLost returning creds does NOT reset this โ that would break
+ // oscillation protection when the new env immediately dies.)
+ environmentRecreations = 0
+
+ // Reset error tracking on successful poll
+ if (consecutiveErrors > 0) {
+ logForDebugging(
+ `[bridge:repl] Poll recovered after ${consecutiveErrors} consecutive error(s)`,
+ )
+ consecutiveErrors = 0
+ firstErrorTime = null
+ lastPollErrorTime = null
+ onStateChange?.('ready')
+ }
+
+ if (!work) {
+ // Read-and-clear: after a detected suspension, skip the at-capacity
+ // branch exactly once. The pollForWork above already refreshed the
+ // server's BRIDGE_LAST_POLL_TTL; this fast cycle gives any
+ // re-dispatched work item a chance to land before we go back under.
+ const skipAtCapacityOnce = suspensionDetected
+ suspensionDetected = false
+ if (isAtCapacity?.() && capacitySignal && !skipAtCapacityOnce) {
+ const atCapMs = pollConfig.poll_interval_ms_at_capacity
+ // Heartbeat loops WITHOUT polling. When at-capacity polling is also
+ // enabled (atCapMs > 0), the loop tracks a deadline and breaks out
+ // to poll at that interval โ heartbeat and poll compose instead of
+ // one suppressing the other. Breaks out when:
+ // - Poll deadline reached (atCapMs > 0 only)
+ // - Auth fails (JWT expired โ poll refreshes tokens)
+ // - Capacity wake fires (transport lost โ poll for new work)
+ // - Heartbeat config disabled (GrowthBook update)
+ // - Loop aborted (shutdown)
+ if (
+ pollConfig.non_exclusive_heartbeat_interval_ms > 0 &&
+ getHeartbeatInfo
+ ) {
+ logEvent('tengu_bridge_heartbeat_mode_entered', {
+ heartbeat_interval_ms:
+ pollConfig.non_exclusive_heartbeat_interval_ms,
+ })
+ // Deadline computed once at entry โ GB updates to atCapMs don't
+ // shift an in-flight deadline (next entry picks up the new value).
+ const pollDeadline = atCapMs > 0 ? Date.now() + atCapMs : null
+ let needsBackoff = false
+ let hbCycles = 0
+ while (
+ !signal.aborted &&
+ isAtCapacity() &&
+ (pollDeadline === null || Date.now() < pollDeadline)
+ ) {
+ const hbConfig = getPollIntervalConfig()
+ if (hbConfig.non_exclusive_heartbeat_interval_ms <= 0) break
+
+ const info = getHeartbeatInfo()
+ if (!info) break
+
+ // Capture capacity signal BEFORE the async heartbeat call so
+ // a transport loss during the HTTP request is caught by the
+ // subsequent sleep.
+ const cap = capacitySignal()
+
+ try {
+ await api.heartbeatWork(
+ info.environmentId,
+ info.workId,
+ info.sessionToken,
+ )
+ } catch (err) {
+ logForDebugging(
+ `[bridge:repl:heartbeat] Failed: ${errorMessage(err)}`,
+ )
+ if (err instanceof BridgeFatalError) {
+ cap.cleanup()
+ logEvent('tengu_bridge_heartbeat_error', {
+ status:
+ err.status as unknown as AnalyticsMetadata_I_VERIFIED_THIS_IS_NOT_CODE_OR_FILEPATHS,
+ error_type: (err.status === 401 || err.status === 403
+ ? 'auth_failed'
+ : 'fatal') as AnalyticsMetadata_I_VERIFIED_THIS_IS_NOT_CODE_OR_FILEPATHS,
+ })
+ // JWT expired (401/403) or work item gone (404/410).
+ // Either way the current transport is dead โ SSE
+ // reconnects and CCR writes will fail on the same
+ // stale token. If the caller gave us a recovery hook,
+ // tear down work state and skip backoff: isAtCapacity()
+ // flips to false, next outer-loop iteration fast-polls
+ // for the server's re-dispatched work item. Without
+ // the hook, backoff to avoid tight poll+heartbeat loop.
+ if (onHeartbeatFatal) {
+ onHeartbeatFatal(err)
+ logForDebugging(
+ `[bridge:repl:heartbeat] Fatal (status=${err.status}), work state cleared โ fast-polling for re-dispatch`,
+ )
+ } else {
+ needsBackoff = true
+ }
+ break
+ }
+ }
+
+ hbCycles++
+ await sleep(
+ hbConfig.non_exclusive_heartbeat_interval_ms,
+ cap.signal,
+ )
+ cap.cleanup()
+ }
+
+ const exitReason = needsBackoff
+ ? 'error'
+ : signal.aborted
+ ? 'shutdown'
+ : !isAtCapacity()
+ ? 'capacity_changed'
+ : pollDeadline !== null && Date.now() >= pollDeadline
+ ? 'poll_due'
+ : 'config_disabled'
+ logEvent('tengu_bridge_heartbeat_mode_exited', {
+ reason:
+ exitReason as AnalyticsMetadata_I_VERIFIED_THIS_IS_NOT_CODE_OR_FILEPATHS,
+ heartbeat_cycles: hbCycles,
+ })
+
+ // On auth_failed or fatal, backoff before polling to avoid a
+ // tight poll+heartbeat loop. Fall through to the shared sleep
+ // below โ it's the same capacitySignal-wrapped sleep the legacy
+ // path uses, and both need the suspension-overrun check.
+ if (!needsBackoff) {
+ if (exitReason === 'poll_due') {
+ // bridgeApi throttles empty-poll logs (EMPTY_POLL_LOG_INTERVAL=100)
+ // so the once-per-10min poll_due poll is invisible at counter=2.
+ // Log it here so verification runs see both endpoints in the debug log.
+ logForDebugging(
+ `[bridge:repl] Heartbeat poll_due after ${hbCycles} cycles โ falling through to pollForWork`,
+ )
+ }
+ continue
+ }
+ }
+ // At-capacity sleep โ reached by both the legacy path (heartbeat
+ // disabled) and the heartbeat-backoff path (needsBackoff=true).
+ // Merged so the suspension detector covers both; previously the
+ // backoff path had no overrun check and could go straight back
+ // under for 10 min after a laptop wake. Use atCapMs when enabled,
+ // else the heartbeat interval as a floor (guaranteed > 0 on the
+ // backoff path) so heartbeat-only configs don't tight-loop.
+ const sleepMs =
+ atCapMs > 0
+ ? atCapMs
+ : pollConfig.non_exclusive_heartbeat_interval_ms
+ if (sleepMs > 0) {
+ const cap = capacitySignal()
+ const sleepStart = Date.now()
+ await sleep(sleepMs, cap.signal)
+ cap.cleanup()
+ // Process-suspension detector. A setTimeout overshooting its
+ // deadline by 60s means the process was suspended (laptop lid,
+ // SIGSTOP, VM pause) โ even a pathological GC pause is seconds,
+ // not minutes. Early aborts (wakePollLoop โ cap.signal) produce
+ // overrun < 0 and fall through. Note: this only catches sleeps
+ // that outlast their deadline; WebSocketTransport's ping
+ // interval (10s granularity) is the primary detector for shorter
+ // suspensions. This is the backstop for when that detector isn't
+ // running (transport mid-reconnect, interval stopped).
+ const overrun = Date.now() - sleepStart - sleepMs
+ if (overrun > 60_000) {
+ logForDebugging(
+ `[bridge:repl] At-capacity sleep overran by ${Math.round(overrun / 1000)}s โ process suspension detected, forcing one fast-poll cycle`,
+ )
+ logEvent('tengu_bridge_repl_suspension_detected', {
+ overrun_ms: overrun,
+ })
+ suspensionDetected = true
+ }
+ }
+ } else {
+ await sleep(pollConfig.poll_interval_ms_not_at_capacity, signal)
+ }
+ continue
+ }
+
+ // Decode before type dispatch โ need the JWT for the explicit ack.
+ let secret
+ try {
+ secret = decodeWorkSecret(work.secret)
+ } catch (err) {
+ logForDebugging(
+ `[bridge:repl] Failed to decode work secret: ${errorMessage(err)}`,
+ )
+ logEvent('tengu_bridge_repl_work_secret_failed', {})
+ // Can't ack (needs the JWT we failed to decode). stopWork uses OAuth.
+ // Prevents XAUTOCLAIM re-delivering this poisoned item every cycle.
+ await api.stopWork(envId, work.id, false).catch(() => {})
+ continue
+ }
+
+ // Explicitly acknowledge to prevent redelivery. Non-fatal on failure:
+ // server re-delivers, and the onWorkReceived callback handles dedup.
+ logForDebugging(`[bridge:repl] Acknowledging workId=${work.id}`)
+ try {
+ await api.acknowledgeWork(envId, work.id, secret.session_ingress_token)
+ } catch (err) {
+ logForDebugging(
+ `[bridge:repl] Acknowledge failed workId=${work.id}: ${errorMessage(err)}`,
+ )
+ }
+
+ if (work.data.type === 'healthcheck') {
+ logForDebugging('[bridge:repl] Healthcheck received')
+ continue
+ }
+
+ if (work.data.type === 'session') {
+ const workSessionId = work.data.id
+ try {
+ validateBridgeId(workSessionId, 'session_id')
+ } catch {
+ logForDebugging(
+ `[bridge:repl] Invalid session_id in work: ${workSessionId}`,
+ )
+ continue
+ }
+
+ onWorkReceived(
+ workSessionId,
+ secret.session_ingress_token,
+ work.id,
+ secret.use_code_sessions === true,
+ )
+ logForDebugging('[bridge:repl] Work accepted, continuing poll loop')
+ }
+ } catch (err) {
+ if (signal.aborted) break
+
+ // Detect permanent "environment deleted" error โ no amount of
+ // retrying will recover. Re-register a new environment instead.
+ // Checked BEFORE the generic BridgeFatalError bail. pollForWork uses
+ // validateStatus: s => s < 500, so 404 is always wrapped into a
+ // BridgeFatalError by handleErrorStatus() โ never an axios-shaped
+ // error. The poll endpoint's only path param is the env ID; 404
+ // unambiguously means env-gone (no-work is a 200 with null body).
+ // The server sends error.type='not_found_error' (standard Anthropic
+ // API shape), not a bridge-specific string โ but status===404 is
+ // the real signal and survives body-shape changes.
+ if (
+ err instanceof BridgeFatalError &&
+ err.status === 404 &&
+ onEnvironmentLost
+ ) {
+ // If credentials have already been refreshed by a concurrent
+ // reconnection (e.g. WS close handler), the stale poll's error
+ // is expected โ skip onEnvironmentLost and retry with fresh creds.
+ const currentEnvId = getCredentials().environmentId
+ if (envId !== currentEnvId) {
+ logForDebugging(
+ `[bridge:repl] Stale poll error for old env=${envId}, current env=${currentEnvId} โ skipping onEnvironmentLost`,
+ )
+ consecutiveErrors = 0
+ firstErrorTime = null
+ continue
+ }
+
+ environmentRecreations++
+ logForDebugging(
+ `[bridge:repl] Environment deleted, attempting re-registration (attempt ${environmentRecreations}/${MAX_ENVIRONMENT_RECREATIONS})`,
+ )
+ logEvent('tengu_bridge_repl_env_lost', {
+ attempt: environmentRecreations,
+ } as AnalyticsMetadata_I_VERIFIED_THIS_IS_NOT_CODE_OR_FILEPATHS)
+
+ if (environmentRecreations > MAX_ENVIRONMENT_RECREATIONS) {
+ logForDebugging(
+ `[bridge:repl] Environment re-registration limit reached (${MAX_ENVIRONMENT_RECREATIONS}), giving up`,
+ )
+ onStateChange?.(
+ 'failed',
+ 'Environment deleted and re-registration limit reached',
+ )
+ onFatalError?.()
+ break
+ }
+
+ onStateChange?.('reconnecting', 'environment lost, recreating session')
+ const newCreds = await onEnvironmentLost()
+ // doReconnect() makes several sequential network calls (1-5s).
+ // If the user triggered teardown during that window, its internal
+ // abort checks return false โ but we need to re-check here to
+ // avoid emitting a spurious 'failed' + onFatalError() during
+ // graceful shutdown.
+ if (signal.aborted) break
+ if (newCreds) {
+ // Credentials are updated in the outer scope via
+ // reconnectEnvironmentWithSession โ getCredentials() will
+ // return the fresh values on the next poll iteration.
+ // Do NOT reset environmentRecreations here โ onEnvLost returning
+ // creds only proves we tried to fix it, not that the env is
+ // healthy. A successful poll (above) is the reset point; if the
+ // new env immediately dies again we still want the limit to fire.
+ consecutiveErrors = 0
+ firstErrorTime = null
+ onStateChange?.('ready')
+ logForDebugging(
+ `[bridge:repl] Re-registered environment: ${newCreds.environmentId}`,
+ )
+ continue
+ }
+
+ onStateChange?.(
+ 'failed',
+ 'Environment deleted and re-registration failed',
+ )
+ onFatalError?.()
+ break
+ }
+
+ // Fatal errors (401/403/404/410) โ no point retrying
+ if (err instanceof BridgeFatalError) {
+ const isExpiry = isExpiredErrorType(err.errorType)
+ const isSuppressible = isSuppressible403(err)
+ logForDebugging(
+ `[bridge:repl] Fatal poll error: ${err.message} (status=${err.status}, type=${err.errorType ?? 'unknown'})${isSuppressible ? ' (suppressed)' : ''}`,
+ )
+ logEvent('tengu_bridge_repl_fatal_error', {
+ status: err.status,
+ error_type:
+ err.errorType as AnalyticsMetadata_I_VERIFIED_THIS_IS_NOT_CODE_OR_FILEPATHS,
+ })
+ logForDiagnosticsNoPII(
+ isExpiry ? 'info' : 'error',
+ 'bridge_repl_fatal_error',
+ { status: err.status, error_type: err.errorType },
+ )
+ // Cosmetic 403 errors (e.g., external_poll_sessions scope,
+ // environments:manage permission) โ suppress user-visible error
+ // but always trigger teardown so cleanup runs.
+ if (!isSuppressible) {
+ onStateChange?.(
+ 'failed',
+ isExpiry
+ ? 'session expired ยท /remote-control to reconnect'
+ : err.message,
+ )
+ }
+ // Always trigger teardown โ matches bridgeMain.ts where fatalExit=true
+ // is unconditional and post-loop cleanup always runs.
+ onFatalError?.()
+ break
+ }
+
+ const now = Date.now()
+
+ // Detect system sleep/wake: if the gap since the last poll error
+ // greatly exceeds the max backoff delay, the machine likely slept.
+ // Reset error tracking so we retry with a fresh budget instead of
+ // immediately giving up.
+ if (
+ lastPollErrorTime !== null &&
+ now - lastPollErrorTime > POLL_ERROR_MAX_DELAY_MS * 2
+ ) {
+ logForDebugging(
+ `[bridge:repl] Detected system sleep (${Math.round((now - lastPollErrorTime) / 1000)}s gap), resetting poll error budget`,
+ )
+ logForDiagnosticsNoPII('info', 'bridge_repl_poll_sleep_detected', {
+ gapMs: now - lastPollErrorTime,
+ })
+ consecutiveErrors = 0
+ firstErrorTime = null
+ }
+ lastPollErrorTime = now
+
+ consecutiveErrors++
+ if (firstErrorTime === null) {
+ firstErrorTime = now
+ }
+ const elapsed = now - firstErrorTime
+ const httpStatus = extractHttpStatus(err)
+ const errMsg = describeAxiosError(err)
+ const wsLabel = getWsState?.() ?? 'unknown'
+
+ logForDebugging(
+ `[bridge:repl] Poll error (attempt ${consecutiveErrors}, elapsed ${Math.round(elapsed / 1000)}s, ws=${wsLabel}): ${errMsg}`,
+ )
+ logEvent('tengu_bridge_repl_poll_error', {
+ status: httpStatus,
+ consecutiveErrors,
+ elapsedMs: elapsed,
+ } as AnalyticsMetadata_I_VERIFIED_THIS_IS_NOT_CODE_OR_FILEPATHS)
+
+ // Only transition to 'reconnecting' on the first error โ stay
+ // there until a successful poll (avoid flickering the UI state).
+ if (consecutiveErrors === 1) {
+ onStateChange?.('reconnecting', errMsg)
+ }
+
+ // Give up after continuous failures
+ if (elapsed >= POLL_ERROR_GIVE_UP_MS) {
+ logForDebugging(
+ `[bridge:repl] Poll failures exceeded ${POLL_ERROR_GIVE_UP_MS / 1000}s (${consecutiveErrors} errors), giving up`,
+ )
+ logForDiagnosticsNoPII('info', 'bridge_repl_poll_give_up')
+ logEvent('tengu_bridge_repl_poll_give_up', {
+ consecutiveErrors,
+ elapsedMs: elapsed,
+ lastStatus: httpStatus,
+ } as AnalyticsMetadata_I_VERIFIED_THIS_IS_NOT_CODE_OR_FILEPATHS)
+ onStateChange?.('failed', 'connection to server lost')
+ break
+ }
+
+ // Exponential backoff: 2s โ 4s โ 8s โ 16s โ 32s โ 60s (cap)
+ const backoff = Math.min(
+ POLL_ERROR_INITIAL_DELAY_MS * 2 ** (consecutiveErrors - 1),
+ POLL_ERROR_MAX_DELAY_MS,
+ )
+ // The poll_due heartbeat-loop exit leaves a healthy lease exposed to
+ // this backoff path. Heartbeat before each sleep so /poll outages
+ // (the VerifyEnvironmentSecretAuth DB path heartbeat was introduced to
+ // avoid) don't kill the 300s lease TTL.
+ if (getPollIntervalConfig().non_exclusive_heartbeat_interval_ms > 0) {
+ const info = getHeartbeatInfo?.()
+ if (info) {
+ try {
+ await api.heartbeatWork(
+ info.environmentId,
+ info.workId,
+ info.sessionToken,
+ )
+ } catch {
+ // Best-effort โ if heartbeat also fails the lease dies, same as
+ // pre-poll_due behavior (where the only heartbeat-loop exits were
+ // ones where the lease was already dying).
+ }
+ }
+ }
+ await sleep(backoff, signal)
+ }
+ }
+
+ logForDebugging(
+ `[bridge:repl] Work poll loop ended (aborted=${signal.aborted}) env=${getCredentials().environmentId}`,
+ )
+}
+
+// Exported for testing only
+export {
+ startWorkPollLoop as _startWorkPollLoopForTesting,
+ POLL_ERROR_INITIAL_DELAY_MS as _POLL_ERROR_INITIAL_DELAY_MS_ForTesting,
+ POLL_ERROR_MAX_DELAY_MS as _POLL_ERROR_MAX_DELAY_MS_ForTesting,
+ POLL_ERROR_GIVE_UP_MS as _POLL_ERROR_GIVE_UP_MS_ForTesting,
+}
diff --git a/original-source-code/src/bridge/replBridgeHandle.ts b/original-source-code/src/bridge/replBridgeHandle.ts
new file mode 100644
index 0000000..f04d745
--- /dev/null
+++ b/original-source-code/src/bridge/replBridgeHandle.ts
@@ -0,0 +1,36 @@
+import { updateSessionBridgeId } from '../utils/concurrentSessions.js'
+import type { ReplBridgeHandle } from './replBridge.js'
+import { toCompatSessionId } from './sessionIdCompat.js'
+
+/**
+ * Global pointer to the active REPL bridge handle, so callers outside
+ * useReplBridge's React tree (tools, slash commands) can invoke handle methods
+ * like subscribePR. Same one-bridge-per-process justification as bridgeDebug.ts
+ * โ the handle's closure captures the sessionId and getAccessToken that created
+ * the session, and re-deriving those independently (BriefTool/upload.ts pattern)
+ * risks staging/prod token divergence.
+ *
+ * Set from useReplBridge.tsx when init completes; cleared on teardown.
+ */
+
+let handle: ReplBridgeHandle | null = null
+
+export function setReplBridgeHandle(h: ReplBridgeHandle | null): void {
+ handle = h
+ // Publish (or clear) our bridge session ID in the session record so other
+ // local peers can dedup us out of their bridge list โ local is preferred.
+ void updateSessionBridgeId(getSelfBridgeCompatId() ?? null).catch(() => {})
+}
+
+export function getReplBridgeHandle(): ReplBridgeHandle | null {
+ return handle
+}
+
+/**
+ * Our own bridge session ID in the session_* compat format the API returns
+ * in /v1/sessions responses โ or undefined if bridge isn't connected.
+ */
+export function getSelfBridgeCompatId(): string | undefined {
+ const h = getReplBridgeHandle()
+ return h ? toCompatSessionId(h.bridgeSessionId) : undefined
+}
diff --git a/original-source-code/src/bridge/replBridgeTransport.ts b/original-source-code/src/bridge/replBridgeTransport.ts
new file mode 100644
index 0000000..2a844f9
--- /dev/null
+++ b/original-source-code/src/bridge/replBridgeTransport.ts
@@ -0,0 +1,370 @@
+import type { StdoutMessage } from 'src/entrypoints/sdk/controlTypes.js'
+import { CCRClient } from '../cli/transports/ccrClient.js'
+import type { HybridTransport } from '../cli/transports/HybridTransport.js'
+import { SSETransport } from '../cli/transports/SSETransport.js'
+import { logForDebugging } from '../utils/debug.js'
+import { errorMessage } from '../utils/errors.js'
+import { updateSessionIngressAuthToken } from '../utils/sessionIngressAuth.js'
+import type { SessionState } from '../utils/sessionState.js'
+import { registerWorker } from './workSecret.js'
+
+/**
+ * Transport abstraction for replBridge. Covers exactly the surface that
+ * replBridge.ts uses against HybridTransport so the v1/v2 choice is
+ * confined to the construction site.
+ *
+ * - v1: HybridTransport (WS reads + POST writes to Session-Ingress)
+ * - v2: SSETransport (reads) + CCRClient (writes to CCR v2 /worker/*)
+ *
+ * The v2 write path goes through CCRClient.writeEvent โ SerialBatchEventUploader,
+ * NOT through SSETransport.write() โ SSETransport.write() targets the
+ * Session-Ingress POST URL shape, which is wrong for CCR v2.
+ */
+export type ReplBridgeTransport = {
+ write(message: StdoutMessage): Promise
+ writeBatch(messages: StdoutMessage[]): Promise
+ close(): void
+ isConnectedStatus(): boolean
+ getStateLabel(): string
+ setOnData(callback: (data: string) => void): void
+ setOnClose(callback: (closeCode?: number) => void): void
+ setOnConnect(callback: () => void): void
+ connect(): void
+ /**
+ * High-water mark of the underlying read stream's event sequence numbers.
+ * replBridge reads this before swapping transports so the new one can
+ * resume from where the old one left off (otherwise the server replays
+ * the entire session history from seq 0).
+ *
+ * v1 returns 0 โ Session-Ingress WS doesn't use SSE sequence numbers;
+ * replay-on-reconnect is handled by the server-side message cursor.
+ */
+ getLastSequenceNum(): number
+ /**
+ * Monotonic count of batches dropped via maxConsecutiveFailures.
+ * Snapshot before writeBatch() and compare after to detect silent drops
+ * (writeBatch() resolves normally even when batches were dropped).
+ * v2 returns 0 โ the v2 write path doesn't set maxConsecutiveFailures.
+ */
+ readonly droppedBatchCount: number
+ /**
+ * PUT /worker state (v2 only; v1 is a no-op). `requires_action` tells
+ * the backend a permission prompt is pending โ claude.ai shows the
+ * "waiting for input" indicator. REPL/daemon callers don't need this
+ * (user watches the REPL locally); multi-session worker callers do.
+ */
+ reportState(state: SessionState): void
+ /** PUT /worker external_metadata (v2 only; v1 is a no-op). */
+ reportMetadata(metadata: Record): void
+ /**
+ * POST /worker/events/{id}/delivery (v2 only; v1 is a no-op). Populates
+ * CCR's processing_at/processed_at columns. `received` is auto-fired by
+ * CCRClient on every SSE frame and is not exposed here.
+ */
+ reportDelivery(eventId: string, status: 'processing' | 'processed'): void
+ /**
+ * Drain the write queue before close() (v2 only; v1 resolves
+ * immediately โ HybridTransport POSTs are already awaited per-write).
+ */
+ flush(): Promise
+}
+
+/**
+ * v1 adapter: HybridTransport already has the full surface (it extends
+ * WebSocketTransport which has setOnConnect + getStateLabel). This is a
+ * no-op wrapper that exists only so replBridge's `transport` variable
+ * has a single type.
+ */
+export function createV1ReplTransport(
+ hybrid: HybridTransport,
+): ReplBridgeTransport {
+ return {
+ write: msg => hybrid.write(msg),
+ writeBatch: msgs => hybrid.writeBatch(msgs),
+ close: () => hybrid.close(),
+ isConnectedStatus: () => hybrid.isConnectedStatus(),
+ getStateLabel: () => hybrid.getStateLabel(),
+ setOnData: cb => hybrid.setOnData(cb),
+ setOnClose: cb => hybrid.setOnClose(cb),
+ setOnConnect: cb => hybrid.setOnConnect(cb),
+ connect: () => void hybrid.connect(),
+ // v1 Session-Ingress WS doesn't use SSE sequence numbers; replay
+ // semantics are different. Always return 0 so the seq-num carryover
+ // logic in replBridge is a no-op for v1.
+ getLastSequenceNum: () => 0,
+ get droppedBatchCount() {
+ return hybrid.droppedBatchCount
+ },
+ reportState: () => {},
+ reportMetadata: () => {},
+ reportDelivery: () => {},
+ flush: () => Promise.resolve(),
+ }
+}
+
+/**
+ * v2 adapter: wrap SSETransport (reads) + CCRClient (writes, heartbeat,
+ * state, delivery tracking).
+ *
+ * Auth: v2 endpoints validate the JWT's session_id claim (register_worker.go:32)
+ * and worker role (environment_auth.py:856). OAuth tokens have neither.
+ * This is the inverse of the v1 replBridge path, which deliberately uses OAuth.
+ * The JWT is refreshed when the poll loop re-dispatches work โ the caller
+ * invokes createV2ReplTransport again with the fresh token.
+ *
+ * Registration happens here (not in the caller) so the entire v2 handshake
+ * is one async step. registerWorker failure propagates โ replBridge will
+ * catch it and stay on the poll loop.
+ */
+export async function createV2ReplTransport(opts: {
+ sessionUrl: string
+ ingressToken: string
+ sessionId: string
+ /**
+ * SSE sequence-number high-water mark from the previous transport.
+ * Passed to the new SSETransport so its first connect() sends
+ * from_sequence_num / Last-Event-ID and the server resumes from where
+ * the old stream left off. Without this, every transport swap asks the
+ * server to replay the entire session history from seq 0.
+ */
+ initialSequenceNum?: number
+ /**
+ * Worker epoch from POST /bridge response. When provided, the server
+ * already bumped epoch (the /bridge call IS the register โ see server
+ * PR #293280). When omitted (v1 CCR-v2 path via replBridge.ts poll loop),
+ * call registerWorker as before.
+ */
+ epoch?: number
+ /** CCRClient heartbeat interval. Defaults to 20s when omitted. */
+ heartbeatIntervalMs?: number
+ /** ยฑfraction per-beat jitter. Defaults to 0 (no jitter) when omitted. */
+ heartbeatJitterFraction?: number
+ /**
+ * When true, skip opening the SSE read stream โ only the CCRClient write
+ * path is activated. Use for mirror-mode attachments that forward events
+ * but never receive inbound prompts or control requests.
+ */
+ outboundOnly?: boolean
+ /**
+ * Per-instance auth header source. When provided, CCRClient + SSETransport
+ * read auth from this closure instead of the process-wide
+ * CLAUDE_CODE_SESSION_ACCESS_TOKEN env var. Required for callers managing
+ * multiple concurrent sessions โ the env-var path stomps across sessions.
+ * When omitted, falls back to the env var (single-session callers).
+ */
+ getAuthToken?: () => string | undefined
+}): Promise