Extracted from npm package
@anthropic-ai/claude-codeversion 2.1.88. The published package ships a single bundledcli.js(~12MB). Thesrc/directory in this repo contains the unbundled TypeScript source extracted from the npm tarball.
This source is incomplete. 108 modules referenced by
feature()-gated branches are not included in the npm package. They exist only in Anthropic's internal monorepo and were dead-code-eliminated at compile time. They cannot be recovered fromcli.js,sdk-tools.d.ts, or any published artifact.
These modules have no source files anywhere in the npm package. They are internal Anthropic infrastructure.
| Module | Purpose | Feature Gate |
|---|---|---|
daemon/main.js |
Background daemon supervisor | DAEMON |
daemon/workerRegistry.js |
Daemon worker registry | DAEMON |
proactive/index.js |
Proactive notification system | PROACTIVE |
contextCollapse/index.js |
Context collapse service (experimental) | CONTEXT_COLLAPSE |
contextCollapse/operations.js |
Collapse operations | CONTEXT_COLLAPSE |
contextCollapse/persist.js |
Collapse persistence | CONTEXT_COLLAPSE |
skillSearch/featureCheck.js |
Remote skill feature check | EXPERIMENTAL_SKILL_SEARCH |
skillSearch/remoteSkillLoader.js |
Remote skill loader | EXPERIMENTAL_SKILL_SEARCH |
skillSearch/remoteSkillState.js |
Remote skill state | EXPERIMENTAL_SKILL_SEARCH |
skillSearch/telemetry.js |
Skill search telemetry | EXPERIMENTAL_SKILL_SEARCH |
skillSearch/localSearch.js |
Local skill search | EXPERIMENTAL_SKILL_SEARCH |
skillSearch/prefetch.js |
Skill prefetch | EXPERIMENTAL_SKILL_SEARCH |
coordinator/workerAgent.js |
Multi-agent coordinator worker | COORDINATOR_MODE |
bridge/peerSessions.js |
Bridge peer session management | BRIDGE_MODE |
assistant/index.js |
Kairos assistant mode | KAIROS |
assistant/AssistantSessionChooser.js |
Assistant session picker | KAIROS |
compact/reactiveCompact.js |
Reactive context compaction | CACHED_MICROCOMPACT |
compact/snipCompact.js |
Snip-based compaction | HISTORY_SNIP |
compact/snipProjection.js |
Snip projection | HISTORY_SNIP |
compact/cachedMCConfig.js |
Cached micro-compact config | CACHED_MICROCOMPACT |
sessionTranscript/sessionTranscript.js |
Session transcript service | TRANSCRIPT_CLASSIFIER |
commands/agents-platform/index.js |
Internal agents platform | ant (internal) |
commands/assistant/index.js |
Assistant command | KAIROS |
commands/buddy/index.js |
Buddy system notifications | BUDDY |
commands/fork/index.js |
Fork subagent command | FORK_SUBAGENT |
commands/peers/index.js |
Multi-peer commands | BRIDGE_MODE |
commands/proactive.js |
Proactive command | PROACTIVE |
commands/remoteControlServer/index.js |
Remote control server | DAEMON + BRIDGE_MODE |
commands/subscribe-pr.js |
GitHub PR subscription | KAIROS_GITHUB_WEBHOOKS |
commands/torch.js |
Internal debug tool | TORCH |
commands/workflows/index.js |
Workflow commands | WORKFLOW_SCRIPTS |
jobs/classifier.js |
Internal job classifier | TEMPLATES |
memdir/memoryShapeTelemetry.js |
Memory shape telemetry | MEMORY_SHAPE_TELEMETRY |
services/sessionTranscript/sessionTranscript.js |
Session transcript | TRANSCRIPT_CLASSIFIER |
tasks/LocalWorkflowTask/LocalWorkflowTask.js |
Local workflow task | WORKFLOW_SCRIPTS |
protectedNamespace.js |
Internal namespace guard | ant (internal) |
protectedNamespace.js (envUtils) |
Protected namespace runtime | ant (internal) |
coreTypes.generated.js |
Generated core types | ant (internal) |
devtools.js |
Internal dev tools | ant (internal) |
attributionHooks.js |
Internal attribution hooks | COMMIT_ATTRIBUTION |
systemThemeWatcher.js |
System theme watcher | AUTO_THEME |
udsClient.js / udsMessaging.js |
UDS messaging client | UDS_INBOX |
systemThemeWatcher.js |
Theme watcher | AUTO_THEME |
These tools have type signatures in sdk-tools.d.ts but their implementations were stripped at compile time.
| Tool | Purpose | Feature Gate |
|---|---|---|
REPLTool |
Interactive REPL (VM sandbox) | ant (internal) |
SnipTool |
Context snipping | HISTORY_SNIP |
SleepTool |
Sleep/delay in agent loop | PROACTIVE / KAIROS |
MonitorTool |
MCP monitoring | MONITOR_TOOL |
OverflowTestTool |
Overflow testing | OVERFLOW_TEST_TOOL |
WorkflowTool |
Workflow execution | WORKFLOW_SCRIPTS |
WebBrowserTool |
Browser automation | WEB_BROWSER_TOOL |
TerminalCaptureTool |
Terminal capture | TERMINAL_PANEL |
TungstenTool |
Internal perf monitoring | ant (internal) |
VerifyPlanExecutionTool |
Plan verification | CLAUDE_CODE_VERIFY_PLAN |
SendUserFileTool |
Send files to users | KAIROS |
SubscribePRTool |
GitHub PR subscription | KAIROS_GITHUB_WEBHOOKS |
SuggestBackgroundPRTool |
Suggest background PRs | KAIROS |
PushNotificationTool |
Push notifications | KAIROS |
CtxInspectTool |
Context inspection | CONTEXT_COLLAPSE |
ListPeersTool |
List active peers | UDS_INBOX |
DiscoverSkillsTool |
Skill discovery | EXPERIMENTAL_SKILL_SEARCH |
These are internal prompt templates and documentation, never published.
| File | Purpose |
|---|---|
yolo-classifier-prompts/auto_mode_system_prompt.txt |
Auto-mode system prompt for classifier |
yolo-classifier-prompts/permissions_anthropic.txt |
Anthropic-internal permission prompt |
yolo-classifier-prompts/permissions_external.txt |
External user permission prompt |
verify/SKILL.md |
Verification skill documentation |
verify/examples/cli.md |
CLI verification examples |
verify/examples/server.md |
Server verification examples |
Anthropic Internal Monorepo Published npm Package
────────────────────────── ─────────────────────
feature('DAEMON') → true ──build──→ feature('DAEMON') → false
↓ ↓
daemon/main.js ← INCLUDED ──bundle─→ daemon/main.js ← DELETED (DCE)
tools/REPLTool ← INCLUDED ──bundle─→ tools/REPLTool ← DELETED (DCE)
proactive/ ← INCLUDED ──bundle─→ (referenced but absent from src/)
Bun's feature() is a compile-time intrinsic:
- Returns
truein Anthropic's internal build → code is kept in the bundle - Returns
falsein the published build → code is dead-code-eliminated - The 108 modules simply do not exist anywhere in the published artifact
Copyright (c) Anthropic PBC. All rights reserved.
All source code in this repository is the intellectual property of Anthropic.
This repository is provided strictly for technical research and educational purposes.
Commercial use is strictly prohibited.
If you are the copyright owner and believe this repository infringes your rights,
please contact the repository owner for immediate removal.
| Item | Count |
|---|---|
| Source files (.ts/.tsx) | ~1,884 |
| Lines of code | ~512,664 |
| Largest single file | query.ts (~785KB) |
| Built-in tools | ~40+ |
| Slash commands | ~80+ |
| Dependencies (node_modules) | ~192 packages |
| Runtime | Bun (compiled to Node.js >= 18 bundle) |
THE CORE LOOP
=============
User --> messages[] --> Claude API --> response
|
stop_reason == "tool_use"?
/ \
yes no
| |
execute tools return text
append tool_result
loop back -----------------> messages[]
That is the minimal agent loop. Claude Code wraps this loop
with a production-grade harness: permissions, streaming,
concurrency, compaction, sub-agents, persistence, and MCP.
src/
├── main.tsx # REPL bootstrap, 4,683 lines
├── QueryEngine.ts # SDK/headless query lifecycle engine
├── query.ts # Main agent loop (785KB, largest file)
├── Tool.ts # Tool interface + buildTool factory
├── Task.ts # Task types, IDs, state base
├── tools.ts # Tool registry, presets, filtering
├── commands.ts # Slash command definitions
├── context.ts # User input context
├── cost-tracker.ts # API cost accumulation
├── setup.ts # First-run setup flow
│
├── bridge/ # Claude Desktop / remote bridge
│ ├── bridgeMain.ts # Session lifecycle manager
│ ├── bridgeApi.ts # HTTP client
│ ├── bridgeConfig.ts # Connection config
│ ├── bridgeMessaging.ts # Message relay
│ ├── sessionRunner.ts # Process spawning
│ ├── jwtUtils.ts # JWT refresh
│ ├── workSecret.ts # Auth tokens
│ └── capacityWake.ts # Capacity-based wakeup
│
├── cli/ # CLI infrastructure
│ ├── handlers/ # Command handlers
│ └── transports/ # I/O transports (stdio, structured)
│
├── commands/ # ~80 slash commands
│ ├── agents/ # Agent management
│ ├── compact/ # Context compaction
│ ├── config/ # Settings management
│ ├── help/ # Help display
│ ├── login/ # Authentication
│ ├── mcp/ # MCP server management
│ ├── memory/ # Memory system
│ ├── plan/ # Plan mode
│ ├── resume/ # Session resume
│ ├── review/ # Code review
│ └── ... # 70+ more commands
│
├── components/ # React/Ink terminal UI
│ ├── design-system/ # Reusable UI primitives
│ ├── messages/ # Message rendering
│ ├── permissions/ # Permission dialogs
│ ├── PromptInput/ # Input field + suggestions
│ ├── LogoV2/ # Branding + welcome screen
│ ├── Settings/ # Settings panels
│ ├── Spinner.tsx # Loading indicators
│ └── ... # 40+ component groups
│
├── entrypoints/ # Application entry points
│ ├── cli.tsx # CLI main (version, help, daemon)
│ ├── sdk/ # Agent SDK (types, sessions)
│ └── mcp.ts # MCP server entry
│
├── hooks/ # React hooks
│ ├── useCanUseTool.tsx # Permission checking
│ ├── useReplBridge.tsx # Bridge connection
│ ├── notifs/ # Notification hooks
│ └── toolPermission/ # Tool permission handlers
│
├── services/ # Business logic layer
│ ├── api/ # Claude API client
│ │ ├── claude.ts # Streaming API calls
│ │ ├── errors.ts # Error categorization
│ │ └── withRetry.ts # Retry logic
│ ├── analytics/ # Telemetry + GrowthBook
│ ├── compact/ # Context compression
│ ├── mcp/ # MCP connection management
│ ├── tools/ # Tool execution engine
│ │ ├── StreamingToolExecutor.ts # Parallel tool runner
│ │ └── toolOrchestration.ts # Batch orchestration
│ ├── plugins/ # Plugin loader
│ └── settingsSync/ # Cross-device settings
│
├── state/ # Application state
│ ├── AppStateStore.ts # Store definition
│ └── AppState.tsx # React provider + hooks
│
├── tasks/ # Task implementations
│ ├── LocalShellTask/ # Bash command execution
│ ├── LocalAgentTask/ # Sub-agent execution
│ ├── RemoteAgentTask/ # Remote agent via bridge
│ ├── InProcessTeammateTask/ # In-process teammate
│ └── DreamTask/ # Background thinking
│
├── tools/ # 40+ tool implementations
│ ├── AgentTool/ # Sub-agent spawning + fork
│ ├── BashTool/ # Shell command execution
│ ├── FileReadTool/ # File reading (PDF, image, etc)
│ ├── FileEditTool/ # String-replace editing
│ ├── FileWriteTool/ # Full file creation
│ ├── GlobTool/ # File pattern search
│ ├── GrepTool/ # Content search (ripgrep)
│ ├── WebFetchTool/ # HTTP fetching
│ ├── WebSearchTool/ # Web search
│ ├── MCPTool/ # MCP tool wrapper
│ ├── SkillTool/ # Skill invocation
│ ├── AskUserQuestionTool/ # User interaction
│ └── ... # 30+ more tools
│
├── types/ # Type definitions
│ ├── message.ts # Message discriminated unions
│ ├── permissions.ts # Permission types
│ ├── tools.ts # Tool progress types
│ └── ids.ts # Branded ID types
│
├── utils/ # Utilities (largest directory)
│ ├── permissions/ # Permission rule engine
│ ├── messages/ # Message formatting
│ ├── model/ # Model selection logic
│ ├── settings/ # Settings management
│ ├── sandbox/ # Sandbox runtime adapter
│ ├── hooks/ # Hook execution
│ ├── memory/ # Memory system utils
│ ├── git/ # Git operations
│ ├── github/ # GitHub API
│ ├── bash/ # Bash execution helpers
│ ├── swarm/ # Multi-agent swarm
│ ├── telemetry/ # Telemetry reporting
│ └── ... # 30+ more util groups
│
└── vendor/ # Native module source stubs
├── audio-capture-src/ # Audio input
├── image-processor-src/ # Image processing
├── modifiers-napi-src/ # Native modifiers
└── url-handler-src/ # URL handling
┌─────────────────────────────────────────────────────────────────────┐
│ ENTRY LAYER │
│ cli.tsx ──> main.tsx ──> REPL.tsx (interactive) │
│ └──> QueryEngine.ts (headless/SDK) │
└──────────────────────────────┬──────────────────────────────────────┘
│
▼
┌─────────────────────────────────────────────────────────────────────┐
│ QUERY ENGINE │
│ submitMessage(prompt) ──> AsyncGenerator<SDKMessage> │
│ │ │
│ ├── fetchSystemPromptParts() ──> assemble system prompt │
│ ├── processUserInput() ──> handle /commands │
│ ├── query() ──> main agent loop │
│ │ ├── StreamingToolExecutor ──> parallel tool execution │
│ │ ├── autoCompact() ──> context compression │
│ │ └── runTools() ──> tool orchestration │
│ └── yield SDKMessage ──> stream to consumer │
└──────────────────────────────┬──────────────────────────────────────┘
│
┌────────────────┼────────────────┐
▼ ▼ ▼
┌──────────────────┐ ┌─────────────────┐ ┌──────────────────┐
│ TOOL SYSTEM │ │ SERVICE LAYER │ │ STATE LAYER │
│ │ │ │ │ │
│ Tool Interface │ │ api/claude.ts │ │ AppState Store │
│ ├─ call() │ │ API client │ │ ├─ permissions │
│ ├─ validate() │ │ compact/ │ │ ├─ fileHistory │
│ ├─ checkPerms() │ │ auto-compact │ │ ├─ agents │
│ ├─ render() │ │ mcp/ │ │ └─ fastMode │
│ └─ prompt() │ │ MCP protocol │ │ │
│ │ │ analytics/ │ │ React Context │
│ 40+ Built-in: │ │ telemetry │ │ ├─ useAppState │
│ ├─ BashTool │ │ tools/ │ │ └─ useSetState │
│ ├─ FileRead │ │ executor │ │ │
│ ├─ FileEdit │ │ plugins/ │ └──────────────────┘
│ ├─ Glob/Grep │ │ loader │
│ ├─ AgentTool │ │ settingsSync/ │
│ ├─ WebFetch │ │ cross-device │
│ └─ MCPTool │ │ oauth/ │
│ │ │ auth flow │
└──────────────────┘ └─────────────────┘
│ │
▼ ▼
┌──────────────────┐ ┌─────────────────┐
│ TASK SYSTEM │ │ BRIDGE LAYER │
│ │ │ │
│ Task Types: │ │ bridgeMain.ts │
│ ├─ local_bash │ │ session mgmt │
│ ├─ local_agent │ │ bridgeApi.ts │
│ ├─ remote_agent │ │ HTTP client │
│ ├─ in_process │ │ workSecret.ts │
│ ├─ dream │ │ auth tokens │
│ └─ workflow │ │ sessionRunner │
│ │ │ process spawn │
│ ID: prefix+8chr │ └─────────────────┘
│ b=bash a=agent │
│ r=remote t=team │
└──────────────────┘
USER INPUT (prompt / slash command)
│
▼
processUserInput() ← parse /commands, build UserMessage
│
▼
fetchSystemPromptParts() ← tools → prompt sections, CLAUDE.md memory
│
▼
recordTranscript() ← persist user message to disk (JSONL)
│
▼
┌─→ normalizeMessagesForAPI() ← strip UI-only fields, compact if needed
│ │
│ ▼
│ Claude API (streaming) ← POST /v1/messages with tools + system prompt
│ │
│ ▼
│ stream events ← message_start → content_block_delta → message_stop
│ │
│ ├─ text block ──────────────→ yield to consumer (SDK / REPL)
│ │
│ └─ tool_use block?
│ │
│ ▼
│ StreamingToolExecutor ← partition: concurrent-safe vs serial
│ │
│ ▼
│ canUseTool() ← permission check (hooks + rules + UI prompt)
│ │
│ ├─ DENY ────────────────→ append tool_result(error), continue loop
│ │
│ └─ ALLOW
│ │
│ ▼
│ tool.call() ← execute the tool (Bash, Read, Edit, etc.)
│ │
│ ▼
│ append tool_result ← push to messages[], recordTranscript()
│ │
└─────────┘ ← loop back to API call
│
▼ (stop_reason != "tool_use")
yield result message ← final text, usage, cost, session_id
TOOL INTERFACE
==============
buildTool(definition) ──> Tool<Input, Output, Progress>
Every tool implements:
┌────────────────────────────────────────────────────────┐
│ LIFECYCLE │
│ ├── validateInput() → reject bad args early │
│ ├── checkPermissions() → tool-specific authz │
│ └── call() → execute and return result │
│ │
│ CAPABILITIES │
│ ├── isEnabled() → feature gate check │
│ ├── isConcurrencySafe() → can run in parallel? │
│ ├── isReadOnly() → no side effects? │
│ ├── isDestructive() → irreversible ops? │
│ └── interruptBehavior() → cancel or block on user? │
│ │
│ RENDERING (React/Ink) │
│ ├── renderToolUseMessage() → input display │
│ ├── renderToolResultMessage() → output display │
│ ├── renderToolUseProgressMessage() → spinner/status │
│ └── renderGroupedToolUse() → parallel tool groups │
│ │
│ AI FACING │
│ ├── prompt() → tool description for LLM │
│ ├── description() → dynamic description │
│ └── mapToolResultToAPI() → format for API response │
└────────────────────────────────────────────────────────┘
FILE OPERATIONS SEARCH & DISCOVERY EXECUTION
═════════════════ ══════════════════════ ══════════
FileReadTool GlobTool BashTool
FileEditTool GrepTool PowerShellTool
FileWriteTool ToolSearchTool
NotebookEditTool INTERACTION
═══════════
WEB & NETWORK AGENT / TASK AskUserQuestionTool
════════════════ ══════════════════ BriefTool
WebFetchTool AgentTool
WebSearchTool SendMessageTool PLANNING & WORKFLOW
TeamCreateTool ════════════════════
MCP PROTOCOL TeamDeleteTool EnterPlanModeTool
══════════════ TaskCreateTool ExitPlanModeTool
MCPTool TaskGetTool EnterWorktreeTool
ListMcpResourcesTool TaskUpdateTool ExitWorktreeTool
ReadMcpResourceTool TaskListTool TodoWriteTool
TaskStopTool
TaskOutputTool SYSTEM
════════
SKILLS & EXTENSIONS ConfigTool
═════════════════════ SkillTool
SkillTool ScheduleCronTool
LSPTool SleepTool
TungstenTool
TOOL CALL REQUEST
│
▼
┌─ validateInput() ──────────────────────────────────┐
│ reject invalid inputs before any permission check │
└────────────────────┬───────────────────────────────┘
│
▼
┌─ PreToolUse Hooks ─────────────────────────────────┐
│ user-defined shell commands (settings.json hooks) │
│ can: approve, deny, or modify input │
└────────────────────┬───────────────────────────────┘
│
▼
┌─ Permission Rules ─────────────────────────────────┐
│ alwaysAllowRules: match tool name/pattern → auto │
│ alwaysDenyRules: match tool name/pattern → deny │
│ alwaysAskRules: match tool name/pattern → ask │
│ Sources: settings, CLI args, session decisions │
└────────────────────┬───────────────────────────────┘
│
no rule match?
│
▼
┌─ Interactive Prompt ───────────────────────────────┐
│ User sees tool name + input │
│ Options: Allow Once / Allow Always / Deny │
└────────────────────┬───────────────────────────────┘
│
▼
┌─ checkPermissions() ───────────────────────────────┐
│ Tool-specific logic (e.g. path sandboxing) │
└────────────────────┬───────────────────────────────┘
│
APPROVED → tool.call()
MAIN AGENT
==========
│
┌───────────────┼───────────────┐
▼ ▼ ▼
┌──────────────┐ ┌──────────┐ ┌──────────────┐
│ FORK AGENT │ │ REMOTE │ │ IN-PROCESS │
│ │ │ AGENT │ │ TEAMMATE │
│ Fork process │ │ Bridge │ │ Same process │
│ Shared cache │ │ session │ │ Async context│
│ Fresh msgs[] │ │ Isolated │ │ Shared state │
└──────────────┘ └──────────┘ └──────────────┘
SPAWN MODES:
├─ default → in-process, shared conversation
├─ fork → child process, fresh messages[], shared file cache
├─ worktree → isolated git worktree + fork
└─ remote → bridge to Claude Code Remote / container
COMMUNICATION:
├─ SendMessageTool → agent-to-agent messages
├─ TaskCreate/Update → shared task board
└─ TeamCreate/Delete → team lifecycle management
SWARM MODE (feature-gated):
┌─────────────────────────────────────────────┐
│ Lead Agent │
│ ├── Teammate A ──> claims Task 1 │
│ ├── Teammate B ──> claims Task 2 │
│ └── Teammate C ──> claims Task 3 │
│ │
│ Shared: task board, message inbox │
│ Isolated: messages[], file cache, cwd │
└─────────────────────────────────────────────┘
CONTEXT WINDOW BUDGET
═════════════════════
┌─────────────────────────────────────────────────────┐
│ System Prompt (tools, permissions, CLAUDE.md) │
│ ══════════════════════════════════════════════ │
│ │
│ Conversation History │
│ ┌─────────────────────────────────────────────┐ │
│ │ [compacted summary of older messages] │ │
│ │ ═══════════════════════════════════════════ │ │
│ │ [compact_boundary marker] │ │
│ │ ─────────────────────────────────────────── │ │
│ │ [recent messages — full fidelity] │ │
│ │ user → assistant → tool_use → tool_result │ │
│ └─────────────────────────────────────────────┘ │
│ │
│ Current Turn (user + assistant response) │
└─────────────────────────────────────────────────────┘
THREE COMPRESSION STRATEGIES:
├─ autoCompact → triggers when token count exceeds threshold
│ summarizes old messages via a compact API call
├─ snipCompact → removes zombie messages and stale markers
│ (HISTORY_SNIP feature flag)
└─ contextCollapse → restructures context for efficiency
(CONTEXT_COLLAPSE feature flag)
COMPACTION FLOW:
messages[] ──> getMessagesAfterCompactBoundary()
│
▼
older messages ──> Claude API (summarize) ──> compact summary
│
▼
[summary] + [compact_boundary] + [recent messages]
┌─────────────────────────────────────────────────────────┐
│ MCP ARCHITECTURE │
│ │
│ MCPConnectionManager.tsx │
│ ├── Server Discovery (config from settings.json) │
│ │ ├── stdio → spawn child process │
│ │ ├── sse → HTTP EventSource │
│ │ ├── http → Streamable HTTP │
│ │ ├── ws → WebSocket │
│ │ └── sdk → in-process transport │
│ │ │
│ ├── Client Lifecycle │
│ │ ├── connect → initialize → list tools │
│ │ ├── tool calls via MCPTool wrapper │
│ │ └── disconnect / reconnect with backoff │
│ │ │
│ ├── Authentication │
│ │ ├── OAuth 2.0 flow (McpOAuthConfig) │
│ │ ├── Cross-App Access (XAA / SEP-990) │
│ │ └── API key via headers │
│ │ │
│ └── Tool Registration │
│ ├── mcp__<server>__<tool> naming convention │
│ ├── Dynamic schema from MCP server │
│ ├── Permission passthrough to Claude Code │
│ └── Resource listing (ListMcpResourcesTool) │
│ │
└─────────────────────────────────────────────────────────┘
Claude Desktop / Web / Cowork Claude Code CLI
══════════════════════════ ═════════════════
┌───────────────────┐ ┌──────────────────┐
│ Bridge Client │ ←─ HTTP ──→ │ bridgeMain.ts │
│ (Desktop App) │ │ │
└───────────────────┘ │ Session Manager │
│ ├── spawn CLI │
PROTOCOL: │ ├── poll status │
├─ JWT authentication │ ├── relay msgs │
├─ Work secret exchange │ └── capacityWake │
├─ Session lifecycle │ │
│ ├── create │ Backoff: │
│ ├── run │ ├─ conn: 2s→2m │
│ └─ stop │ └─ gen: 500ms→30s│
└─ Token refresh scheduler └──────────────────┘
SESSION STORAGE
══════════════
~/.claude/projects/<hash>/sessions/
└── <session-id>.jsonl ← append-only log
├── {"type":"user",...}
├── {"type":"assistant",...}
├── {"type":"progress",...}
└── {"type":"system","subtype":"compact_boundary",...}
RESUME FLOW:
getLastSessionLog() ──> parse JSONL ──> rebuild messages[]
│
├── --continue → last session in cwd
├── --resume <id> → specific session
└── --fork-session → new ID, copy history
PERSISTENCE STRATEGY:
├─ User messages → await write (blocking, for crash recovery)
├─ Assistant msgs → fire-and-forget (order-preserving queue)
├─ Progress → inline write (dedup on next query)
└─ Flush → on result yield / cowork eager flush
DEAD CODE ELIMINATION (Bun compile-time)
══════════════════════════════════════════
feature('FLAG_NAME') ──→ true → included in bundle
──→ false → stripped from bundle
FLAGS (observed in source):
├─ COORDINATOR_MODE → multi-agent coordinator
├─ HISTORY_SNIP → aggressive history trimming
├─ CONTEXT_COLLAPSE → context restructuring
├─ DAEMON → background daemon workers
├─ AGENT_TRIGGERS → cron/remote triggers
├─ AGENT_TRIGGERS_REMOTE → remote trigger support
├─ MONITOR_TOOL → MCP monitoring tool
├─ WEB_BROWSER_TOOL → browser automation
├─ VOICE_MODE → voice input/output
├─ TEMPLATES → job classifier
├─ EXPERIMENTAL_SKILL_SEARCH → skill discovery
├─ KAIROS → push notifications, file sends
├─ PROACTIVE → sleep tool, proactive behavior
├─ OVERFLOW_TEST_TOOL → testing tool
├─ TERMINAL_PANEL → terminal capture
├─ WORKFLOW_SCRIPTS → workflow tool
├─ CHICAGO_MCP → computer use MCP
├─ DUMP_SYSTEM_PROMPT → prompt extraction (ant-only)
├─ UDS_INBOX → peer discovery
├─ ABLATION_BASELINE → experiment ablation
└─ UPGRADE_NOTICE → upgrade notifications
RUNTIME GATES:
├─ process.env.USER_TYPE === 'ant' → Anthropic-internal features
└─ GrowthBook feature flags → A/B experiments at runtime
┌──────────────────────────────────────────────────────────┐
│ AppState Store │
│ │
│ AppState { │
│ toolPermissionContext: { │
│ mode: PermissionMode, ← default/plan/etc │
│ additionalWorkingDirectories, │
│ alwaysAllowRules, ← auto-approve │
│ alwaysDenyRules, ← auto-reject │
│ alwaysAskRules, ← always prompt │
│ isBypassPermissionsModeAvailable │
│ }, │
│ fileHistory: FileHistoryState, ← undo snapshots │
│ attribution: AttributionState, ← commit tracking │
│ verbose: boolean, │
│ mainLoopModel: string, ← active model │
│ fastMode: FastModeState, │
│ speculation: SpeculationState │
│ } │
│ │
│ React Integration: │
│ ├── AppStateProvider → creates store via createContext │
│ ├── useAppState(sel) → selector-based subscriptions │
│ └── useSetAppState() → immer-style updater function │
└──────────────────────────────────────────────────────────┘
This source code demonstrates 12 layered mechanisms that a production AI agent harness needs beyond the basic loop. Each builds on the previous:
s01 THE LOOP "One loop & Bash is all you need"
query.ts: the while-true loop that calls Claude API,
checks stop_reason, executes tools, appends results.
s02 TOOL DISPATCH "Adding a tool = adding one handler"
Tool.ts + tools.ts: every tool registers into the dispatch
map. The loop stays identical. buildTool() factory provides
safe defaults.
s03 PLANNING "An agent without a plan drifts"
EnterPlanModeTool/ExitPlanModeTool + TodoWriteTool:
list steps first, then execute. Doubles completion rate.
s04 SUB-AGENTS "Break big tasks; clean context per subtask"
AgentTool + forkSubagent.ts: each child gets fresh messages[],
keeping the main conversation clean.
s05 KNOWLEDGE ON DEMAND "Load knowledge when you need it"
SkillTool + memdir/: inject via tool_result, not system prompt.
CLAUDE.md files loaded lazily per directory.
s06 CONTEXT COMPRESSION "Context fills up; make room"
services/compact/: three-layer strategy:
autoCompact (summarize) + snipCompact (trim) + contextCollapse
s07 PERSISTENT TASKS "Big goals → small tasks → disk"
TaskCreate/Update/Get/List: file-based task graph with
status tracking, dependencies, and persistence.
s08 BACKGROUND TASKS "Slow ops in background; agent keeps thinking"
DreamTask + LocalShellTask: daemon threads run commands,
inject notifications on completion.
s09 AGENT TEAMS "Too big for one → delegate to teammates"
TeamCreate/Delete + InProcessTeammateTask: persistent
teammates with async mailboxes.
s10 TEAM PROTOCOLS "Shared communication rules"
SendMessageTool: one request-response pattern drives
all negotiation between agents.
s11 AUTONOMOUS AGENTS "Teammates scan and claim tasks themselves"
coordinator/coordinatorMode.ts: idle cycle + auto-claim,
no need for lead to assign each task.
s12 WORKTREE ISOLATION "Each works in its own directory"
EnterWorktreeTool/ExitWorktreeTool: tasks manage goals,
worktrees manage directories, bound by ID.
| Pattern | Where | Purpose |
|---|---|---|
| AsyncGenerator streaming | QueryEngine, query() |
Full-chain streaming from API to consumer |
| Builder + Factory | buildTool() |
Safe defaults for tool definitions |
| Branded Types | SystemPrompt, asSystemPrompt() |
Prevent string/array confusion |
| Feature Flags + DCE | feature() from bun:bundle |
Compile-time dead code elimination |
| Discriminated Unions | Message types |
Type-safe message handling |
| Observer + State Machine | StreamingToolExecutor |
Tool execution lifecycle tracking |
| Snapshot State | FileHistoryState |
Undo/redo for file operations |
| Ring Buffer | Error log | Bounded memory for long sessions |
| Fire-and-Forget Write | recordTranscript() |
Non-blocking persistence with ordering |
| Lazy Schema | lazySchema() |
Defer Zod schema evaluation for performance |
| Context Isolation | AsyncLocalStorage |
Per-agent context in shared process |
This source is not directly compilable from this repo alone:
- Missing
tsconfig.json, build scripts, and Bun bundler config feature()calls are Bun compile-time intrinsics — resolved during bundlingMACRO.VERSIONis injected at build timeprocess.env.USER_TYPE === 'ant'sections are Anthropic-internal- The compiled
cli.jsis a self-contained 12MB bundle requiring only Node.js >= 18 - Source maps (
cli.js.map, 60MB) map back to these source files for debugging
See QUICKSTART.md for build instructions and workarounds.
All source code in this repository is copyright Anthropic PBC. This repository is for technical research and education only. See the original npm package for full license terms.