claude-code-working

A working reverse-engineered Claude Code CLI codebase rebuilt from source analysis to reproduce the original terminal workflow, architecture, and engineering behavior.

Start with the architecture docs

Current focus:

• The codebase runs and passes baseline type checks.
• The monorepo and engineering scaffolding are in place.
  • Biome formatting is intentionally conservative to avoid massive diff churn.
  • The build pipeline produces runnable Node and Bun output.
• The documentation set is available in English.
• More tests can still be added to improve confidence and stability.

Quick start

Environmental requirements

You must have the latest version of bun, otherwise there will be a bunch of weird bugs!!! bun upgrade!!!

• Bun >= 1.3.11
• Conventional way to configure CC, each major provider has its own configuration method

Installation

bun install

Run

# Development mode, if you see the version number 888, it means it is correct
bun run dev

# Build
bun run build

The build uses code splitting multi-file packaging (build.ts), and the product is output to the dist/ directory (entry dist/cli.js + about 450 chunk files).

The built version can be started by both bun and node. You can start it directly by publishing it to a private source.

If you encounter a bug, please raise an issue directly and we will resolve it first.

Related documentation

• Architecture overview
• Conversation model
• Tools and execution
• Safety and permissions
• Extensibility

Capability list

✅ = implemented ⚠️ = partially implemented / conditionally enabled ❌ = stub / removed / feature flag closed

Core System

Capabilities	Status	Description
REPL interactive interface (Ink terminal rendering)	✅	Home screen 5000+ lines, complete interaction
API Communication — Anthropic Direct	✅	Support API Key + OAuth
API Communication — AWS Bedrock	✅	Supports credential refresh, Bearer Token
API Communication — Google Vertex	✅	Support GCP credential refresh
API Communication — Azure Foundry	✅	Supports API Key + Azure AD
Streaming dialogue and tool call loop (query.ts)	✅	1700+ lines, including automatic compression and token tracking
Conversation Engine (QueryEngine.ts)	✅	1300+ lines to manage conversation state and attribution
Context build (git status / CLAUDE.md / memory)	✅	context.ts complete implementation
Permission system (plan/auto/manual mode)	✅	6300+ lines, including YOLO classifier, path verification, rule matching
Hook system (pre/post tool use)	✅	Support settings.json configuration
Session Resume (/resume)	✅	Standalone ResumeConversation screen
Doctor Diagnosis (/doctor)	✅	Version, API, plugin, sandbox check
Automatic compression (compaction)	✅	auto-compact / micro-compact / API compact

Tools — always available

Tools	Status	Description
BashTool	✅	Shell execution, sandbox, permission check
FileReadTool	✅	File / PDF / Picture / Notebook Reading
FileEditTool	✅	String replacement editing + diff tracking
FileWriteTool	✅	File creation/overwriting + diff generation
NotebookEditTool	✅	Jupyter Notebook cell editing
AgentTool	✅	Subagent fork (fork/async/background/remote)
WebFetchTool	✅	URL Fetching → Markdown → AI Summary
WebSearchTool	✅	Web search + domain name filtering
AskUserQuestionTool	✅	Multiple question interactive prompts + preview
SendMessageTool	✅	Message sending (peers / teammates / mailbox)
SkillTool	✅	Slash command / Skill call
EnterPlanModeTool	✅	Enter planning mode
ExitPlanModeTool (V2)	✅	Exit plan mode
TodoWriteTool	✅	Todo List v1
BriefTool	✅	Short message + attachment sending
TaskOutputTool	✅	Background task output reading
TaskStopTool	✅	Background task stop
ListMcpResourcesTool	⚠️	MCP resource list (filtered by specialTools, added under specific conditions)
ReadMcpResourceTool	⚠️	MCP resource reading (same as above)
SyntheticOutputTool	⚠️	Created only in non-interactive sessions (SDK/pipe mode)
CronCreateTool	✅	Scheduled task creation (AGENT_TRIGGERS gate has been removed)
CronDeleteTool	✅	Scheduled task deletion
CronListTool	✅	Scheduled task list
EnterWorktreeTool	✅	Enter Git Worktree (isWorktreeModeEnabled() is hardcoded to true)
ExitWorktreeTool	✅	Exit Git Worktree

Tools — Conditional Enablement

Tools	Status	Enablement Conditions
GlobTool	✅	Enabled when bfs/ugrep is not embedded (enabled by default)
GrepTool	✅	Same as above
TaskCreateTool	⚠️	When isTodoV2Enabled() is true
TaskGetTool	⚠️	Same as above
TaskUpdateTool	⚠️	Same as above
TaskListTool	⚠️	Same as above
TeamCreateTool	⚠️	isAgentSwarmsEnabled()
TeamDeleteTool	⚠️	Same as above
ToolSearchTool	⚠️	isToolSearchEnabledOptimistic()
PowerShellTool	⚠️	Windows Platform Detection
LSPTool	⚠️	ENABLE_LSP_TOOL environment variable
ConfigTool	❌	USER_TYPE === 'ant' (always false)

Tools — Feature Flag Off (all disabled)

Tools	Feature Flag
SleepTool	PROACTIVE / KAIROS
RemoteTriggerTool	AGENT_TRIGGERS_REMOTE
MonitorTool	MONITOR_TOOL
SendUserFileTool	KAIROS
OverflowTestTool	OVERFLOW_TEST_TOOL
TerminalCaptureTool	TERMINAL_PANEL
WebBrowserTool	WEB_BROWSER_TOOL
SnipTool	HISTORY_SNIP
WorkflowTool	WORKFLOW_SCRIPTS
PushNotificationTool	KAIROS / KAIROS_PUSH_NOTIFICATION
SubscribePRTool	KAIROS_GITHUB_WEBHOOKS
ListPeersTool	UDS_INBOX
CtxInspectTool	CONTEXT_COLLAPSE

Tools — Stub / Not available

Tools	Instructions
TungstenTool	ANT-ONLY stub
REPLTool	ANT-ONLY，isEnabled: () => false
SuggestBackgroundPRTool	ANT-ONLY，isEnabled: () => false
VerifyPlanExecutionTool	Requires CLAUDE_CODE_VERIFY_PLAN=true environment variable and is stub
ReviewArtifactTool	stub, not registered to tools.ts
DiscoverSkillsTool	stub, not registered to tools.ts

Slash command — available

Command	Status	Description
/add-dir	✅	Add directory
/advisor	✅	Advisor Configuration
/agents	✅	Agent list/management
/branch	✅	Branch management
/btw	✅	Quick Notes
/chrome	✅	Chrome integration
/clear	✅	Clear screen
/color	✅	Agent color
/compact	✅	Compact conversation
/config (/settings)	✅	Configuration management
/context	✅	Context information
/copy	✅	Copy last message
/cost	✅	Session cost
/desktop	✅	Claude Desktop Integration
/diff	✅	show diff
/doctor	✅	Health Check
/effort	✅	Set effort level
/exit	✅	Exit
/export	✅	Export conversation
/extra-usage	✅	Extra usage information
/fast	✅	Switch fast mode
/feedback	✅	Feedback
/loop	✅	Timed loop execution (bundled skill, can be turned off via CLAUDE_CODE_DISABLE_CRON)
/heapdump	✅	Heap dump (debug)
/help	✅	Help
/hooks	✅	Hook Management
/ide	✅	IDE connection
/init	✅	Initialization project
/install-github-app	✅	Install GitHub App
/install-slack-app	✅	Install Slack App
/keybindings	✅	Shortcut key management
/login / /logout	✅	Login / Logout
/mcp	✅	MCP Service Management
/memory	✅	Memory / CLAUDE.md Management
/mobile	✅	Mobile QR code
/model	✅	Model selection
/output-style	✅	Output style
/passes	✅	Recommendation code
/permissions	✅	Permission management
/plan	✅	Plan mode
/plugin	✅	Plug-in management
/pr-comments	✅	PR comments
/privacy-settings	✅	Privacy Settings
/rate-limit-options	✅	Rate limit options
/release-notes	✅	Update log
/reload-plugins	✅	Reload plugins
/remote-env	✅	Remote environment configuration
/rename	✅	Rename session
/resume	✅	Resume session
/review	✅	Code review (local)
/ultrareview	✅	Cloud review
/rewind	✅	Rewind conversation
/sandbox-toggle	✅	Toggle sandbox
/security-review	✅	Security review
/session	✅	Session information
/skills	✅	Skill Management
/stats	✅	Session statistics
/status	✅	Status information
/statusline	✅	Status bar UI
/stickers	✅	Stickers
/tasks	✅	Task management
/theme	✅	Terminal theme
/think-back	✅	Year in Review
/upgrade	✅	Upgrade CLI
/usage	✅	Usage information
/insights	✅	Usage analysis report
/vim	✅	Vim mode

Slash Command — Feature Flag Off

Command	Feature Flag
/voice	VOICE_MODE
/proactive	PROACTIVE / KAIROS
/brief	KAIROS / KAIROS_BRIEF
/assistant	KAIROS
/remote-control (alias rc)	BRIDGE_MODE
/remote-control-server	DAEMON + BRIDGE_MODE
/force-snip	HISTORY_SNIP
/workflows	WORKFLOW_SCRIPTS
/web-setup	CCR_REMOTE_SETUP
/subscribe-pr	KAIROS_GITHUB_WEBHOOKS
/ultraplan	ULTRAPLAN
/torch	TORCH
/peers	UDS_INBOX
/fork	FORK_SUBAGENT
/buddy	BUDDY

Slash command — ANT-ONLY (not available)

/files /tag /backfill-sessions /break-cache /bughunter /commit /commit-push-pr /ctx_viz /good-claude /issue /init-verifiers /mock-limits /bridge-kick /version /reset-limits /onboarding /share /summary /teleport /ant-trace /perf-issue /env /oauth-refresh /debug-tool-call /agents-platform /autofix-pr

CLI subcommands

Subcommand	Status	Description
claude (default)	✅	Main REPL / interactive / print mode
claude mcp serve/add/remove/list/get/...	✅	MCP service management (7 subcommands)
claude auth login/status/logout	✅	Authentication management
claude plugin validate/list/install/...	✅	Plug-in management (7 subcommands)
claude setup-token	✅	Long-term Token configuration
claude agents	✅	Agent list
claude doctor	✅	Health Check
claude update / upgrade	✅	Automatic update
claude install [target]	✅	Native installation
claude server	❌	DIRECT_CONNECT flag
claude ssh <host>	❌	SSH_REMOTE flag
claude open <cc-url>	❌	DIRECT_CONNECT flag
claude auto-mode	❌	TRANSCRIPT_CLASSIFIER flag
claude remote-control	❌	BRIDGE_MODE + DAEMON flag
claude assistant	❌	KAIROS flag
claude up/rollback/log/error/export/task/completion	❌	ANT-ONLY

Service layer

Service	Status	Description
API Client (services/api/)	✅	3400+ lines, 4 providers
MCP (services/mcp/)	✅	34 files, 12000+ lines
OAuth (services/oauth/)	✅	Complete OAuth process
Plugins (services/plugins/)	✅	Complete infrastructure, no built-in plugins
LSP (services/lsp/)	⚠️	Implementation exists, closed by default
Compression (services/compact/)	✅	auto / micro / API compression
Hook system (services/tools/toolHooks.ts)	✅	pre/post tool use hooks
Session Memory (services/SessionMemory/)	✅	Session Memory Management
Memory extraction (services/extractMemories/)	✅	Automatic memory extraction
Skill Search (services/skillSearch/)	✅	Local/remote skill search
Policy Limits (services/policyLimits/)	✅	Policy Limit Enforcement
Analysis / GrowthBook / Sentry	⚠️	The frame exists, the actual sink is empty
Voice (services/voice.ts)	❌	VOICE_MODE flag off

Internal packages (packages/)

Package	Status	Description
color-diff-napi	✅	1006 lines of complete TypeScript implementation (syntax highlighting diff)
audio-capture-napi	✅	151 lines of complete implementation (cross-platform audio recording, using SoX/arecord)
image-processor-napi	✅	125 lines of complete implementation (macOS clipboard image reading, using osascript + sharp)
modifiers-napi	✅	67 lines of complete implementation (macOS modifier key detection, bun:ffi + CoreGraphics)
url-handler-napi	❌	stub, waitForUrlEvent() returns null
@ant/claude-for-chrome-mcp	❌	stub, createServer() returns null
@ant/computer-use-mcp	⚠️	Type-safe stub (line 265, complete type definition but function returns null)
@ant/computer-use-input	✅	183 lines of complete implementation (macOS keyboard and mouse simulation, AppleScript/JXA/CGEvent)
@ant/computer-use-swift	✅	388 lines of complete implementation (macOS display/application management/screenshot, JXA/screencapture)

Feature Flags (31, all return false)

ABLATION_BASELINE AGENT_MEMORY_SNAPSHOT BG_SESSIONS BRIDGE_MODE BUDDY CCR_MIRROR CCR_REMOTE_SETUP CHICAGO_MCP COORDINATOR_MODE DAEMON DIRECT_CONNECT EXPERIMENTAL_SKILL_SEARCH FORK_SUBAGENT HARD_FAIL HISTORY_SNIP KAIROS KAIROS_BRIEF KAIROS_CHANNELS KAIROS_GITHUB_WEBHOOKS LODESTONE MCP_SKILLS PROACTIVE SSH_REMOTE TORCH TRANSCRIPT_CLASSIFIER UDS_INBOX ULTRAPLAN UPLOAD_USER_SETTINGS VOICE_MODE WEB_BROWSER_TOOL WORKFLOW_SCRIPTS

Project structure

claude-code/
├── src/
│   ├── entrypoints/
│ │ ├── cli.tsx # Entry file (including MACRO/feature polyfill)
│ │ └── sdk/ # SDK submodule stub
│ ├── main.tsx # Main CLI logic (Commander definition)
│   └── types/
│ ├── global.d.ts # Global variable/macro declaration
│ └── internal-modules.d.ts # Internal npm package type declaration
├── packages/ # Monorepo workspace package
│ ├── color-diff-napi/ # Complete implementation (terminal color diff)
│ ├── modifiers-napi/ # stub (macOS modifier key detection)
│   ├── audio-capture-napi/  # stub
│   ├── image-processor-napi/# stub
│   ├── url-handler-napi/    # stub
│ └── @ant/ # Anthropic internal package stub
│       ├── claude-for-chrome-mcp/
│       ├── computer-use-mcp/
│       ├── computer-use-input/
│       └── computer-use-swift/
├── scripts/ # Automatic stub generation script
├── build.ts # Build script (Bun.build + code splitting + Node.js compatible post-processing)
├── dist/ # Build output (entry cli.js + ~450 chunk files)
└── package.json # Bun workspaces monorepo configuration

Technical description

Runtime Polyfill

The necessary polyfill is injected at the top of the entry file src/entrypoints/cli.tsx:

• feature() — return false for all feature flags, skip unimplemented branches
• globalThis.MACRO — emulates build-time macro injection (VERSION, etc.)

Monorepo

The project uses Bun workspaces to manage internal packages. The stubs that were originally placed manually under node_modules/ have been uniformly moved into packages/ and parsed through workspace:*.

Feature Flags detailed explanation

The original Claude Code injects feature flags during build through feature() of bun:bundle, and the grayscale release is controlled by A/B experiment platforms such as GrowthBook. In this project, feature() is polyfilled to always return false, so the following 30 flags are all turned off.

Autonomous Agent

Flag	Purpose
KAIROS	Assistant mode - long-running autonomous Agent (including brief, push notification, file sending)
KAIROS_BRIEF	Kairos Brief — Send briefing summary to user
KAIROS_CHANNELS	Kairos Channels — Multi-Channel Communication
KAIROS_GITHUB_WEBHOOKS	GitHub Webhook subscription — PR events are pushed to Agent in real time
PROACTIVE	Active mode - Agent actively performs tasks, including SleepTool scheduled wake-up
COORDINATOR_MODE	Coordinator mode - multi-Agent orchestration and scheduling
BUDDY	Buddy pairing programming function
FORK_SUBAGENT	Fork subagent — fork an independent subagent from the current session

Remote/Distributed

Flag	Purpose
BRIDGE_MODE	Remote control bridge — allows external clients to remotely control Claude Code
DAEMON	Daemon process — background resident service, supports worker and supervisor
BG_SESSIONS	Background session — ps/logs/attach/kill/--bg and other background process management
SSH_REMOTE	SSH remote — claude ssh <host> connect to remote host
DIRECT_CONNECT	Direct connection mode - cc:// URL protocol, server command, open command
CCR_REMOTE_SETUP	Web remote configuration - configure through browser Claude Code
CCR_MIRROR	Claude Code Runtime Mirroring - Session State Sync/Replication

Communication

Flag	Purpose
UDS_INBOX	Unix Domain Socket Inbox - Local communication between Agents (/peers)

Enhancement Tools

Flag	Purpose
CHICAGO_MCP	Computer Use MCP — Computer operation (screenshot, mouse and keyboard control)
WEB_BROWSER_TOOL	Web browser tool - embedded browser interaction in the terminal
VOICE_MODE	Voice mode - voice input and output, microphone push-to-talk
WORKFLOW_SCRIPTS	Workflow scripts — user-defined automated workflows
MCP_SKILLS	MCP-based Skill loading mechanism

Dialogue management

Flag	Purpose
HISTORY_SNIP	History clipping — manually clip snippets in conversation history (/force-snip)
ULTRAPLAN	Ultra Plan — Large-scale planning capabilities for remote Agent collaboration
AGENT_MEMORY_SNAPSHOT	Memory snapshot function when Agent is running

Infrastructure/Experiments

Flag	Purpose
ABLATION_BASELINE	Science Experiment — Baseline ablation test for A/B experiment control group
HARD_FAIL	Hard failure mode - interrupt directly instead of downgrading when an error occurs
TRANSCRIPT_CLASSIFIER	Conversation classifier — auto-mode command, automatically analyzes and classifies conversation records
UPLOAD_USER_SETTINGS	Setting synchronization upload — synchronize local configuration to the cloud
LODESTONE	Deep link protocol handler - jump from external application to the location specified by Claude Code
EXPERIMENTAL_SKILL_SEARCH	Experimental Skill Search Index
TORCH	Torch function (specific use unknown, may be some kind of highlighting/tracking mechanism)

License

This project is for study and research purposes only. All rights to Claude Code belong to Anthropic.