Claude Code reverse-engineering wiki

This wiki reverse-engineers @anthropic-ai/claude-code to explain how Claude Code works, with source-anchored implementation notes. The primary readable artifact for this analysis is:

claude-code-pkg/src/entrypoints/cli.js

The wiki is organized around reverse-engineered internals questions: how the Bun standalone runtime starts, routes command-line and accessibility/recovery/termination modes, shapes model context, exposes local and hosted tools, controls approved macOS applications, persists sessions and Project/onboarding knowledge, supports remote control, runs subagents and observers, and exposes operational/debug/native surfaces.

Because cli.renamed.js is a renamed view of a bundled/minified artifact, symbol names are unstable across builds. Source anchors are intended for searching the analyzed bundle, not as public API names. Implementation pages use Semantic alias as the first source-anchor table column, followed by the anchor string or symbol grep target and a one-line meaning.

Semantic alias and minified anchor mapping

This home page is a navigation index, not a direct cli.renamed.js implementation analysis. Concrete topic pages map stable semantic aliases to version-specific minified anchors.

Semantic alias	Minified anchor	Scope
Wiki home	N/A — navigation page	Orients readers to the canonical sections and reading paths.
Section indexes	N/A — see linked section README pages	Curated reader routes through source-anchored implementation pages.
Topic implementation pages	See page-level source anchor tables	Bundle-specific anchors live in focused implementation documents.

Internals wiki map

flowchart TD
  Home[Wiki home] --> Start[00 Start here]
  Home --> Runtime[01 Runtime lifecycle]
  Home --> Loop[02 Context and model loop]
  Home --> Tools[03 Tools, integrations, and security]
  Home --> Sessions[04 Sessions, persistence, and remote]
  Home --> Hosted[05 Operations and native support]
  Home --> Agents[06 Agents and automation]
  Home --> Atlas[99 Research atlas]

  Start --> Runtime
  Runtime --> Loop
  Loop --> Tools
  Runtime --> Sessions
  Sessions --> Hosted
  Loop --> Agents
  Atlas --> Start

  click Home "./" "Open wiki home"
  click Start "./00-start-here/" "Open Start here"
  click Runtime "./01-runtime-lifecycle/" "Open Runtime lifecycle"
  click Loop "./02-context-model-loop/" "Open Context and model loop"
  click Tools "./03-tools-integrations-security/" "Open Tools, integrations, and security"
  click Sessions "./04-sessions-persistence-remote/" "Open Sessions, persistence, and remote"
  click Hosted "./05-hosted-agent-ops/" "Open Operations and native support"
  click Agents "./06-agents-automation/" "Open Agents and automation"
  click Atlas "./99-research-atlas/" "Open Research atlas"

Canonical sections

Section	Purpose
Start here	Minimal orientation: what the bundle is, how to read anchors, and the first path through the runtime.
Runtime lifecycle	Package/Bun bootstrap, CLI flags, root command routing, headless/interactive modes, terminal renderer/input, accessibility, model-ended conversation termination, and subcommands.
Context and model loop	Prompt/context sources, memory/settings, model/provider/auth selection, and headless stream-JSON turns.
Tools, integrations, and security	Built-in and hosted tools, permissions, MCP, plugin lifecycle/configuration, Chrome automation, IDE/LSP, hooks, status-line commands, computer-use MCP, settings policy, Artifact, and Claude Design.
Sessions, persistence, and remote	Local JSONL transcripts, resume/continue/fork/rewind, attached Project knowledge, team onboarding/share flows, hosted sessions, teleport, Remote Control, and remote-runner egress/file staging.
Operations and native support	Debug logs, telemetry/traffic policy, enterprise gateway, doctor/update, safe-mode recovery, hosted review signals, and native image/audio modules.
Agents and automation	Custom/background/observer agents, worktree isolation, cross-agent messaging, experimental Agent Teams, task/subagent tools, lifecycle hooks, slash commands, and auto-mode.
Research atlas	Generated module/string inventories, source-anchor methodology, domain ledgers, full-system coverage review, documentation-structure decisions, and promotion rules.

Goal	Read this path
Get oriented quickly	Start here → Glossary and aliases → Main feature map → System architecture → CLI main paths
Answer protocol and boundary questions	System architecture → Runtime communication protocols → Remote control and teleport
Understand one complete local run	CLI main paths → Terminal UI renderer and input lifecycle → Prompt, context, and memory → Prompt assembly scenarios
Understand context budgets and model billing	Context, memory, compaction, checkpoints, and rewind → Model selection, calls, usage, quota, and billing → Headless streaming and resilience
Understand scriptable/headless mode	Command-line reference → Headless streaming and resilience → Context and model loop architecture
Understand accessibility and recovery modes	Accessibility and screen-reader mode → Safe mode and recovery → Settings, policy, and integrations
Use canonical reference tables	Command-line reference → Tool inventory and schemas → Hooks and events reference → Settings schema reference → Environment variables reference
Understand durable sessions and remote control	Sessions, persistence, and remote → Session resume and transcripts → Conversation termination → Remote control and teleport → Session API, events, and storage
Understand hosted-runner networking and files	Remote control and teleport → Remote-environment egress and file staging → Sandbox and isolation
Operate the enterprise gateway	Enterprise gateway server → Models, providers, and auth → Settings, policy, and integrations → Telemetry and tracing
Review trust boundaries	Tools, integrations, and security → Tool runtime, events, and integration flows → Built-in tools and permissions → Sandbox and isolation → MCP, plugins, and hooks → Plugin lifecycle and configuration → Browser automation and Claude in Chrome
Understand hosted knowledge and creation	Hosted Projects and knowledge → Team onboarding and share flows → Artifact publishing and live pages → Claude Design and design-system sync
Understand voice/audio dictation	Operations and native support → Media native modules → Audio capture and voice mode
Study agents and automation	Agents and automation → Agents, tasks, and subagents → Worktree isolation and handoffs → Agent messaging and communication → Agent Teams → Observer agents → Agent steering, interruption, and completion → Dynamic workflows

Cross-cutting implementation matrix

Concern	Primary internals section	Supporting sections
Runtime mode selection	Runtime lifecycle	Command-line reference, Accessibility and screen-reader mode, Safe mode and recovery
Terminal rendering and input	Terminal UI renderer and input lifecycle	Accessibility and screen-reader mode, Command-line reference
Communication protocols	Runtime communication protocols	Agent messaging and communication, MCP, plugins, and hooks, Remote control and teleport
Context engineering	Context and model loop	Context, memory, compaction, checkpoints, and rewind, Agents and automation, Tools, integrations, and security
Model usage, quota, and billing	Model selection, calls, usage, quota, and billing	Models, providers, and auth, Headless streaming and resilience
Tool execution	Tool inventory and schemas	Built-in tools and permissions, Artifact publishing and live pages, Claude Design and design-system sync
Native application control	Computer-use MCP	MCP, plugins, and hooks, Built-in tools and permissions
Browser automation	Browser automation and Claude in Chrome	Built-in tools and permissions, MCP, plugins, and hooks, Computer-use MCP
IDE and language servers	IDE integration and LSP diagnostics	Plugin lifecycle and configuration, Tool runtime, events, and integration flows
Plugin extension lifecycle	Plugin lifecycle and configuration	MCP, plugins, and hooks, Settings, policy, and integrations, Skills system
Session/event lifecycle	Session API, events, and storage	Data models and frame schemas, Hooks and events reference, Operations and native support
Conversation termination	Conversation termination	Session resume and transcripts, Tool inventory and schemas
Remote/hosted operation	Hosted Projects and knowledge	Artifact publishing and live pages, Claude Design and design-system sync, Remote control and teleport
Remote-runner egress and files	Remote-environment egress and file staging	Sandbox and isolation, MCP, plugins, and hooks
Enterprise inference gateway	Enterprise gateway server	Models, providers, and auth, Telemetry and tracing, Settings, policy, and integrations
Permissions and policy	Tools, integrations, and security	Context and model loop, Operations and native support
Settings, hooks, status line, and environment references	Settings schema reference	Hooks and events reference, Status line runtime and command protocol, Environment variables reference, Settings, policy, and integrations, Plugin lifecycle and configuration
Sandbox and isolation	Sandbox and isolation	Built-in tools and permissions, Settings, policy, and integrations
Voice/audio dictation	Audio capture and voice mode	Media native modules, Prompt, context, and memory
Subagents and automation	Agents, tasks, and subagents	Agent steering, interruption, and completion, Tools, integrations, and security, Sessions, persistence, and remote
Worktree isolation	Worktree isolation and handoffs	Agents, tasks, and subagents, Session resume and transcripts, Hooks and events reference
Agent messaging and replies	Agent messaging and communication	Runtime communication protocols, Agent Teams, Agent steering, interruption, and completion
Experimental Agent Teams	Agent Teams	Agents, tasks, and subagents, Runtime communication protocols, Sandbox and isolation
Observer supervision	Observer agents	Agents, tasks, and subagents, Data models and frame schemas
Artifact/source triage	Research atlas	Bundle module map, generated data under `99-research-atlas/data/`

Mechanism-level deep dives

Mechanism	Start here
Semantic aliases, minified anchors, and terminology	Glossary and aliases
System module boundaries and data/control flow	System architecture
Runtime protocol families across modules, agents, bridges, and remote servers	Runtime communication protocols
Command-line flags, subcommands, and aliases	Command-line reference
Classic/fullscreen rendering, raw input, resize, suspend, and `/tui` relaunch	Terminal UI renderer and input lifecycle
Terminal accessibility renderer selection	Accessibility and screen-reader mode
Model-ended conversation lifecycle and `/clear` recovery	Conversation termination
Configuration-isolation recovery	Safe mode and recovery
Per-module architecture deep dives	Runtime lifecycle architecture, Context and model loop architecture, Tool runtime and security architecture, Session and remote-control architecture, Operations and native-support architecture, Agent and automation architecture
Headless stream/control event loop	Headless streaming and resilience
Prompt/template extraction catalog	Prompt template catalog
Major prompt assembly scenarios	Prompt assembly scenarios
Context compaction, memory selection, checkpoints, and rewind	Context, memory, compaction, checkpoints, and rewind
Dynamic model selection, provider calls, usage, quota, and billing	Model selection, calls, usage, quota, and billing
Tool runtime, events, shell execution, SDK/LSP/Web, context exclusion, settings, and persistence	Tool runtime, events, and integration flows
Tool names, schema owners, and tool-family reference	Tool inventory and schemas
Tool authorization and execution boundary	Built-in tools and permissions
Hosted page publication, versioning, and live updates	Artifact publishing and live pages
Collaborative design RPC and design-system synchronization	Claude Design and design-system sync
Hook names, lifecycle events, and stream/control frames	Hooks and events reference
Status-line JSON, refresh, local-command, and rendering lifecycle	Status line runtime and command protocol
Known settings roots, keys, and policy groups	Settings schema reference
Command sandbox design and platform isolation	Sandbox and isolation
MCP runtime connection and elicitation	MCP, plugins, and hooks
Plugin manifests, dependencies, enablement, defaults, user options, and reload	Plugin lifecycle and configuration
Chrome extension/native-host bridge, browser tools, tab groups, and permissions	Browser automation and Claude in Chrome
IDE discovery/dynamic MCP and plugin LSP diagnostic delivery	IDE integration and LSP diagnostics
macOS computer control, app tiers, and session lock	Computer-use MCP
Resume/continue/fork state restoration	Session resume and transcripts
Session API, event, storage, and remote-control reference	Session API, events, and storage
Session-bound shared Project knowledge	Hosted Projects and knowledge
Local usage-derived onboarding guide and hosted share lifecycle	Team onboarding and share flows
Session data models, transcript records, and frame schemas	Data models and frame schemas
Remote-runner HTTPS proxy, working sync, filestore, and staged MCP lanes	Remote-environment egress and file staging
Diagnostics and debug logs	Diagnostics and debug logs
Telemetry and tracing	Telemetry and tracing
Feature gates	Feature gates reference
Updater and doctor	Updater and doctor
Environment variable reference	Environment variables reference
Enterprise OIDC/inference/policy/spend/OTLP gateway	Enterprise gateway server
Task/subagent runtime behavior	Agents, tasks, and subagents
Session/agent worktree creation, locking, persistence, and safe cleanup	Worktree isolation and handoffs
Agent message routing, transport, receive timing, replies, and completion	Agent messaging and communication
Agent Teams spawn, storage, messaging, permissions, and cleanup	Agent Teams
Observer pairing, digesting, resume, and reports	Observer agents
Agent steering, interruption, cancellation, and completion	Agent steering, interruption, and completion
Timed prompt scheduling and cron	Cron and scheduled tasks
Deterministic multi-agent orchestration	Dynamic workflows
Voice capture, transcription stream, and prompt injection	Audio capture and voice mode
Repository-wide coverage decisions and rejected false gaps	Full-system documentation coverage review
Executable-string discoveries, promotions, aliases, noise, and native limits	Disassembled string-surface review
Canonical page ownership, filename decisions, duplicate merges, and intentional non-merges	Documentation structure and duplication review

Source policy

Treat raw cli.js as artifact identity and cli.renamed.js as the primary behavioral reading surface; every semantic rename remains traceable to the same retained bundle.
Use approximate line numbers plus byte offsets because the file is bundled and many logical modules sit on very long lines.
Preserve exact strings, flags, environment variables, and symbols in source-anchor tables.
Do not treat .jsc bytecode dumps as recoverable JavaScript source.
Keep broad synthesis pages, implementation pages, and reference pages distinct; promote confirmed findings into the focused section page that owns the behavior.

Important caveat

These pages are reverse-engineering notes over a bundled/minified production artifact, not recovered clean source. Semantic names in prose and diagrams are explanatory aliases. Generated/minified symbols are retained only when useful as search anchors for the analyzed bundle.

Created and maintained by Yingting Huang.