Claude Reviews Claude Architecture

---
name: claude-reviews-claude-architecture
description: Skill for navigating and referencing the Claude Code architectural analysis — 17-part deep dive written by Claude about its own source code
triggers:
  - "explain claude code architecture"
  - "how does claude code work internally"
  - "claude code query engine explanation"
  - "how does claude code handle tools"
  - "claude code multi-agent coordination"
  - "claude code permission system"
  - "claude code context management"
  - "read the claude code architecture analysis"
---

# Claude Reviews Claude — Architecture Analysis Skill

> Skill by [ara.so](https://ara.so) — Daily 2026 Skills collection.

## What This Project Is

**claude-reviews-claude** is a 17-part architectural analysis of Claude Code (v2.1.88) written *by Claude itself* after reading 1,902 files and 477,439 lines of TypeScript source code. It is not a library or CLI — it is a structured reference document covering every major subsystem of Claude Code.

Use this skill when a developer asks about Claude Code internals, wants to understand design patterns used in Claude Code, or needs to navigate the analysis documents.

---

## Repository Layout

architecture/ ├── zh-CN/ # Chinese versions (primary) │ ├── 00-overview.md # Full system map, 17 subsystems │ ├── 01-query-engine.md # Core LLM loop (1296 lines) │ ├── 02-tool-system.md # 42+ tools, schema-driven registry │ ├── 03-coordinator.md # Parallel worker orchestration │ ├── 04-plugin-system.md # Plugin load/validate/integrate │ ├── 05-hook-system.md # PreToolUse / PostToolUse hooks │ ├── 06-bash-engine.md # Sandboxed command execution │ ├── 07-permission-pipeline.md # 7-layer defense-in-depth │ ├── 08-agent-swarms.md # Multi-agent mailbox IPC │ ├── 09-session-persistence.md # Append-only JSONL storage │ ├── 10-context-assembly.md # 3-layer context construction │ ├── 11-compact-system.md # 4-tier compression cascade │ ├── 12-startup-bootstrap.md # Fast-path init, dynamic imports │ ├── 13-bridge-system.md # Remote control / IDE bridge │ ├── 14-ui-state-management.md # Ink + React 19, 140+ components │ ├── 15-services-api-layer.md # API client, MCP, OAuth │ ├── 16-infrastructure-config.md # Settings, feature flags, telemetry │ └── 17-telemetry-privacy-operations.md # Dual-channel telemetry └── (English mirrors at same filenames, root level)

---

## The Six Architectural Pillars

### 1. Query Engine (`01-query-engine.md`)
The core `while(true)` loop that drives everything:

User Input → QueryEngine.query() → Claude API (streaming) → stop_reason = end_turn? → output → stop_reason = tool_use? → Permission Check → Execute Tool → Inject Result → loop

Key facts:
- 1,296 lines, 12-step state machine
- Manages LLM queries, tool loops, session state
- All intelligence lives in the LLM; the scaffold is just a loop

### 2. Tool System (`02-tool-system.md`)
42+ tools registered as self-contained modules:

```typescript
// Each tool implements this contract (~30 methods)
interface Tool {
  name: string;
  description: string;
  inputSchema: JSONSchema;
  execute(input: unknown, context: ToolContext): Promise<ToolResult>;
  isEnabled(): boolean;
  userFacingName(): string;
}

Tool categories:

• File I/O: Read, Write, Edit, MultiEdit
• Shell: Bash (sandboxed)
• Search: Grep, Glob, WebSearch
• Agent: Task, SubAgent
• MCP: dynamically registered external tools

3. Permission Pipeline (07-permission-pipeline.md)

7-layer defense-in-depth:

Layer 1: Config rules (allowlist/denylist patterns)
Layer 2: Tool-level checks
Layer 3: AST-based command analysis
Layer 4: YOLO classifier (risk scoring)
Layer 5: User confirmation prompts
Layer 6: OS sandbox (macOS Seatbelt / Linux seccomp)
Layer 7: Audit logging

4. Context Assembly & Compression (10, 11)

4-tier cascade for 200K context window:

Tier 1: Micro-compression  (inline dedup, whitespace)
Tier 2: Truncation         (drop oldest non-pinned turns)
Tier 3: LLM summarization  (AI-generated session summary)
Tier 4: Emergency compact  (nuclear option, keep system prompt only)

5. Multi-Agent Swarms (08-agent-swarms.md)

Three backends, mailbox IPC:

Backends:
  - iTerm2  (macOS terminal panes)
  - tmux    (cross-platform multiplexer)
  - In-process (SubAgent tool, no shell)

Coordination:
  - Parent assigns task → writes to worker mailbox
  - Worker polls mailbox → executes → writes result
  - Parent aggregates results → continues loop

7 task types: file editing, code search, test running, web search, sub-planning, validation, summarization.

6. Terminal UI (14-ui-state-management.md)

• Forked Ink + React 19
• 140+ components
• Vim keybinding mode
• Computer Use support
• IDE bridge via 13-bridge-system.md (polling dispatch loop, crash recovery)

Session Persistence (09-session-persistence.md)

~/.claude/projects/<hash>/
  └── <session-uuid>.jsonl   # append-only, one JSON object per line

Each line:
{
  "type": "message" | "tool_use" | "tool_result" | "summary",
  "uuid": "...",
  "parent_uuid": "...",      # linked list for replay
  "timestamp": "...",
  "content": { ... }
}

Resume: read last 64KB → reconstruct minimal context → continue

Startup & Bootstrap (12-startup-bootstrap.md)

Fast-path cascade:

1. Parse CLI args (no config read yet)
2. Detect --print / --dangerously-skip-permissions flags → skip heavy init
3. Dynamic import() of subsystems (avoids paying startup cost upfront)
4. Pre-connect to Anthropic API (parallel with config load)
5. Initialize global state singletons
6. Render first UI frame

Hook System (05-hook-system.md)

// Three hook points
type HookPoint =
  | "PreToolUse"    // before any tool executes
  | "PostToolUse"   // after tool result available
  | "SessionStart"; // once per session init

// Hook registration (in settings or plugin)
{
  "hooks": {
    "PreToolUse": [
      { "matcher": "Bash", "command": "echo 'bash called'" }
    ]
  }
}

Hooks receive full tool input/output via stdin as JSON and can block execution by exiting non-zero.

Plugin System (04-plugin-system.md)

6 plugin sources, loaded in priority order:

1. Enterprise policy (highest, cannot be overridden)
2. User global config (~/.claude/settings.json)
3. Project config (.claude/settings.json)
4. CLAUDE.md memory files (per-directory)
5. MCP servers (dynamic tool injection)
6. CLI flags (session-scoped overrides)

Source Code References

The analysis is based on Claude Code v2.1.88. Community-restored source:

Navigation Guide for AI Agents

Key Design Patterns (Reusable)

1. Dumb loop, smart model — Keep orchestration simple; push intelligence into the LLM
2. Schema-driven tool registry — Tools self-describe via JSON Schema; no manual dispatch tables
3. Append-only event log — Session state is a JSONL log, not mutable DB rows
4. Layered permission defense — Static rules → dynamic analysis → user prompt → OS enforcement
5. 4-tier context compression — Graceful degradation instead of hard context-exceeded errors
6. Mailbox IPC for agents — Filesystem-based message passing; survives process crashes

Cloning and Reading Locally

git clone https://github.com/openedclaude/claude-reviews-claude.git
cd claude-reviews-claude

# English analysis
ls architecture/*.md

# Chinese analysis
ls architecture/zh-CN/*.md

# Start with the overview
open architecture/zh-CN/00-overview.md
# or
cat architecture/00-overview.md

No build step, no dependencies. Pure Markdown.

Troubleshooting

"I can't find the English version of a specific episode" → English files live at architecture/<NN>-<name>.md (root level). → Chinese files live at architecture/zh-CN/<NN>-<name>.md.

"The source code links don't work" → This repo contains analysis only, not the source. See instructkr/claw-code or ChinaSiro/claude-code-sourcemap for source.

"Episode numbers don't match the table of contents" → Episodes are 0-indexed (00-overview through 17-telemetry). The roadmap section lists them 1-indexed in prose — Episode 1 = file 01-query-engine.md.

"I want the latest Claude Code analysis" → Season 1 covers v2.1.88 and is complete (17 episodes). Watch the repo for Season 2 announcements.