Agent skill · research science · aradotso
claude-reviews-claude-architecture
claude-reviews-claude-architecture — an agent skill for coding agents.
Why this skill is useful
Provides a comprehensive architectural analysis of Claude Code, detailing subsystems and design patterns that enhance understanding of its internal workings.
What it needs
About 5k tokens when loaded. Last updated 2026-07-09. 70 stars on the source repository.
What this skill does
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) User Input → QueryEngine.query() → Claude API (streaming) → stopreason = endturn? → output → stopreason = tooluse? → Permission Check → Execute Tool → Inject Result → loop 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; } 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 Tier 1: Micro-compression (inline dedup, whitespace) Tier 2: Truncation (drop oldest non-pinned turns) Tier 3: LLM summarization (AI-genera …
How to use it
Reference it in AdaL, Claude Code, Cursor or any coding agent — nothing to install:
@skills aradotso/claude-reviews-claude-architecture