Claude Code — Source Tree Analysis

An exploration of the Claude Code source at /workspaces/src/ — Anthropic's official CLI coding agent — with an emphasis on what is interesting to learn from: how the agent loop is structured, how the LLM is invoked, what guardrails are in the prompt, and how the many subsystems coordinate.

Documents

Overview series

1. 01-purpose-and-stack.md — what this project is, who it is for, languages, runtime, notable dependencies.
2. 02-architecture-overview.md — the 30,000-foot view: components, dependency layout, data flow, diagrams.
3. 03-entry-and-execution.md — startup path, fast-path flags, boot-to-TUI, SDK headless, bridge / remote / daemon entry points.
4. 04-agent-loop-and-llm.md — the core generator-based query loop, streaming, retry, thinking, compaction. Focus doc for LLM usage.
5. 05-tool-system.md — Tool.ts contract, permissions, deferred tools + ToolSearch, a tour of every tool.
6. 06-agents-skills-commands.md — subagents (built-in + custom), skills, slash commands, fork subagents, MCP.
7. 07-prompts-guardrails.md — verbatim excerpts of the system prompt, agent prompts, compaction prompts, and memory prompts, with commentary on the guardrails and techniques used.
8. 08-memory-and-context.md — CLAUDE.md, auto-memory, session memory, file history, attribution.
9. 09-state-ui-session.md — React/Ink TUI, AppState store, transcript persistence, resume.
10. 10-design-patterns-and-tradeoffs.md — patterns used, tradeoffs, clever tricks, pitfalls to watch for.
11. 11-key-files-map.md — a mapping from essential file → responsibility, in priority-read order.

Deep-dive series

12. 12-btw-command.md — how /btw works: fork agent, cache-safe params, no-tools policy, verbatim wrapper prompt.
13. 13-multi-agent-coordination.md — parallel tool dispatch, StreamingToolExecutor, fork vs fresh subagents, background agents, teams, SendMessageTool, coordinator mode.
14. 14-interruption-and-queue.md — the command queue, query guard, per-tool interruptBehavior, Escape, mailbox, how mid-turn messages work end-to-end.
15. 15-memory-deep-dive.md — extraction prompt verbatim, trigger thresholds, fork agent shape, recall scoring via Sonnet, team memory scope and security, cache implications.
16. 16-agentic-loop-schematic.md — a line-by-line walkthrough of query.ts's queryLoop(): every phase, every transition.reason, invariants, complexity hotspots.

How to read this analysis

If you only have an hour: read 02, 04, 07, 10 in that order. If you want to implement a similar system: also read 05 and 06. If you are porting / extending: read 11 first to find the files you will edit.

If you want the deep mechanics: 16 is the detailed loop schematic; 13 and 14 cover the multi-agent and concurrency parts that overlap with 4; 15 is the full memory story; 12 is the focused /btw case study.

Every major claim cites file:line. Prompt strings are quoted as they appear in source; code snippets are verbatim. Where the source uses feature-flag conditionals, I note the gate rather than inventing a dominant branch.