Claude Code source code leaked, some say 'shit mountain', some say 'really fragrant'. After reviewing 1902 documents, objectively speaking: Not a shit mountain, but a prototype of smart people who don't have time to clean. On the positive side: Clear architecture layering, coordinators/agents/memdir each perform their own duties Multi agent using tmux and email files, simple but effective Memory retrieval is isolated using independent sub agents, and the thinking is correct Rough side: USER_TYPE=='ant' determines that 154 scattered areas have not been cleaned up • Feature flags are used for configuration, with historical flags piled up The same problem was committed once in 2025 and again in 2026 ─── 512K lines of TypeScript can achieve this level, and the team's engineering ability is not a problem. The problem lies in the pace: the feature iteration is too fast, and the "closing" process was not done after the release. ─── For those who write code: it is worth studying the source code, but it is not a benchmark for learning - it demonstrates "how to quickly make good things", not "how to maintain good things". Anthropic may need their own Claude Code to help with code review.