Tragic mistake... Anthropic leaks Claude's source code

TL;DR

Anthropic accidentally shipped Claude Code's full 57MB source map in an npm release, exposing 500,000 lines of TypeScript and revealing hidden features.

Key Points

• 1.Anthropic accidentally leaked Claude Code's entire source code via npm. Version 2.1.88 of the Claude Code npm package included a 57MB source map file with 500,000+ lines of TypeScript, likely caused by a bug in Bun.js, which Anthropic acquired and which had a known issue serving source maps in production.
• 2.The leak spawned two major open-source forks within hours. 'Claw Code' rewrote the TypeScript into Python using OpenAI Codex and became the fastest repo to hit 50,000 GitHub stars; 'OpenClaw' forked the code to work with any AI model, making projects like OpenCode obsolete.
• 3.Claude Code's architecture is 11-step prompt engineering, not advanced AI magic. The codebase is described as a 'dynamic prompt sandwich' with massive hard-coded instruction strings begging Claude to behave, plus a regex-based frustration detector scanning for keywords like 'balls' to log bad user experiences.
• 4.Anti-distillation poison pills were built in to sabotage competitors. Claude Code references fake tools that don't exist, so any model trained on Claude's outputs would learn to hallucinate nonexistent capabilities — but the leak now exposes exactly which ~25 real tools Claude uses, including a 1,000-line bash tool.
• 5.Hidden roadmap features were exposed, threatening Anthropic's IPO plans. Code references include model names 'Opus 4.7' and 'Capiara,' plus features named 'Buddy' (a Tamagotchi-style dev companion), 'Kairos' (a background scheduling agent with dream-mode memory), 'Ultra Plan,' 'Coordinator Mode,' and 'Demon Mode.'