Claude Code's Leaked Source: A Real-World Masterclass in Harness Engineering

A leaked Claude Code repository reveals about 512K lines of TypeScript and detailed harness engineering practices.

The codebase tracks 14 prompt cache invalidation vectors and uses sticky latches to decide cache keep-or-break behavior.

Claude Code implements a multi-agent subsystem called swarms and coordinates agents via natural language prompts.

Shell command execution passes through 23 security checks, including defenses against zero-width characters and zsh expansion.

The code uses a Zig HTTP layer client hash to authenticate native clients.