Claude Code Source Code Leak Analysis

Anthropic's Claude Code source code was leaked, revealing important technical insights into coding agent design. Through a codebase of approximately 500,000 lines of code (LOC), it is possible to see how the agent performs complex coding tasks.

The main technical features are as follows:

• Repo state contextualization: Includes recent commit and Git branch information in the context.
• Efficient tool utilization: Uses about 20-60 custom tools (AgentTool, BashTool, etc.) such as Grep, Glob, and LSP, performing file read deduplication and aggressive cache reuse.
• Subagents: Adopts a Fork-join model leveraging Prompt Caching, sharing the entire context and performing parallel processing very efficiently.

The memory system is designed with a sophisticated 3-tier structure:

1. MEMORY.md, which serves as an index of knowledge
2. Topic files, which are loaded only when needed
3. Searchable full session transcripts

Additionally, memory is optimized through an 'autoDream' mode that merges memories, removes duplicates, and resolves contradictions.