claude-code-reverse

Claude Code Reverse is a JavaScript developer tool that intercepts and visualizes the actual API traffic exchanged between Claude Code and Anthropic's LLM backend. The core technique involves locating Claude Code's installed `cli.js` file, formatting it with `js-beautify`, then applying a monkey patch to the bundled Anthropic TypeScript SDK's `beta.messages.create` method, which causes every outbound request and inbound response to be written to a `messages.log` file. Two companion tools process that output: `parser.js` strips repeated boilerplate prompts by detecting frequency and position, and `visualize.html` renders the cleaned conversation flow in a browser. The repository also publishes extracted results in a dedicated directory, covering reverse-engineered prompts and tool definitions for Claude Code behaviors including quota checks (which use Claude Haiku 3.5 against a single `quota` input), topic detection, context compaction, the core agent loop, and sub-agent task management. Researchers studying agent architecture, prompt engineering, and how Anthropic structures tool-use workflows in a production coding agent are the primary audience.

What people ask about claude-code-reverse