**What I built:**A self-hosted intercepting proxy and VS Code extension that acts as a "git blame" for AI-generated code. Instead of scraping text, it intercepts terminal traffic, parses native tool_use and apply_patch protocols, and tracks exactly which prompt, model, and tool wrote a piece of code. It logs everything to a local SQLite/Postgres dashboard so you can actually audit your AI agents.
AI tools used: claude,codex,copilot
What worked well: Scaffolding the FastAPI backend, SQLAlchemy models, and the React dashboard boilerplate was incredibly fast with Cursor. Claude Code was surprisingly good at writing the initial regex and string-parsing logic for extracting Anthropic tool_use blocks from streaming SSE connections.
What I had to fix manually: The AI completely failed at understanding the nuances of OpenAI's newer Responses API apply_patch DSL) versus standard Chat Completions. I also had to manually write the state machine that correlates a proposed AI edit with the next turn's tool_result to figure out if the code was actually applied or rejected. AI also struggled heavily with the Neo4j graph lineage logic for tracking code evolution—it required full manual intervention
No responses yet.