PatchPilot
Turn a GitHub issue into a tested, human-owned fix that maintainers want to review.
Built at AgentX Hackathon 2026
Team idk bro: Niranjan Iyer, Swarali Patil, Devansh Narlawar, Sarthak Viche
Built with Python 3.12, FastAPI, LangGraph, LiteLLM, SQLite, Docker, React 19, TypeScript, Vite, Tailwind CSS v4, Playwright, GitHub REST API + OAuth
About the project
Problem: AI coding tools flood open-source maintainers with pull requests that look plausible but are unproven, and the person submitting often can't explain the change.
Solution: PatchPilot is a bug-fix agent for public Python repositories. Paste a GitHub issue and it reproduces the bug in a locked-down Docker sandbox (no credentials, no network after setup), writes a test that captures it, and makes the smallest fix it can.
How it works: a durable LangGraph pipeline runs intake, setup, baseline, reproduce, diagnose, fix, verify and review. Every gate decision comes from code, never from the model: the new test must fail twice on the original code and pass twice in a brand-new sandbox, with no existing tests broken. You then review the exact diff, explain it in your own words and approve it. The approval is tied to a hash of that diff and its evidence. A draft PR opens only after a maintainer has agreed on the issue. Once the PR is open, PatchPilot tracks maintainer reviews and CI, explains each request, and turns the requests you pick into a new revision pushed to the same PR.
Impact: maintainers get fewer, better PRs that come with proof and a human who owns them. We've already opened a real contribution: wheelgap/wheelgap#133.