Inspiration Most coding agents either auto-apply changes or leave the developer guessing what was done. We wanted a review-first agent that proposes a precise patch + tests, keeps execution completely in the developer’s hands, and never runs generated code itself. What it does PatchPilot is a test-first coding agent:
Developer pastes buggy code + a short goal. It sends the request to NVIDIA Nemotron (via Nebius Token Factory). Returns structured findings, a corrected patch, suggested unit tests, and a clean unified diff. Explicitly states that generated code is never executed inside the app. Developer downloads the patch and runs it in their own sandbox.
How we built it
Side panel for model configuration (Nebius API key as secret, Token Factory URL, Nemotron model). Example loader (“Python: broken lookup”). Structured prompt asking for findings + replacement code + tests. Three tabs: Patch, Tests, Diff. Download button for the patched code. Multiple safety banners reminding users that the app never executes generated code.
Challenges we ran into
Getting consistent, structured output (findings / patch / tests) from the model. Making the safety guarantees visible and non-bypassable. Keeping the UI focused on review rather than auto-fixing.
Accomplishments that we’re proud of
Clean “review first, ship with confidence” workflow. Accurate syntax-bug fix + useful unit-test suggestions on the demo example. Zero code execution inside the application.
What we learned The most valuable coding agent is often the one that stays out of the way of the developer’s own verification loop. Structured output + clear safety boundaries beat flashy auto-apply features.
What’s next for PatchPilot
Support for multi-file patches and larger codebases. Optional local test-runner integration (still under developer control). Additional language examples and model choices.
Built with
NVIDIA Nemotron · Nebius Token Factory · Python · Modern web UI
Log in or sign up for Devpost to join the conversation.