RepoPilot is an autonomous repo-maintainer agent: point it at a GitHub repository with a failing issue, and it produces a structured repair plan, patches the code in a sandbox, runs the project's own test suite, and opens a pull request carrying both the fix and the test that proves it.
How it works
- RepoPilot clones the target repository and reads the open issue.
- The planner produces a structured repair plan: the files to touch, the exact patch, and the test strategy. In live mode the planner is an NVIDIA Nemotron model served by Nebius Token Factory through the OpenAI-compatible chat API; in offline mode the plan is loaded from a versioned artifact (plan.json). The plan is stored either way, so every run is auditable.
- A sandbox applies the patch, installs the project, and runs the test suite.
- On green tests, RepoPilot pushes a fix branch and opens a pull request. GitHub Actions re-runs the tests on the PR as an independent check.
The demo
The public demo repository, github.com/fo-builds/textstats, carries a seeded bug (median() returns the wrong value for even-length lists) and a failing test. A GitHub Actions workflow (demo.yml) runs RepoPilot end to end in offline plan-file mode (REPOPILOT_PLAN_FILE=plan.json with --no-call): no network LLM call is made, so the run is fully reproducible and verifiable by any judge. It opened PR #2, "Fix #1: median() returns wrong value for even-length lists" with the passing test, and all checks are green. The live Token Factory planning path runs through the same planner interface and is the next verification step.
How we used Nebius and NVIDIA
The planning step is the LLM call. RepoPilot speaks the OpenAI-compatible chat API against Nebius Token Factory with NVIDIA open source Nemotron models, keeping fallbacks across open endpoints for resilience. An offline-safe replay mode re-runs a stored plan so the full pipeline stays verifiable and reproducible in CI even between live runs.
Challenges
The hardest constraint was making the demo honest and reproducible: every claim had to be backed by a public artifact. That pushed the design toward plan files as durable evidence, a workflow that runs unattended in CI, and a PR that any judge can open and re-run themselves.
What we learned
Watching an agent patch code is easy; proving the patch is correct is the real product. The test suite is the agent's conscience, and CI is the agent's second pair of eyes.
Log in or sign up for Devpost to join the conversation.