Inspiration: CI failures are part of software development, but diagnosing them can consume time developers would rather spend building. A failed check means digging through logs, finding the relevant code, forming a fix and checking that the fix didn’t cause another failure.

We wanted to explore whether AI agents could take on that investigation while keeping developers in control. That idea became Patchwork: a system that turns a failed GitHub check into a proposed, tested repair for a person to review.

What We Built: Patchwork connects to GitHub and can respond to completed workflow or check failures in monitored repositories. It gathers relevant failure output, source files, tests and dependency information, then checks for an exact match in its verified-fix cache.

Our coding differentiator is Millisa, our own AI coding intelligence. We pair it with NVIDIA Nemotron 3 Super on Nebius Token Factory. Nemotron analyzes the failure and helps plan a focused repair, while Millisa generates the proposed code change.

Patchwork checks the proposed patch against repository policies and runs it in an isolated Docker environment with network access disabled and resource limits. When validation passes, it can create a draft pull request with the change and its validation details. A developer still decides whether to merge it. Patchwork never merges or deploys automatically.

Challenges and What We Learned: Connecting a GitHub App involves more than a button: installation, OAuth redirects, repository permissions and webhooks all need to line up. We also had to treat logs, repository files and AI-generated patches as untrusted input.

The biggest lesson is that a plausible patch is not enough. An agent needs relevant context, clear boundaries and repeatable tests. We also learned to distinguish code failures from flaky or infrastructure problems, where changing application code may be the wrong response.

What’s Next: We’re continuing to verify the complete live GitHub-to-draft-PR flow. The repair pipeline and isolated validation have been exercised with a deterministic failing repository fixture. The next milestone is demonstrating the full flow against a connected GitHub repository.