Inspiration

I build webapps. Users kept reporting "bugs" that weren't bugs: they just couldn't find a button or misread a screen. That's a usability problem, not a code problem. I wanted something that tells them apart, answers people, and only touches code when a bug is real.

What it does

auto-debugger reads every bug report and crash. GitLab Duo agents decide if it's a real bug, a misunderstanding, a feature request or missing info, and reply to the user. For real bugs, an agent writes a test that fails, another fixes it, and CI proves it with no AI. A human approves the merge request. No proof, no patch.

How I built it

Three GitLab Duo agents (Triager, Reproducer, Fixer) plus a Retry agent, GitLab issues as the inbox, GitLab CI/CD for the proof gate, security scans, release and live checks, and a small FastAPI intake on Google Cloud Run. SDKs for web, iOS, Android, Unity and Unreal.

Challenges I ran into

Making "is this a bug?" a decision based on evidence, not a guess. And stopping an agent from loosening its own rules: the policy is always read from main.

Accomplishments that we're proud of

In the live demo, 3 of 5 reports were answered without touching code, and both real bugs were fixed with a failing test as proof.

What we learned

Most of the work in a bug report happens before the fix. Proof matters more than speed.

What's next for auto-debugger

Run it on real apps in the stores, and spot recurring usability confusion to suggest design changes.

Built With

Share this project:

Updates

Submission history