We won the Microsoft prize at ShellHacks 2026
Repro, the project we built in 36 hours at ShellHacks 2026, won the Microsoft prize. Here is what it is and why we built it the way we did.
This post is a draft for the team to review. It isn't final yet.
We’re Caffeine Driven: Alexander Gese, Adrian Morton and Brandon Delgado, three college students who build at hackathons together. At ShellHacks 2026 we built Repro in 36 hours, and it won the Microsoft prize.
What Repro is
Repro’s one-line pitch is the principle we built it around: proof, not promises.
Repro reproduces every scanner warning in a sandbox and fixes only what it proves. Then a person merges. It never decides on its own that a change is ready to ship.
Why we built it this way
A scanner warning is a claim that something might be wrong. We wanted a tool that treats every warning as unproven until it can show the problem, and that only spends effort fixing what it has shown to be real.
That idea shaped the whole pipeline:
- Finding is deterministic. Scanners decide what counts as a finding. No model makes that call.
- Nothing gets fixed without a reproduction. Each finding’s reproduction command is re-run in a sandbox, and only findings that show the problem move forward.
- Fixes are checked, then handed to a person. Tests and the reproduction run again after the fix, an adversarial check tries to break it, and a person decides whether to merge.
We wrote up how each stage works in How Repro proves a bug before it fixes it.
The numbers
- 0% of model tokens are spent finding bugs, because detection is deterministic.
- ~68% fewer tokens to fix them than repairing every finding separately. That one is an estimate, and we explain it in Finding bugs with 0% of our model tokens.
Links
- Live site: repro.miami
- Code: github.com/Bomoga/repro
- Devpost: Repro on Devpost
- The full case study: Repro