Repro
Proof, not promises.
Repro reproduces every scanner warning in a sandbox and fixes only what it proves; a person merges.
- Event
- ShellHacks 2026
- Result
- Winner · Microsoft prize
- Built in
- 36 hours
A warning is a claim.
Scanners report warnings. Repro treats each one as unproven until it can be reproduced, and it only fixes what it has proven.
No model decides what counts as a finding, nothing is repaired without a reproduction, and no fix is merged by Repro itself. A person decides whether to merge.
Reproduce first. Verify after.
- Prove the problem
Ingest
Pinned commitClones the target into a fresh workspace, pinned to one commit.
- Prove the problem
Detect
DeterministicDeterministic scanners; no model decides what counts as a finding.
- Prove the problem
Reproduce
Proof gateRe-runs each finding's reproduction command in a sandbox, keeps only what shows the problem.
- Prove the fix
Diagnose
By root causeGroups findings by root cause, proposes a fix per group.
- Prove the fix
Repair
Diff from gitEdits files through sandboxed tools; the diff comes from git.
- Prove the fix
Verify
A person mergesTests + reproduction re-run, an adversarial check tries to break the fix; a person decides whether to merge.
Results
- Finding bugs
- 0%of model tokens spent finding bugs. Scanners find; no model decides what counts as a finding.
- Fixing them
- ~68%*fewer tokens to fix them than repairing every finding separately, by fixing one root cause per group.
- Build time
- 36hBuilt in 36 hours at ShellHacks 2026.
*The ~68% figure is an estimate: it compares grouping findings by root cause with repairing every finding separately. It is not a measured benchmark.
Microsoft prize.
Repro won the Microsoft prize at ShellHacks 2026. The project page is on Devpost.