Reproduce before you answer
Paste a public GitHub issue. An agent on NVIDIA Nemotron writes a pytest reproduction, runs it in a Docker sandbox at the reported version and again on main, and fixed rules, not the model, decide the verdict.
Works on pure-Python projects tested with pytest. An issue from a repository GitHub lists as another language is turned away before a run is counted.
Progress
Report
Example reports
Real runs on real issues: an open bug with a fix verified in the sandbox, backports that went upstream and to a fork, and a bug already fixed on main. Or browse every saved report.
How the verdict is decided
| At the reported version | On main | Verdict |
|---|---|---|
| test fails | test passes | Fixed on main |
| test fails | test fails | Open bug |
| test passes | test passes | Not reproduced |
| test passes | test fails | Inconsistent |
| test fails | test fails with a different error | Inconsistent |
| setup or collection error | any | Unsupported |
| no runnable snippet in the issue | Needs info | |
Each side runs the test twice. Confidence is high only when every run agrees and neither side errored. Issue text is treated as data, and all code runs in a sandbox.