OSS-Forge
Issue to verified patch, with a proof-of-fix capsule
- Status
- Prototype
- Year
- 2026
- Rooms
- AI & Agents
$ oss-forge solve <repo> --issue <n> map → context → hypotheses → reproduce → patch → verify verdict: VERIFIED (test fails on base, passes with patch, no new suite failures) capsule: proof-of-fix.json + proof-of-fix.md
Problem
Maintainers don't need more AI patches; they need patches they can trust quickly.
Approach
Maps the repository, packs only the definitions that matter into a token budget, has Gemma 4 write a regression test and a minimal patch, and marks a fix VERIFIED only when the test fails before and passes after in a fresh worktree.
Engineering decisions
- Unclear issues end as INCONCLUSIVE with the missing information listed, not a speculative patch.
- Python does everything deterministic; the model makes only the judgement calls.
What it does
- AST repo map & context ranking
- Isolated reproduction
- Guarded patching
- Draft PR only with --approve
Stack
My part
All commits in the repository are mine. Built for MLH Hacktoberfest 2026; the prize categories it targeted are not claimed as won.
Limits, stated plainly
- Python and JS/TS repositories; typical runs make 3–5 model calls.
Evidence
- OSS-Forge README