An agent team picks the issue, builds the fix, scores its own PR, and iterates until a maintainer would merge it.
Plus merges into llama_index, mem0, langchain4j, lima, LibreChat, and mocha.
Open-loop tools generate once and hope. superhuman grades its own work before a maintainer ever sees it.
A 10-dimension rubric grades the PR like a maintainer would.
The single weakest dimension. Nothing else.
Routed to the matching language specialist.
Until 95% merge probability — two runs in a row.
every outcome → merge_outcomes.jsonl → a sharper prior for the next repo
Needs gh (authenticated), git, jq · Forks only, --force-with-lease · MIT licensed