1code

ThePolisher: AI writes fixes. ThePolisher proves them. The problem: open source maintainers are flooded with AI pull requests and can't tell which ones work without checking each one. Writing a fix is now easy. Checking it is slow, and it falls on volunteers. What it does: paste a GitHub issue from a public Python repo and get a pull request whose tests prove the fix. Muse Spark writes three fixes at once, and tests in a Modal sandbox pick the best one. Four agents (CodePolish, SecurityPolish, DocPolish, WritePolish) polish the code, security, docs and PR description, and a change is kept only if it breaks nothing. A fix counts as proven only when its new tests fail on the original code and pass with the fix. Then Muse Wonder, a stronger Meta model that didn't write the fix, writes its own tests to try to break it. You can also paste anyone's pull request to have it tested, or add a GitHub Action that checks each new PR and comments with the result. Results: on 16 real open issues from 9 popular Python projects, which I chose before running, 13 and 14 of 16 fixes were proven with no new failures across two runs, for under $0.50 per run. Muse Wonder's tests failed on a merged Rich pull request and passed on the code before it, exposing a real bug that is still open (Rich#4090), and ThePolisher then fixed it in a sandbox for $0.04. It opened two real pull requests: invoke#1099, where real mypy confirms the fix and all project CI passes, and icecream#266, where Muse Wonder found a case the fix gets wrong and the PR says so. Neither has been reviewed yet. Meta models: Muse Spark 1.3 does the work, using vision for issue screenshots and reasoning effort set per task. Muse Wonder is the independent adversary, and gets one chance to fix its own wrong guesses, which cut its noise from 43% to 11%. Limits I publish: the check doesn't predict what gets merged, Muse Wonder doesn't catch every bug, and "verified" means a fix's own tests pass, not that it's right. Every result is public, including my mistakes. I built it entirely during the hackathon: the first file was created on Oct 27, with no earlier codebase. 398 automated tests, CI on every push, open source under MIT. Live app: https://thepolisher.pages.dev Results: https://github.com/momoikura/The-Polisher/blob/main/benchmarks/results/repeat-2026-11-02.md Real PRs: https://github.com/pyinvoke/invoke/pull/1099 and https://github.com/gruns/icecream/pull/266 GitHub Action demo: https://github.com/momoikura/thepolisher-action-demo/pull/1