roast-fixlisted
Install: claude install-skill vivekbhange-hub/repo-roast-skill
# roast-fix: from punchline to pull request
A roast is only worth something if the repo gets better. This skill takes the panel's top fixes,
turns each into a step anyone could follow, and then works through the quick wins one by one with
the test suite as the referee. Nothing is changed before the user sees the plan, and nothing is
committed at all.
## Inputs
- `roast/roast.json` from roast-repo (default), or a `roast/PR-<id>.md` from roast-pr.
- `--quick-wins`: only fixes with effort S. `--only N`: only fix number N from the plan.
## Outputs
- `roast/FIXPLAN.md`: the plan, with a status column that this skill updates as it goes.
- Code changes in the working tree, uncommitted, one fix at a time.
## Steps
1. **Load the roast.** Read `roast/roast.json`: `top_fixes` (ranked), `fix_today`, and each
judge's roasts. If there is no roast, suggest `roast-repo` and stop. If the roast is older than
the latest commit touching a cited file (`git log -1 --format=%cs -- <file>`), warn that some
evidence may have moved and re-check those lines first.
2. **Find the referee.** Work out how to run the tests and linters from the repo itself:
`package.json` scripts, `Makefile`, `pyproject.toml`/`tox.ini`, `Cargo.toml`, `go.mod`, CI
workflow steps. Run the test command once now and record the baseline (pass, fail, or none).
If the baseline already fails, tell the user and ask whether to continue; fixes cannot be
judged against a red suite.
3. **Write roast/FIXPLAN.