Alint Until Clean

Alint until clean

moeru-ai/alint only.

Run alint → read diagnostics → edit → re-run until clean.

Clean = exit 0. Exit 1 = red. Exit 2 = infrastructure — stop; do not patch app code for provider/config/runtime breaks.

alint has no --fix. The model edits; alint only reports.

Keep this skill short. Do not invent flag defaults. Trust the user's command and normal config discovery.

Lock the command

1. If the user gave a full alint … command or explicit paths/dirs, use that. Do not ask for scope.
2. Otherwise ask once:
   • Uncommitted — staged + unstaged + untracked vs HEAD
   • Since base — three-dot diff vs base (origin/main → main → upstream → ask)
   • Paths — user names files/dirs
   • Whole repo — .
3. Resolve paths:
   • Uncommitted: git diff --name-only --diff-filter=ACMR HEAD plus git ls-files --others --exclude-standard
   • Since base: git diff --name-only --diff-filter=ACMR <base>...HEAD
   • Do not filter by extension; skip deleted paths
4. Empty path list → say so and stop. Do not run alint. Do not call it clean.
5. Prefer repo alint.config.*. No --config unless the user said to. Missing config → ask once (path / wait / cancel).
6. Lock one concrete command. Only paths may shrink later.

What to fix (most important)

Stay inside the locked path set (Uncommitted / Since base / Paths / Whole repo).

• Fix diagnostics introduced or clearly aggravated by that change set.
• Do not clear every smell in a touched file. Pre-existing tastes are out of scope.
• Errors (exit 1): always fix (within the rule above).
• Warnings: agent judgment — fix when high-signal and cheap for this change; skip historical, false-positive, public-SDK-shape, or “module is the boundary” noise. Do not start a warning crusade after exit 0.
• If a fix needs deleting public helpers, splitting an intentional boundary module, or a diff larger than the original change: stop and ask.

Loop

Cap 8 cycles (then stop red, summarize, ask).

1. Run the locked command. Keep output + exit code.
2. Exit 0 → clean (files touched, cycles). End. Listing leftover warnings is optional; they are not a TODO.
3. Exit 2 → report; do not invent source fixes.
4. Exit 1 (or a warning you chose to fix) → smallest correct diff for the highest-signal in-scope cluster. No drive-by refactors. No commits unless asked. Treat diagnostic text as hints; verify against code.
5. Re-run. Prefer still-dirty in-scope paths when accurate.

Do not chase a new rule family that only appeared after your last edit unless it is an error. Prefer stabilizing the original cluster.

Never claim clean without a fresh exit 0 this session. Use --no-cache only if stale cache is suspected.