fix-cilisted
Install: claude install-skill Ruyue213/fix-ci
# fix-ci — on-call for your CI
Diagnose red CI, reproduce it locally, fix the root cause minimally, and prove the fix with verification output. A fix must change the reason CI is red — never the way red is reported.
## Hard rules
1. Never make CI green by weakening checks: no skipped/deleted tests, no `continue-on-error`, no `--no-verify`, no assertions edited to match buggy output.
2. Never raise a timeout, retry count, or memory limit as the "fix" before proving the root cause.
3. Never push, rerun workflows, or open PRs without explicit user confirmation.
4. Never claim a fix without verification output from `scripts/verify-cmd.sh` or a green rerun id.
5. When the root cause cannot be determined, deliver a triage report — never guess-edit.
Full rationale and edge cases: `references/anti-cheat.md`.
## Workflow
### 1. Locate the failed run
Do not ask the user to paste logs — fetch them yourself:
```bash
scripts/analyze-run.sh # latest failed run on current branch
scripts/analyze-run.sh --run-id 12345678 # a specific run
scripts/analyze-run.sh --branch feature/x --max-lines 300
```
It prints failed jobs/steps, a keyword-filtered error excerpt, and the path to the full raw log. Read the excerpt first; drill into the raw log with `grep -n` only for specifics. Never `cat` the full log into context.
### 2. Classify the failure
| Class | Typical symptoms in the excerpt |
|---|---|
| build | `error TS####`, `SyntaxError`, compiler/lint errors |
| de