| name | ci-fix |
| description | Iteratively diagnose and fix failing GitHub Actions runs until CI passes or a safe escalation point is reached. Use when a branch has failing CI and an automated fix loop is acceptable. |
| disable-model-invocation | true |
| allowed-tools | Bash, Read, Write, ApplyPatch, Grep, Glob |
| argument-hint | [--max-attempts N] [--dry-run] [--classify-only] [--no-push] |
CI Fix
Purpose
Automatically diagnose and fix GitHub Actions CI failures in a bounded loop until tests pass or the workflow must safely escalate to a human.
Inputs
-
--max-attempts N: maximum fix attempts, default 3
-
--dry-run: show what would be done without making changes
-
--classify-only: classify failures without attempting fixes
-
--no-push: allow local fixes/commits without pushing
-
Template paths are resolved from ${AGENTS_SKILLS_ROOT}/ci-fix/.
Steps
- Resolve the current failing run on the active branch, or stop if CI is already green.
- Use
${AGENTS_SKILLS_ROOT}/ci-logs/commands/ci-logs.md behavior to gather structured failure details.
- Classify failures into lint, type, test, build, dependency, config, or infrastructure buckets.
- Apply only bounded, explainable fixes. Retry infrastructure/transient failures instead of inventing code changes.
- If
--classify-only is set, return the taxonomy and suggested next action without editing files.
- If changes are made, stage the eligible change set, create a descriptive
fix(ci): ... commit, and push unless --no-push is set.
- Use
ci-watch to wait for the rerun result.
- Stop when CI passes, the attempt budget is exhausted, or the failure is unsafe to automate.
Output
- attempt-by-attempt summary of failures, fixes, reruns, and outcomes
- final status: fixed, escalated, or classify-only
- commit hash and run URL when applicable
