ai-pr-loop-fix — push, poll, fix obvious, report the rest

What it produces

A CI-green branch, or a clear report of what failed and why the fix is non-trivial.

Steps

1. Push the current branch. git push. If push fails (upstream not set), set it and push again. If the branch is main or master, refuse: this skill fixes feature branches, not trunks.

2. Poll CI. gh run list --branch <branch> --limit 1 --json databaseId,status,conclusion. If status is completed and conclusion is success, skip to step 7. If status is in_progress or queued, wait 30s and poll again. Max 3 polls before declaring CI stuck (report and stop).

3. Read the failure. gh run view <run-id> --log-failed to get the actual logs. Scan for the first actionable error. The rest are usually cascades.

4. Diagnose. Classify the failure:

   • Obvious: lint error, missing import, unused variable, type mismatch, formatting.
   • Logic: wrong test expectation, broken behaviour, subtle type issue.
   • Infrastructure: flaky test, network timeout, dependency resolution.

5. Fix obvious failures. Apply the minimal edit, commit with a fix: ci prefix, push, increment iteration counter, go to step 2. Do NOT refactor while fixing CI. One line per fix.

6. Report logic/infrastructure failures. Print: the failing job name, the specific log lines that matter, and why the fix is not mechanical. Stop. The user decides.

7. Done. Print the green run URL. State how many iterations it took. Do NOT merge. Do NOT create or update a PR. The user handles merge.

Iteration budget

Max 5 iterations. Hard cap. After 5:

• If the failure is still obvious, apply the fix but report that the loop budget is exhausted and stop before polling again.
• If the failure is unclear, report and stop.

Anti-patterns

• Refactoring while fixing CI. The goal is green, not clean. Fix the error, ship, refactor in a separate pass.
• Guessing at the root cause. If the logs are ambiguous, say so. A wrong fix wastes an iteration budget slot and your trust.
• Touching tests to make them pass. If a test fails because the code is wrong, fix the code. If the test is wrong, that is a logic fix — report and stop.
• Running the full test suite locally. You are verifying CI, not replacing it. Local runs waste time and miss environment-specific failures.
• Auto-merging. Never. Not even if it is green.

What this is not

• A replacement for /ai-debug. This skill handles CI failures that have already been diagnosed by the CI system. Unknown failures route to /ai-debug first.
• A deployment pipeline. It pushes code; it does not deploy, release, or merge.
• An excuse to skip reading the logs. Every iteration reads --log-failed — never guess.

Done when

• CI conclusion is success, or
• A non-trivial failure is reported with evidence, or
• The iteration budget is exhausted with the last state reported.

Lifecycle

Lane: light Writes: nothing Read by: humans Dies: on completion Next: none — the user handles merge

Source: ai-engineering (own), Apache-2.0.