Babysit

Drive one pull request through repeated NorbiAI reviews until no valid issue remains. Fix each valid finding. Rebut each false positive with evidence. Ask for a new review after each round.

Use this skill only for the NorbiAI review on the current pull request. Do not use it for human reviews or for other bots.

Before you start

Do this once, not in each loop.

• Read AGENTS.md and each nested AGENTS.md for a directory you change.
• Read .github/norbiai-review-prompt.md. It defines [NEW], [REMAINS], [RESOLVED], [WITHDRAWN] and what counts as a valid rebuttal.
• Find the pull request number and the branch. Check git status. Commit or stash unrelated work first.
• Confirm gh can read the pull request. You need the latest review comment and the NorbiAI review commit status on the head SHA.

Step 1, Read the latest review

1. Load only the newest comment from github-actions[bot] with <!-- norbiai-review --> in the body. Filter with gh api --jq; do not print all PR comments.
2. Record Reviewed commit from the comment. Compare it with the head SHA. Stop when the review is stale. Wait for a fresh review instead.
3. Read ## Findings, ## Resolved Since Previous Review, and ## Withdrawn Findings as one set. Note the [NEW] and [REMAINS] tags and each P0 to P3 priority.
4. Read the NorbiAI review commit status on the head SHA. The gate blocks merge on unresolved P0 and P1 findings.

Do not act on an old review. Do not edit the review comment.

Step 2, Triage each finding

Check each finding against the full current diff and the surrounding code. Prove the failure mode. Never trust the title alone.

Mark a finding as valid when it shows all of the following: