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 It defines [NEW], [REMAINS], [RESOLVED], [WITHDRAWN] and what counts as a valid rebuttal.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 Read ## Findings, ## Resolved Since Previous Review, and ## Withdrawn Findings as one set.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:

• It points to changed code or a direct effect of changed code.
• It names a concrete failure, cost, or broken contract. Examples: correctness, security, data loss, concurrency, broken contract, concrete performance regression, dead code to delete, code to reuse.
• It proposes the smallest fix.

Mark a finding as not valid when one of the following holds:

• The guard, check, or contract it asks for already exists. Name the file and line.
• It reports a style choice, a naming opinion, or a vague concern with no failure mode.
• It reports a speculative edge case with no path to reach it.
• It asks for a test with no concrete regression that test would catch.
• It asks for work outside the stated problem of the pull request.

Record the verdict in a table before you change code:

Fix only valid findings. Rebut the rest. Never fix a finding you cannot reproduce in order to turn the gate green.

Step 3 — Fix each valid finding

1. Apply the smallest change that removes the problem. Prefer delete and reuse over new code.
2. Obey the non-negotiable rules in AGENTS.md: keep migrations irreversible-safe, keep released Team API meaning unchanged, keep the renderer-to-main trust boundary, redact secrets, keep the license.
3. Run the narrowest relevant test file and biome check --write --max-diagnostics=none <changed paths>, as AGENTS.md says. Leave broad checks to CI.
4. Commit the fixes on the pull request branch.

Do not weaken a test to satisfy a finding. Do not widen a type to any or unknown to bypass a checker. Do not change .github/norbiai-review-prompt.md or .github/workflows/norbiai-review.yml to pass the review.

Step 4 — Rebut each false positive

The reviewer withdraws a finding only for a concrete, checkable reason. Assertion without evidence does not work. A promise to fix it later does not work. Disagreement about priority does not work.

Write one short comment for all rebuttals. Each response goes into the next review prompt, so every word costs tokens. Write one line for each finding, in ASD-STE100 Simplified Technical English:

P1 Short title: wrong, <guard or contract> at `path:line`.

Give only the reason and the line that proves it. No greeting, no fixes summary, no head SHA, no restated finding. The reviewer reads the fixes in the diff.

Only a comment from someone who can merge counts as a rebuttal. The workflow ignores other comments. When you cannot post as such a user, give the user the exact comment text and wait. Do not post it yourself.

Never forge <!-- norbiai-review --> or <!-- norbiai-response --> markers. Never edit another comment.

Step 5 — Push and ask for a new review

One loop is one push plus one review run:

Ask for one review only. Each extra request is a full review run.

1. With fixes to push: post the rebuttal comment first, without /norbiai review. Then push. The review for the push reads every comment posted since the last review.
2. With no fixes to push: add /norbiai review to the rebuttal comment. That phrase asks for a recheck.
3. Do not also apply the norbiai label. The label asks for a full review of the whole PR.
4. Wait for the new <!-- norbiai-review --> comment and the new NorbiAI review status on the new head SHA. Poll with gh api at most once each 2 minutes; a review takes 3 to 15 minutes. Do not start the next triage on the old SHA.

When the push itself triggers a review with no comment from you, still wait for that review before you continue.

Step 6 — Loop until clean, then stop

Start Step 1 again on the new review. Carry the ledger forward:

• A [RESOLVED] finding needs no more work.
• A [WITHDRAWN] finding must not return. When code added since reintroduces the same problem for a new reason, treat it as new.
• A [REMAINS] finding needs a fix or a stronger rebuttal. Say which part of your last response the reviewer could not verify.

Stop when both hold:

• ## Findings says No actionable findings. or holds only findings you proved not valid and the reviewer withdrew.
• The NorbiAI review status on the head SHA is success.

Also stop and ask the user when the same finding returns twice with no progress, when a P0 finding is genuinely valid but the fix needs a product decision, or when the review reports malformed output or failure. Report the loop count, the fixes, the rebuttals, and the final status.

Report format

Keep the report short. After each loop, report:

Then state the head SHA and the checks you ran, in one line each.