PR Fix

You are entering an autonomous loop that drives a pull request to a clean state. Each round: read CI status + all reviewer feedback, fix what's blocking, commit, push, wait for new signal, repeat.

Reading and waiting on the bots goes through prbots (from the watching-pr-bot-reviews skill: on PATH where skillsync installed it, otherwise that skill's bin/prbots). Never write a sleep / until poller.

Prerequisites

The current branch MUST have an open PR. Verify with gh pr view --json number,state,headRefName. If no PR exists, stop and tell the user to open one first.

Process

Step 1: Identify the PR

Capture once and reuse throughout the loop:

PR_NUMBER=$(gh pr view --json number -q .number)
REPO=$(gh repo view --json nameWithOwner -q .nameWithOwner)
NEED=codex,ci
gh api repos/$REPO/actions/workflows --jq '.workflows[].path' | grep -qi archit && NEED=$NEED,sar

NEED is what each round waits for: Codex and CI always, the SAR only when the repo has an architecture-review workflow (waiting on a SAR that never runs would sit out the timeout).

Step 2: Run the Fix Loop

Execute up to 5 iterations. Each iteration is one full read → fix → push → wait cycle.

2a. Read PR state

1. Bots and CI, one call:

   prbots $PR_NUMBER --need $NEED --full

   Codex findings, the SAR, CI checks, and a CLEAN / FIX / BLOCK / PENDING verdict for HEAD. For a failing check, fetch its log:

   gh run view <run-id> --log-failed

2. Human feedback (prbots covers only the bots):

   gh api repos/$REPO/pulls/$PR_NUMBER/comments --paginate --jq '.[] | select(.user.type != "Bot") | {user: .user.login, path, line, created_at, body}'
   gh api repos/$REPO/pulls/$PR_NUMBER/reviews --paginate --jq '.[] | select(.user.type != "Bot") | {user: .user.login, state, submitted_at, body}'
   gh api repos/$REPO/issues/$PR_NUMBER/comments --paginate --jq '.[] | select(.user.type != "Bot") | {user: .user.login, created_at, body}'

3. Last-commit timestamp — anchor for distinguishing addressed vs. unaddressed human feedback:

   git log -1 --format=%cI

2b. Classify the signal

• Blocking — everything behind a BLOCK verdict (Codex P0/P1, SAR blockers or severity ≥ 3 of 4, failing CI), a human CHANGES_REQUESTED review, human comments flagging correctness/security/architecture. Must fix.
• Fix or answer — everything behind a FIX verdict (Codex P2, SAR advisories). Fix it, or reply on the thread with the reason; a non-bot reply marks the finding handled.
• Nice-to-have — P3, style nits, optional suggestions. Defer and report at the end.
• Already addressed — bot findings prbots shows as [replied] / [resolved]; human feedback older than the last commit whose diff visibly resolves it. Skip.

A PENDING verdict is not a failure — proceed to 2d (wait).

2c. Apply fixes

For each blocking item:

1. Make the code change.
2. Reference the specific comment/check being addressed in the commit message body.

When all blocking items in this round are fixed, commit and push:

git add <files>
git commit -m "$(cat <<'EOF'
fix: address PR feedback (round N)

- [item 1 — link to comment or check]
- [item 2 — link to comment or check]
EOF
)"
git push

End the message with whatever attribution trailer your environment asks for.

2d. Wait for new signal

After pushing (or if no fixes were needed because everything is pending), start the wait with Bash run_in_background: true:

prbots wait $PR_NUMBER --need $NEED --trigger

Do not poll it: no sleep, no tailing its output, no re-checking with gh. You are notified when it exits, and its output is the digest for the next round. --trigger matters because neither bot re-reviews a push on its own. Exit codes:

• 0 CLEAN → go to 2e.
• 1 FIX / BLOCK → next iteration.
• 2 PENDING (timed out after 30 min, review never started, or PR closed) → stop and report; read its x … line for the reason.
• 3 usage / API error → stop and report.

2e. Exit conditions

Stop the loop when the wait exits 0 (every needed bot reviewed HEAD, CI green, nothing open) and no blocking human feedback newer than the most recent commit remains.

If you hit 5 iterations without converging, stop and report the remaining blockers.

IMPORTANT: Work silently through iterations — do NOT spam the user with per-round status. Only emit the final summary in Step 3.

Step 3: Report

After the loop exits, present the report below. When the outcome is clean and the user wants it merged, land it with the pr-land skill (prland).

Final Status

• PR: #$PR_NUMBER ($PR_URL)
• Outcome: clean | blocked after N rounds | pending: <reason from prbots>
• Iterations: N of 5
• Final digest: the last prbots output

Fixes Applied

• Round 1: [list of commits + what they addressed]
• Round 2: ...

Deferred (Nice-to-Haves Not Addressed)

• [suggestion]: [why deferred — usually "non-blocking" or "needs user judgment"]

Remaining Blockers (if outcome ≠ clean)

• [item]: [why it couldn't be auto-resolved — e.g., needs human decision, ambiguous, requires infra change]

Important Notes

• Bots through prbots, humans through the three gh api calls. prbots reads every bot surface for HEAD; it does not report human comments.
• Anchor on commit timestamp to avoid re-fixing already-addressed human feedback in later rounds.
• Never merge on PENDING. The SAR scores severity out of 4; 3 or more blocks.
• Never force-push unless the user explicitly asks. Normal pushes only.
• Never skip hooks (--no-verify). If a pre-commit hook fails, fix the underlying issue.
• If feedback is ambiguous or requires a product/architecture decision, defer it and surface in the final report — do not guess.