github-review

/harness-plugin:github-review — Fix Codex PR Review Findings (Self-Driving Loop)

Fetch the latest Codex code review from the current branch's PR, fix every finding, push, then poll for the next review and repeat — until the review comes back clean or the loop hits the max rounds limit.

Loop Control

• Max rounds: 10 (hard cap to prevent runaway loops)
• Poll interval: 60 seconds between checks for a new review
• Poll timeout: 10 minutes per round (if no new review appears, stop)
• Track the comment ID of the last processed review to detect new ones

Step 1: Get PR Number and Baseline

Run these commands once at the start of the loop:

# Get current branch's PR number
PR_NUMBER=$(gh pr view --json number --jq '.number' 2>/dev/null)

# Get the repo name
REPO=$(gh repo view --json nameWithOwner --jq '.nameWithOwner')

If no PR exists, tell the user and stop.

Step 1.5: Check CI Status

Before looking at Codex review comments, check if the PR's CI checks are passing:

gh pr checks $PR_NUMBER

If any checks other than Codex Code Review are failing (e.g., Code Quality Check, Unit Tests):

1. Read the failing check's log: gh run view <run_id> --log-failed
2. Fix the issue (formatting, lint, type errors, test failures)
3. Commit and push the fix
4. Re-run this step until CI is green

Common CI failures:

• Code Quality Check (Prettier): run npx prettier --write <file> on the flagged files
• Code Quality Check (ESLint): run npm run lint:fix
• Unit Tests: run npm run test to reproduce, then fix

Only proceed to Step 2 once all non-Codex checks are passing.

Step 2: Fetch Latest Codex Review

# Fetch the latest Codex review comment ID and body
gh api "repos/$REPO/issues/$PR_NUMBER/comments" \
  --jq '[.[] | select(.body | contains("## Codex Code Review"))] | last | {id, body}'

• If no Codex review comment found, tell the user and stop.
• Store the comment ID so you can detect when a NEW review arrives after pushing.

Step 3: Parse Findings

The Codex review comment has this structure:

## Codex Code Review

### [SEVERITY_ICON] [SEVERITY] Title

📍 `file_path` L{start}-{end}
🎯 Confidence: N%

Description of the issue and how to fix it.

---

**Overall: verdict** (confidence: N%)
> Summary

Extract each finding: severity, file path, line range, description.

If the review says "No issues found" or "patch is correct" — tell the user the review is clean and exit the loop.

Step 4: Fix Each Finding

For each finding:

1. Read the file at the specified path and line range
2. Understand the issue in context
3. Decide:
   • FIX — make the minimal change to resolve the issue
   • SKIP — explain why (false positive, intentional pattern, out of scope)
4. If fixing: make the change, verify with npm run lint and npm run test

Rules:

• Fix ONLY what the review identified — no drive-by refactoring
• Never introduce new issues while fixing existing ones
• Run tests after ALL fixes, not after each one

Step 5: Commit and Push

git add -A
git commit -m "fix: address Codex review round N findings

- [list what was fixed]
- [list what was skipped and why]"

git push

Step 6: Report Round Results

Tell the user:

• Round number (e.g., "Round 1 of 10")
• Which findings were fixed
• Which were skipped (with reasons)

Step 7: Poll for Next Review

After pushing, the Codex GitHub Action will run on the new commit. Poll for a NEW review comment (different comment ID than the one just processed):

# Poll every 60s, up to 10 minutes
# Look for a comment with a DIFFERENT id than the last processed one
gh api "repos/$REPO/issues/$PR_NUMBER/comments" \
  --jq '[.[] | select(.body | contains("## Codex Code Review"))] | last | {id, body}'

• If a new comment appears (different ID): go back to Step 3
• If no new comment after 10 minutes: tell the user the poll timed out, they can re-run /harness-plugin:github-review later
• If max rounds (10) reached: tell the user the loop hit its cap

Exit Conditions

The loop exits when ANY of these is true:

1. Clean review — Codex says no issues found / patch is correct
2. Poll timeout — no new review appeared within 10 minutes after push
3. Max rounds reached — 10 fix-push-poll cycles completed
4. No PR / no review — nothing to process
5. All findings skipped — nothing was actually changed, no point pushing