pr-review-loop

Skill: PR Review Loop

When to Use

Invoke this skill after EVERY commit to a PR, or when the user says:

• "address PR reviews"
• "fix AI reviewer comments"
• "clean up PR"
• "make sure no comments are left"

Workflow

Phase 1 — Wait for CI and Async Reviewers

# 1. Wait for all GitHub Actions checks to complete
gh pr checks <pr> --repo erxes/erxes --json name,bucket
# Poll until no check has bucket="pending"

# 2. Buffer for async AI reviewers (CodeRabbit, Sourcery post 2-5 min after checks finish)
sleep 180

# 3. Poll inline-comment count until stable across 2 consecutive 60s polls
# Stability means: same comment count on two consecutive polls

Why: AI reviewers (CodeRabbit, Sourcery, Claude Code Action) post asynchronously after pull_request.synchronize. Reading comments immediately after push will miss late-arriving reviews.

Phase 2 — Fetch Full Review State

Gather these four data sources:

1. Inline comments (gh api /repos/erxes/erxes/pulls/<pr>/comments) — file:line level
2. Review summaries (gh api /repos/erxes/erxes/pulls/<pr>/reviews) — top-level approvals/requests
3. Issue/walkthrough comments (gh api /repos/erxes/erxes/issues/<pr>/comments) — Kimi, SonarCloud, Claude Code Action overview
4. Unanswered bot threads — the TRUE actionable signal (see Lesson below)

Phase 3 — Triage Every Unresolved Item

Classify each comment by reviewer + body markers:

Critical Lesson — Author-Engagement is the Truth:

Do NOT trust isResolved or isOutdated. GitHub auto-resolves threads when later commits move line numbers, hiding genuinely unaddressed AI findings. The correct rule:

"The thread has at least one reviewer-bot comment AND the PR author has NEVER posted in the thread" → actionable.

GraphQL query for unanswered threads:

query($n:Int!){
  repository(owner:"erxes",name:"erxes"){
    pullRequest(number:$n){
      reviewThreads(first:100){
        nodes{
          id
          comments(first:50){
            nodes{
              author{login}
              body
            }
          }
        }
      }
    }
  }
}

Filter: at least one comment from coderabbitai|sourcery-ai|github-advanced-security|sonarqubecloud|sonarcloud|kimi-review, AND no comment from PR author.

Phase 4 — Apply Fixes (Consolidated Commit)

For every triage entry with action="fix":

1. Read the file at path:line, understand the comment
2. Apply the minimum fix via Edit
3. Run pnpm nx affected:build --base=origin/main and pnpm nx affected:lint --base=origin/main
4. If build/lint fails → revert that fix, retry or downgrade to reply

Commit all fixes as ONE consolidated commit:

git add <changed-files>
git commit -m "fix: address PR #<pr> review (round <n>)

- <path>:<line> — <reviewer> — <action>: <rationale>"
git push

Why one commit: Prevents the race condition where fix Sourcery → push → CodeRabbit re-reviews mid-flight → posts overlapping comments.

Phase 5 — Post Replies

For action="reply" items:

• Inline threads: gh api -X POST /repos/erxes/erxes/pulls/<pr>/comments/<id>/replies
• Top-level walkthrough: gh pr comment <pr> --repo erxes/erxes --body "..."

Reply template: Noted — out of scope for this PR. or False positive: <one-line reason>.

Phase 6 — Verify and Loop

After push + replies, re-fetch the full review state. The PR is SETTLED only when ALL of:

1. Zero unanswered bot threads (author-engagement filter)
2. Zero failing relevant checks (excluding known-flaky: Analyze (go), DeepSource: Error)
3. Zero walkthrough findings (Kimi, SonarCloud top-level comments)
4. Comment count stable across 2 consecutive polls
5. Working tree clean (git status --porcelain empty)

If any condition fails → increment round counter and loop (up to round_cap, default 5).

If round cap reached without settling → write /tmp/BLOCKED-pr-<num>.md with unresolved state and stop.

Stop Conditions

Self-Evaluation Checklist

Run after every round:

• Zero unanswered bot threads (author-engagement filter)
• Zero walkthrough findings (Kimi/SonarCloud top-level)
• All relevant checks green
• Zero GAS regressions
• Comment count stable (2 consecutive polls)
• Build + lint pass locally
• Working tree clean
• No Anthropic branding in commits or comments

Final Exit Evaluation

Before declaring the PR "done", run one final comprehensive check:

# Returns 0 only if ALL conditions pass
.agents/scripts/pr-review-check.sh --final <pr_number>

This verifies:

1. Zero open issues — no unresolved review threads from any bot
2. All CI green — every check passing (excluding known-flaky)
3. Zero walkthrough findings — Kimi/SonarCloud report clean
4. No GAS regressions — no new security alerts

If any condition fails:

• DO NOT declare the task done
• DO continue to next round (if under round cap)
• DO report blocker to user in ≤20 words, e.g.:
  • "Blocked: 3 unresolved CodeRabbit comments."
  • "Blocked: CI check payment_ui-ci failing."
  • "Blocked: SonarCloud quality gate failed."
  • "Blocked: 2 GAS regressions need fixing."

Critical Rules

1. Wait for checks BEFORE reading reviews. Always poll gh pr checks until complete, sleep 180s buffer, THEN read comments.
2. One consolidated commit per round. Never one commit per reviewer.
3. Author-engagement > isResolved. GitHub auto-resolves on outdated lines. Use the bot-presence + author-absence filter.
4. Triage before fixing. Not every comment needs a code change. Style/docs get replies.
5. No scope creep. If a reviewer says "while you're in here...", reply "out of scope" and move on.
6. No Anthropic branding. Strip Co-Authored-By: Claude... and "Generated with Claude Code" from all commits and comments.
7. Round cap = 5. If not settled after 5 rounds, the fix approach is wrong. Stop and report.