Reviewloop

Reviewloop

Drive a PR to all-clear across every reviewer — each bot AND each human — then loop until nothing actionable remains.

Core principle: A bot is a pure function you can re-invoke (re-trigger → poll → read verdict). A human is async and uncontrollable. So bots loop; humans get one pass and then you hand back. Never block the loop waiting on a person.

When to use

• A PR has comments from several reviewers and you want them all resolved in one driven pass.
• The reviewers are a mix you can't name up front — "address whatever's on the PR."
• You want bot reviews re-triggered after each fix, not just read once.

When NOT to use: You just want a one-shot read of what's outstanding with no loop and no re-triggering — that's a plain review-check, not this.

The reviewer registry (the spine)

No uniform code path covers all reviewers. Each needs four things — detect it, re-trigger it, read its verdict, know when it's done:

Detection is structural, not an allowlist. Classify every reviewer by GitHub's __typename (Bot vs User) first; only then look up known bots in the table for their re-trigger command. An unknown bot is still handled (read + report) — it is never silently skipped.

Inputs

• PR number (optional): default to the PR for the current branch.
• --max-iterations N (optional, default 5): cap on the bot loop.

The loop

0. Identify the PR

gh pr view --json number,headRefOid -q '{number: .number, sha: .headRefOid}'

Switch to the PR branch if not already on it. Capture <PR> and <HEAD_SHA>.

1. Enumerate reviewers (one query — the heart of the skill)

Pull verdicts, inline threads, and each author's __typename in a single GraphQL call:

gh api graphql -f query='
query($owner:String!,$repo:String!,$pr:Int!,$cursor:String) {
  repository(owner:$owner, name:$repo) {
    pullRequest(number:$pr) {
      reviews(first:50) { nodes { author { login __typename } state submittedAt body } }
      reviewThreads(first:100, after:$cursor) {
        pageInfo { hasNextPage endCursor }
        nodes {
          id isResolved isOutdated
          comments(first:1) { nodes { author { login __typename } path body } }
        }
      }
    }
  }
}' -F owner=OWNER -F repo=REPO -F pr=<PR>

Paginate reviewThreads via endCursor until hasNextPage is false. Also fetch:

• Bot summaries / scores (Greptile score, CodeRabbit walkthrough) — issue comments and PR body:

  gh pr view <PR> --json body -q .body
  gh api repos/{owner}/{repo}/issues/<PR>/comments
  

• CI-based bots (Greptile and others run as checks):

  gh api repos/{owner}/{repo}/commits/<HEAD_SHA>/check-runs --jq '.check_runs[] | {name, status, conclusion}'
  

Now build the work list: for each reviewer present, classify (known bot / unknown bot / human) and collect its unresolved threads (isResolved == false) and verdict.

2. Check the exit condition

Stop the loop when all hold:

• Every bot reviewer is at its done-condition (table) and its checks are success.
• Zero unresolved actionable threads from any reviewer (bot or human).
• The human pass (step 6) has been completed once this run.

Or when max iterations is hit (report remaining state).

3. Fix actionable comments

For each unresolved thread, across all reviewers:

1. Read the file and understand the comment in context.
2. Decide actionable (code change) vs informational / false positive.
3. If actionable, make the fix. If not, note why — you'll still resolve/reply.

Treat human comments with the same rigor as bot comments; do not down-prioritize them.

4. Resolve threads

Resolve every thread you've addressed (and informational ones), batching with GraphQL aliases. For the exact resolveReviewThread mutation and thread-fetch pagination, see references/github-mechanics.md. For human threads, prefer a brief reply explaining the fix before resolving.

5. Commit, push, re-trigger bots

git add -A
git commit -m "address review feedback (reviewloop iteration N)"
git push
sleep 5

Then re-trigger each bot reviewer by its registry command — but only if it isn't already running (guard on check-runs status / its last comment timestamp first; see references/github-mechanics.md):

# Greptile (only if its check isn't already PENDING/IN_PROGRESS)
gh pr comment <PR> --body "@greptile review"
# CodeRabbit
gh pr comment <PR> --body "@coderabbitai full review"
# Copilot
gh pr edit <PR> --add-reviewer @copilot

For an unknown bot, do not invent a command — log "no known re-trigger for ; left as-is" and carry its current verdict forward.

6. Handle humans — ONE pass, then hand back

For each human reviewer with outstanding feedback:

1. Fix their actionable comments (step 3) and reply + resolve threads (step 4).
2. Re-request their review: gh pr edit <PR> --add-reviewer <login>.
3. Do not poll or wait for them. Record them as "waiting on @" and continue.

The loop's job for humans ends at "addressed + re-requested." Blocking on a person is out of scope.

7. Poll bot completion, then loop

Poll each re-triggered bot's check run / review to a terminal state (the check-run polling loop is in references/github-mechanics.md). When all bots are terminal, go back to step 1.

Common mistakes

Report

Reviewloop complete.
  PR:             #5876
  Iterations:     3
  Bots cleared:   greptile-apps (5/5), coderabbitai (0 open), Copilot (0 open)
  Humans:         re-requested @neeraj-lightwork-ai (was CHANGES_REQUESTED)
  Resolved:       11 threads
  Remaining:      0 bot / waiting on 1 human

If stopped at max iterations, list remaining unresolved threads per reviewer and suggest next steps.