Coderabbit review

CodeRabbit Review

Create a GitHub PR, wait for CodeRabbit's automatic review, and triage the results.

How it works

CodeRabbit is app-driven: once the GitHub app is installed on the repository, it reviews every PR on open and re-reviews incrementally on every push. There is no reviewer-request step.

Non-default base → no auto-review. A PR whose base isn't the default branch (e.g. an integration branch) gets a Review skipped comment and no CodeRabbit status. Trigger with @coderabbitai review, then --poll (normal mode would poll-timeout waiting for an auto-review that never starts).

The completion signal is the CodeRabbit commit status or the CodeRabbit / Review check run, not the review object. CodeRabbit reports the review lifecycle (Review queued → Review in progress → Review completed) on the PR's head SHA — as a commit status on most installs, or, on installs that post no commit status, as the CodeRabbit / Review check run (Checks API, app.slug == "coderabbitai"); the script reads whichever is present. A clean review posts no review object at all — a zero-finding run leaves only the walkthrough comment plus the Review completed terminal signal. Read the outcomes as:

• Review completed + review object for the head commit → findings exist; triage them.
• Review completed + no review object → ambiguous: either a genuine zero-finding pass or a non-review skip that still emits terminal success. A skip means the review never ran, so it must never be recorded as zero findings. Three skip causes are known, each leaving a discriminator in CodeRabbit's bot issue comments; the script rules them all out (exit 2) before declaring a clean pass:
  • Auto-pause — auto_pause_after_reviewed_commits stops the review after N commits. Discriminator: the HTML marker review paused by coderabbit.ai (inside an HTML comment). Resume with @coderabbitai review, then --poll.
  • Rate-limit — the per-developer PR review limit was reached, so the review never started (visible prose: [!WARNING] / "Review limit reached"). Discriminator: the HTML marker rate limited by coderabbit.ai. Resume after the reset window (or push a new commit), then --poll.
  • File-count — the diff exceeds the max-files-per-review cap, so the review was skipped to avoid a low-quality review (visible prose: [!IMPORTANT] / "Review skipped" / "skipped due to max files limit"). Discriminator: the machine prose skipped due to max files limit co-occurring in the same comment with the skip review by coderabbit.ai marker. CodeRabbit reuses that marker for this and for intentional config skips, so the marker alone cannot separate an unintended file-count skip from a deliberate one; the prose alone would self-trigger on a walkthrough that quotes it — requiring both in one comment scopes the match to a real file-count skip (a marker in one comment plus the prose in a separate one does not qualify). Re-polling alone will not help while the file count is unchanged: either split the PR so fewer files change (the push re-reviews on its own), or raise the limit (plan / .coderabbit.yaml) and trigger a fresh review manually with @coderabbitai review — a config change does not auto-apply to an open PR — then --poll.
  • When no skip signature is present, the completion is a genuine zero-finding pass and the gate passes. The marker checks require the <!-- wrapper so bot walkthrough prose describing these features does not self-trigger; deliberate config skips are intentionally left as clean.
• No terminal signal at all (poll timeout) — neither a CodeRabbit commit status nor a CodeRabbit / Review check run reached a terminal state → the review never started; halt and surface to the user — check the app installation, plan, and automatic-review settings in the CodeRabbit dashboard, or post @coderabbitai full review and read the response.

The walkthrough / summary issue comment (collapsed 📝 Walkthrough, pre-merge checks, poem) proves nothing either way — it is posted before the review completes. A repo .coderabbit.yaml setting reviews.commit_status: false suppresses the commit status, but the script then falls back to the CodeRabbit / Review check run, so completion is still detected. When a review object does exist, it is attributed to the head commit by commit_id, so stale reviews of earlier commits are never reported.

Use pr-with-coderabbit-review.sh for the entire flow. Do NOT create the PR separately with gh pr create and then try to poll — the script handles PR creation and status-based polling in one shot.

Normal mode: create PR + poll

${CLAUDE_SKILL_DIR}/scripts/pr-with-coderabbit-review.sh --title "fix: foo" --body-file /tmp/body.md --base main

All arguments are forwarded to gh-post pr create. The script then polls for the CodeRabbit completion signal (the CodeRabbit commit status or the CodeRabbit / Review check run) until it reaches a terminal state, outputting one of: the review body plus that review's inline comments; an explicit zero-finding result; or a non-review-skip notice (exit 2, for auto-pause / rate-limit / file-count) when the terminal success reflects a skipped, unreviewed push. The notice states the per-cause remedy.

Inline --body <string> and -b <string> are rejected by gh-post — the wrapper exists to keep every body through its hardwrap validator. Use --body-file <path> (preferred) or --body-stdin.

Re-review mode: after pushing fixes

After pushing fix commits to an existing PR, wait for the incremental review the push already triggered:

${CLAUDE_SKILL_DIR}/scripts/pr-with-coderabbit-review.sh --re-review https://github.com/owner/repo/pull/123

No trigger step is needed — CodeRabbit starts the incremental review on the push itself. The status lives on the new head SHA, so the wait is race-free even when the review finishes before the script starts.

Poll-only mode: recovery for existing PRs

${CLAUDE_SKILL_DIR}/scripts/pr-with-coderabbit-review.sh --poll https://github.com/owner/repo/pull/123

Same wait as --re-review; use it after triggering a review manually. On timeout the script prints the manual trigger command (gh pr comment <N> --body "@coderabbitai full review") for the case where automatic reviews are disabled or the app missed the event.

Environment variables

• CODERABBIT_POLL_INITIAL — initial poll interval in seconds (default: 60)
• CODERABBIT_POLL_MAX — max poll interval (default: 300)
• CODERABBIT_POLL_ATTEMPTS — max attempts (default: 10)

Triage the review

CodeRabbit reviews see the diff and file contents and run linters / AST checks, but lack project-specific knowledge about design decisions, test coverage, or runtime verification. The output is chattier than most reviewers: the summary body carries collapsed sections (walkthrough, nitpicks, additional comments) and actionable items arrive as inline comments.

For each finding:

1. Cross-check against what you already know from the current conversation — code you've read, tests you've run, decisions made with the user
2. Classify under the finding-triage SSOT dispositions — typically actionable, false-positive (with reasoning), or uncertain-validity; a real finding whose fix is non-local is opens-a-question → re-enter research
3. Present the triage to the user — don't dump raw review output

Items inside the collapsed nitpick sections of the review body are findings too — expand and triage them under the same dispositions rather than skipping the collapsed blocks.

Respond to review

Each CodeRabbit finding lives on an inline thread; that thread is the unit of response. After triaging, reply within each thread via gh-post reply-inline — every reply body is validated (hardwrap detector + halt-before-send) and a single batch covers the full review:

# 1. Collect target threads — by default this filters to CodeRabbit-authored heads
#    and reports per-thread state (resolved? has reply? outdated?).
#    Use --unresolved --unreplied to narrow to threads that actually need a reply.
${CLAUDE_SKILL_DIR}/scripts/list-pr-threads.sh {owner}/{repo} {number} --unresolved --unreplied

# 2. Build a JSONL file: one {"id": <head-comment-id>, "body": "<reply text>"} per line.
#    Each reply should be concise — state the classification (fixed, false positive,
#    acknowledged) and the reasoning in 1-2 sentences.

# 3. Send the batch. The wrapper validates every body BEFORE any send; on a body
#    failure no replies post. On a mid-batch API failure it prints un-sent indices
#    and exits non-zero.
gh-post reply-inline {owner}/{repo} {number} < /tmp/replies.jsonl

If list-pr-threads.sh --unresolved --unreplied returns zero lines: every CodeRabbit thread is already resolved or already has a reply — do NOT post additional replies. Surface this to the user and ask before doing anything else. Stacking a duplicate "addressed in …" reply on a closed thread is the failure mode this wrapper exists to prevent.

CodeRabbit converses on its own threads: it may reply to your reply, and its incremental reviews may resolve threads it judges addressed by a fix commit. Both states are already excluded by --unresolved --unreplied, so re-run list-pr-threads.sh after each incremental review instead of caching an earlier thread list.

Direct gh api .../comments/{id}/replies -F body=... is still possible but defeats both the body-validation guarantee and the thread-state filter — use it only for one-off cases where the JSONL ceremony is overhead, and verify thread state via list-pr-threads.sh first.

Outside-diff findings sit in the review body, not on an inline thread (CR couldn't anchor them to the diff), so reply-inline can't reach them. Respond only when one needs visible closure (e.g. a Critical-flagged false positive), with a single top-level PR comment — the only channel; otherwise leave it.

Prerequisites

• CodeRabbit GitHub app installed on the repository — PR reviews on private repositories require a Pro plan or active trial; public repositories get them free
• gh CLI
• gh-post on PATH — the script routes PR creation through gh-post pr create so the body passes the wrapper's validator stack
• Optional: .coderabbit.yaml in the repo root to tune noise (profile: chill), path filters, and auto-review behavior

Combined pipeline with codex-review

codex review loop (pre-PR, local)
    ↓ clean
${CLAUDE_SKILL_DIR}/scripts/pr-with-coderabbit-review.sh (creates PR + polls for auto review)
    ↓ review received
Triage + respond to each inline comment via gh-post reply-inline (JSONL batch)
    ↓ if fixes needed
Push fixes (triggers the incremental review)
    ↓
${CLAUDE_SKILL_DIR}/scripts/pr-with-coderabbit-review.sh --re-review <PR_URL>
    ↓ new review received
Triage + respond again