Diagnostic de pipeline d'examen PR

---

name: debug-review description: Trace a specific PR through the review pipeline to diagnose failures. user-invocable: true

You are a debugging assistant for diagnosing PR review issues in the claude-code-reviewer service. You trace a specific PR through the entire pipeline and produce an actionable diagnosis.

Input

Required: A PR identifier — owner/repo#N or just N (PR number).

If no argument is provided, ask the user for a PR number.

Step 1: Load PR State

Read data/state.json and find the PR entry. If only a number is given, search all entries for a matching number field. If not found, report that the PR has no state entry (may never have been seen by the service).

Display a brief status summary: key, status, headSha (7 chars), lastReviewedSha (7 chars), title.

Step 2: Error Analysis

If lastError is set:

1. Show the error details: phase, message, occurredAt
2. Explain what the phase means:
   • diff_fetch — failed to fetch the PR diff via gh pr diff
   • clone_prepare — failed to clone/fetch the repo or create a worktree
   • claude_review — Claude CLI returned an error or timed out
   • comment_post — failed to post the review comment/PR review to GitHub
3. Show consecutiveErrors vs maxRetries (default: 3)
4. Calculate current backoff: 60s * 2^(consecutiveErrors - 1)
5. Show when the backoff expires: lastError.occurredAt + backoff

If no errors, note that the error tracking is clean.

Step 3: Review History

List all ReviewRecord entries:

• sha (7 chars) | reviewedAt | verdict | posted | findings breakdown | commentId/reviewId

Highlight:

• Reviews with posted: false (review ran but comment failed to post)
• Reviews with blocking findings (severity: "issue" or blocking: true)
• The most recent review's verdict and whether it matches the current state

Step 4: Skip Analysis

If status is skipped:

• Show skipReason and explain:
  • draft — PR is marked as draft; clears when marked ready_for_review
  • wip_title — title starts with "WIP"; clears when title is edited
  • diff_too_large — diff exceeds maxDiffLines; show skipDiffLines vs limit
• Whether the skip condition could clear automatically: draft clears on ready_for_review event; wip_title clears when title is edited to remove WIP prefix; diff_too_large clears when new commits are pushed (SHA changes), then the diff is re-checked on next review attempt

Step 5: shouldReview Simulation

Walk through the gating logic in src/state/decisions.ts with the current PR state:

1. Terminal state check (merged/closed)
2. In-progress lock (reviewing)
3. Config-based skips (skipDrafts + isDraft, skipWip + title starts with "wip")
4. Skipped state (still skipped after evaluateTransitions clears conditions)
5. Already reviewed this SHA (status === "reviewed" AND lastReviewedSha === headSha). Bypass: forceReview skips this gate.
6. Debounce check (lastPushAt age vs debouncePeriodSeconds). Bypasses: skipped when forceReview is true, or when last review verdict was REQUEST_CHANGES and SHA has changed (author is fixing comments).
7. Error backoff: first checks consecutiveErrors >= maxRetries (permanent block), then timed exponential backoff. Bypass: forceReview skips both checks.
8. Ready states (pending_review, changes_pushed, error, reviewed with new SHA)

For each gate, report pass or blocked with the specific values. Note whether forceReview (from /review comment trigger) would change the outcome. Stop at the first blocking gate.

Read config.yaml to get the actual config values for debounce, maxRetries, skipDrafts, skipWip.

Step 6: Clone/Worktree Check

If codebase access is enabled (review.codebaseAccess: true in config):

• Check if a bare clone exists: ls data/clones/<owner>/<repo> (or configured cloneDir). Note: the directory has no .git suffix despite being a bare clone.
• Check for worktree: ls data/clones/<owner>/<repo>--pr-<N> (double dash --pr-)
• Report existence and age (last modified time)
• If missing, note that it will be created on next review

Step 7: Diff Size Check

Run: gh pr diff <N> --repo <owner>/<repo> | wc -l

Compare against maxDiffLines from config. Report whether the diff would be skipped for size.

If gh fails, report the error (may indicate auth issues or PR not found).

Step 8: Diagnosis Summary

Produce a summary with:

1. Root cause — the primary issue (e.g., "stuck in error state due to repeated claude_review failures")
2. Current blocking gate — which shouldReview gate is preventing the next review
3. Recommendations — actionable steps:
   • For errors: check Claude CLI auth, check GH_TOKEN, retry manually
   • For skipped: un-draft the PR, rename title, reduce diff size
   • For stale review: push new commits or force re-review with /review comment
   • For stuck: manually edit state.json to reset consecutiveErrors, or delete the entry
   • For missing state: the poller should pick it up on next cycle, or send a test webhook

Notes

• If gh CLI is not available or not authenticated, skip Step 7 and note it
• All timestamps should be shown in human-readable format with relative age (e.g., "2h ago")
• Keep the output focused and actionable — this is for debugging, not exploration