| name | get-pr-code-review |
| description | Pull Codex and CodeRabbit code review feedback from a GitHub PR, analyze each finding against the actual code, rank by priority, and label every item with emoji verdicts (✅ agree, ❌ disagree, 🤔 needs human review). Use this skill whenever the user wants to triage PR review comments, understand which review findings matter, or get a prioritized summary of automated code review feedback. Also triggers when the user mentions "review feedback", "codex review", "coderabbit review", "triage PR comments", or "what did the reviewers say". |
Fix PR Review
Sync automated review feedback from Codex and CodeRabbit on a GitHub PR, analyze
every finding against the current branch, rank by priority, and label each one
with an emoji verdict so the user instantly knows what to act on.
Prerequisites
gh CLI must be installed and authenticated (gh auth status).
- If the user provides a PR URL, use that. Otherwise, use the PR for the current branch.
Step 1: Sync Feedback
Run the bundled sync script to pull review comments from GitHub:
python3 .claude/skills/get-pr-code-review/scripts/sync_review_feedback.py --pr <pr-url> --debug
Or for the current branch:
python3 .claude/skills/get-pr-code-review/scripts/sync_review_feedback.py --debug
Always use --debug so you get the structured items.json and full context needed
for analysis. Read the generated files from .review/.
Step 2: Analyze Each Finding
For every actionable finding from the sync output:
- Read the actual code at the referenced file and line range. Don't trust the
bot's description blindly — check whether the issue actually exists in the
current branch.
- Assess validity — is the finding correct? Has it already been fixed? Is it
a false positive?
- Assess priority using this scale:
- P0 — Critical: Security vulnerabilities, data loss, crashes, broken auth
- P1 — Bug: Logic errors, race conditions, incorrect behavior
- P2 — Quality: Missing error handling, poor patterns, maintainability issues
- P3 — Nitpick: Style, naming, minor suggestions, cosmetic
- Assign an emoji verdict:
- ✅ Agree — The finding is valid and should be fixed
- ❌ Disagree — The finding is incorrect, already addressed, or not applicable
- 🤔 Needs Human Review — The finding involves a product decision, architecture
tradeoff, or is ambiguous enough that a human should weigh in
Step 3: Write the Report to the Review File
Append the analysis directly to the same .review/ markdown file that was generated
by the sync script in Step 1. Do NOT output the report to the conversation only — it
must be persisted in the file.
Use the Edit tool to append a ## Triage Analysis section to the end of the synced
review file. The format:
## Triage Analysis
| Verdict | Count |
|---------|-------|
| ✅ Agree | N |
| ❌ Disagree | N |
| 🤔 Needs Human | N |
**Priority breakdown**: P0: N · P1: N · P2: N · P3: N
### <emoji> [P<n>] <short title>
**Source**: <codex|coderabbit> · **File**: `<path>:<line>`
**Verdict**: <Agree|Disagree|Needs Human Review>
<1-2 sentence rationale explaining why you agree, disagree, or are unsure>
> <original reviewer comment, abbreviated if long>
Group findings by priority (P0 first, P3 last). Within each priority group, list
agree items first, then needs-human, then disagree.
Collapsing Duplicates
When Codex and CodeRabbit flag the same issue, merge them into a single entry.
Note both sources. Don't inflate the count with duplicate findings.
Step 4: Show Summary and Offer to Fix
After writing the report to the file, show the user a brief summary in the
conversation (verdict counts + priority breakdown) and tell them the full
analysis was written to the review file.
Then ask:
"Want me to fix the ✅ agreed items? I can work through them by priority."
If the user says yes, fix them starting from P0 down. Commit each fix atomically
with a clear message referencing the finding.
References