| name | pr-review |
| description | Perform thorough pull request reviews covering correctness, security, performance, observability, backward compatibility, data migrations, and developer ergonomics. Use when the user asks to review a PR, examine a pull request, or requests a code review of a branch or GitHub PR link. |
PR Review
Perform a structured, multi-dimensional review of a pull request. The review is advisory, not authoritative -- the human reviewer verifies all claims.
Workflow
1. Gather the PR Context
Determine the source of the PR:
GitHub PR URL or number provided:
gh pr view <number-or-url> --json title,body,additions,deletions,changedFiles,baseRefName,headRefName
gh pr diff <number-or-url>
Local branch (no PR link):
git log --oneline main..HEAD
git diff main...HEAD
git diff main...HEAD --stat
Collect:
- PR title & description (the "why")
- Full diff (the "what")
- File-level stats (additions / deletions / files changed)
2. Assess Scope & Reviewability
Before diving into the code, evaluate the PR holistically:
- Is the PR small enough to review confidently? If it touches > ~400 lines or > ~10 files, flag it and suggest splitting.
- Does every change align with the stated goal? Call out any "drive-by" refactors that are not justified in the description.
- Is there a clear rollout plan and expected user impact? If the description is silent on this, ask.
3. Run the Review Checklist
Walk through each category below. For every finding, cite the exact file and line range (or patch hunk) so the author can locate it instantly.
For the detailed checklist items under each category, see CHECKLIST.md.
Correctness & Edge Cases
- Boundary cases: nulls, empty collections, timezones, retries, partial failures
- Explicit failure modes: what happens when a dependency is down?
- Idempotency: are retries safe (no double effects)?
Security & Privacy
- No secrets logged or committed
- Input validation + output encoding where applicable
- Least privilege for new permissions/roles
- PII handling: logging/redaction, retention, access controls
Observability
- Structured, meaningful logs (not noisy)
- Key metrics: counters/timers relevant to the change (SLO/SLA impact)
- New flows tied into existing traces (if tracing is used)
- Alerts are actionable (no "FYI" alerts)
Performance & Cost
- Hot-path code measured or at least reasoned about
- Query/index checks: N+1, missing indexes, pagination
- Caching strategy and cache invalidation story
- Cloud cost impact: extra calls, bigger payloads, heavier compute
API & Backward Compatibility
- Non-breaking changes where possible
- Versioning / deprecation strategy if breaking is unavoidable
- Migration steps for schema/config changes
Data & Migrations
- Migrations are backward compatible (expand -> migrate -> contract)
- Rollback strategy for data changes (not just code rollback)
- Safety for long-running migrations: batching, locks, timeouts
Docs & Developer Ergonomics
- README / runbook / ADR updates when behavior changes
- "How to test locally" is clear
- Notable operational steps called out (env vars, toggles, backfills)
4. Format the Review
Structure the output as follows:
## PR Review: <PR title>
### Summary
<1-2 sentence overview of what the PR does and overall impression>
### Scope & Reviewability
<Assessment of PR size, focus, rollout plan>
### Findings
#### Critical (must fix before merge)
- **[file:line]** <description of issue and suggested fix>
#### Suggestions (consider improving)
- **[file:line]** <description and rationale>
#### Nits (optional, low priority)
- **[file:line]** <minor style or clarity improvement>
### Checklist Summary
| Category | Status | Notes |
|-----------------------------|--------|--------------------|
| Correctness & Edge Cases | pass/warn/fail | ... |
| Security & Privacy | pass/warn/fail | ... |
| Observability | pass/warn/fail | ... |
| Performance & Cost | pass/warn/fail | ... |
| API & Backward Compat | pass/warn/fail | ... |
| Data & Migrations | pass/warn/fail | ... |
| Docs & Developer Ergonomics | pass/warn/fail | ... |
### Verdict
<approve / request-changes / comment-only>
<Brief rationale>
5. Review Principles
- Cite exact locations: every finding references a file and line range or patch hunk.
- Explain the "why": don't just say "this is wrong" -- explain the risk or impact.
- Be proportional: a one-line typo fix doesn't need a performance analysis. Skip categories that are clearly not applicable.
- Separate blocking from non-blocking: use the Critical / Suggestion / Nit tiers so the author knows what must change.
- Acknowledge good work: if something is well-done, say so briefly.
