PR Reviewer

Receipts, not polish. A review that says "looks good" is not a review. A review that says "line 47 will panic on a nil pointer and here is the fix" is a review.

This skill runs a structured code review and produces a verdict: PASS or FAIL. Every finding is actionable. No findings without evidence.

When to use

• Before merging any PR
• When reviewing agent-generated code before committing
• When a significant refactor or new feature is complete
• As part of the release-readiness checklist

Procedure

1. Scope the review — fetch the diff. Identify:
   • Files changed and their purpose
   • The stated intent of the PR (what problem does it solve?)
   • Any linked spec, issue, or ADR
2. Correctness pass — look for bugs, not style:
   • Null / nil / undefined access on values that could be absent
   • Off-by-one errors in loops or index operations
   • Error returns that are silently ignored
   • Race conditions if the code is concurrent
   • Missing input validation at trust boundaries (user input, external API responses)
   • Logic that diverges from the stated intent of the PR
3. Reuse pass — look for duplication:
   • Does this code reimplement something that already exists in the codebase?
   • Is there a stdlib or already-installed dependency that does this?
   • Can the new code be expressed using an existing abstraction?
4. Simplification pass — look for unnecessary complexity:
   • Can this be fewer lines without losing clarity?
   • Is there an abstraction that has only one implementation? (remove it)
   • Is there a dependency added for something a few lines of code would do?
   • Is there "scaffolding for later" that should not exist yet?
5. Security and data boundary pass — look for:
   • Secrets or credentials in code or logs
   • User-controlled input reaching a shell, SQL query, or file path without validation
   • Data crossing a tenant or trust boundary without an explicit check
   • Missing authentication or authorisation checks on new endpoints
6. Classify findings:
   • BLOCKING: must be fixed before merge (correctness bugs, security issues)
   • SUGGESTED: improvement worth making but not a blocker (simplification, reuse)
   • NOTE: observation for awareness, no action required
7. Produce the verdict:
   • PASS: no unresolved conditions; optional suggestions are explicitly nonblocking
   • FAIL: one or more unresolved conditions

Outputs

• Finding list: location | severity | description | suggested fix
• Verdict: PASS / FAIL; conditions never count as approval

Independent review and Council

The independent review must include an adversarial subagent activity: attempt to refute requirement acceptance, exercise failure/recovery assumptions and identify tests that miss the real user path. One fresh reviewer may cover these lenses; extra agents are not a quota. Give it requirement links and the exact candidate, not the author's desired verdict.

The writer never approves its own work. Use a fresh, read-only reviewer context on a permitted host and retain a substantive verdict bound to the exact candidate. A provider change is optional and never grants authority. Missing review is incomplete.

Read independent Council review before dispatch or when a review loop stalls. It separates the shared contract, host adapters and local subscription policy, preserves unresolved findings across reviewers, and defines bounded escalation. It also documents the optional fail-closed local receipt helper.

Guardrails

• Every BLOCKING finding must have a suggested fix. "This is wrong" is not actionable.
• Style is not a BLOCKING finding. Style enforcement belongs in the linter, not the review.
• Reuse requires evidence. "This might already exist" is not a finding. Find it or drop the comment.
• The review is not a rewrite. Surface the issues; let the author fix them.

See also: REFERENCE.md for a quick lookup of common patterns and their preferred alternatives.