plugin/skills/github-pr-review/SKILL.md

---
name: github-pr-review
description: >
  This skill should be used when the user asks to "review this PR", "review before I merge",
  "is this safe to merge", "check this pull request", "code review", pastes a GitHub PR URL,
  or asks for a merge gate / pre-merge check. Also triggers on "review the diff", "audit this
  change", "any bugs in this PR", and requests to post review comments on GitHub.
metadata:
  version: "0.1.0"
---

# PR Review + Merge Gate

Produce a review that catches real defects and posts them as line comments on GitHub — not a summary of what the diff does.

## Gather before judging

1. `pull_request_read` for metadata: base branch, head SHA, author, mergeable state, requested reviewers.
2. Pull the diff. For anything over ~15 files, clone the repo into the session workspace and diff locally instead of paging the API.
3. Check CI first. A red build makes most of the review moot — report the failure and stop unless the user wants the review anyway.
4. Read the PR body against the repo's template. A missing test plan is itself a finding.

## Review dimensions

Run every dimension. Do not stop at the first finding.

**Correctness** — off-by-one, null/undefined paths, unhandled promise rejections, wrong operator precedence, timezone and float arithmetic, missing early returns, state mutated during iteration.

**Security** — injection (SQL, command, prompt), missing authz on a new endpoint, secrets committed in plaintext, overly permissive CORS, unvalidated user input reaching a query or file path, dependency added with a known CVE, `dangerouslySetInnerHTML`.

**Data layer (Supabase/Postgres)** — N+1 query patterns, missing index on a new filter column, RLS policy absent on a new table, migration without a rollback path, `select *` in a hot path, missing transaction around multi-write.

**Solidity (pair with `solidity-foundry`)** — reentrancy, unchecked external call return, missing access modifier, integer truncation on casts, oracle staleness not validated, unbounded loop over storage, `tx.origin` used for auth, missing event on state change, upgrade-unsafe storage layout change.

**Frontend (React/TS)** — missing dependency in a hook array, state derived in render without memo on an expensive path, `any` escaping a public type boundary, uncontrolled re-renders from an inline object prop, missing error boundary on a new route.

**Tests** — new behavior with no test, test that asserts the mock instead of the behavior, deleted test with no replacement, flake risk from timing or ordering assumptions.

## Severity, and what earns each level

| Level | Bar | Merge |
|---|---|---|
| **Blocking** | Reproducible incorrect behavior, security hole, or data loss. State the concrete input → wrong output. | No |
| **Should fix** | Real defect with narrow blast radius, or a missing test on new logic. | Author's call |
| **Note** | Style, naming, minor simplification. | Yes |

If a finding cannot be expressed as "input X produces wrong result Y", it is a Note at most. Do not inflate.

## Posting to GitHub

Use the pending-review flow so all comments land as one review, not fifteen notifications:

1. `pull_request_review_write` with method `create` — opens a pending review.
2. `add_comment_to_pending_review` per finding, anchored to file and line, each stating the defect and the fix.
3. `pull_request_review_write` with method `submit_pending` — event `REQUEST_CHANGES` if anything is Blocking, otherwise `COMMENT`.

Never `APPROVE` on the user's behalf unless they explicitly ask for an approval.

## Verify before posting

For each Blocking finding, try to refute it before it goes on the record. Re-read the surrounding code for a guard you may have missed, check whether a caller already validates the input, and confirm the line numbers match the current head SHA. Drop anything that survives on vibes alone. A confidently wrong blocking comment costs more credibility than a missed nit.

## Output to the user

Lead with the verdict — **Merge / Merge after fixes / Do not merge** — then findings, most severe first, each with file:line and the concrete failure. Close with the review URL. No recap of what the PR does; the author wrote it.