Pull Request Code Review

Structured GitHub PR review with severity-ranked, citation-bound findings and a clear verdict.

Goal

Structured code review of a GitHub Pull Request — correctness, security, performance, maintainability. Findings ranked by severity, each tied to a verbatim diff citation, plus a verdict (Approve / Request Changes / Comment).

When to Use

• The user names a PR by number, URL, or branch and asks for a review.
• The user asks for "feedback", "thoughts", or "concerns" on a pull request.
• The user explicitly invokes /github-pr-review or /valarmindskills:github-pr-review.

Do not use when

• The user asks about code-review methodology, not a specific GitHub PR → use @code-review.
• The user wants to author a commit or PR description, not review one → use @github-commit or @github-release-note.
• The PR contains no code diff (docs only, lockfile bumps, config-only) and the user wants security/perf analysis — refuse with out of scope: no code diff to analyze and hand off.

Inputs you must collect before starting

Procedure

Step 1 — Fetch PR metadata

gh pr view <number> --json title,body,author,baseRefName,headRefName,files,additions,deletions,commits,labels,state,isDraft

Record: title, author, base/head branch, files changed, lines added/removed, commits, state (open/closed/merged), draft flag.

If state != "OPEN" or isDraft == true, warn the user and ask whether to continue. Do not refuse — historical reviews are valid use cases.

Step 2 — Get the full diff

gh pr diff <number>

If the PR has more than 1000 changed lines, inform the user and ask whether to (a) continue full review, (b) prioritize a file subset, or (c) abort. Do not silently truncate.

Step 3 — Read description and linked issues

• Read the PR description for context and intent.
• Look up linked issues via gh issue view <number> for Closes #N, Fixes #N, Refs #N.
• Identify acceptance criteria, if any.

Step 4 — Analyze the changes

Analyze the diff across four dimensions. Only review changed code — not surrounding unchanged code.

Logic & Correctness

• Business requirements met?
• Edge cases (null, empty, boundaries, concurrency)?
• Error flows correct?
• Loop / recursion termination?
• Type consistency?

Security

• Inputs validated and sanitized?
• Injection risk (SQL, XSS, command, SSRF, path traversal)?
• AuthN / AuthZ checks?
• Sensitive data exposure (logs, responses, env)?
• OWASP Top 10 alignment?

Performance

• N+1 queries or inefficient DB access?
• Unbounded iteration?
• Allocation in hot loops?
• Algorithmic complexity vs. expected volume?
• Blocking calls where async fits?

Maintainability

• Naming clarity (variables, functions, classes)?
• Single responsibility, function size?
• Module coupling?
• Test coverage for the changes?
• Style consistency with the rest of the codebase?

Step 5 — Categorize each finding

Each finding must include Severity, Confidence, and a verbatim citation:

Citation requirement. Cite path:line AND quote the exact code (3–8 lines from the diff). A finding without both is rejected. Never paraphrase the code; never invent paths, line numbers, or function names.

Step 6 — Write the executive summary

2–4 sentences with:

• What the PR does (paraphrased from the diff, not the description).
• Quality assessment.
• Key concerns, if any.
• Verdict: Approve, Request Changes, or Comment.

The verdict must match the highest finding severity:

• Any Critical or Major → Request Changes.
• Only Minor / Nitpick → Comment or Approve (author discretion).
• Zero findings after Step 8 verification → Approve with LGTM.

Step 7 — Present the review

Follow the template in EXAMPLE.md. For zero-findings PRs, follow the LGTM template in the same file.

If the user explicitly approves, publish via:

gh pr review <number> --approve --body "<message>"
gh pr review <number> --request-changes --body "<message>"
gh pr review <number> --comment --body "<message>"

Never publish without explicit user approval — one approval is scoped to one publish action.

Step 8 — Verify before presenting

Before returning the review, check:

1. Every cited path:line appears in the diff from Step 2.
2. Every code quote is verbatim from the diff (no paraphrase, no fabricated lines).
3. Every finding has a Severity and a Confidence field.
4. The verdict in the executive summary matches the highest finding severity per Step 6.
5. If zero findings, the output uses the LGTM template — not a synthesized Minor.
6. No finding was promoted from Minor to Major to fill the report.

If any check fails, repair before returning.

Refusal hooks

• Out of scope (non-code PR): if the PR is docs-only, lockfile-only, or config-only and the user asked for security / performance analysis, respond out of scope: no code diff to analyze — switching to <appropriate skill> and hand off.
• Ambiguous PR identifier: if multiple PRs match (e.g., a branch name with multiple open PRs), ask exactly one clarifying question listing the candidates, then stop.
• Base branch missing locally: if gh pr diff fails because the base ref is not fetched, request git fetch <remote> <base> from the user and stop.
• Publish without approval: never call gh pr review --approve|--request-changes|--comment without an explicit, in-conversation approval scoped to that PR.

Constraints

• Only review changed code — do not review surrounding unchanged code.
• Do not assume context not visible in the diff; read referenced files via Read to understand the change, but do not flag them.
• Be constructive — propose a fix alongside every finding.
• Prioritize by severity; do not bury Critical issues under Nitpicks.
• Cite path:line AND a verbatim 3–8 line code quote per finding. Never invent paths, line numbers, function names, CVE IDs, or RFC numbers.
• Never fabricate findings to fill the report. Zero findings is a valid outcome — emit the LGTM template and stop.
• Never promote a Minor to Major to look thorough. Severity is bound to impact × likelihood, not to report length.
• If the PR is > 1000 lines, ask before proceeding (Step 2).
• Never publish the review on GitHub without explicit, scoped user approval.

Output format

Structured review following the template in EXAMPLE.md:

• Executive summary (2–4 sentences) ending with the verdict.
• Findings grouped by severity (Critical → Major → Minor → Nitpick), each with Confidence, path:line, code quote, suggested fix.
• Summary by severity table.
• Zero-findings path: LGTM template from EXAMPLE.md.

Example request

• "Review PR #42"
• "Analyze this pull request: https://github.com/org/repo/pull/123"
• "Do a quick review of the payment feature PR"
• "Revisar PR #87 desse repo"
• "Feedback no PR mais recente"

Related Skills

• @code-review — lifecycle code-review skill (multi-language Go/Rust/TS). Use when the user wants methodology-driven review not tied to a GitHub PR (audit, pre-merge gate, refactor review).
• @code-debugger — when a finding here uncovers a runtime bug that needs root-cause analysis.
• @github-commit — when the review concludes Approve and the user wants to author the merge commit.
• @code-security-review — hand off security findings into a stack-aware audit (Go branch — references/golang/; Next branch — references/nextjs/).

References