code-review

name	code-review
description	Multi-agent deep review for code PRs in any repo. Use when asked to "deep review this PR," "multi-agent review," "review

Code Review — Multi-Agent PR Review

User-invokable orchestrator for code-PR review across any repo. Validated on a Vercel-deployed Next.js app PR across 5 review rounds: Copilot caught ~10 line-level issues; parallel subagents caught the P0 architectural miss (Vercel maxDuration) Copilot missed across all 5 rounds, plus a persistent-replay vector and 4 operational issues.

If you have access to a paid multi-agent review service (e.g. /ultrareview in Claude Code), this skill is the in-flight free version of the same idea — useful when you want this depth without spinning up a billed cloud review.

---

Step 1: Assess PR Scope

gh pr view N --json files,additions,deletions,title

Categorize:

Tier	Examples	Treatment
Trivial	Typo, copy edit, single-line deps bump, README polish	Skip multi-agent. Copilot alone is enough.
Standard	Component refactor, single-feature bug fix, contained logic change	Copilot + 1 subagent (pick the angle that matches the risk shape)
High-stakes	Payment, auth, crypto/wallet, RPC, deployment configs, scale-relevant changes, new external SDK integrations	Copilot + 2 parallel subagents minimum (always include adversarial + operational)

If trivial → stop reading this skill, just request Copilot review and ship.

---

Step 2: Spawn Subagents in Parallel

Three angles that empirically work. Pick by risk shape:

• Adversarial — "Try to break it. What's the worst attack?" Finds security holes, abuse vectors, replay/race conditions, malformed-input handling.
• Operational — "What fails in production at scale? Latency, cost, observability, timeouts, env config, retries, cold-start?" Finds the architectural P0s Copilot misses (Vercel maxDuration was the canonical one).
• Reference-comparison — "Does this match the upstream reference impl exactly? Where does it diverge and why?" Use for third-party SDK / spec integration (Phantom deep-link, Stripe, OAuth providers, RPC clients).

Spawn all chosen subagents in a single tool-call batch so they run in parallel. Prompt templates live in patterns/subagent-prompts.md.

Every subagent prompt must specify:

1. Repo path (absolute) + branch + commit hash at HEAD
2. Files to focus on (paste the changed-files list from Step 1)
3. Confidence rating + critical findings first
4. Cap report length at ~500 words

---

Step 3: Request Copilot Review (in parallel with subagents)

gh pr edit N --add-reviewer copilot-pull-request-reviewer

Don't wait for Copilot to start before spawning subagents. They run in parallel.

---

Step 4: Triage Findings

Three buckets:

Real findings — fix now. Push fix to the PR branch.

Stale re-flags — Copilot will sometimes re-flag findings against unmoved diff lines after fixes ship at HEAD. Always read the file at HEAD before re-fixing. If the fix is already there, drop a one-line reply on the comment ("Fixed at ") and move on. Do not re-fix.

Deferred (filed) — tests, refactors, or work outside scope. File as a GitHub issue, link the issue number in the merge commit message, move on. Don't expand the PR.

---

Step 5: Iteration Cap

Maximum 2 Copilot review rounds. Empirically, by round 3 the signal-to-noise ratio collapses (~50% stale re-flags). Ship after round 2. If subagents flag genuinely new issues after round 2, those are issue-tracker items, not PR-blockers.

---

Step 6: When NOT to Use This Skill

• Already-merged PRs → use Claude Code's built-in /security-review if you suspect a problem.
• Trivial-tier PRs (per Step 1) → just request Copilot, skip the multi-agent overhead.
• PRs where a paid multi-agent service was already run → don't double-spend.
• Skill / prompt / instruction-file repos (where the "code" is markdown) → multi-agent is overkill; a single review pass against the trigger-quality and instruction-quality dimensions is enough.

---

Empirical Evidence

Validation case (5 rounds on a high-stakes PR):

• Copilot caught ~10 real line-level issues (null checks, error handling, type mismatches).
• Subagents caught what Copilot missed across all 5 rounds:
  • P0: Vercel maxDuration not set — function would time out under realistic load. Architectural miss invisible at the line level.
  • Persistent-replay vector in the auth flow.
  • Transient poll tolerance (no retry budget on flaky upstream).
  • Env-configurable timeout (hardcoded value).
  • Structured logging (string-concat logs, unparseable in production).
  • bodyParser size limit (DoS surface).
• Round 3+ pattern: Copilot re-flagged the same already-fixed lines. Confirmed the round-2 cap.

Subagents are not redundant with Copilot. They cover a different layer.

---

Cross-references

• patterns/subagent-prompts.md — paste-ready prompt templates for the three angles.