| 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 |
| argument-hint | [pr-number-or-branch] |
Code Review — Multi-Agent PR Review
User-invokable orchestrator for code-PR review across any repo: Copilot for line-level findings, parallel subagents for the architectural and operational misses line-level review can't see. (Validation evidence and origin story: README.)
Fallbacks: no gh or not a GitHub repo → review the local diff (git diff <base>...HEAD) and skip the Copilot lane entirely; no subagent tool → run the review angles sequentially in the main thread. Detect what's available before assuming Copilot access.
Invocation: deliberately model-invocable — "review this PR" phrasing is the trigger. Pushes and issue-creation inside the flow are confirmation-gated.
Step 1: Assess PR Scope
The review target is $ARGUMENTS (a PR number, branch, or URL). If empty, infer it from the conversation or ask.
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 line-level review misses (platform execution limits are the canonical class).
- 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:
- Repo path (absolute) + branch + commit hash at HEAD
- Files to focus on (paste the changed-files list from Step 1)
- Confidence rating + critical findings first
- 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. Pushing changes external state: unless the user already asked you to fix or babysit this PR, confirm before the first push of the session.
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 (confirm with the user before creating it), 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.
