/review

Review Loop (Phase 5)

You are reviewing work before it ships. Your job is to verify correctness, run specialized review agents, fix critical issues, and produce a clean report for the developer.

Step 0: Verify Agent Dispatch

Before running review agents, confirm the Task tool works:

1. Launch a trivial Task agent — Dispatch a general-purpose agent with the prompt: "Reply with the single word: pong". Set max_turns to 1.
2. If it completes and returns "pong" (or any response) → proceed to Step 1.
3. If it fails or times out → Stop with: "Agent dispatch failed. Cannot run review agents. Check Task tool availability."

This catches the case where you'd wait for 3 parallel agents that all silently fail.

Step 1: Self-Verification

Before launching review agents, verify your own work against the execution plan:

1. Check each plan step — Was it completed? Does the implementation match what was planned?
2. Run the test suite — Execute the project's test command (check package.json scripts, CLAUDE.md, or common patterns like npm test, npx vitest, npx jest). If no test suite exists, note it and proceed.
3. Verify the build — Run npm run build (or equivalent) to catch type errors and build failures.
4. Review your own diff — Run git diff and read through every change. Look for:
   • Files you changed that you didn't mean to
   • Debug code, console.logs, or TODO comments left behind
   • Incomplete implementations or placeholder code

If self-verification reveals issues, fix them before proceeding to agent review.

Step 2: Launch Review Agents

Dispatch three specialized review agents in parallel using the Task tool. Each agent reviews the current diff against the codebase.

Prepare the review context first:

• Run git diff main...HEAD (or appropriate base branch) to capture all changes
• Identify which files were modified, added, or deleted

Launch all three simultaneously:

1. code-reviewer agent — "Review the code changes on this branch for bugs, logic errors, and quality issues. The diff is from git diff main...HEAD. Use P1/P2/P3 severity."

2. security-reviewer agent — "Review the code changes on this branch for security vulnerabilities. The diff is from git diff main...HEAD. Use P1/P2/P3 severity."

3. typescript-reviewer agent — "Review the code changes on this branch for TypeScript, React, and Next.js issues. The diff is from git diff main...HEAD. Use P1/P2/P3 severity."

Wait for all three to complete.

Step 3: Collect & Classify Findings

Merge findings from all three agents into a single report, deduplicated and sorted by severity:

## Review Findings

### P1 — Must Fix
- [Finding from agent] — [file:line]
- ...

### P2 — Should Fix
- [Finding from agent] — [file:line]
- ...

### P3 — Nit
- [Finding from agent] — [file:line]
- ...

---
**Totals**: X P1, Y P2, Z P3
**Sources**: code-reviewer (A findings), security-reviewer (B findings), typescript-reviewer (C findings)

Step 4: Fix Loop (P1s Only)

If there are P1 findings:

1. Fix each P1 issue.
2. Re-run the test suite and build to verify fixes don't break anything.
3. Re-launch only the relevant review agent(s) to verify the P1 is resolved.
4. Max 3 loops. If a P1 persists after 3 fix attempts, flag it for human review with full context on what was tried.

If there are no P1 findings, skip to Step 5.

Step 5: Final Report

Present the final state to the developer:

## Review Complete

**P1 (fixed)**: [list what was fixed, or "None"]
**P2 (your call)**: [list remaining P2s with context]
**P3 (FYI)**: [list P3s briefly]

**Tests**: Passing / Failing (details)
**Build**: Clean / Errors (details)

**Verdict**: Ready to ship / Needs your input on P2s / Blocked on P1

If all P1s are fixed and tests pass, suggest: "Ready for /britenites:ship when you are."

If P2s need decisions, ask the developer which to fix and which to accept.

Rules

• Always self-verify before launching agents — catch the obvious stuff yourself.
• Launch all 3 agents in parallel — don't wait for one before starting another.
• Only fix P1s automatically. P2s require developer approval.
• Never suppress or downgrade a P1 finding. If you disagree with an agent's classification, present both perspectives to the developer.
• If no test suite exists, flag that as a P2 finding ("no automated tests").
• The fix loop is for P1s only — don't enter a fix loop for P2s or P3s.