Codebase review

Codebase Review

Full codebase review using two waves of parallel agents. Tested at scale: 65 findings → 5 audit agents → 2 FP removed, 14 corrected → 18 work items captured for triage.

Argument: optional scope constraint (directory, layer, or "recent changes"). Default: entire codebase.

Progress Checklist

Copy this checklist and track your progress:

Codebase Review Progress:
- [ ] Phase 1: Launch 6 review agents (security, perf, arch, testing, data, UX)
- [ ] Phase 2: Launch audit agents (one per review agent with findings)
- [ ] Phase 3: Consolidate — remove FPs, correct partial findings, present table
- [ ] Phase 4: Capture confirmed findings via capture_raw

Phase 1: Review Wave

Launch 6 review agents in parallel. In Claude Code, use the Agent tool with subagent_type=Explore. In Codex, spawn 6 explorer agents explicitly for the read-heavy review wave. Each agent has a specific focus area and produces findings as a numbered list. If your harness cannot spawn subagents (e.g. Gemini CLI, Copilot CLI, or plain chat), do this work yourself sequentially, using each agent's brief above as a checklist.

Agent 1 — Security:

Review {scope} for security issues. Check for: injection vulnerabilities (SQL, command, XSS), authentication/authorization gaps, secrets in code, insecure deserialization, SSRF, path traversal, timing attacks, missing input validation at system boundaries. For each finding: file, line, severity (critical/high/medium/low), description, suggested fix.

Agent 2 — Performance:

Review {scope} for performance issues. Check for: N+1 queries, missing indexes, unbounded queries, unnecessary re-renders, memory leaks, missing pagination, expensive operations in hot paths, missing caching opportunities. For each finding: file, line, severity, description, suggested fix.

Agent 3 — Architecture:

Review {scope} for architecture issues. Check for: layer violations, circular dependencies, god classes/functions, missing abstractions, inconsistent patterns, violation of project conventions (check CLAUDE.md or AGENTS.md), dead code, orphaned files. For each finding: file, line, severity, description, suggested fix.

Agent 4 — Testing:

Review {scope} for testing gaps. Check for: untested public functions, missing edge case tests, flaky test patterns (time-dependent, ordering-dependent), test isolation issues, missing integration tests for critical paths, tests that don't assert meaningful behavior. For each finding: file, line, severity, description, suggested fix.

Agent 5 — Data Integrity:

Review {scope} for data integrity issues. Check for: race conditions, missing transactions, inconsistent state updates, missing validation on data mutations, silent data corruption risks, missing audit trails for important operations, schema drift. For each finding: file, line, severity, description, suggested fix.

Agent 6 — Client/UX (if applicable):

Review {scope} for client-side and UX issues. Check for: accessibility gaps, broken responsive layouts, missing loading/error states, inconsistent UI patterns, unhandled edge cases in user flows, missing keyboard navigation. If no client-side code exists, report "N/A — no client code in scope." For each finding: file, line, severity, description, suggested fix.

Collect all findings into a consolidated list, deduplicating across agents.

Phase 2: Audit Wave

Launch audit agents in parallel — one per review agent that produced findings. Skip agents that found nothing. If your harness cannot spawn subagents (e.g. Gemini CLI, Copilot CLI, or plain chat), do this work yourself sequentially, using each agent's brief above as a checklist.

Each audit agent receives the findings from its corresponding review agent:

You are auditing {focus area} findings from a codebase review. For each finding below, verify it against the actual code. Classify as:

• CONFIRMED — the finding is accurate and actionable
• PARTIALLY CORRECT — the finding identifies a real issue but the description or severity is wrong. Provide correction.
• FALSE POSITIVE — the finding is incorrect. Explain why.

Findings to audit: {findings list}

Phase 3: Consolidate

1. Remove all FALSE POSITIVE findings

2. Update PARTIALLY CORRECT findings with corrections

3. Keep all CONFIRMED findings as-is

4. Present the final verified findings to the user with a summary table:

   #	Area	Severity	File	Description	Verdict

5. Ask user to confirm which findings should be captured for triage

Phase 4: Capture Findings

For each confirmed finding, call capture_raw with the following format. Get today's date by running date +%Y-%m-%d — do not assume you know it — and use it for the source tag:

capture_raw("**<brief description>.** <details>. <severity> priority. Source: codebase-review_YYYYMMDD.")

The adapter creates .maestro/inbox.md if it does not already exist.

Suggest running /triage in Claude Code or $triage in Codex next to process the new work items.