Review
Thorough, interactive review grounded in project principles. Do NOT make changes — the review is the deliverable.
Autonomous Session Mode
When this skill runs in a non-interactive Noodle execution session (for example Cook, Oops, or Repair):
- Do not use AskUserQuestion. For each issue found, state the recommended action directly instead of presenting options.
- Write findings to a report file at
brain/audits/review-<date>-<subject>.mdin addition to session output. - For high-severity issues, file a todo in
brain/todos.mdusing the/todoskill. - For low-severity issues, note them in the report but don't file separate todos.
- Conclude by committing the review report. Do not wait for direction.
Use Tasks to track progress. Create a task for each step below (TaskCreate), mark each in_progress when starting and completed when done (TaskUpdate). Check TaskList after each step.
Step 1 — Load Principles
Read brain/principles.md. Follow every [[wikilink]] and read each linked principle file. These principles govern review judgments — refer back to them when evaluating issues.
Do NOT skip this. Do NOT use memorized principle content — always read fresh.
Step 2 — Determine Scope
Infer what to review from context — the user's message, recent diffs, or referenced plans/PRs. If genuinely ambiguous (nothing to infer), ask.
Auto-detect review mode from change size:
- BIG CHANGE (50+ lines changed, 3+ files, or new architecture) — all sections, at most 4 top issues per section
- SMALL CHANGE (under those thresholds) — one issue per section
Step 3 — Gather Context
For SMALL CHANGE reviews, read files directly in the main context — delegation overhead exceeds the cost of reading a few files. For BIG CHANGE reviews, delegate exploration to subagents via the Task tool.
Spawn exploration agents (subagent_type: Explore) to:
- Read the code or plan under review
- Identify dependencies, callers, and downstream effects
- Map relevant types, tests, and infrastructure
Run multiple agents in parallel when investigating independent areas.
Step 4 — Gather Domain Skills
Check installed skills (.claude/skills/) for any that match the review's domain. Common matches:
| Domain | Skill | When |
|---|---|---|
| Frontend UI | frontend-design | Web UI components, layouts, visual design |
| Codex delegation | codex | Tasks delegated to Codex workers |
For domains not listed above, use find-skills to search for a relevant skill.
Invoke matched skills now — read their output and use domain guidance to inform your review.
Step 5 — Review Sections
Work through all sections, then present the full review. The user can redirect mid-stream if needed.
1. Architecture
- System design and component boundaries
- Dependency graph and coupling
- Data flow patterns and bottlenecks
- Scaling characteristics and single points of failure
- Security architecture (auth, data access, API boundaries)
2. Code Quality
- Code organization and module structure
- DRY violations — be aggressive
- Error handling patterns and missing edge cases (call out explicitly)
- Technical debt hotspots
- Over-engineering or under-engineering relative to foundational-thinking principles; consider redesign-from-first-principles
3. Tests
- Coverage gaps (unit, integration, e2e)
- Test quality and assertion strength
- Missing edge case coverage — be thorough
- Untested failure modes and error paths
4. Performance
- N+1 queries and database access patterns
- Memory-usage concerns
- Caching opportunities
- Slow or high-complexity code paths
Step 6 — Issue Format
NUMBER each issue (1, 2, 3...). For every issue:
- Describe the problem concretely with file and line references
- Present 2–3 options with LETTERS (A, B, C), including "do nothing" where reasonable
- For each option: implementation effort, risk, impact on other code, maintenance burden
- Give a recommended option and why, mapped to foundational-thinking principles
- Ask whether the user agrees or wants a different direction
When using AskUserQuestion, label each option with issue NUMBER and option LETTER. Recommended option is always first.
Interaction Rules
- Do not assume priorities on timeline or scale
- Do not make changes — present findings and wait for direction
- Present all sections together, then ask for feedback once at the end
- Per prove-it-works: if something can be tested, note how in the issue description