review（审查）

<tool_restrictions>

MANDATORY Tool Restrictions

BANNED TOOLS — calling these is a skill violation:

• EnterPlanMode — BANNED. Do NOT call this tool. This skill has its own structured process. Execute the steps below directly.
• ExitPlanMode — BANNED. You are never in plan mode. </tool_restrictions>

<arc_runtime> This workflow requires the full Arc bundle, not a prompts-only install. Resolve the Arc install root from this skill's location and refer to it as ${ARC_ROOT}. Use ${ARC_ROOT}/... for Arc-owned files such as references/, disciplines/, agents/, templates/, and scripts/. Use project-local paths such as .ruler/ or rules/ for the user's repository. </arc_runtime>

<required_reading> Read these reference files NOW:

1. ${ARC_ROOT}/references/review-patterns.md
2. ${ARC_ROOT}/disciplines/dispatching-parallel-agents.md
3. ${ARC_ROOT}/disciplines/receiving-code-review.md </required_reading>

<rules_context> Check for project coding rules:

Use Glob tool: .ruler/*.md

Determine rules source:

• If .ruler/ exists: Read rules from .ruler/
• If .ruler/ doesn't exist: Read rules from rules/

Pass relevant core rules to each reviewer:

<progress_context> Use Read tool: docs/arc/progress.md (first 50 lines)

Check for context on what led to the plan being reviewed. </progress_context>

<scope_discipline>

Scope Discipline

Reviewers must respect the plan's scope. This is non-negotiable:

• Do not silently argue for less work. If you think the plan is overbuilt, raise it once in a "Scope Check" early in the review. After the user responds, commit to their decision.
• Do not sneak in additional scope. Don't suggest features, enhancements, or "while you're at it" additions beyond what the plan covers.
• Your job is to make this plan succeed, not to lobby for a different plan. Once scope is agreed, optimize within it — find bugs, catch edge cases, improve the architecture — but don't re-litigate what gets built.
• Include this principle in every reviewer prompt: "Respect the plan's scope. Flag scope concerns once, then commit to making the plan succeed." </scope_discipline>

If --diff argument provided:

• Switch to diff review mode — skip Phase 1 (plan search) entirely
• Jump to Phase 1D (Diff Review) below

If argument provided (e.g., daniel-product-engineer):

• Look for ${ARC_ROOT}/agents/review/{argument}.md
• If found → use only this reviewer, skip Phase 2 detection
• If not found → list available reviewers from ${ARC_ROOT}/agents/review/ and ask user to pick

Available reviewers:

• daniel-product-engineer — Type safety, UI completeness, React patterns
• lee-nextjs-engineer — Next.js App Router, server-first architecture
• senior-engineer — Asymmetric strictness, review discipline
• architecture-engineer — System design, component boundaries
• performance-engineer — Bottlenecks, scalability
• security-engineer — Vulnerabilities, OWASP
• data-engineer — Migrations, transactions
• designer — Visual design quality, UX fundamentals, AI slop detection

Phase 1: Find the Plan

Check if plan file path provided as argument:

• If yes → read that file and proceed to Phase 2
• If no → search for plans

Search strategy:

1. Check conversation context first — Look for Claude Code plan mode output

- Look back through recent conversation messages - Search for plan structure markers: - "# Plan" or "## Plan" headings - "Implementation Steps" sections - Task lists with implementation details - Step-by-step procedures - If found → extract the plan content and proceed to Phase 2

1. Search Arc plan folders — Look for plan files Use Glob tool: docs/arc/plans/*.md Fallback: docs/plans/*.md

- Sort results by modification time (newest first) - Show all plan files (design, implementation, etc.)

1. Present options if multiple found:

- List up to 5 most recent plans - Show: filename, modification date, brief preview - Ask user: "Which plan should I review?"

1. If no plans found:

- Check if the current branch has changes vs main: git fetch origin main --quiet && git diff origin/main --stat - If branch has changes: Offer to review the diff instead: "No plans found, but this branch has changes against main. Want me to review the diff?" - If yes → switch to Phase 1D (Diff Review) - If no → "Can you point me to a plan file, or paste the plan you'd like me to review?" - If no changes: "I couldn't find any plans or branch changes to review."

Once plan located:

• Store the plan content
• Note the source (conversation, file path, or user-provided)
• Proceed to Phase 2

Phase 1D: Diff Review

This phase runs instead of plan review when --diff is passed or the user opts into diff review from Phase 1.

1. Check branch state: git branch --show-current If on main with no changes: "Nothing to review — you're on main with no changes." Stop.
2. Read the checklist: Read: ${ARC_ROOT}/references/diff-review-checklist.md
3. Get the diff: git fetch origin main --quiet git diff origin/main
4. Run two-pass review applying the checklist against the diff:

- Pass 1 (CRITICAL): Race conditions, trust boundaries, data safety - Pass 2 (INFORMATIONAL): Conditional side effects, stale references, test gaps, dead code, performance

1. Present findings using the checklist's output format.
2. If CRITICAL issues found: For each critical issue, present as a Socratic question:

- "This pattern reads then writes without a transaction — what happens if two requests hit this simultaneously?" - Wait for user to decide: fix now, acknowledge, or mark as false positive

1. Skip to Phase 6 (Summary and Next Steps) — diff review doesn't need the full plan review pipeline.

Phase 2: Detect Project Type

Skip if specific reviewer provided in Phase 0.

Detect project type for reviewer selection:

Use Grep tool on package.json:

• Pattern: "next" → nextjs
• Pattern: "react" → react

Use Glob tool:

• requirements.txt, pyproject.toml → python

Select reviewers based on project type:

TypeScript/React:

• ${ARC_ROOT}/agents/review/daniel-product-engineer.md
• ${ARC_ROOT}/agents/review/senior-engineer.md
• ${ARC_ROOT}/agents/review/architecture-engineer.md

Next.js:

• ${ARC_ROOT}/agents/review/lee-nextjs-engineer.md
• ${ARC_ROOT}/agents/review/daniel-product-engineer.md
• ${ARC_ROOT}/agents/review/senior-engineer.md

Python:

• ${ARC_ROOT}/agents/review/senior-engineer.md
• ${ARC_ROOT}/agents/review/performance-engineer.md
• ${ARC_ROOT}/agents/review/architecture-engineer.md

General/Unknown:

• ${ARC_ROOT}/agents/review/senior-engineer.md
• ${ARC_ROOT}/agents/review/architecture-engineer.md

Conditional addition (all UI project types):

• If plan involves UI components, forms, or user-facing features → add ${ARC_ROOT}/agents/review/accessibility-engineer.md
• If plan involves UI components, pages, or visual design → add ${ARC_ROOT}/agents/review/designer.md

Phase 2.5: Team Mode Check

<team_mode_check> Skip if specific reviewer was provided in Phase 0 (single reviewer, no team needed).

Check if agent teams are available by attempting to detect team support in the current environment.

If teams are available, offer the user a choice:

Execution mode:
1. Team mode — Reviewers challenge each other's findings before you see them (higher quality, 3-5x token cost)
2. Standard mode — Independent reviewers, findings consolidated by skill (faster, lower cost)

Use the AskUserQuestion interaction pattern with:

• "Team mode" — Reviewers cross-review and debate findings. Questions that survive peer scrutiny are stronger. Best when reviewing complex or high-stakes plans.
• "Standard mode (Recommended)" — Independent reviewers run in parallel. Faster and cheaper. Good default for most reviews.

If teams are NOT available, proceed silently with standard mode. Do not mention teams to the user.

If team mode selected, read the team reference:

${ARC_ROOT}/references/agent-teams.md

</team_mode_check>

Phase 3: Run Expert Review

If specific reviewer from Phase 0: Spawn single reviewer agent.

If team mode selected: Run team review (see Team Execution below).

Otherwise: Spawn 3 reviewer agents in parallel:

Task [reviewer-1] model: sonnet: "Review this plan for [specialty concerns].
Plan:
[plan content]

Focus on: [specific area based on reviewer type]"

Task [reviewer-2] model: sonnet: "Review this plan for [specialty concerns]..."

Task [reviewer-3] model: sonnet: "Review this plan for [specialty concerns]..."

Team Execution (Agent Teams mode)

Only if user opted into team mode in Phase 2.5.

Create team arc-review-[plan-slug] with the 3 selected reviewers as teammates.

Round 1 — Initial Review:

Each reviewer performs their standard analysis independently (same prompts as standard mode).

Create team: arc-review-[plan-slug]
Teammates: [reviewer-1], [reviewer-2], [reviewer-3]

Each teammate reviews the plan through their domain lens.
Same prompts and focus areas as standard mode.

Round 2 — Cross-Review:

Each reviewer reads the others' findings and responds:

• "My analysis supports this" — Confirms another reviewer's concern with additional evidence
• "My analysis addresses that" — Points out that their recommendation already handles the concern
• "I disagree because" — Challenges a finding with code-level evidence or domain reasoning

Resolution rules (from agent-teams reference):

• Code-level evidence wins over principle-based reasoning
• Domain authority wins within domain
• Every challenge must include explicit rationale

Round 2 output: Pre-debated findings where each concern has either been confirmed by peers, refined through challenge, or dropped with stated rationale.

Wait for team to complete.

If team creation fails, fall back silently to standard parallel dispatch above.

Phase 4: Consolidate and Present

<team_consolidation> If team mode was used, consolidation is simpler:

• Findings already survived peer scrutiny — false positives were caught and removed during debate
• Conflicting recommendations already resolved with rationale from both sides
• Socratic questions derived from team-debated findings carry more weight: "Two reviewers independently flagged this — what if we..."
• Focus on transforming debated findings into questions (skip deduplication and conflict resolution) </team_consolidation>

Transform findings into Socratic questions:

See ${ARC_ROOT}/references/review-patterns.md for approach.

Instead of presenting critiques:

• Turn findings into exploratory questions
• "What if we..." not "You should..."
• Collaborative spirit, not adversarial

Example transformations:

• Reviewer: "This is overengineered" → "We have three layers here. What if we started with one?"
• Reviewer: "Missing error handling" → "What happens if the API call fails? Should we handle that now or later?"
• Reviewer: "Security concern" → "This stores the token in localStorage. Is that acceptable for this use case?"

Present questions one at a time:

• Wait for user response
• If user wants to keep something, they probably have context
• Track decisions as you go

Phase 5: Apply Decisions

For each decision:

• Note what was changed
• Note what was kept and why

If plan came from a file:

• Update the file with changes
• Commit: git commit -m "docs: update <plan> based on review"

Phase 6: Summary and Next Steps

## Review Summary

**Reviewed:** [plan name/source]
**Reviewers:** [list]

### Changes Made
- [Change 1]
- [Change 2]

### Kept As-Is
- [Decision 1]: [reason]

### Open Questions
- [Any unresolved items]

Show remaining arc:

/arc:ideate     → Design doc (on main) ✓
     ↓
/arc:review     → Review ✓ YOU ARE HERE
     ↓
/arc:implement  → Plan + Execute task-by-task

Offer next steps based on what was reviewed:

If reviewed a design doc:

• "Ready to implement?" → /arc:implement (which will create the plan internally)
• "Done for now" → end

If reviewed an implementation plan:

• "Ready to implement?" → /arc:implement
• "Done for now" → end

Phase 7: Cleanup

Kill orphaned subagent processes:

After spawning reviewer agents, some may not exit cleanly. Run cleanup:

${ARC_ROOT}/scripts/cleanup-orphaned-agents.sh

<arc_log> After completing this skill, append to the activity log. See: ${ARC_ROOT}/references/arc-log.md

Entry: /arc:review — [Plan name] reviewed </arc_log>

<success_criteria> Plan review is complete when:

• Plan located (conversation, file, or user-provided)
• Project type detected and reviewers selected
• Parallel expert review completed (3 agents)
• All findings presented as Socratic questions
• User made decisions on each finding
• Plan updated (if from file)
• Summary presented
• Remaining arc shown (based on plan type)
• User chose next step (detail/implement or done)
• Progress journal updated
• Orphaned agents cleaned up

Diff review is complete when:

• Branch has changes vs main
• Checklist loaded from ${ARC_ROOT}/references/diff-review-checklist.md
• Full diff read before flagging anything
• Two-pass review applied (critical then informational)
• Findings presented (critical as Socratic questions)
• User decided on each critical finding
• Summary presented </success_criteria>