Review

Review

Orchestrated code review that launches 5 specialized agents in parallel.

Input

• $ARGUMENTS — Review scope (optional)
• Empty/omitted: unstaged changes (git diff)
• staged: staged changes (git diff --cached)
• HEAD: last commit (git show HEAD)
• ``: specific commit
• ...: specific files
• ``: changes on branch vs main

Agents

| Agent | Focus | |-------|-------| | orc:code-reviewer | CLAUDE.md compliance, bugs, quality | | orc:silent-failure-hunter | Error handling, silent failures | | orc:test-analyzer | Test coverage, gaps, quality | | orc:type-design-analyzer | Type invariants, encapsulation | | orc:frontend-verifier | UI verification via browser automation |

Process

Step 1: Enter Plan Mode

Use EnterPlanMode tool to enter planning mode. The review findings will form the plan for fixes.

Step 2: Determine Review Scope

Parse $ARGUMENTS to determine what to review:

if $ARGUMENTS is empty:
  scope = "unstaged changes"
  diff_cmd = "git diff"
elif $ARGUMENTS == "staged":
  scope = "staged changes"
  diff_cmd = "git diff --cached"
elif $ARGUMENTS == "HEAD":
  scope = "last commit"
  diff_cmd = "git show HEAD"
elif $ARGUMENTS matches commit hash:
  scope = "commit $ARGUMENTS"
  diff_cmd = "git show $ARGUMENTS"
else:
  scope = "$ARGUMENTS"
  diff_cmd = "git diff -- $ARGUMENTS" (or read files directly)

Run the diff command to get the changes to review.

Step 3: Launch 5 Parallel Review Agents

Launch ALL agents simultaneously using Task tool with subagent_type="orc:".

Critical: Send a SINGLE message with 5 Task tool calls to run them in parallel.

Review scope: $scope

$diff_output

Review these changes against project guidelines (CLAUDE.md).
Focus on: compliance, bugs, code quality.
Report issues with confidence ≥80 only.

Review scope: $scope

$diff_output

Audit error handling in these changes.
Focus on: silent failures, empty catches, poor feedback.
Flag CRITICAL/HIGH/MEDIUM severity issues.

Review scope: $scope

$diff_output

Analyze test coverage for these changes.
Focus on: critical gaps, edge cases, test quality.
Rate gaps 1-10 (10 = must have).

Review scope: $scope

$diff_output

Evaluate type designs in these changes.
Focus on: invariants, encapsulation, enforcement.
Rate each dimension 1-10.

Review scope: $scope

$diff_output

Verify the frontend still works correctly after these changes.
IMPORTANT: Backend API changes can break frontend even if no frontend files changed.

For backend changes: Test that frontends consuming these APIs still work.
For frontend changes: Test rendering, interactions, visual regressions.

Always:
1. Start the relevant services (backend + frontend)
2. Test key user flows end-to-end
3. Capture screenshots at each major step
4. Report any errors, broken flows, or unexpected behavior

Step 4: Collect and Synthesize Results

Wait for all 5 agents to complete. Synthesize findings into the plan file.

Step 5: Exit Plan Mode

Use ExitPlanMode tool to present the review findings as a plan for user approval.

The plan should list all issues found, prioritized by severity, with specific file paths and line numbers.

Step 6: Implement Fixes

After user approves the plan, implement the fixes:

1. Address Critical Issues first (must fix)
2. Address Important Issues (should fix)
3. Skip suggestions unless user requests them

For each fix:

• Read the file
• Make the specific change
• Verify the fix doesn't break anything

Plan Format (written to plan file)

# Code Review: $scope

## Summary

| Agent | Status | Issues |
|-------|--------|--------|
| code-reviewer | ✓/✗ | N critical, M important |
| silent-failure-hunter | ✓/✗ | N critical, M high |
| test-analyzer | ✓/✗ | N critical gaps |
| type-design-analyzer | ✓/✗ | N concerns |
| frontend-verifier | ✓/✗ | N UI issues |

## Critical Issues (will fix)

For each issue:
- [ ] **File**: `path/to/file.ts:123`
- **Issue**: Description
- **Fix**: Specific change to make

## Important Issues (will fix)

[Same format as critical]

## Suggestions (optional, user decides)

[Lower priority items - won't fix unless approved]

## Positive Findings

[What's done well across agents]

---
Total: N issues to fix, M suggestions available.

Execution Rules

1. Enter plan mode first — use EnterPlanMode before starting
2. Always run all 5 agents — don't skip any
3. Run in parallel — single message with 5 Task calls
4. Include full diff — each agent needs the context
5. Synthesize results — don't just concatenate outputs
6. Prioritize by severity — critical first, suggestions last
7. Exit plan mode — present findings for user approval
8. Implement fixes — after approval, fix critical and important issues

CRITICAL: Never Skip Frontend Verification

NEVER skip frontend-verifier based on file types in the diff.

Backend and frontend are linked systems. Backend API changes (routes, middleware, state management, auth) can and do break frontend functionality even when no frontend files are modified. Examples:

• API response shape changes → frontend parsing fails
• Auth middleware changes → frontend auth flow breaks
• State management changes → frontend data loading fails
• Route parameter changes → frontend API calls fail

The frontend-verifier MUST run for:

• Any backend API changes (routes, controllers, middleware)
• Any auth/session changes
• Any database schema changes
• Any state management changes
• ANY change that could affect API contracts

Only skip frontend-verifier if:

• User explicitly passes --no-ui flag
• Changes are purely documentation (README, comments only)
• Changes are purely test files with no implementation changes
• Changes are purely CI/CD configuration

When in doubt, RUN THE FRONTEND VERIFIER.

Skip Conditions

Skip specific agents only if explicitly requested by user:

• --no-tests: skip test-analyzer
• --no-types: skip type-design-analyzer
• --no-ui: skip frontend-verifier (user must explicitly request this)

Do NOT skip agents based on your own judgment about file types. The user knows their system better than you do.

Examples

# Review unstaged changes (default)
/review

# Review staged changes
/review staged

# Review last commit
/review HEAD

# Review specific files
/review src/auth.ts src/login.tsx

# Review branch changes
/review feature-branch

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: qGolem
• Source: qGolem/orc
• License: MIT

Install and usage instructions live in the source repository linked above.