review

name	review
description	Multi-agent code review with quality scoring. Trigger: review, code review, check quality, review code, review changes

/review -- Multi-Agent Code Review

Philosophy

Code review catches what tests miss -- architectural drift, security gaps, readability problems, and design violations. Multiple specialized reviewers find more issues than one generalist because each reviewer brings deep focus to a single dimension. Reviews are read-only operations: reviewers report findings, they never edit code. The combination of independent perspectives produces a thorough, unbiased assessment.

---

Anti-Pattern: Rubber-Stamp Review

WRONG (rubber stamp):                  RIGHT (multi-dimensional):
1. Glance at diff                      1. Read changed files fully
2. "Looks good to me"                  2. Spawn specialized reviewers in parallel
3. Merge                               3. Each reviewer focuses on one dimension
                                       4. Synthesize findings by severity
                                       5. Score against quality gates

---

Workflow

Step 1: Determine Scope

/review              --> all uncommitted changes (git diff)
/review [file|dir]   --> specific files or directory
/review --staged     --> staged changes only
/review --plan [file]--> review a plan file (delegates to /review-plan)

• Scope identified
• Files exist and are readable

Step 2: Identify Changes

• Read all files in scope
• Categorize each changed file:

Category	Examples
Application logic	Business rules, domain models, services
API	Endpoints, request/response schemas, middleware
Data layer	Queries, migrations, repository implementations
Config	Environment, build, deployment configuration
Tests	Unit, integration, e2e test files
Infrastructure	Docker, CI/CD, scripts

• All files read
• Changes categorized

Step 3: Select and Launch Reviewers

Based on change categories, read agent definitions from .claude/agents/ and spawn reviewers in parallel:

Agent File	subagent_type	When to Spawn
code-reviewer.md	senior-code-reviewer	Always
test-reviewer.md	senior-code-reviewer	Always
security-reviewer.md	security-code-auditor	Auth, input handling, config, deps
architecture-reviewer.md	senior-code-reviewer	New modules, boundary changes
performance-reviewer.md	senior-code-reviewer	Data access, algorithms, hot paths
doc-reviewer.md	senior-code-reviewer	Public APIs, README changes

For plan review (/review --plan):

Agent File	subagent_type	When to Spawn
architecture-reviewer.md	senior-code-reviewer	Always for plans
test-reviewer.md	senior-code-reviewer	Always for plans
security-reviewer.md	security-code-auditor	If plan touches auth/data

Spawning pattern -- all selected reviewers in the same response:

Task(subagent_type="senior-code-reviewer",
     prompt="<content of .claude/agents/code-reviewer.md>\n\nReview: {changed_files}")

Task(subagent_type="senior-code-reviewer",
     prompt="<content of .claude/agents/test-reviewer.md>\n\nReview: {test_files}")

Task(subagent_type="security-code-auditor",
     prompt="<content of .claude/agents/security-reviewer.md>\n\nAudit: {changed_files}")

• Agent definitions read
• Reviewers spawned in parallel
• All reviewer reports collected

Step 4: Run Verification

Run make check (build + test + lint + typecheck) to confirm the codebase is healthy independent of reviewer opinions.

• make check executed
• Results captured

Step 5: Compile Results

Merge all agent reports into a single summary:

## Combined Code Review

**Files Reviewed:** [list]
**Review Agents Used:** [list]
**Overall Quality:** [Excellent / Good / Needs Work / Significant Issues]
**Quality Score:** [N]/100

### Critical Issues (block commit)
[Merged from all agents, deduplicated]

### Major Issues (block PR)
[Merged from all agents, deduplicated]

### Minor Issues (polish)
[Merged from all agents, deduplicated]

### Verification
| Check | Status |
|-------|--------|
| Build | PASS/FAIL |
| Tests | PASS/FAIL |
| Lint  | PASS/FAIL |

### Agent Summaries
- **Code Review:** [one-line summary]
- **Test Review:** [one-line summary]
- **Security:** [one-line summary]
- **Architecture:** [one-line summary]
- **Performance:** [one-line summary]
- **Documentation:** [one-line summary]

### Recommended Actions
1. [Highest priority action]
2. [Next priority]

• Reports merged and deduplicated
• Issues grouped by severity
• Recommended actions prioritized

Step 6: Score

Apply the quality-gates rubric to compute a score:

Score	Gate Status
>= 95	Release quality
>= 90	Ready for PR
>= 80	Ready to commit
< 80	Needs work before commit

---

Rules

• Reviewers are read-only -- they report findings, they never edit files
• Always spawn independent reviewers in parallel for speed
• Deduplicate findings when multiple agents flag the same issue
• Critical issues from ANY agent block the commit regardless of overall score
• The code-reviewer and test-reviewer agents are always included; others are conditional
• When reviewing a plan (--plan), focus on architecture, testability, and security -- not code style