/code-review — Structured Multi-Leg Code Review

Parallel specialist review agents, each in a fresh subagent context with a narrow focus and concrete checklist. Adapts to the codebase via /skill-discovery — loads domain skills (ruby-on-rails, vue, api-integration, etc.) based on changed files to inject project-specific review criteria.

Usage

/code-review                                          # all legs, current branch diff to default branch
/code-review current branch diff to master             # explicit target description
/code-review --legs=security,wiring,test-quality       # specific legs only
/code-review --branch=feature-branch                   # explicit branch
/code-review src/components/                           # specific directory

Default behavior: Run ALL legs in parallel against current branch diff to default branch. To limit: Use --legs= with comma-separated leg names.

Process

Step 1: Determine Review Target

Parse $ARGUMENTS for:

• Legs filter — value after --legs= (comma-separated list). If omitted, run ALL legs.
• Branch — value after --branch= (diff against default branch)
• Target — remaining args (file path, directory, or description like "current branch diff to master")

Get the diff:

• If --branch or "branch" mentioned: git diff {default-branch}...{branch}
• If target is a path: git diff -- {path} or read files directly
• If no specific target: git diff {default-branch}...HEAD (current branch vs default)
• Fallback: git diff (uncommitted changes)

Also gather: git diff --stat for file change summary, count of files/insertions/deletions.

If no changes found, report "No changes to review" and exit.

Step 2: Skill Discovery

Invoke /skill-discovery via the Skill tool. This scans the installed skill catalog, evaluates relevance to the changed files, and outputs reasoning. It handles:

• Identifying which domain skills are relevant (ruby-on-rails, vue, unit-testing, etc.)
• Loading those skills so their guidance is available in context
• Outputting the reasoning for transparency

The loaded domain skill content becomes available for injection into leg prompts.

Step 3: Launch ALL Leg Subagents in Parallel

Every leg runs in its own separate subagent (via the Agent tool, run_in_background: true). Fresh context per leg ensures focused, independent analysis. The overhead is worth it for review quality.

If your environment has a specialized code-review subagent type configured, use it for sharper, review-tuned analysis. Otherwise use the general-purpose agent type — the leg prompt's narrow focus and checklist do most of the work regardless of agent type.

For each leg (all legs unless --legs= filter was provided), spawn an Agent with:

1. Base preamble — from resources/leg-prompts.md (output format, file:line reference requirement)
2. Leg-specific prompt — the "Look for" checklist and "Questions to answer" for that leg
3. The full diff — each subagent gets the complete diff to analyze from its perspective
4. Domain context — relevant project conventions from loaded skills and CLAUDE.md

Launch ALL legs in a single message with multiple Agent tool calls.

Step 4: Synthesize Findings

After all subagents return, synthesize into a unified report:

1. Read all leg outputs
2. Deduplicate — same issue found by multiple legs → merge into single entry, note all legs that found it
3. Cross-reference confidence — issues found by 3+ legs → auto-promote to next severity
4. Resolve conflicts — if legs disagree, flag as "Needs discussion" with both perspectives
5. Prioritize — P0 (must fix) → P1 (should fix) → P2 (nice to fix)

Step 5: Output Report

@resources/synthesis-template.md

Review Legs

Each leg has a narrow focus, concrete "Look for" checklist, and judgment-forcing "Questions to answer." Full leg definitions are in resources/leg-prompts.md.

By default ALL legs run. Use --legs= to run a subset.

Analysis Legs

Verification Legs

NOT Responsible For

These are separate skills with their own scope:

• Spec compliance — /spec-check (per-requirement PASS/FAIL against spec)
• QA classification — /qa-plan (classifying verification items by execution channel)
• Handoff compliance — checking code against task handoff directives is a separate concern this skill does not cover
• Browser verification — a companion skill using the browser-testing capability (visual QA in real browser)

This skill reviews code quality, correctness, and implementation patterns — not whether the implementation matches a spec or handoff.

If no browser-testing binding is configured, note in the report that visual/browser QA was out of scope for this review and should be requested separately.

Interoperates With

• /skill-discovery (Step 2) — loads domain skills relevant to the changed files. If /skill-discovery is not installed, skip domain skill loading and proceed with the generic "Look for" checklists only; note in the report that no domain skills were loaded.
• /spec-check, /qa-plan — complementary review skills with disjoint scope (see NOT Responsible For above). No shared configuration is required between them; each reads the same diff independently.
• Browser-testing companion skill — if the environment binds the browser-testing capability, a companion skill can pick up visual QA where this skill's scope ends. The binding lives in CLAUDE.md / CLAUDE.local.md so any skill that needs it (this one included) resolves to the same tool.

Project and User Configuration

Load configuration in this order. When multiple sources define the same binding, the first one found wins — stop checking lower sources for that binding:

1. Read resources/personal-skill.md if it exists — personal review preferences (e.g., preferred leg subset, notification routing).
2. Read resources/domain-skill.md if it exists — project-specific review configuration (e.g., extra legs, project conventions to weight heavily).
3. Check the project's CLAUDE.md / CLAUDE.local.md for environment-level capability bindings (especially browser-testing).
4. For any capability still unbound, use the defaults specified in this skill (browser QA reported as out of scope; general-purpose agent type for legs).

Extension Points

domain-skill.md

Provide project-specific review configuration:

• Additional or project-specific review legs beyond the standard set
• Project conventions the elegance and style legs should weight heavily (beyond what /skill-discovery already loads)
• Known false-positive patterns to deprioritize (e.g., a pattern that looks like a smell but is intentional in this codebase)
• Default --legs= subset for this project, if the full set is usually overkill

personal-skill.md

Provide personal review preferences:

• Preferred severity threshold for what gets surfaced (e.g., suppress P2s by default)
• Notification routing for completed reviews, if any
• Whether to auto-run /skill-discovery or skip it for speed