Agents Reference
6 core agents and 6 pack agents: roles, models, allowed tools, and when to use each.
Overview
AI Craftsman Superpowers includes 12 agents: 6 core agents, plus the 6 agents of the Symfony, React and AI/ML packs. Since 4.12.0 the pack agents also ship as ordinary copies under agents/, kept byte for byte identical to their pack’s own. Agents are invoked automatically by /craftsman:team or explicitly using the Agent tool.
Model tiers, as declared in each agent’s own frontmatter:
- Opus:
team-lead,architect,legacy-surgeon, the three whose decisions span files - Haiku:
doc-writer, the one repetitive-writing agent - Sonnet: the other eight
Read the frontmatter, not the role name
Each agent declares its own tool grant in tools:. Most craftsmen hold Write and Edit because their job is to change code. architect and team-lead hold neither. A role name does not imply read-only.
Cost optimization
The agent hooks (a DDD verifier after each Write or Edit, a final review at Stop) run as headless Haiku calls on Claude Code, roughly $0.15-0.30 per session of 50 writes. They are on by default; the agents on this page run only when a skill or you invoke them.
Core Agents
team-lead - CTO-Level Orchestrator
| Property | Value |
|---|---|
| Model | Opus |
| Effort | High |
| Memory | User |
| Max turns | 50 |
| Allowed tools | Read, Glob, Grep, Bash, Agent, TaskCreate, TaskList, TaskUpdate, SendMessage, craftsman:challenge. No Write, no Edit |
The team-lead is the top-level orchestrator for complex multi-step tasks. It:
- Manages other agents via delegation
- Maintains project context across sessions
- Makes architectural decisions
- Challenges suboptimal approaches
When to use:
/craftsman:team feature # team-lead orchestrates the full feature
/craftsman:plan execute-with-agentsarchitect - DDD/Clean Architecture Validator
| Property | Value |
|---|---|
| Model | Opus |
| Effort | High |
| Memory | Project |
| Max turns | 60 |
| Allowed tools | Read, Glob, Grep, Bash, Agent. No Write, no Edit |
The architect validates dependency direction, bounded contexts, aggregate boundaries, and Clean Architecture layer violations. Read-only - never writes or edits code.
Validates:
- Layer imports (Domain → nothing, Application → Domain, Infrastructure → both)
- Aggregate boundaries (no cross-aggregate direct references)
- Domain model purity (no infrastructure in Domain layer)
- DDD tactical patterns (Entity, Value Object, Aggregate, Repository)
When to use:
/craftsman:challenge Review the current architecture
/craftsman:team code-review # architect runs as part of the teamdoc-writer - Technical Documentation Specialist
| Property | Value |
|---|---|
| Model | Haiku |
| Memory | Project |
| Max turns | 20 |
| Allowed tools | Read, Glob, Grep, Write, Edit |
Generates and reviews API docs (OpenAPI), ADRs, READMEs, runbooks, and user guides. Verifies documentation accuracy against actual code.
When to use:
/craftsman:team code-review # doc-writer checks documentation gapssecurity-pentester - OWASP Top 10 Auditor
| Property | Value |
|---|---|
| Model | Sonnet |
| Memory | Project |
| Max turns | 20 |
| Allowed tools | Read, Glob, Grep, Bash, Write. No Edit |
Performs automated + manual security analysis focused on OWASP Top 10, SQL injection, XSS, authentication flaws, and API security.
Checks:
- SQL injection (parameterized queries)
- XSS (output encoding)
- Authentication/authorization flaws
- Insecure direct object references
- Security misconfiguration
- Sensitive data exposure
When to use:
/craftsman:team security-audit # dedicated security pass
/craftsman:team code-review # security included in full reviewui-ux-director - Design System Guardian
| Property | Value |
|---|---|
| Model | Sonnet |
| Memory | Project |
| Max turns | 20 |
| Allowed tools | Read, Glob, Grep, Write. No Bash, no Edit |
Deep expertise in SaaS dashboard UX, data visualization, accessibility (WCAG 2.1 AA), and design token systems.
Validates:
- WCAG 2.1 AA compliance
- Design token usage (no hardcoded colors)
- Component consistency
- Keyboard navigation
- Screen reader compatibility
When to use:
/craftsman:team code-review # ui-ux reviews frontend componentslegacy-surgeon - Inherited Code Rescuer
| Property | Value |
|---|---|
| Model | Opus |
| Effort | High |
| Memory | Project |
| Isolation | Worktree |
| Max turns | 25 |
| Allowed tools | Read, Glob, Grep, Bash, Write, Edit |
Brings untested, tangled, inherited code under control without breaking it. Characterizes behavior first, breaks dependencies with seams, refactors under a net, and migrates with strangler fig. Never rewrites from scratch.
When to use:
/craftsman:legacy audit # rank hotspots, produce LEGACY-AUDIT.md
/craftsman:legacy cover # characterization net before touching anything
/craftsman:legacy untangle # break dependencies
/craftsman:legacy migrate # strangler figRuns in its own git worktree, so a rescue in progress never blocks your working tree.
Pack Agents
Pack agents live in their pack (packs/<pack>/agents/) and ship as copies under agents/ as well, which the plugin’s scripts/native-manifests.py checks byte for byte.
backend-craftsman (Symfony Pack)
| Property | Value |
|---|---|
| Model | Sonnet |
| Pack | symfony |
| Expertise | Symfony 7/8, API Platform 4, Doctrine ORM, Messenger, PHPUnit |
Deep PHP/Symfony expertise. Validates DDD tactical patterns, Symfony service configuration, Doctrine entity design, and API Platform resources.
Use for: PHP and Symfony implementation, through /craftsman:team or the Agent tool. The review that runs after each write is the agent hooks’ headless verifier (Haiku on Claude Code), not this agent.
frontend-craftsman (React Pack)
| Property | Value |
|---|---|
| Model | Sonnet |
| Pack | react |
| Expertise | React 19, TypeScript, Tailwind, shadcn/ui, TanStack Query |
Deep React/TypeScript expertise. Validates component design, hook patterns, type safety, and accessibility.
Use for: React and TypeScript implementation, through /craftsman:team or the Agent tool. The review that runs after each write is the agent hooks’ headless verifier, not this agent.
ai-engineer (AI/ML Pack)
| Property | Value |
|---|---|
| Model | Sonnet |
| Pack | ai-ml |
| Expertise | RAG pipelines, pgvector, embeddings, LLM integration, MLOps |
AI/ML specialist for RAG pipeline design, embedding strategies, vector database integration, prompt engineering, and MLOps practices.
Its frontmatter binds the three AI/ML skills: craftsman:rag, craftsman:agent-design and craftsman:mlops.
api-craftsman (Symfony Pack)
| Property | Value |
|---|---|
| Model | Sonnet |
| Effort | Medium |
| Pack | symfony |
| Max turns | 30 |
| Allowed tools | Read, Glob, Grep, Bash, Agent, Edit, Write, craftsman:test |
| Expertise | API Platform 4, REST and HATEOAS, OpenAPI, JSON-LD and Hydra, OAuth2 and JWT |
Senior API architect. Use for API design reviews, API Platform configuration, and RESTful architecture decisions. Holds Write and Edit: it changes code.
symfony-reviewer (Symfony Pack)
| Property | Value |
|---|---|
| Model | Sonnet |
| Effort | Medium |
| Pack | symfony |
| Max turns | 15 |
| Allowed tools | Read, Glob, Grep, Bash, Edit, craftsman:challenge |
Review specialist for Symfony applications: PHP code, Doctrine entities, Symfony services. Despite the name, it holds Edit and can patch what it finds.
react-reviewer (React Pack)
| Property | Value |
|---|---|
| Model | Sonnet |
| Effort | Medium |
| Pack | react |
| Max turns | 15 |
| Allowed tools | Read, Glob, Grep, Bash, Edit, craftsman:challenge |
Review specialist for React and TypeScript: components, hooks, type safety. Despite the name, it holds Edit and can patch what it finds.
Agent Teams
Use /craftsman:team to orchestrate multiple agents together:
/craftsman:team code-review
# → architect + security-pentester + doc-writer run in parallel
# → team-lead synthesizes findings
/craftsman:team feature
# → team-lead orchestrates: architect (design) → backend/frontend (implement) → verify
/craftsman:team security-audit
# → security-pentester does a comprehensive OWASP deep diveParallel execution
When using /craftsman:team code-review, architect, security-pentester, and doc-writer run in parallel using separate agent contexts.