Agents

Agents are roles: who does the work. Each one returns its result to you, or to the main conversation; roles never call each other, which keeps cost predictable.

Model tiers. "Strong" means your best reasoning model; "everyday" means a fast, inexpensive one. See model tiers per tool.

Plan and design

Role	Does	Edits	Tier
analyst	Scans the project, and turns a ticket and notes into the discovery part of a plan.	Docs and plans	Everyday
research-collector	Gathers sources on a question into notes, for the researcher or a person. Doesn't judge.	Research notes	Everyday
researcher	Answers a question that needs outside knowledge: options, a recommendation, confidence and sources.	Research notes	Strong
architect	Designs the approach, contract first, writes steps, diagrams and decisions into the plan, and decides blockers.	Plans and design docs	Strong

Build

Role	Does	Edits	Tier
relational-database-developer	Schema changes, migrations, indexes, queries and data fixes, with strict stop rules for anything irreversible.	Migrations, queries, tests	Everyday
terraform-developer	Terraform, OpenTofu and Terragrunt modules and units, checked with plans. Never applies: hands you the command.	Infrastructure code, tests, docs	Everyday
flask-developer	Python Flask endpoints, services, jobs and integrations.	Server code, tests, docs	Everyday
spring-developer	Java Spring Boot controllers, services, repositories and configuration.	Server code, tests, docs	Everyday
react-developer	React pages, components, hooks, forms and API calls.	UI code, tests, docs	Everyday
developer	Anything no framework developer covers: scripts, tools, libraries, infrastructure, other stacks.	Code, tests, docs	Everyday

All developers share the implementation skill's rules, and each framework developer adds its framework's skill.

Check and finish

Role	Does	Edits	Tier
tester	Finds untested behavior that matters, writes the tests, reports bugs.	Tests only	Everyday
reviewer	Checks that a plan's goal was actually met, and reviews a change for bugs, broken contracts, weakened tests, stand-ins, stale docs and over-engineering.	Nothing	Strong
security-reviewer	Reviews a change for exploitable security problems, reporting only those with a plausible attack path.	Nothing	Strong
documenter	Keeps CONTEXT.md and docs true to the code, closes finished plans, and drafts ticket updates for handover.	Docs only	Everyday

Restrictions are instructions, not enforcement

"Edits tests only" is written in the agent's text. To enforce it, add a tools line to the agent. See Customizing.