Skip to content

Routing, Experts & Skills

Roles, topics, experts, and skills solve different configuration problems:

ConceptPurpose
RoleResponsibility, prompt, and available tool containers
TopicModel binding for a type of work
ExpertAdditional prompt guidance, deliverable expectations, skills, and optional model preference
SkillDomain instructions loaded into model context
PersonaUser-facing identity and tone

The root runs as general, answers with its own tools, and may delegate through spawn_child. Classification provides context hints rather than a mandatory routing hop. Specialist roles include coding, research, architecture, review, QA, communication, design, DevOps, security, data, AI, finance, automation, project management, and writing.

Use GET /api/roles and GET /api/tools/role-map to inspect the current role bindings. Tool access is separate from permission to execute an action.

For a planned child, a configured lane executor takes precedence. Otherwise an expert’s explicit model preference precedes the topic primary. A missing specialist binding fails with a configuration error. The root alone can select the configured default. Backup bindings apply to failure recovery rather than as a proactive cheaper tier; worker paths have their own fallback behavior.

Configure topics through GET /api/topics, PUT /api/topics/:topic/binding, and PATCH /api/topics/:topic/config. Text topics usually match role names; embedding, OCR, vision, and voice have separate capability requirements.

Assigning the same model to every text topic is supported through POST /api/topics/assign-all with a registered model name. It does not create embedding or vision capabilities in a text-only model.

Select an expert with /expert or the expert controls in your client. The worker receives the expert’s role tools and prompt guidance. Critical rules and success metrics are prompt instructions, not enforced output schemas or independent verification of the result.

System skills are database-backed; filesystem SKILL.md documents are discovered from configured directories. Skill distillation creates proposals for review; a proposal is not automatically a trusted new capability.

Model examples and role suggestions are starting points. Verify tool calling, output quality, and cost using the models and tasks you actually configure.