Core concepts
Skills & agent patterns
How user requests route to behaviour.
A skill is the unit a request routes to: trigger phrases and keywords for the classifier, an execution target, allowed tools, a confirmation rule, and knowledge scopes. The base agent routes via a cascade — semantic centroid classifier (when enabled) → keyword classifier → fallback pattern below the confidence threshold.
| Pattern | Use for |
|---|---|
simple_chat | Basic local assistant |
retrieval | Grounded Q&A over local knowledge, with sources |
structured_extraction | Schema-first data capture with auto-repair |
summarisation | Documents, conversations, records |
planner_executor / deep_task | Bounded goal decomposition; deep task adds durable per-step checkpoints |
react_tool_use | Model-driven tool loops, hard-capped at 4 iterations, every call policy-gated |
Enable semantic routing (paraphrase-tolerant)
// Rust host: rt.enable_semantic_routing(None) // CLI: lybo studio --semantic