Skip to main content
The Intent System classifies the operator’s purpose so Lucia can use the correct reasoning path and response shape.

Purpose

The intent system prevents Lucia from treating every message as generic chat.
It may distinguish requests such as:

Current runtime ownership

Dated: the Focus intent router is retired. Until 2026-09-04 the served Focus route classified the owner’s prompt with a regex intent table and an intent-assist model call before answering. Focus answers from one brain (ALGO-40) removed both from the route; Fable now reads the whole house and composes the answer and its doors in one call, and the purpose of the prompt is understood in that composition. The intent families on this page remain doctrine for what Lucia must understand, and the Focus bank pins them as expected subjects and doors. Guest-facing intent tasks are unchanged: guest_intent and guest_entity_extraction remain separate, bounded model tasks in the Engine (GPT-5.6 Sol by default at e2166b86ac6794a1423c7b2ce1b293a3dba505c4). Intent understanding is Engine-owned. Admin supplies the owner’s words and where the owner is standing, then presents the Engine response. Admin does not choose the model or infer provenance. The package value 0.1.3.6 is metadata, not a release identity.

Classification rules

Intent classification should determine:
  • what mode to use
  • which verified data matters
  • what should be ignored
  • what response structure fits the moment
  • which named entity and action boundary must survive routing
It classifies by purpose and semantic family, not one lucky exact phrase.

Action-routing doctrine

The Resolver Matrix of the router era is not a served concept on the one-brain path; the pack’s keyed door table is what maps Lucia’s answer to a finite destination (LUCI-214, 2026-09-06). If the operator names a guest, service, request, or issue, that entity must be preserved through classification and routing.

Truth-state boundaries

Payment intent must preserve the difference between:
A note does not change payment state. A payment item closes only when the ledger says confirmed_paid; a charge, a refund, a payment link or a recorded manual payment happens only on the owner’s tap of a real door (LUCI-215, LUCI-219). Workspace context may scope the request before generic ranking. An explicit subject still wins over page context. Operator output must not expose raw plumbing such as:
Short follow-ups may resolve against verified prior recommendation or offer context. If the context is absent or ambiguous, Lucia must clarify instead of guessing.

Example intent families

priority_triage

Pick the single highest-priority verified operational issue.

deferral

Identify safe-to-defer work without forcing the operator to inspect everything.

human_utility

Respond naturally, briefly, and calmly to social or emotionally soft input while preserving Lucia’s operator boundary.

semantic_conversational_assist

Recognize lightweight utility, social momentum, and scoped context prompts by meaning. Answer or clarify within Lucia’s operator purpose rather than drifting into open-domain chat.

maintenance_focus

Treat guest impact, arrival and departure timing, stay windows, payment context, and task state as ranking signals only when those facts are verified.

Correct behavior

Lucia should:
  • classify quietly
  • preserve named entities
  • route cleanly
  • protect semantic families by meaning
  • maintain a stable response shape
  • remain conservative when evidence is incomplete
Lucia should not:
  • expose classification machinery
  • phrase-patch individual prompts as a substitute for intent understanding
  • force every message into operational triage
  • invent action or truth from a classification
  • become an open-domain assistant

Truth-state rule

Intent classification is a routing decision, not truth. If intent or context is uncertain, Lucia must avoid overclaiming.

See also