Community · E2 · artifact verified

Decide Home Assistant commands with typed answers

A Home Assistant conversation agent - installable via HACS - that decides with a TypeSafe System One model instead of an LLM: your spoken or typed commands become typed choices and Nouls that drive devices, with metered costs documented.

01 · Role in the system

What Jev does here

The custom component registers as a Home Assistant conversation agent: commands arrive through the normal assistant pipeline, and system_one.py turns each into typed questions - Choice answers pick the intent and target, Noul answers gate whether the request was understood - before any device acts. Because the deciding model never generates text, responses are composed from the typed answers and cost is a known quantity per command rather than an open-ended generation; the README documents the metering explicitly and links a cost section. Tests and CI cover the answer parsing (choice and noul accessors with typed answer classes), and distribution is the standard HACS custom-repository flow, so it installs into existing Home Assistant setups without forking the core.

A conversation agent that never generates its replies

The custom component registers as a Home Assistant conversation agent through the standard assistant pipeline, then does something structurally different from the LLM default: each spoken or typed command becomes typed questions — Choice answers pick the intent and target, Noul answers gate whether the request was even understood — and the reply is composed from those typed answers, never generated.

Composed replies make cost a known quantity per command instead of an open-ended generation bill, which the README documents with a dedicated cost section. Tests and CI cover the answer parsing with typed answer classes, and distribution is the standard HACS flow, so it installs into existing setups without forking anything.

Early, small, and clear about it

The record self-describes as early with rough edges expected, requires a metered TypeSafe API key, and carries 4 stars at review time as an unofficial HACS install. Intent coverage is bounded by the typed question set rather than open-vocabulary understanding — commands outside the declared choices will not be divined.

Within the home cluster, this record and the official routing demo form a before-and-after pair: the official example shows speculative parallel classification as architecture, and this one shows it shipped as an installable component with metered costs. The limitation both share is the same: the question set defines the ceiling of understanding.

02 · Control boundary

Where Jev sits

Jev as the conversation agent behind a smart home assistant: each command becomes typed choices and Nouls that map to intents and devices, answers are composed - never generated - and cost per command is metered and documented.

Code owns the loop, permissions, thresholds, validation, and side effects. Jev owns only the bounded judgments described above.

03 · Known limits

What this evidence does not prove

  • Self-described early - expect rough edges; requires a metered TypeSafe API key.
  • Intent coverage is bounded by the typed question set, not open-vocabulary understanding.
  • Small project (4 stars at review time); HACS custom install, not official.

04 · Attribution

Public sources

This is a Community record: the project was published by a third-party community author.

  • the-sof ↗Community · github · public · checked 2026-09-30