The agent asks; a cheaper model drafts your answer

Mid-turn, the agent parks a question: three viable ways to split the monolith, which one? The turn stops until a human answers.

Your product calls POST /api/v1/agent/sessions/{id}/answer:assist with a small, cheap model named in the body. It reads the parked question and emits a question_suggestion event on the session stream: a drafted answer, with reasoning. The suggestion lands on the customer's phone as a tap-to-accept card. Tapping it POSTs the draft to /answer.

The daemon never answers itself. The suggestion is an event, not a decision. A human either accepts the draft or types their own, and the expensive model that did the work stays paused until the reply arrives.

The hard question gets a draft answer before it gets a human.