On Auto, PO picks the model for each task, each delegation or, in full mode, each turn of a conversation, within hard limits on tools, context size, budget and consent. Every decision is recorded with a readable reason, and PO learns from the results.
A model you chose yourself is never replaced, even when a conversation resumes. In a conversation you can also tick the models you allow instead of using Auto.
The honest limit: routing has only run on two stand-in providers, never on a real model yet, and a pool of models that spans several providers is planned, not available. If a learned choice does worse than the declared one, routing falls back by itself to only recording what it would have chosen.