agent workflow
every turn checks if you're done. if not, the model runs with gmail tools, then the reply streams back as text and audio in parallel.
voice activity detection
speech gets cut on the phone before upload. it starts above -30 db, stops below -40 db, checked every 100 ms.