the app shipped audio cues and it's the small thing that changes the whole feel. three quiet synthesized tones generated in app, no audio files. a rising two note when listening opens, a single soft note while it thinks, and a falling two note when the reply is done so i know it's my turn without looking at the screen. there's a toggle under voice settings, on by default.
the first cue never played though. it was routed to the system sonification stream instead of media volume, so any muted state killed it. moved it, and while in there fixed the weird screen off flow where the app stayed over the lock screen, which was the session window still being alive, not a lock screen bug.
also got the full map of what runs where and it's better than i assumed: wake word and speech to text are on device, the language model is my own server, and screenshots only get sent when i explicitly ask to look at the screen, not on every request. that's why the fast path stays fast.
the changelog pill lives in settings, quick voice replies run on a leaner profile, and the app finally feels like it knows the difference between a conversation and a command.