Agent Status and Avatar Presentation
Desktop provides lightweight feedback from the current Conversation result and local recording, input, and playback state. Avatar is a separate embodiment App responsible for Live2D or VRM models, activity, local audio playback, and lipsync analysis.
Data Sources
| Surface | Input and responsibility |
|---|---|
| Desktop chat | Reads admitted Conversation results and combines them with local input/playback state |
| Avatar | Uses its own protected App session for activity, emotion, posture, voice-timing, and final voice artifacts |
Each surface obtains data admitted to its current session. One window's cache is not another surface's state source, and Apps cannot reuse session-scoped Agent handles as cross-session operation selectors.
Voice and Lipsync
Recorded input, Realtime, and message playback use separate paths; see Voice Interaction. Avatar's ordinary Conversation playback reads the final voice artifact, then analyzes it locally to produce mouth weights.
Common Runtime presentation events do not carry viseme/phoneme sequences or renderer mouth parameters. A recording or playback status does not establish that another window has opened Realtime or is displaying word-highlighted subtitles.
Readiness and Failure
Rendering a model and enabling product interaction have different prerequisites. A development preview can show a model without establishing a live session. Follow the surface's diagnostic reason when session or presentation conditions fail. See Avatar Instance Lifecycle.