Setup
- Gemini Developer API (AI Studio key, paid tier), model Gemini 3.1 Flash Live
- BidiGenerateContent over WebSocket, Python, LiveKit Agents (google realtime plugin)
- Production telephony voice agent, Arabic audio
Problem
Mid-call, the Live session silently stops generating. The WebSocket stays open and the server accepts our realtime audio and turn requests without error, but nothing comes back: no serverContent, no modelTurn, no turnComplete, no interrupted, no GoAway, no error, no close frame. The caller hears silence.
We detect it with a 10s watchdog, close the session and open a new one with the conversation history re-seeded. That sometimes recovers, but it can recur more than once in the same call.
Example: request sent around 2026-09-30 12:17:38 UTC, watchdog fired at 12:17:48 UTC (second forced reconnect on that call).
What we have ruled out: network is healthy, sessions are well under the duration limit, and no quota or rate-limit errors are returned.
Questions
- Is anyone else seeing Live sessions stall silently like this?
- Is there a signal we should watch for, or a recommended timeout?
- Better recovery: new session with re-seeded history, or session resumption with a handle?
- Any known triggers (long context, tool calls, VAD settings, a specific model version)?
Thanks.