Is there a way in the Gemini Live API to detect “near end of generation” (not just final completion)?

Gemini API
1

Hey all, I’m working with the Gemini Live API (via WebSockets) and using it for streaming LLM output in real-time. I understand that the API exposes signals like generationComplete and turnComplete, which tell me when the model has finished its current output. I can react to those cleanly in my client or backend.

What I need is something a bit different:
Instead of waiting until the model is done, I want a way to detect when the model is getting close to done — so I can call a function, update my UI, prep the next turn, or transition my state before the final completion happens.

Right now my model pipeline looks like this:

  1. Client opens a gemini.live.connect session.

  2. I stream text/audio and receive chunks back from the model.

  3. I watch for server_content.generation_complete or server_content.turn_complete to know the reply is finished.

But there doesn’t seem to be any built-in “N tokens left” or “almost done” event in the Gemini Live spec that gets emitted before generationComplete. The API docs only define the normal completion flags — no progress percentage or remaining token info.

Before I build a heuristic (like counting streamed tokens/chars and calling my callback when some threshold is met), I wanted to check:

  • Has anyone seen undocumented or hidden signals that signal “approaching end of generation”?

  • Are there better client-side heuristics people use in Gemini Live if they need early notice of an ending?

  • Or is the community just using generationComplete as the de facto only reliable signal?

For context: I’m aware this isn’t about end of turn detection or voice activity detection — I’m talking strictly about approaching end of the model’s text/audio generation while it’s streaming.

Thanks!