Is your feature request related to a problem? Please describe.
Runner.runLive() and GeminiLlmConnection currently only support Gemini 2.0 Flash Live. Gemini 2.5 Flash and 3.1 Flash Live preview models are not supported — attempts to use them result in silent failures or incompatible connection behaviour. This blocks production use of newer, higher-quality models in voice/realtime agent applications.
Describe the solution you'd like
Full support for gemini-2.5-flash-preview-native-audio and gemini-3.1-flash-live-preview (and future Live-capable models) in the ADK JS live bridge — including sendRealtimeInput, receive(), and turn/interruption handling compatible with those models' protocol. An ETA or tracking issue for this would be very helpful.
Describe alternatives you've considered
Forking @google/adk and patching GeminiLlmConnection directly (as we are currently doing) to fix audio blob sending and keep receive() open across turns. This is not sustainable long-term as it requires manually rebasing against upstream releases.
Additional context
We are building a production voice agent using ADK JS with Runner.runLive(). Gemini 2.5 and 3.1 Live models offer significantly better quality for Czech-language voice interactions, which is critical for our use case. The lack of official support is a blocker for us moving off the fork.
Is your feature request related to a problem? Please describe.
Runner.runLive()andGeminiLlmConnectioncurrently only support Gemini 2.0 Flash Live. Gemini 2.5 Flash and 3.1 Flash Live preview models are not supported — attempts to use them result in silent failures or incompatible connection behaviour. This blocks production use of newer, higher-quality models in voice/realtime agent applications.Describe the solution you'd like
Full support for
gemini-2.5-flash-preview-native-audioandgemini-3.1-flash-live-preview(and future Live-capable models) in the ADK JS live bridge — includingsendRealtimeInput,receive(), and turn/interruption handling compatible with those models' protocol. An ETA or tracking issue for this would be very helpful.Describe alternatives you've considered
Forking
@google/adkand patchingGeminiLlmConnectiondirectly (as we are currently doing) to fix audio blob sending and keepreceive()open across turns. This is not sustainable long-term as it requires manually rebasing against upstream releases.Additional context
We are building a production voice agent using ADK JS with
Runner.runLive(). Gemini 2.5 and 3.1 Live models offer significantly better quality for Czech-language voice interactions, which is critical for our use case. The lack of official support is a blocker for us moving off the fork.