Session resumption crashes on non-Vertex AI backend — blocks usage of Gemini 3.1 Flash Live

[amitanshupanigrahi2704]

Labels: bug, live-api

Description

Problem

When using the Live API (runner.run_live()) with a non-Vertex AI backend (GOOGLE_GENAI_USE_VERTEXAI=FALSE), session resumption always crashes on reconnect. This makes it impossible to use Gemini 3.1 Flash Live Preview (gemini-3.1-flash-live-preview) with reliable session resumption, since that model is only available on the direct Gemini API — not on Vertex AI.

Root Cause

In google/adk/flows/llm_flows/base_llm_flow.py, the ConnectionClosed exception handler unconditionally sets transparent=True when reconnecting with a session resumption handle (around line 517 in ADK 1.25.1):

# base_llm_flow.py — ConnectionClosed handler
session_resumption_config = types.SessionResumptionConfig(
    handle=invocation_context.live_session_resumption_handle,
    transparent=True,  # <— always set to True
)

However, in google/adk/models/google_llm.py (around lines 448-451), there is a guard that raises a ValueError when transparent=True is used with a non-Vertex AI backend:

if (session_resumption_config and session_resumption_config.transparent
    and self._api_backend == GoogleLLMVariant.GEMINI_API):
    raise ValueError(
        "Transparent session resumption is only supported with Vertex AI backend"
    )

This creates a deadlock:

• ADK always sets transparent=True on reconnect
• Non-Vertex AI backends always reject transparent=True
• Therefore, session resumption always fails on non-Vertex AI

Why This Matters

	Direct Gemini API	Vertex AI
gemini-3.1-flash-live-preview	Available	Not available
Transparent session resumption	Rejected (ValueError)	Supported

Users who want to use Gemini 3.1 Flash Live (only available on Direct Gemini API) cannot have working session resumption, because ADK forces transparent=True which is rejected on that backend.

Environment

• ADK version: 1.25.1
• google-genai SDK: 1.59.0
• Backend: GOOGLE_GENAI_USE_VERTEXAI=FALSE
• Model: gemini-3.1-flash-live-preview
• OS: Linux / Windows

Expected Behavior

Session resumption should work on non-Vertex AI backends, either by:

1. Setting transparent=False (non-transparent resumption) when the backend is GEMINI_API, or
2. Making transparent configurable by the caller rather than hardcoded to True, or
3. Making Gemini 3.1 Flash Live available on Vertex AI

Steps to Reproduce

1. Set GOOGLE_GENAI_USE_VERTEXAI=FALSE
2. Use runner.run_live() with gemini-3.1-flash-live-preview
3. Configure SessionResumptionConfig with a handle
4. Wait for the server to send a GoAway or for the connection to drop
5. ADK attempts to reconnect → sets transparent=True → ValueError is raised → session dies

Workaround Attempted

Setting transparent=False manually on initial connection does not help because ADK's internal ConnectionClosed handler overrides it to True on every reconnect attempt.

[he-yufeng]

I reproduced the reconnect path in the current code. The narrow fix is to preserve an existing session_resumption.transparent=False on reconnect and only default transparent=True when the reconnect path has to create a fresh SessionResumptionConfig.

That keeps the current transparent default for callers that did not configure the field, but avoids overwriting an explicit non-transparent config used with the direct Gemini API backend. I have a small patch with a focused BaseLlmFlow.run_live regression test and will open a PR for it.

[llalitkumarrr]

Hello @amitanshupanigrahi2704 ,

Could you please check if you experience the same issue with earlier models when using the Gemini API? Additionally, we noticed you are using ADK version 1.25.1. We recommend updating to the latest version (1.33.0). Please let us know if the issue is resolved after the update.