No output from gemini-3-flash-preview with Tools + Streaming

[wltnrhkd123]

Describe the bug When using the gemini-3-flash-preview model with tools enabled and streaming set to True, the LLM response is consistently empty/blank.

This issue does not occur under the following conditions:

Using gemini-2.5-flash (works perfectly).

Setting stream=False with gemini-3-flash-preview.

Using gemini-3-flash-preview without any tools.

Repeated testing with the same prompt confirms that this issue is specific to the combination of the Gemini 3 Flash Preview model + Tools + Streaming.

To Reproduce Steps to reproduce the behavior:

Define an LlmAgent using gemini-3-flash-preview.

Attach a tool (e.g., a simple calculation tool).

Run the agent with a prompt that triggers the tool, ensuring stream=True.

Minimal Reproducible Code:

Python

import asyncio
from google.adk.agents import LlmAgent

async def calculation_tool(input: str) -> str:
    try:
        result = str(eval(input))
        return result 
    except Exception as e:
        return f"Error: {str(e)}"

# Setup the agent with gemini-3-flash-preview
root_agent = LlmAgent(
    name="calculation_agent",
    description="This agent performs basic arithmetic calculations based on user input.",
    instruction="""
    Use the provided calculation_tool to perform arithmetic operations.
    Only use the calculation_tool for calculations.
    Provide the final result directly to the user without additional explanations.
    """,
    model="gemini-3-flash-preview", # Bug occurs with this version
    tools=[
        calculation_tool
    ]
)

Expected behavior The model should invoke the tool, process the result, and stream the text response back to the user, similar to how gemini-2.5-flash behaves under the same configuration.

Screenshots N/A

Desktop (please complete the following information):

OS: window

Python version:3.12

ADK version: 1.21.0

Model Information:

Using Gemini API (Not vertexai)

Which model is being used: gemini-3-flash-preview

Additional context

Isolation Test 1: Changed model to gemini-2.5-flash -> Works.

Isolation Test 2: Kept gemini-3-flash-preview but set stream=False -> Works.

Isolation Test 3: Kept gemini-3-flash-preview and stream=True but removed tools -> Works.

The issue appears to be a specific breakdown in handling the streaming response payload when tool calls are involved in the new Gemini 3 preview model.

[wukath]

when you say stream=True does that refer to the token streaming toggle in the adk web ui? just trying to reproduce the bug

[bokang-ugent]

This bug can also be reproduced programmatically using Runner.run_async() with RunConfig(streaming_mode=StreamingMode.SSE) (ADK version 1.21.0), not just via adk web:

import asyncio
from google.adk.agents import LlmAgent
from google.adk.agents.run_config import RunConfig, StreamingMode
from google.adk.runners import Runner
from google.adk.sessions import InMemorySessionService
from google.genai import types


async def calculate(expression: str) -> str:
    """Evaluate arithmetic expression."""
    return str(eval(expression))


agent = LlmAgent(
    name="calc_agent",
    model="gemini-3-flash-preview",  # Bug: empty response with this model
    # model="gemini-2.5-flash",       # Workaround: this model works
    instruction="Use the calculate tool for math questions.",
    tools=[calculate],
)


async def main():
    session_service = InMemorySessionService()
    runner = Runner(agent=agent, app_name="test", session_service=session_service)
    session = await session_service.create_session(app_name="test", user_id="user")
    
    # Enable streaming
    run_config = RunConfig(streaming_mode=StreamingMode.SSE)
    # run_config = RunConfig() # disable streaming (default)
    
    # new_message must be a Content object, not a string
    user_message = types.Content(
        role="user",
        parts=[types.Part.from_text(text="What is 2 + 2?")],
    )
    
    response = ""
    async for event in runner.run_async(
        user_id="user",
        session_id=session.id,
        new_message=user_message,
        run_config=run_config,
    ):
        if event.content and event.content.parts:
            for part in event.content.parts:                
                if hasattr(part, "text") and part.text:
                    response += part.text                    
                    print(f"Part text: {part.text!r}")        
    print(f"FinalResponse: {response!r}")

if __name__ == "__main__":
    asyncio.run(main())

When I ran this script with gemini-2.5-flash, I got a successful response:

Part text: '2'
Part text: ' + 2 = 4.'
Part text: '2 + 2 = 4.'
FinalResponse: '2 + 2 = 4.2 + 2 = 4.'

When I ran this script with gemini-3-flash-preview, I got a empty response:

FinalResponse: ''

This appears to be related to #3754 (streaming=True in /run_sse returns empty text after AgentTool calls), which reports the same issue with Fastapi server using google.adk.cli.fast_api import get_fast_api_app: empty text response after tool execution when streaming is enabled.

[xuanyang15]

@wltnrhkd123 @bokang-ugent Thanks for reporting this issue! There is a known issue with SSE Streaming + Gemini 3 + tools, which should be addressed by the newly added progressive SSE streaming feature:

• If you are using v1.21.0, you can enable the feature by setting env variable ADK_ENABLE_PROGRESSIVE_SSE_STREAMING=1.
• If you upgrade to our new version v1.22.0 released today, the feature is enabled by default

Could you please try if it works for you?

[bokang-ugent]

@xuanyang15 Thanks for the response. I tested with longer tool response, both v1.21.0 + ADK_ENABLE_PROGRESSIVE_SSE_STREAMING=1 and upgrade to v1.22.0 solution works!

[wukath]

nice!