Empty model message after tool call

[CommonGibbon]

Describe the bug
After calling a tool, the final model message in a session will randomly have no content parts.
For example:
Event(model_version='gemini-2.5-flash-lite', content=Content(
role='model'
)

To Reproduce
Please share a minimal code and data to reproduce your problem.
Steps to reproduce the behavior:

1. Install google-adk
2. Run:

###############################################
# SETUP CODE
from typing import Any, Dict

from google.adk.agents import LlmAgent
from google.adk.models.google_llm import Gemini
from google.adk.sessions import InMemorySessionService
from google.adk.runners import Runner
from google.adk.tools.tool_context import ToolContext
from google.genai import types


retry_config = types.HttpRetryOptions(
    attempts=5,  # Maximum retry attempts
    exp_base=7,  # Delay multiplier
    initial_delay=1,
    http_status_codes=[429, 500, 503, 504],  # Retry on these HTTP errors
)

USER_NAME_SCOPE_LEVELS = ("temp", "user", "app")

# This demonstrates how tools can read from session state.
def retrieve_userinfo(tool_context: ToolContext) -> Dict[str, Any]:
    """
    Tool to retrieve user name and country from session state.
    """
    # Read from session state
    user_name = tool_context.state.get("user:name", "Username not found")
    country = tool_context.state.get("user:country", "Country not found")

    return {"status": "success", "user_name": user_name, "country": country}


    # Configuration
APP_NAME = "default"
USER_ID = "default"

# Create an agent with session state tools
root_agent = LlmAgent(
    model=Gemini(model="gemini-2.5-flash-lite", retry_options=retry_config),
    name="text_chat_bot",
    description="""A text chatbot.
    Tools for managing user context:
    * To fetch username and country when required use `retrieve_userinfo` tool.
    """,
    tools=[retrieve_userinfo],  # Provide the tools to the agent
)

# Set up session service and runner
session_service = InMemorySessionService()
runner = Runner(agent=root_agent, session_service=session_service, app_name="default")

# Inject user info into session state
session = await runner.session_service.create_session(
    app_name=APP_NAME,
    user_id=USER_ID,
    state = {"user:name": "Will", "user:country":"USA"}
)
idx=0 # we'll use this below
###############################################
# RUN CODE 
# You may need to run the following code multiple times in order to reproduce the issue. 
# I've built in session cycling so you don't accidentally contaminate new attempts
# It easiest to run this in a jupyter notebook with the setup code in one cell and the run code in another
idx+=1
session = await session_service.create_session(
    app_name=APP_NAME, user_id=USER_ID, session_id=f"new-session-{idx}" # this may take several attempts
)

query = types.Content(role="user", parts=[types.Part(text="Could you use a tool call to recall my name?")])

async for event in runner.run_async(
    user_id=USER_ID, session_id=session.id, new_message=query
):
    print(f"{event.content.role}:")
    if event.content.parts is not None:
        for part in event.content.parts:
            if part.function_call is not None:
                print(f"FUNCTION CALL: {part.function_call.name}: {part.function_call.args}")
            elif part.function_response is not None:
                print(f"FUNCTION RESPONSE: {part.function_response.name}: {part.function_response.response}")
            else:
                print(f"{part.text}")
    else:
        print("TARGET BUG BEHAVIOR IDENTIFIED: NO CONTENT GENERATED")
        print(event.content)
    print("-"*15)
###############################################

Expected behavior
We expect the final message from the model to output the saved user name and country:

Sending out request, model: gemini-2.5-flash-lite, backend: GoogleLLMVariant.GEMINI_API, stream: False
HTTP Request: POST https://generativelanguage.googleapis.com/v1beta/models/gemini-2.5-flash-lite:generateContent "HTTP/1.1 200 OK"
Response received from the model.
Warning: there are non-text parts in the response: ['function_call'], returning concatenated text result from text parts. Check the full candidates.content.parts accessor to get the full model response.
Sending out request, model: gemini-2.5-flash-lite, backend: GoogleLLMVariant.GEMINI_API, stream: False
model:
FUNCTION CALL: retrieve_userinfo: {}
---------------
user:
FUNCTION RESPONSE: retrieve_userinfo: {'status': 'success', 'user_name': 'Will', 'country': 'USA'}
---------------
HTTP Request: POST https://generativelanguage.googleapis.com/v1beta/models/gemini-2.5-flash-lite:generateContent "HTTP/1.1 200 OK"
Response received from the model.
model:
I'm sorry, I cannot recall your name, but I can see that you are Will and you are from the USA.
---------------

Actual behavior
Maybe 50% of the time, the final message in the event stream has no parts. See example output below:

Sending out request, model: gemini-2.5-flash-lite, backend: GoogleLLMVariant.GEMINI_API, stream: False
HTTP Request: POST https://generativelanguage.googleapis.com/v1beta/models/gemini-2.5-flash-lite:generateContent "HTTP/1.1 200 OK"
Response received from the model.
Warning: there are non-text parts in the response: ['function_call'], returning concatenated text result from text parts. Check the full candidates.content.parts accessor to get the full model response.
Sending out request, model: gemini-2.5-flash-lite, backend: GoogleLLMVariant.GEMINI_API, stream: False
model:
FUNCTION CALL: retrieve_userinfo: {}
---------------
user:
FUNCTION RESPONSE: retrieve_userinfo: {'status': 'success', 'user_name': 'Will', 'country': 'USA'}
---------------
HTTP Request: POST https://generativelanguage.googleapis.com/v1beta/models/gemini-2.5-flash-lite:generateContent "HTTP/1.1 200 OK"
Response received from the model.
model:
TARGET BUG BEHAVIOR IDENTIFIED: NO CONTENT GENERATED
parts=None role='model'
---------------

Desktop (please complete the following information):

• OS: Mac
• Python version(python -V): 3.13
• ADK version(pip show google-adk): 1.18

Model Information:

• Are you using LiteLLM: No? Is this what the 'lite' means in gemini-2.5-flash-lite?
• Which model is being used(e.g. gemini-2.5-pro) gemini-2.5-flash-lite

Additional context
If I switch from gemini-2.5-flash-lite, the issue appears to resolve to happen either at a much lower rate or not at all.

[zmy1116]

You probably want to reformat your message... the cap letters are burning my eyes.

This is a serious problem I am facing as well. From what I've seen, this is not an adk problem, this is a gemini-2.5 problem:
LLM decide to return an empty response when in tool turn, I have seen the 3 following behaviours:

• LLM return message "I will use the tool xxx to....", then terminates the turn. No tool call is dispatched.
• LLM does tool call, get function response, terminate (I think this is what you experience)
• LLM returns empty directly.

This problem exists for both gemini-2.5-flash and gemini-2.5-flash-lite.

I've noticed the third scenarios when I started testing with adk 1.18. For the exact same system prompt + function definition + chat history sequence, the empty response can be reproduced 100% of times. Upon closely examination, there was a change in 1.18 that fixed how function definition is passed to LLM context if the function contains optional args with default value. In ealier version, all args including the optional ones are passed in as required args for function definition. 1.18 fix this error... but somehow is triggering empty response.

I did not create a bug report with this example because there are client data involved, and I cannot reproduce the same issue with a dummy agent. I can share the data to google devs privately if there is an interest...

In gemini-cli I rememer there is a module that explicitly create an user message "please continue" if the model return empty and there is no error, so I don't think this is an uncommon issue.

Not sure what is the best way to handle it, for now we also have a please continue when receiving an empty response. That solves scenario 2 and 3. But for scenario 1, we don't have a solution.

[Jacksunwei]

Thanks for the detailed reproduction case. Based on the symptoms and @zmy1116's observations, this appears to be a Gemini 2.5 model-level behavior rather than an ADK bug—the model is choosing to return empty responses in certain tool call scenarios.

Debugging recommendation:

Run with verbose logging to inspect the actual LLM request being sent:

adk web -v

Or in code:

import logging
logging.basicConfig(level=logging.DEBUG)

Look at the outgoing LlmRequest (especially the function definitions and chat history) in the logs. This will show you exactly what context the model is receiving when it decides to return empty.

Potential workarounds:

1. Adjust your agent's instruction/description - Sometimes rephrasing the system prompt or tool descriptions can reduce this behavior
2. Add explicit instructions about tool use (e.g., "After calling a tool, always explain the result to the user")
3. Switch to gemini-2.5-flash if the issue persists with -lite

The "please continue" approach mentioned by @zmy1116 is a known pattern for handling this, though ideally the prompt tuning should prevent it.

Let us know if you discover any prompt patterns that reliably trigger or avoid this behavior.

[stfines-clgx]

We can reliably replicate this error with longer message sizes - messages of >5000 tokens seem to trigger the behavior.