Empty final output after successful tool call when functionCallingConfig.mode=AUTO

[hapiiiiGit]

🔴 Required Information

Please ensure all items in this section are completed to allow for efficient
triaging. Requests without complete information may be rejected / deprioritized.
If an item is not applicable to you - please mark it as N/A

Describe the Bug:
I am testing Google ADK with a simple HumanEval-style Python code generation agent.
When functionCallingConfig.mode is set to AUTO, I sometimes get an empty final output after the model decides to call a tool.

The issue does not happen when the model directly returns text in the first LLM response. It happens when the first model response contains a function_call, the tool executes successfully, and then the second model response returns an empty content.parts list with finish_reason: "STOP".

As a result, ADK treats the final model event as empty and the final agent output becomes an empty string.

Steps to Reproduce:
Please provide a numbered list of steps to reproduce the behavior:

1. Model: gemini-2.5-flash-lite
2. Agent task: HumanEval-style Python code generation
3. Agent Config
4. Yaml

agent_class: LlmAgent

name: my_agent
model: gemini-2.5-flash-lite
description: Agent for HumanEval-style Python code generation experiments.

instruction: |
  You are a Python coding assistant. Your task is to complete the following Python code from the HumanEval benchmark.
  Use tools when they help, and give concise, evidence-aware answers.
  Note:
    1. Output only final valid Python code. The final answer must be executable Python code.

tools:
  - name: my_agent.tools.python_executor

generate_content_config:
  temperature: 0.0
  maxOutputTokens: 8024
  responseMimeType: text/plain
  toolConfig:
    functionCallingConfig:
      mode: AUTO
  automaticFunctionCalling:
    disable: false
    maximumRemoteCalls: 8
    ignoreCallHistory: true

Expected Behavior:
After a successful tool call, the model should continue and produce the final answer text, for example the completed Python code.

Or, if the model returns an empty response after a tool call, ADK should probably retry, surface an error, or avoid treating the empty model event as the final output.

Observed Behavior:
The tool call succeeds, but the second model response is empty:

{
  "role": "model",
  "text": "",
  "content": {
    "parts": [],
    "role": "model"
  },
  "raw_response": {
    "finish_reason": "STOP",
    "usage_metadata": {
      "candidates_token_count": 0
    }
  }
}

Then ADK emits a final model event with empty text:

{
  "author": "my_agent",
  "role": "model",
  "text": "",
  "function_calls": [],
  "function_responses": []
}

And the final agent output is:

"final_output": ""

Environment Details:

• ADK Library Version (pip show google-adk): 1.32.0
• Desktop OS: Windows
• Python Version : 3.14.4

Model Information:

• Are you using LiteLLM: Yes
• Which model is being used: gemini-2.5-flash-lite

---

🟡 Optional Information

Providing this information greatly speeds up the resolution process.

How often has this issue occurred?:

• Often (50%+)

[sanketpatil06]

Hi @hapiiiiGit , When using functionCallingConfig.mode="AUTO" alongside strict formatting constraints (such as "output only valid code"), the agent may occasionally complete the tool execution but halt before generating the final synthesized text.
Workaround:
To ensure the agent consistently outputs the final code after the tool runs, you can use In-Context Prompt Forcing. By adding explicit continuation instructions to both the agent's prompt and the tool's return payload, you can successfully guide the agent to complete the final generation step.

1. Update the Agent Instructions (agent.yaml):

instruction: 
  You are a Python coding assistant. Your task is to complete the following Python code from the HumanEval benchmark.
  Use tools when they help, and give concise, evidence-aware answers.
  Note:
    1. Output only final valid Python code. The final answer must be executable Python code.
    2. CRITICAL: After receiving the tool execution results, do not stop. Immediately output the final synthesized Python code.

2. Update the Tool Output (python_executor):

def python_executor(code: str) -> str:
    # ... your existing execution logic ...
    
    # Example for handling failures:
    # return f"Execution failed:\n{error_output}\n\nSYSTEM DIRECTIVE: Fix the error and output the final corrected Python code."
    
    # Example for handling success:
    directive = "\n\nSYSTEM DIRECTIVE: Tool execution is complete. You must now generate the final valid Python code."
    
    if not output.strip():
        return f"Execution completed successfully, but no output was printed.{directive}"
        
    return f"Execution completed successfully. Output:\n{output}{directive}"

Could you please give it a try and let us know if this works for you? Thanks.

[hapiiiiGit]

Hi @sanketpatil06 , thanks for the suggestion.
I tried your suggestions by updating instructions and Tool Output. Based on my tests, this does reduce the frequency of the empty final response issue, but it does not completely eliminate it. I can still occasionally reproduce cases where the tool finishes successfully, but the model returns no final synthesized code.

I also noticed a side effect after updating the agent instruction: the agent trace becomes much longer. Previously, the agent usually called the tool only once or twice. After updating the instructions, I now sometimes see the agent calling the tool around 20–30 times before stopping.

By the way, I also have a question about functionCallingConfig.mode="ANY". My understanding was that ANY should force the LLM to call a tool. However, in my tests, even with mode set to ANY, I still sometimes observe responses where no tool call is made. Is this expected behavior in ADK, or should ANY guarantee that every model turn contains a tool call?