ADK LlmAgent/Gemini: Tool Parameter Error & 400 on Tool Failure

[ilteris]

Describe the bug
An LlmAgent using certain Gemini models (e.g., gemini-2.5-flash-preview-04-17) fails to correctly extract and use a parameter from its input JSON (as described in its instruction) when making a tool call. Instead, it appears to use a hallucinated or example parameter value. When the tool consequently returns an error message string (e.g., "Tool Error: Item not found"), the subsequent LLM call (intended to generate a final natural language response based on the tool's error) fails with a 400 INVALID_ARGUMENT error. The error message is: "Please ensure that function call turn comes immediately after a user turn or after a function response turn." This suggests an issue with how the tool's string error output is handled in the conversation history for the LLM.

To Reproduce
Steps to reproduce the behavior:

1. Define an LlmAgent with:
   • An input_schema (e.g., Pydantic model for user_query and item_id).
   • An instruction explicitly telling the agent to:
     • Extract item_id from the input JSON string.
     • Call a tool (e.g., get_item_details_tool) using this exact extracted item_id.
     • Extract user_query from the input JSON.
     • Analyze the tool's output to answer the user_query.
     • Respond in natural language, or with a specific string (e.g., "$I don't know$") if the tool fails or the answer isn't found.
   • The tool (e.g., get_item_details_tool) should be an async def function decorated with @FunctionTool. It should be designed to return an error string like "Tool Error: Item [item_id] not found." if the item isn't found.
2. Configure the agent to use a model like gemini-2.5-flash-preview-04-17.
3. Invoke the agent using Runner.run_async() with an input JSON containing a specific item_id (e.g., "item_id": "actual_id_123") and a user_query.
4. Observe the logs:
   • The agent's tool call log will show it attempting to use an incorrect item_id (e.g., "sample_id", "test_id") instead of "actual_id_123".
   • The tool will log that it cannot find this incorrect item_id and will return its error string.
   • The httpx logs will show the second call to generateContent for the model failing with a HTTP/1.1 400 Bad Request.
   • The application logs will show the google.genai.errors.ClientError: 400 INVALID_ARGUMENT. {'error': {'code': 400, 'message': 'Please ensure that function call turn comes immediately after a user turn or after a function response turn.', 'status': 'INVALID_ARGUMENT'}}.

Expected behavior

1. The LlmAgent should correctly extract the item_id from the input JSON string as per its instructions.
2. The agent should call the tool using this correct item_id.
3. If the tool returns an error string (e.g., "Tool Error: Item actual_id_123 not found."), the ADK framework should correctly package this string as the "function response" in the conversation history.
4. The subsequent LLM call (to generate the final natural language answer) should proceed without a 400 INVALID_ARGUMENT error. The LLM should then use the tool's error message to formulate its final response according to the agent's instructions (e.g., outputting "$I don't know$").

Desktop :

• OS: macOS
• Python version: 3.11
• ADK version: 0.5.0

Additional context

• The issue seems to be twofold:
  1. The LLM's failure to adhere to instructions to use the specific ID from the input for the tool call.
  2. The 400 INVALID_ARGUMENT error when a tool returns a simple error string, suggesting the model is very strict about the format/content of the "function response" part of the turn history. This error was specifically observed with gemini-2.5-flash-preview-04-17.
• The problem persists despite multiple iterations of increasingly explicit prompt engineering for the agent's instructions.
• The setup involves an async def tool decorated with @FunctionTool, and the agent is invoked via Runner.run_async(). The tool is designed to return a string in all cases (data or error message).

[seanzhou1023]

thanks for reporting the issue @ilteris. Could you provide your agent codes to reproduce, so that we are on the same page ?

[seanzhou1023]

feel free to reopen when you have more details. Thanks :)