[Bug] error in streaming mode if llm call tool does not require parameters

[wusskk]

Description

Communicate with qwen3-32b via streaming, when need to call a tool without arguments, if the model returns the argument as '', wait until the completion of the tool call and then return to the model will report an error.
I think there are 3 main reasons for this problem:

1. the ability of the model itself (no such phenomenon when using deepseek-chat API, but it will be triggered when using locally deployed qwen3-32b)
2. the use of streaming (not triggered when stream=False)
3. tool does not require parameters

Steps to Reproduce

1. select qwen3-32b as model
2. use stream=True
3. selec a tool does not require parameters

Agent Configuration (if applicable)

Minimum Implementation Code:

import random

from agno.agent import Agent
from agno.models.openai.like import OpenAILike
from agno.tools import tool


@tool(show_result=True)
def get_number() -> str:
    """Get the number."""
    # In a real implementation, this would call a number API
    random_number = random.randint(1, 100)

    return f"The number is {random_number}."


model = OpenAILike(
    id="qwen3-32b", base_url="http://127.0.0.1:8080, api_key="sk-xxxxx"
)

agent = Agent(model=model, tools=[get_number], debug_mode=True)
agent.print_response("Give me a random number.", stream=True)

Expected Behavior

Call the tool and output the results

Actual Behavior

PS D:\Code\agent\agno_bug> & D:/Code/agent/agno_bug/.venv/Scripts/python.exe d:/Code/agent/agno_bug/main.py
DEBUG ******************************************* Agent ID: d70c8231-afc4-4770-8ff7-812fc329179d *******************************************
DEBUG ****************************************** Session ID: 4a93145c-b17e-4f0a-aace-08780e3fe147 ******************************************
DEBUG *************************************** Agent Run Start: 4d03b13d-f073-4f03-b938-ffa1b1a242e9 ****************************************
DEBUG Processing tools for model
DEBUG Added tool get_number
DEBUG ---------------------------------------------------- OpenAI Response Stream Start ----------------------------------------------------
DEBUG ---------------------------------------------------------- Model: qwen3-32b ----------------------------------------------------------
DEBUG ================================================================ user ================================================================
DEBUG Give me a random number.
DEBUG messages in stream: [{'role': 'user', 'content': 'Give me a random number.'}]
DEBUG ============================================================= assistant ==============================================================
DEBUG <think>
      Okay, the user is asking for a random number. Let me check the tools available. There's a function called get_number with no parameters required.   
      Since the user wants a random number, I should call that function. But wait, the function's description just says "Get the number." Does it generate
      a random one or return a fixed value? The parameters are empty, so maybe it's designed to return a random number by default. I'll proceed to call   
      get_number without any arguments. I need to make sure the tool call is correctly formatted in JSON within the XML tags. Alright, that should do it. 
      </think>


DEBUG Tool Calls:                                                                                                                                         
        - ID: 'chatcmpl-tool-1077bda4db6941d1936130b9f9d57102'                                                                                            
          Name: 'get_number'                                                                                                                              
DEBUG *************************************************************  METRICS  **************************************************************              
DEBUG * Tokens:                      input=139, output=144, total=283                                                                                     
DEBUG * Time:                        3.3514s                                                                                                              
DEBUG * Tokens per second:           42.9673 tokens/s                                                                                                     
DEBUG * Time to first token:         0.9840s                                                                                                              
DEBUG *************************************************************  METRICS  **************************************************************              
DEBUG Getting function get_number                                                                                                                         
DEBUG Running: get_number()                                                                                                                               
DEBUG ================================================================ tool ================================================================              
DEBUG Tool call Id: chatcmpl-tool-1077bda4db6941d1936130b9f9d57102
DEBUG The number is 67.                                                                                                                                   
DEBUG ***********************************************************  TOOL METRICS  ***********************************************************              
DEBUG * Time:                        0.0022s                                                                                                              
DEBUG ***********************************************************  TOOL METRICS  ***********************************************************              
DEBUG messages in stream: [{'role': 'user', 'content': 'Give me a random number.'}, {'role': 'assistant', 'content': '<think>\nOkay, the user is asking   
      for a random number. Let me check the tools available. There\'s a function called get_number with no parameters required. Since the user wants a    
      random number, I should call that function. But wait, the function\'s description just says "Get the number." Does it generate a random one or      
      return a fixed value? The parameters are empty, so maybe it\'s designed to return a random number by default. I\'ll proceed to call get_number      
      without any arguments. I need to make sure the tool call is correctly formatted in JSON within the XML tags. Alright, that should do                
      it.\n</think>\n\n', 'tool_calls': [{'id': 'chatcmpl-tool-1077bda4db6941d1936130b9f9d57102', 'type': 'function', 'function': {'name': 'get_number',  
      'arguments': ''}}]}, {'role': 'tool', 'content': 'The number is 67.', 'tool_call_id': 'chatcmpl-tool-1077bda4db6941d1936130b9f9d57102'}]            
ERROR    API status error from OpenAI API: Error code: 400 - {'object': 'error', 'message': 'Expecting value: line 1 column 1 (char 0)', 'type':
         'BadRequestError', 'param': None, 'code': 400}
▰▱▱▱▱▱▱ Thinking...
┏━ Message ━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━┓
┃                                                                                                                                                        ┃
┃ Give me a random number.                                                                                                                               ┃
┃                                                                                                                                                        ┃
┗━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━┛
┏━ Tool Calls ━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━┓
┃                                                                                                                                                        ┃
┃ • get_number()                                                                                                                                         ┃
┃                                                                                                                                                        ┃
┗━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━┛
┏━ Response (3.4s) ━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━┓
┃                                                                                                                                                        ┃
┃ <think>                                                                                                                                                ┃
┃ Okay, the user is asking for a random number. Let me check the tools available. There's a function called get_number with no parameters required.      ┃
┃ Since the user wants a random number, I should call that function. But wait, the function's description just says "Get the number." Does it generate a ┃
┃ random one or return a fixed value? The parameters are empty, so maybe it's designed to return a random number by default. I'll proceed to call        ┃
┃ get_number without any arguments. I need to make sure the tool call is correctly formatted in JSON within the XML tags. Alright, that should do it.    ┃
┃ </think>                                                                                                                                               ┃
┃                                                                                                                                                        ┃
┃ The number is 67.                                                                                                                                      ┃
┃                                                                                                                                                        ┃
┗━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━┛
Traceback (most recent call last):
  File "D:\Code\agent\agno_bug\.venv\Lib\site-packages\agno\models\openai\chat.py", line 455, in invoke_stream
    yield from self.get_client().chat.completions.create(
               ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "D:\Code\agent\agno_bug\.venv\Lib\site-packages\openai\_utils\_utils.py", line 287, in wrapper
    return func(*args, **kwargs)
           ^^^^^^^^^^^^^^^^^^^^^
  File "D:\Code\agent\agno_bug\.venv\Lib\site-packages\openai\resources\chat\completions\completions.py", line 925, in create
    return self._post(
           ^^^^^^^^^^^
  File "D:\Code\agent\agno_bug\.venv\Lib\site-packages\openai\_base_client.py", line 1239, in post
    return cast(ResponseT, self.request(cast_to, opts, stream=stream, stream_cls=stream_cls))
                           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "D:\Code\agent\agno_bug\.venv\Lib\site-packages\openai\_base_client.py", line 1034, in request
    raise self._make_status_error_from_response(err.response) from None
openai.BadRequestError: Error code: 400 - {'object': 'error', 'message': 'Expecting value: line 1 column 1 (char 0)', 'type': 'BadRequestError', 'param': None, 'code': 400}

The above exception was the direct cause of the following exception:

Traceback (most recent call last):
  File "d:\Code\agent\agno_bug\main.py", line 24, in <module>
    agent.print_response("Give me a random number.", stream=True)
  File "D:\Code\agent\agno_bug\.venv\Lib\site-packages\agno\agent\agent.py", line 4545, in print_response
    for resp in self.run(
                ^^^^^^^^^
  File "D:\Code\agent\agno_bug\.venv\Lib\site-packages\agno\agent\agent.py", line 655, in _run
    for model_response_chunk in self.model.response_stream(messages=run_messages.messages):
                                ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "D:\Code\agent\agno_bug\.venv\Lib\site-packages\agno\models\base.py", line 520, in response_stream
    yield from self.process_response_stream(
  File "D:\Code\agent\agno_bug\.venv\Lib\site-packages\agno\models\base.py", line 492, in process_response_stream
    for response_delta in self.invoke_stream(messages=messages):
                          ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "D:\Code\agent\agno_bug\.venv\Lib\site-packages\agno\models\openai\chat.py", line 490, in invoke_stream
    raise ModelProviderError(
agno.exceptions.ModelProviderError: Unknown model error

Screenshots or Logs (if applicable)

Environment

• OS: Windows 11
• Agno Version: v1.4.5
• External Dependency Versions: openai-v1.78.0
• Additional Environment Details: Python 3.12.8

Possible Solutions (optional)

# add code at _format_messages in agno.model.opnai.chat.OpenAIChat
if "tool_calls" in message_dict:
            for tool in message_dict["tool_calls"]:
                if tool["type"] == "function":
                    if not tool["function"]["arguments"]:
                        tool["function"]["arguments"] = '{}'

Additional Context

A similar problem arises in pydantic-ai, and they think should fix in agent framework.

[ysolanky]

Hello @wusskk ! Looks like what is happening here is an incompatibility between the model qwen3-32b and the OpenAI spec for function calling.

OpenAI expects the arguments for a tool call with no params to be {} whereas qwen3-32b returns '' resulting in the error. Your suggestion works but is tied to just qwen3-32b. I think a better solution would be to have a separate Qwen model class if this behaviour is consistent across all qwen models. Can you please share how you are hosting this model?

[wusskk]

Hi @ysolanky , I am using the model qwen3-32b, but it is an API provided by others. As far as I know, they use vllm for deployment, and I am not clear about the specific commands.
Now I have switched to other API providers (siliconflow and Aliyun), using the qwen3 series models for testing, and found that this issue did not occur. Therefore, I suspect it might be related to the deployment method.
Moreover, only when the streaming output is enabled, the qwen3 deployed by vllm will encounter this issue; it runs normally when the streaming output is disabled.