[BUG] Tools are not invoked when used with vLLM's OpenAI compatible endpoint

[ssbusc1]

Checks

• I have updated to the lastest minor and patch version of Strands
• I have checked the documentation and this is not expected behavior
• I have searched ./issues and there are no duplicates of my issue

Strands Version

1.12.0

Python Version

3.13.5

Operating System

Ubuntu 22.04.4 LTS

Installation Method

pip

Steps to Reproduce

I have a vLLM server with a Qwen3-coder model and am hitting the OpenAI compatible endpoint.
Steps to reproduce:

• pip install strands-agents[openai]
• Configure the OpenAI model provider with the vLLM base_url and model (see below)
• Configure strands with the OpenAI model provider
• Invoke the agent

from strands import Agent

@tool
def get_weather(city: str) -> str:
    """returns weather info for the specified city."""
    return f"The weather in {city} is sunny with a high of 85F and a low of 58F"

model = OpenAIModel(
    client_args={
        "api_key": "EMPTY",
        "base_url": "<my base url>",
    },
    # **model_config
    model_id="<my Qwen3-coder model path>",
    params={
        "max_tokens": 1000,
        "temperature": 0.7,
    }
)

# Initialize the agent with tools, model, and configuration
agent = Agent(model=model,
              tools=[get_weather],
              system_prompt="You are the chief meteorologist at a TV station. When asked, you should look up the weather in the specified city and provide a colorful response"
              )
result = agent(""What is the weather in San Francisco?"")
print(result.message)

I've also tried replacing the get_weather tool with the calculator tool from the strands-agents-tools package, and that doesn't invoke the tool either.

Expected Behavior

"The weather in San Francisco is sunny with a high of 85F and a low of 58F" or something like that.

FWIW, the same tool when used with the openai-agents SDK (with the relevant annotation) and the same vLLM server+ model combination generates the weather report as expected.

Actual Behavior

The result of running the code prints out:

<tool_call>
<function=get_weather>
<parameter=city>
San Francisco
</parameter>
</function>
</tool_call>

instead of the weather.

Here's the entire result:

AgentResult(stop_reason='end_turn',
            message={'content': [{'text': '<tool_call>\n'
                                          '<function=get_weather>\n'
                                          '<parameter=city>\n'
                                          'San Francisco\n'
                                          '</parameter>\n'
                                          '</function>\n'
                                          '</tool_call>'}],
                     'role': 'assistant'},
            metrics=EventLoopMetrics(cycle_count=1,
                                     tool_metrics={},
                                     cycle_durations=[0.3705911636352539],
                                     traces=[<strands.telemetry.metrics.Trace object at 0x7fc473363460>],
                                     accumulated_usage={'inputTokens': 310,
                                                        'outputTokens': 23,
                                                        'totalTokens': 333},
                                     accumulated_metrics={'latencyMs': 0}),
            state={})

Additional Context

I also issued a direct chat.completions request against the server using the OpenAI client. Here's that response:

ChatCompletion(id='chatcmpl-6f6f42cf28014456893aee0f6fa61b2e', choices=[Choice(finish_reason='tool_calls', index=0, logprobs=None, message=ChatCompletionMessage(content=None, refusal=None, role='assistant', annotations=None, audio=None, function_call=None, tool_calls=[ChatCompletionMessageFunctionToolCall(id='chatcmpl-tool-15fa2e41b38a4a5aadece55724056f4a', function=Function(arguments='{"city": "San Francisco"}', name='get_weather'), type='function')], reasoning_content=None), stop_reason=None, token_ids=None)], created=1760142049, model='<my Qwen3-coder model name>', object='chat.completion', service_tier=None, system_fingerprint=None, usage=CompletionUsage(completion_tokens=23, prompt_tokens=300, total_tokens=323, completion_tokens_details=None, prompt_tokens_details=None), prompt_logprobs=None, prompt_token_ids=None, kv_transfer_params=None)

The finish_reason is set to tool_calls as expected, and the tool call and arguments seem fine to me.

Possible Solution

No response

Related Issues

No response