Reasoning with Ollama loses Host setting

[turicumoperarius]

When using the reasoning option in combination with Ollama as a backend that uses a different endpoint than localhost, the following error occurs.

ERROR Reasoning error: [WinError 10061] No connection could be made because
the target machine actively refused it

I tried the basic plan itinerary example with Ollama.

task = "Plan an itinerary from Los Angeles to Las Vegas" reasoning_agent = Agent(model=Ollama(id="vanilj/Phi-4", host='192.168.1.2'), reasoning=True, markdown=True, structured_outputs=True, debug_mode=False) reasoning_agent.print_response(task, stream=False, show_full_reasoning=True)

After some digging, I found out, that somewhere on the way to the _run() function, the model seems to lose the host setting when its run in reasoning mode. Without reasoning everything works fine. I printed the self.model object from the _run() function with and without reasoning set.

With reasoning:
id='vanilj/Phi-4' name='Ollama' provider='Ollama' metrics={} response_format=<class 'phi.reasoning.step.ReasoningSteps'> tools=None tool_choice=None run_tools=True show_tool_calls=False tool_call_limit=None functions=None function_call_stack=None system_prompt=None instructions=None session_id='e35f4e16-4bfc-47ec-b978-52b4ac9810bf' structured_outputs=True supports_structured_outputs=True format=None options=None keep_alive=None request_params=None host=None timeout=None client_params=None client=None async_client=None

Without reasoning:
id='vanilj/Phi-4' name='Ollama' provider='Ollama' metrics={} response_format=None tools=None tool_choice=None run_tools=True show_tool_calls=False tool_call_limit=None functions=None function_call_stack=None system_prompt=None instructions=None session_id='cb67467c-f9c6-474a-9334-bdc4c64ab34d' structured_outputs=False supports_structured_outputs=True format=None options=None keep_alive=None request_params=None host='192.168.1.2' timeout=None client_params=None client=None async_client=None

After a really dirty fix, by setting the self.model.host parameter manually in the _run() function, everything seems to work fine. So far reasoning with Ollama models does a good job and it would be great if this minor flaw could be fixed.

[ysolanky]

@turicumoperarius that is an excellent find. Thank you so much! I will raise this issue and get a fix out