Ollama integration fails with tool calls: "role" field missing in Message validation

[HECer]

Bug Description

When using OpenCook with an Ollama provider (specifically qwen3.5:9b), the agent crashes during the iterative loop (typically Step 3) when attempting to send a tool call result back to the model. The crash is triggered by a Pydantic validation error because the message object passed to the Ollama client lacks a mandatory role field.

Error Traceback snippet:
pydantic_core._pydantic_core.ValidationError: 1 validation error for Message role Field required [type=missing, input_value={'call_id': 'e8a01082-5c3... 'function_call_output'}, input_type=dict]

Steps to Reproduce

1. Install OpenCook: pip install git+https://github.com/OpenDataBox/OpenCook.git
2. Create a configuration file (config.json) with an Ollama provider:

{
  "provider": "ollama",
  "model": "qwen3.5:9b",
  "model_base_url": "http://localhost:11434",
  "max_steps": 50,
  "console_type": "simple",
  "agent_type": "code_agent"
}

3. Run the agent:
   opencook run "some task" --config-file config.json --provider ollama --model qwen3.5:9b --api-key dummy --working-dir .
4. Observe the crash: The agent successfully initiates a tool call but fails immediately when processing the output of that tool for the next LLM turn.

Expected Behavior

The tool call result should be formatted as a valid message object containing a role field (e.g., "role": "tool") before being sent back to the Ollama API, allowing the model to process the observation.

Actual Behavior

The ollama_client.py constructs a message dictionary containing the call_id and output, but omits the role key. The underlying Ollama Python library uses Pydantic for schema enforcement and rejects the malformed dictionary.

Environment

• OpenCook version: 0.1.0
• Python: 3.10
• OS: Windows 11
• Ollama version: latest
• Model: qwen3.5:9b

Root Cause

The issue is located in code_agent/utils/llm_clients/ollama_client.py at approximately line 83. The method _create_ollama_response (or the logic handling the message history assembly) fails to inject the required role attribute into the message dictionary representing a tool's output.

Suggested Fix

Ensure that the dictionary for tool responses includes the "role": "tool" key and maps the output to the "content" key as expected by the Ollama/OpenAI-compatible schema.

Proposed Message Structure:

{
    "role": "tool",
    "tool_call_id": "...", # or call_id depending on library version
    "content": "..."
}