fix(ollama): set finish_reason to tool_calls and remove broken capability check

[rsp2k]

Fix Ollama tool calling issues

Fixes: #18922

The Problems

Two issues causing tool calls to be ignored:

1. finish_reason hardcoded to "stop" - Clients use finish_reason to determine how to process responses. When tool_calls were present, it should be "tool_calls" not "stop".

2. Broken model capability check - get_model_info() fails when Ollama runs on a remote server, triggering a broken JSON prompt injection fallback.

The Fix

~30 lines removed, ~10 lines added.

# Before: Complex try/except with network call and broken fallback
try:
    model_info = litellm.get_model_info(model=model, custom_llm_provider="ollama")
    if model_info.get("supports_function_calling") is True:
        optional_params["tools"] = value
    else:
        raise Exception
except Exception:
    # Broken fallback with JSON prompt injection...

# After: Just pass tools directly - Ollama 0.4+ handles capability detection
optional_params["tools"] = value

# Before: finish_reason always "stop"
_message = litellm.Message(**response_json_message)
model_response.choices[0].message = _message

# After: Set "tool_calls" when tool_calls present
_message = litellm.Message(**response_json_message)
model_response.choices[0].message = _message
if _message.tool_calls:
    model_response.choices[0].finish_reason = "tool_calls"

Breaking Change

Removes the old JSON prompt injection fallback for pre-0.4 Ollama. That approach was unreliable anyway, and Ollama 0.4 has been out for a while with proper native tool calling.

Testing

• All 14 existing Ollama tests pass
• Added 3 focused tests for the fixes
• Live tested with qwen3 + Ollama - tool_calls work correctly

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

Project	Deployment	Review	Updated (UTC)
litellm	Ready	Preview, Comment	Jan 11, 2026 5:43pm

[CLAassistant]

All committers have signed the CLA.