|
| 1 | +import types |
| 2 | + |
| 3 | +import pytest |
| 4 | + |
| 5 | +import eval_protocol.mcp.execution.policy as policy_mod |
| 6 | +from eval_protocol.mcp.execution.policy import LiteLLMPolicy |
| 7 | + |
| 8 | + |
| 9 | +@pytest.mark.asyncio |
| 10 | +async def test_litellm_policy_surfaces_provider_specific_reasoning_details(monkeypatch): |
| 11 | + """ |
| 12 | + Ensure that provider_specific_fields from the LiteLLM message object are |
| 13 | + preserved on the returned message dict from LiteLLMPolicy._make_llm_call. |
| 14 | + """ |
| 15 | + |
| 16 | + # Define a fake ModelResponse base class and patch the module's ModelResponse |
| 17 | + class FakeModelResponseBase: ... |
| 18 | + |
| 19 | + policy_mod.ModelResponse = FakeModelResponseBase |
| 20 | + |
| 21 | + async def fake_acompletion(*args, **kwargs): |
| 22 | + # This mimics the LiteLLM Message object shape we rely on in policy._make_llm_call |
| 23 | + message_obj = types.SimpleNamespace( |
| 24 | + role="assistant", |
| 25 | + content="", |
| 26 | + tool_calls=[ |
| 27 | + types.SimpleNamespace( |
| 28 | + id="tool_get_reservation_details_123", |
| 29 | + type="function", |
| 30 | + function=types.SimpleNamespace( |
| 31 | + name="get_reservation_details", |
| 32 | + arguments='{"reservation_id":"EHGLP3"}', |
| 33 | + ), |
| 34 | + ) |
| 35 | + ], |
| 36 | + provider_specific_fields={ |
| 37 | + "reasoning_details": [{"id": "tool_get_reservation_details_123", "type": "reasoning.encrypted"}], |
| 38 | + "custom_field": "keep_me", |
| 39 | + }, |
| 40 | + ) |
| 41 | + |
| 42 | + class FakeModelResponse(FakeModelResponseBase): |
| 43 | + def __init__(self) -> None: |
| 44 | + self.choices = [ |
| 45 | + types.SimpleNamespace( |
| 46 | + finish_reason="tool_calls", |
| 47 | + index=0, |
| 48 | + message=message_obj, |
| 49 | + ) |
| 50 | + ] |
| 51 | + self.usage = types.SimpleNamespace( |
| 52 | + prompt_tokens=10, |
| 53 | + completion_tokens=5, |
| 54 | + total_tokens=15, |
| 55 | + ) |
| 56 | + |
| 57 | + return FakeModelResponse() |
| 58 | + |
| 59 | + # Patch acompletion so we don't hit the network |
| 60 | + monkeypatch.setattr(policy_mod, "acompletion", fake_acompletion) |
| 61 | + |
| 62 | + # Use a concrete policy instance; base_url/model_id values don't matter for this unit test |
| 63 | + policy = LiteLLMPolicy(model_id="openrouter/google/gemini-3-pro-preview", use_caching=False) |
| 64 | + |
| 65 | + messages = [ |
| 66 | + { |
| 67 | + "role": "assistant", |
| 68 | + "content": "", |
| 69 | + "tool_calls": [ |
| 70 | + { |
| 71 | + "id": "tool_get_reservation_details_123", |
| 72 | + "type": "function", |
| 73 | + "function": {"name": "get_reservation_details", "arguments": '{"reservation_id":"EHGLP3"}'}, |
| 74 | + } |
| 75 | + ], |
| 76 | + } |
| 77 | + ] |
| 78 | + |
| 79 | + # No tools are needed for this test – we only care about the returned message shape |
| 80 | + result = await policy._make_llm_call(messages, tools=[]) |
| 81 | + |
| 82 | + assert "choices" in result |
| 83 | + assert len(result["choices"]) == 1 |
| 84 | + msg = result["choices"][0]["message"] |
| 85 | + |
| 86 | + # Core fields should be present |
| 87 | + assert msg["role"] == "assistant" |
| 88 | + assert isinstance(msg.get("tool_calls"), list) |
| 89 | + |
| 90 | + # provider_specific_fields should be preserved on the message |
| 91 | + ps = msg.get("provider_specific_fields") |
| 92 | + assert isinstance(ps, dict) |
| 93 | + assert ps["reasoning_details"] == [{"id": "tool_get_reservation_details_123", "type": "reasoning.encrypted"}] |
| 94 | + # Non-core provider_specific_fields should also be preserved |
| 95 | + assert ps.get("custom_field") == "keep_me" |
0 commit comments