Add logprobs to ModelSettings (#971)

Author: gold-crow-997

src/agents/extensions/models/litellm_model.py

@@ -321,6 +321,7 @@ async def _fetch_response(
             stream=stream,
             stream_options=stream_options,
             reasoning_effort=reasoning_effort,
+            top_logprobs=model_settings.top_logprobs,
             extra_headers={**HEADERS, **(model_settings.extra_headers or {})},
             api_key=self.api_key,
             base_url=self.base_url,

src/agents/model_settings.py

@@ -55,6 +55,7 @@ class MCPToolChoice:
 ToolChoice: TypeAlias = Union[Literal["auto", "required", "none"], str, MCPToolChoice, None]
 
 
+
 @dataclass
 class ModelSettings:
     """Settings to use when calling an LLM.
@@ -116,6 +117,10 @@ class ModelSettings:
     """Additional output data to include in the model response.
     [include parameter](https://platform.openai.com/docs/api-reference/responses/create#responses-create-include)"""
 
+    top_logprobs: int | None = None
+    """Number of top tokens to return logprobs for. Setting this will
+    automatically include ``"message.output_text.logprobs"`` in the response."""
+
     extra_query: Query | None = None
     """Additional query fields to provide with the request.
     Defaults to None if not provided."""

src/agents/models/openai_chatcompletions.py

@@ -287,6 +287,7 @@ async def _fetch_response(
             stream_options=self._non_null_or_not_given(stream_options),
             store=self._non_null_or_not_given(store),
             reasoning_effort=self._non_null_or_not_given(reasoning_effort),
+            top_logprobs=self._non_null_or_not_given(model_settings.top_logprobs),
             extra_headers={**HEADERS, **(model_settings.extra_headers or {})},
             extra_query=model_settings.extra_query,
             extra_body=model_settings.extra_body,

src/agents/models/openai_responses.py

@@ -3,7 +3,7 @@
 import json
 from collections.abc import AsyncIterator
 from dataclasses import dataclass
-from typing import TYPE_CHECKING, Any, Literal, overload
+from typing import TYPE_CHECKING, Any, Literal, cast, overload
 
 from openai import NOT_GIVEN, APIStatusError, AsyncOpenAI, AsyncStream, NotGiven
 from openai.types import ChatModel
@@ -247,9 +247,12 @@ async def _fetch_response(
         converted_tools = Converter.convert_tools(tools, handoffs)
         response_format = Converter.get_response_format(output_schema)
 
-        include: list[ResponseIncludable] = converted_tools.includes
+        include_set: set[str] = set(converted_tools.includes)
         if model_settings.response_include is not None:
-            include = list({*include, *model_settings.response_include})
+            include_set.update(model_settings.response_include)
+        if model_settings.top_logprobs is not None:
+            include_set.add("message.output_text.logprobs")
+        include = cast(list[ResponseIncludable], list(include_set))
 
         if _debug.DONT_LOG_MODEL_DATA:
             logger.debug("Calling LLM")
@@ -264,6 +267,10 @@ async def _fetch_response(
                 f"Previous response id: {previous_response_id}\n"
             )
 
+        extra_args = dict(model_settings.extra_args or {})
+        if model_settings.top_logprobs is not None:
+            extra_args["top_logprobs"] = model_settings.top_logprobs
+
         return await self._client.responses.create(
             previous_response_id=self._non_null_or_not_given(previous_response_id),
             instructions=self._non_null_or_not_given(system_instructions),
@@ -286,7 +293,7 @@ async def _fetch_response(
             store=self._non_null_or_not_given(model_settings.store),
             reasoning=self._non_null_or_not_given(model_settings.reasoning),
             metadata=self._non_null_or_not_given(model_settings.metadata),
-            **(model_settings.extra_args or {}),
+            **extra_args,
         )
 
     def _get_client(self) -> AsyncOpenAI:

tests/model_settings/test_serialization.py

@@ -58,6 +58,7 @@ def test_all_fields_serialization() -> None:
         store=False,
         include_usage=False,
         response_include=["reasoning.encrypted_content"],
+        top_logprobs=1,
         extra_query={"foo": "bar"},
         extra_body={"foo": "bar"},
         extra_headers={"foo": "bar"},
@@ -164,6 +165,7 @@ def test_pydantic_serialization() -> None:
         metadata={"foo": "bar"},
         store=False,
         include_usage=False,
+        top_logprobs=1,
         extra_query={"foo": "bar"},
         extra_body={"foo": "bar"},
         extra_headers={"foo": "bar"},

tests/test_logprobs.py

@@ -0,0 +1,50 @@
+import pytest
+from openai.types.responses.response_usage import InputTokensDetails, OutputTokensDetails
+
+from agents import ModelSettings, ModelTracing, OpenAIResponsesModel
+
+
+class DummyResponses:
+    async def create(self, **kwargs):
+        self.kwargs = kwargs
+
+        class DummyResponse:
+            id = "dummy"
+            output = []
+            usage = type(
+                "Usage",
+                (),
+                {
+                    "input_tokens": 0,
+                    "output_tokens": 0,
+                    "total_tokens": 0,
+                    "input_tokens_details": InputTokensDetails(cached_tokens=0),
+                    "output_tokens_details": OutputTokensDetails(reasoning_tokens=0),
+                },
+            )()
+
+        return DummyResponse()
+
+
+class DummyClient:
+    def __init__(self):
+        self.responses = DummyResponses()
+
+
+@pytest.mark.allow_call_model_methods
+@pytest.mark.asyncio
+async def test_top_logprobs_param_passed():
+    client = DummyClient()
+    model = OpenAIResponsesModel(model="gpt-4", openai_client=client)  # type: ignore
+    await model.get_response(
+        system_instructions=None,
+        input="hi",
+        model_settings=ModelSettings(top_logprobs=2),
+        tools=[],
+        output_schema=None,
+        handoffs=[],
+        tracing=ModelTracing.DISABLED,
+        previous_response_id=None,
+    )
+    assert client.responses.kwargs["top_logprobs"] == 2
+    assert "message.output_text.logprobs" in client.responses.kwargs["include"]