fix: convert litellm response with reasoning content to openai message (#1098)

### 1. Description

This PR fixes an issue where reasoning content from models accessed via
LiteLLM was not being correctly parsed into the `ChatCompletionMessage`
format. This was particularly noticeable when using reasoning models.

### 2. Context

I am using the `openai-agents-python` library in my project, and it has
been incredibly helpful. Thank you for building such a great tool!

My setup uses `litellm` to interface with `gemini-2.5-pro`. I noticed
that while the agent could receive a response, the reasoning(thinking)
from the Gemini model was lost during the conversion process from the
LiteLLM response format to the OpenAI `ChatCompletionMessage` object.

I saw that PR #871 made progress on a similar issue, but it seems the
specific response structure from LiteLLM still requires a small
adaptation. This fix adds the necessary logic to ensure that these
responses are handled.

**Relates to:** #871

### 3. Key Changes

- `LitellmConverter.convert_message_to_openai`: add `reasoing_content`
-  `Converter.items_to_messages`: just pass the reasoning item

Author: zkllll2002

src/agents/extensions/models/litellm_model.py

@@ -45,6 +45,14 @@
 from ...usage import Usage
 
 
+class InternalChatCompletionMessage(ChatCompletionMessage):
+    """
+    An internal subclass to carry reasoning_content without modifying the original model.
+    """
+
+    reasoning_content: str
+
+
 class LitellmModel(Model):
     """This class enables using any model via LiteLLM. LiteLLM allows you to acess OpenAPI,
     Anthropic, Gemini, Mistral, and many other models.
@@ -364,13 +372,18 @@ def convert_message_to_openai(
             provider_specific_fields.get("refusal", None) if provider_specific_fields else None
         )
 
-        return ChatCompletionMessage(
+        reasoning_content = ""
+        if hasattr(message, "reasoning_content") and message.reasoning_content:
+            reasoning_content = message.reasoning_content
+
+        return InternalChatCompletionMessage(
             content=message.content,
             refusal=refusal,
             role="assistant",
             annotations=cls.convert_annotations_to_openai(message),
             audio=message.get("audio", None),  # litellm deletes audio if not present
             tool_calls=tool_calls,
+            reasoning_content=reasoning_content,
         )
 
     @classmethod

src/agents/models/chatcmpl_converter.py

@@ -36,6 +36,7 @@
     ResponseOutputRefusal,
     ResponseOutputText,
     ResponseReasoningItem,
+    ResponseReasoningItemParam,
 )
 from openai.types.responses.response_input_param import FunctionCallOutput, ItemReference, Message
 from openai.types.responses.response_reasoning_item import Summary
@@ -210,6 +211,12 @@ def maybe_response_output_message(cls, item: Any) -> ResponseOutputMessageParam
             return cast(ResponseOutputMessageParam, item)
         return None
 
+    @classmethod
+    def maybe_reasoning_message(cls, item: Any) -> ResponseReasoningItemParam | None:
+        if isinstance(item, dict) and item.get("type") == "reasoning":
+            return cast(ResponseReasoningItemParam, item)
+        return None
+
     @classmethod
     def extract_text_content(
         cls, content: str | Iterable[ResponseInputContentParam]
@@ -459,7 +466,11 @@ def ensure_assistant_message() -> ChatCompletionAssistantMessageParam:
                     f"Encountered an item_reference, which is not supported: {item_ref}"
                 )
 
-            # 7) If we haven't recognized it => fail or ignore
+            # 7) reasoning message => not handled
+            elif cls.maybe_reasoning_message(item):
+                pass
+
+            # 8) If we haven't recognized it => fail or ignore
             else:
                 raise UserError(f"Unhandled item type or structure: {item}")