Fix responses API (#8880)

* fix responses API

* comments

Author: chenmoneygithub

dspy/clients/base_lm.py

@@ -217,26 +217,43 @@ def _extract_citations_from_response(self, choice):
 
     def _process_response(self, response):
         """Process the response of OpenAI Response API and extract outputs.
-        
+
         Args:
             response: OpenAI Response API response
                 https://platform.openai.com/docs/api-reference/responses/object
-            
+
         Returns:
-            List of processed outputs
+            List of processed outputs, which is always of size 1 because the Response API only supports one output.
         """
-        outputs = []
+        text_outputs = []
         tool_calls = []
+        reasoning_contents = []
+
         for output_item in response.output:
-            if output_item.type == "message":
+            output_item_type = output_item.type
+            if output_item_type == "message":
                 for content_item in output_item.content:
-                    outputs.append(content_item.text)
-            elif output_item.type == "function_call":
+                    text_outputs.append(content_item.text)
+            elif output_item_type == "function_call":
                 tool_calls.append(output_item.model_dump())
+            elif output_item_type == "reasoning":
+                if getattr(output_item, "content", None) and len(output_item.content) > 0:
+                    for content_item in output_item.content:
+                        reasoning_contents.append(content_item.text)
+                elif getattr(output_item, "summary", None) and len(output_item.summary) > 0:
+                    for summary_item in output_item.summary:
+                        reasoning_contents.append(summary_item.text)
+
+        result = {}
+        if len(text_outputs) > 0:
+            result["text"] = "".join(text_outputs)
+        if len(tool_calls) > 0:
+            result["tool_calls"] = tool_calls
+        if len(reasoning_contents) > 0:
+            result["reasoning_content"] = "".join(reasoning_contents)
+        # All `response.output` items map to one answer, so we return a list of size 1.
+        return [result]
 
-        if tool_calls:
-            outputs.append({"tool_calls": tool_calls})
-        return outputs
 
 
 def inspect_history(n: int = 1):

tests/clients/test_lm.py

@@ -10,6 +10,8 @@
 from litellm.types.llms.openai import ResponseAPIUsage, ResponsesAPIResponse
 from litellm.utils import Choices, Message, ModelResponse
 from openai import RateLimitError
+from openai.types.responses import ResponseOutputMessage, ResponseReasoningItem
+from openai.types.responses.response_reasoning_item import Summary
 
 import dspy
 from dspy.utils.dummies import DummyLM
@@ -505,36 +507,49 @@ def test_disable_history():
                 model="openai/gpt-4o-mini",
             )
 
-def test_responses_api(litellm_test_server):
-    api_base, _ = litellm_test_server
-    expected_text = "This is a test answer from responses API."
-
+def test_responses_api():
     api_response = make_response(
         output_blocks=[
-            {
-                "id": "msg_1",
-                "type": "message",
-                "role": "assistant",
-                "status": "completed",
-                "content": [
-                    {"type": "output_text", "text": expected_text, "annotations": []}
-                ],
-            }
+            ResponseOutputMessage(
+                **{
+                    "id": "msg_1",
+                    "type": "message",
+                    "role": "assistant",
+                    "status": "completed",
+                    "content": [
+                        {"type": "output_text", "text": "This is a test answer from responses API.", "annotations": []}
+                    ],
+                },
+            ),
+            ResponseReasoningItem(
+                **{
+                    "id": "reasoning_1",
+                    "type": "reasoning",
+                    "summary": [Summary(**{"type": "summary_text", "text": "This is a dummy reasoning."})],
+                },
+            ),
         ]
     )
 
     with mock.patch("litellm.responses", autospec=True, return_value=api_response) as dspy_responses:
         lm = dspy.LM(
-            model="openai/dspy-test-model",
-            api_base=api_base,
-            api_key="fakekey",
+            model="openai/gpt-5-mini",
             model_type="responses",
             cache=False,
+            temperature=1.0,
+            max_tokens=16000,
         )
-        assert lm("openai query") == [expected_text]
+        lm_result = lm("openai query")
+
+        assert lm_result == [
+            {
+                "text": "This is a test answer from responses API.",
+                "reasoning_content": "This is a dummy reasoning.",
+            }
+        ]
 
         dspy_responses.assert_called_once()
-        assert dspy_responses.call_args.kwargs["model"] == "openai/dspy-test-model"
+        assert dspy_responses.call_args.kwargs["model"] == "openai/gpt-5-mini"
 
 
 def test_lm_replaces_system_with_developer_role():