From fa8804ad9c3f21dad3143418c7d9f40190f609ae Mon Sep 17 00:00:00 2001
From: Andrew Xia <axia@meta.com>
Date: Mon, 1 Dec 2025 18:11:35 -0800
Subject: [PATCH] [responsesAPI][4] fix responseOutputItem Kimi K2 thinking bug
 (#29555)

Signed-off-by: Andrew Xia <axia@fb.com>
Co-authored-by: Andrew Xia <axia@fb.com>
---
 tests/entrypoints/test_responses_utils.py | 21 +++++++++++++++++++++
 vllm/entrypoints/responses_utils.py       |  7 ++++++-
 2 files changed, 27 insertions(+), 1 deletion(-)

diff --git a/tests/entrypoints/test_responses_utils.py b/tests/entrypoints/test_responses_utils.py
index 893d806b6574..3951bd484008 100644
--- a/tests/entrypoints/test_responses_utils.py
+++ b/tests/entrypoints/test_responses_utils.py
@@ -5,6 +5,8 @@ import pytest
 from openai.types.responses.response_function_tool_call_output_item import (
     ResponseFunctionToolCallOutputItem,
 )
+from openai.types.responses.response_output_message import ResponseOutputMessage
+from openai.types.responses.response_output_text import ResponseOutputText
 from openai.types.responses.response_reasoning_item import (
     Content,
     ResponseReasoningItem,
@@ -101,3 +103,22 @@ class TestResponsesUtils:
         )
         with pytest.raises(ValueError):
             construct_chat_message_with_tool_call(item)
+
+        output_item = ResponseOutputMessage(
+            id="msg_bf585bbbe3d500e0",
+            content=[
+                ResponseOutputText(
+                    annotations=[],
+                    text="dongyi",
+                    type="output_text",
+                    logprobs=None,
+                )
+            ],
+            role="assistant",
+            status="completed",
+            type="message",
+        )
+
+        formatted_item = construct_chat_message_with_tool_call(output_item)
+        assert formatted_item["role"] == "assistant"
+        assert formatted_item["content"] == "dongyi"
diff --git a/vllm/entrypoints/responses_utils.py b/vllm/entrypoints/responses_utils.py
index 07abb80ebc9e..2e01cb038af8 100644
--- a/vllm/entrypoints/responses_utils.py
+++ b/vllm/entrypoints/responses_utils.py
@@ -97,13 +97,18 @@ def construct_chat_message_with_tool_call(
             "role": "assistant",
             "reasoning": reasoning_content,
         }
+    elif isinstance(item, ResponseOutputMessage):
+        return {
+            "role": "assistant",
+            "content": item.content[0].text,
+        }
     elif isinstance(item, ResponseFunctionToolCallOutputItem):
         return ChatCompletionToolMessageParam(
             role="tool",
             content=item.output,
             tool_call_id=item.call_id,
         )
-    elif item.get("type") == "function_call_output":
+    elif isinstance(item, dict) and item.get("type") == "function_call_output":
         # Append the function call output as a tool message.
         return ChatCompletionToolMessageParam(
             role="tool",