[responsesAPI] fix incomplete_messages for simple/parsable context (#31836)

Signed-off-by: Andrew Xia <axia@fb.com>
Co-authored-by: Andrew Xia <axia@fb.com>

tests/entrypoints/openai/test_response_api_parsable_context.py

@@ -58,6 +58,7 @@ async def test_basic(client: OpenAI, model_name: str):
     assert response is not None
     print("response: ", response)
     assert response.status == "completed"
+    assert response.incomplete_details is None
 
 
 @pytest.mark.asyncio
@@ -184,3 +185,17 @@ async def test_mcp_tool_call(client: OpenAI, model_name: str):
     assert len(response.input_messages) == 1
     assert len(response.output_messages) == 3
     assert "312" in response.output_messages[2]["message"]
+
+
+@pytest.mark.asyncio
+@pytest.mark.parametrize("model_name", [MODEL_NAME])
+async def test_max_tokens(client: OpenAI, model_name: str):
+    response = await client.responses.create(
+        model=model_name,
+        input="What is the first paragraph of Moby Dick?",
+        reasoning={"effort": "low"},
+        max_output_tokens=30,
+    )
+    assert response is not None
+    assert response.status == "incomplete"
+    assert response.incomplete_details.reason == "max_output_tokens"

tests/entrypoints/openai/test_response_api_simple.py

@@ -40,6 +40,7 @@ async def test_basic(client: OpenAI, model_name: str):
     assert response is not None
     print("response: ", response)
     assert response.status == "completed"
+    assert response.incomplete_details is None
 
 
 @pytest.mark.asyncio
@@ -132,3 +133,17 @@ async def test_streaming_output_consistency(client: OpenAI, model_name: str):
         f"Streaming: {streaming_text!r}\n"
         f"Final: {final_output_text!r}"
     )
+
+
+@pytest.mark.asyncio
+@pytest.mark.parametrize("model_name", [MODEL_NAME])
+async def test_max_tokens(client: OpenAI, model_name: str):
+    response = await client.responses.create(
+        model=model_name,
+        input="What is the first paragraph of Moby Dick?",
+        reasoning={"effort": "low"},
+        max_output_tokens=30,
+    )
+    assert response is not None
+    assert response.status == "incomplete"
+    assert response.incomplete_details.reason == "max_output_tokens"

vllm/entrypoints/openai/parser/responses_parser.py

@@ -51,7 +51,13 @@ class ResponsesParser:
         if tool_parser_cls is not None:
             self.tool_parser_instance = tool_parser_cls(tokenizer)
 
+        # Store the last finish_reason to determine response status
+        self.finish_reason: str | None = None
+
     def process(self, output: CompletionOutput) -> "ResponsesParser":
+        # Store the finish_reason from the output
+        self.finish_reason = output.finish_reason
+
         reasoning_content, content = self.reasoning_parser_instance.extract_reasoning(
             output.text, request=self.request
         )

vllm/entrypoints/openai/serving_responses.py

@@ -692,6 +692,10 @@ class OpenAIServingResponses(OpenAIServing):
             # TODO: Calculate usage.
             # assert final_res.prompt_token_ids is not None
             num_tool_output_tokens = 0
+
+            # Check finish reason from the parser
+            if context.parser.finish_reason == "length":
+                status = "incomplete"
         else:
             assert isinstance(context, SimpleContext)
             # Use final_output which has accumulated text/token_ids/logprobs
@@ -703,6 +707,10 @@ class OpenAIServingResponses(OpenAIServing):
             # finish_reason='error' indicates retryable internal error
             self._raise_if_error(final_output.finish_reason, request.request_id)
 
+            # Check if generation was stopped due to max_tokens
+            if final_output.finish_reason == "length":
+                status = "incomplete"
+
             output = self._make_response_output_items(request, final_output, tokenizer)
 
             if request.enable_response_messages: