From 2a675a7b9fce6577e8eab937f67c7c1973a54f7f Mon Sep 17 00:00:00 2001 From: yzong-rh Date: Tue, 23 Jun 2026 08:54:46 -0400 Subject: [PATCH] [Bugfix] Responses API assistant EasyInputMessageParam input (#44361) Signed-off-by: Yifan Zong Co-authored-by: Ben Browning --- .../responses/test_function_call_parsing.py | 51 ++++++++++++++++++- .../openai/responses/test_responses_utils.py | 10 ++++ vllm/entrypoints/openai/responses/protocol.py | 31 ++++++----- vllm/entrypoints/openai/responses/utils.py | 14 +++++ 4 files changed, 93 insertions(+), 13 deletions(-) diff --git a/tests/entrypoints/openai/responses/test_function_call_parsing.py b/tests/entrypoints/openai/responses/test_function_call_parsing.py index 8b4d7c7397..f90a641db7 100644 --- a/tests/entrypoints/openai/responses/test_function_call_parsing.py +++ b/tests/entrypoints/openai/responses/test_function_call_parsing.py @@ -5,7 +5,7 @@ import json import pytest -from openai.types.responses import ResponseFunctionToolCall +from openai.types.responses import ResponseFunctionToolCall, ResponseOutputMessage from vllm.entrypoints.openai.responses.protocol import ResponsesRequest @@ -328,3 +328,52 @@ def test_validator_handles_empty_iterator(): request = ResponsesRequest(**mock_data) assert request.input == [] + + +def test_assistant_string_content_stays_easyinput(): + """EasyInput assistant message with plain string content is not + coerced into a ResponseOutputMessage.""" + request_data = { + "model": "test-model", + "input": [ + {"type": "message", "role": "assistant", "content": "hello"}, + ], + } + + request = ResponsesRequest(**request_data) + + item = request.input[0] + assert isinstance(item, dict), ( + "String-content assistant message should remain a dict (EasyInput), " + f"got {type(item)}" + ) + assert item.get("content") == "hello" + assert "id" not in item + assert "status" not in item + + +def test_assistant_output_style_content_coerced(): + """Assistant message whose content is output-message-shaped (list of + output_text items) should be coerced to ResponseOutputMessage.""" + request_data = { + "model": "test-model", + "input": [ + { + "type": "message", + "role": "assistant", + "content": [{"type": "output_text", "text": "world"}], + }, + ], + } + + request = ResponsesRequest(**request_data) + + item = request.input[0] + assert isinstance(item, ResponseOutputMessage), ( + "Output-style assistant message should be coerced to " + f"ResponseOutputMessage, got {type(item)}" + ) + assert item.content[0].text == "world" + assert item.content[0].annotations == [] + assert item.status == "completed" + assert item.id.startswith("msg_") diff --git a/tests/entrypoints/openai/responses/test_responses_utils.py b/tests/entrypoints/openai/responses/test_responses_utils.py index c9ba52b143..efbfb5c07e 100644 --- a/tests/entrypoints/openai/responses/test_responses_utils.py +++ b/tests/entrypoints/openai/responses/test_responses_utils.py @@ -782,6 +782,16 @@ class TestConstructChatMessagesCombinePolicy: ["call_123", "call_456"], id="reasoning-output-tool-call", ), + pytest.param( + [ + make_reasoning_item(content_text="Let me think"), + {"type": "message", "role": "assistant", "content": "Hello"}, + ], + "Hello", + "Let me think", + None, + id="reasoning-easyinput-assistant", + ), ], ) def test_assistant_side_items_merge_until_tool_output( diff --git a/vllm/entrypoints/openai/responses/protocol.py b/vllm/entrypoints/openai/responses/protocol.py index 30a9206636..eb2d66bdd8 100644 --- a/vllm/entrypoints/openai/responses/protocol.py +++ b/vllm/entrypoints/openai/responses/protocol.py @@ -525,23 +525,30 @@ class ResponsesRequest(OpenAIBaseModel): processed_input.append(item) elif item_type == "message" and item.get("role") == "assistant": + content = item.get("content") + if not isinstance(content, list): + # String content is a valid EasyInputMessageParam, + # do not coerce it to ResponseOutputMessage + processed_input.append(item) + continue + + original_item = item item = dict(item) if "id" not in item: item["id"] = f"msg_{random_uuid()}" if "status" not in item: item["status"] = "completed" # ResponseOutputText requires annotations - if isinstance(item.get("content"), list): - new_content = [] - for c in item["content"]: - if ( - isinstance(c, dict) - and c.get("type") == "output_text" - and "annotations" not in c - ): - c = {**c, "annotations": []} - new_content.append(c) - item["content"] = new_content + new_content = [] + for c in content: + if ( + isinstance(c, dict) + and c.get("type") == "output_text" + and "annotations" not in c + ): + c = {**c, "annotations": []} + new_content.append(c) + item["content"] = new_content try: processed_input.append(ResponseOutputMessage(**item)) except ValidationError: @@ -549,7 +556,7 @@ class ResponsesRequest(OpenAIBaseModel): "Failed to parse assistant message to ResponseOutputMessage, " "leaving for Pydantic validation" ) - processed_input.append(item) + processed_input.append(original_item) else: processed_input.append(item) diff --git a/vllm/entrypoints/openai/responses/utils.py b/vllm/entrypoints/openai/responses/utils.py index 81f60b0663..15b6fa88ab 100644 --- a/vllm/entrypoints/openai/responses/utils.py +++ b/vllm/entrypoints/openai/responses/utils.py @@ -306,6 +306,20 @@ def _construct_message_from_response_item( content=item.get("output"), tool_call_id=item.get("call_id"), ) + elif isinstance(item, dict) and item.get("role") == "assistant": + content = item.get("content") + text: str | None = None + if isinstance(content, str): + text = content + elif isinstance(content, list) and content: + text = content[0].get("text") + if text is not None: + if prev_assistant_msg: + previous_content = prev_assistant_msg.get("content") + if previous_content is None: + prev_assistant_msg["content"] = text + return None + return {"role": "assistant", "content": text} return item # type: ignore[arg-type]