[Sunt-ing]

Purpose

OpenAI's Responses streaming reference defines response.completed as the event emitted when the model response is complete, and response.incomplete as the event emitted when a response finishes as incomplete. vLLM already builds the final Responses payload with response.status == "incomplete" and incomplete_details.reason == "max_output_tokens" when generation stops on the output-token limit, but the streaming path still wraps that payload in a response.completed terminal event.

This PR makes the terminal event match the final response status: incomplete emits response.incomplete, and other successful responses keep emitting response.completed. It also adds the vLLM ResponseIncompleteEvent wrapper, updates the streaming event-stack helper so both completed and incomplete are valid terminal response events, and treats response.incomplete as terminal when replaying stored background stream events.

Scope: this only handles max-token truncation after the final payload is already status="incomplete". It does not change generation-error streaming; #40418 and #38190 cover response.failed paths.

Test Plan

• Run a real Responses streaming truncation E2E with Qwen3-8B and max_output_tokens=30.
• Run the same truncation path with background=True, then replay it with retrieve(..., stream=True).
• Revert only the source change while keeping the new E2E assertion.
• Run the full Responses simple E2E file.
• Run related Responses protocol, serving, and error tests.
• Run changed-file pre-commit.

Test Result

Check	Result
Real truncation E2E with fix	Terminal event is response.incomplete; payload status is incomplete; reason is max_output_tokens
Background truncation replay E2E with fix	Direct stream and stored replay both end with response.incomplete
Same E2E after source-only revert	Failed with AssertionError: assert 'response.completed' == 'response.incomplete'
Full tests/entrypoints/openai/responses/test_simple.py	12 passed
Related Responses protocol, serving, and error tests	26 passed
Changed-file pre-commit	Passed

Real server E2E details

Environment:

• GPU: NVIDIA A800-SXM4-80GB, single GPU
• Model: Qwen/Qwen3-8B
• Server harness: RemoteOpenAIServer
• Server args: --reasoning-parser qwen3 --max_model_len 5000
• Server env: VLLM_SYSTEM_START_DATE=2023-09-12, VLLM_ENABLE_RESPONSES_API_STORE=1
• Hardware workaround for this A800 setup: VLLM_USE_FLASHINFER_SAMPLER=0

The committed E2E starts a real OpenAI-compatible server and sends a streaming Responses request that is intentionally truncated:

stream = await client.responses.create(
    model="Qwen/Qwen3-8B",
    input="What is the first paragraph of Moby Dick?",
    reasoning={"effort": "low"},
    max_output_tokens=30,
    temperature=0.0,
    stream=True,
)
events = [event async for event in stream]

terminal_event = events[-1]
assert terminal_event.type == "response.incomplete"
assert terminal_event.response.status == "incomplete"
assert terminal_event.response.incomplete_details.reason == "max_output_tokens"

The background replay E2E uses the same truncation trigger with background=True, then replays the stored stream:

stream = await client.responses.create(
    model="Qwen/Qwen3-8B",
    input="What is the first paragraph of Moby Dick?",
    reasoning={"effort": "low"},
    max_output_tokens=30,
    temperature=0.0,
    stream=True,
    background=True,
)
events = [event async for event in stream]

async with await client.responses.retrieve(
    response_id=events[0].response.id,
    stream=True,
) as replay_stream:
    replay_events = [event async for event in replay_stream]

assert replay_events == events
assert replay_events[-1].type == "response.incomplete"

With the fix:

tests/entrypoints/openai/responses/test_simple.py::test_streaming_max_tokens_terminal_event
tests/entrypoints/openai/responses/test_simple.py::test_background_streaming_max_tokens_terminal_event
2 passed, 16 warnings in 73.61s

After reverting only the source change and keeping the same test:

AssertionError: assert 'response.completed' == 'response.incomplete'

Additional checks

Commands:

PYTHONPATH=$PWD VLLM_USE_FLASHINFER_SAMPLER=0 python -m pytest tests/entrypoints/openai/responses/test_simple.py::test_streaming_max_tokens_terminal_event tests/entrypoints/openai/responses/test_simple.py::test_background_streaming_max_tokens_terminal_event -q --tb=short --disable-warnings
PYTHONPATH=$PWD VLLM_USE_FLASHINFER_SAMPLER=0 python -m pytest tests/entrypoints/openai/responses/test_simple.py -q --tb=short --disable-warnings
CUDA_VISIBLE_DEVICES="" PYTHONPATH=$PWD python -m pytest tests/entrypoints/openai/responses/test_protocol.py tests/entrypoints/openai/responses/test_serving_responses.py tests/entrypoints/openai/responses/test_errors.py -q --tb=short --disable-warnings
pre-commit run --files tests/entrypoints/openai/responses/conftest.py tests/entrypoints/openai/responses/test_simple.py vllm/entrypoints/openai/responses/protocol.py vllm/entrypoints/openai/responses/serving.py

Results:

tests/entrypoints/openai/responses/test_simple.py
12 passed, 16 warnings in 169.64s

tests/entrypoints/openai/responses/test_protocol.py
tests/entrypoints/openai/responses/test_serving_responses.py
tests/entrypoints/openai/responses/test_errors.py
26 passed, 16 warnings in 4.27s

pre-commit run --files <changed files>
Passed

AI assistance was used to prepare this PR.

cc @sfeng33 @chaunceyjiang

[Sunt-ing]

Closing in favor of #59324, which covers this fix, including the response.incomplete terminal event and background replay handling.