[heyemlee]

Overview

Fixes #47200. /v1/responses no longer returns null for output item and tool fields that nothing set, so clients that validate the response strictly can parse it.

Claims

• Output items no longer carry unset fields such as namespace, caller, async, phase, or a reasoning item's status and encrypted_content, in the JSON response or in SSE events.
• Echoed tools contain what the client sent, plus parameters and strict, which the schema requires but allows to be null. Unset defer_loading, allowed_callers, async and output_schema are gone.
• Values that were set are unchanged, including explicit nulls from the client and nulls inside a tool's JSON schema.

Validation

• Added a regression test through RemoteOpenAIServer, streaming and non-streaming, that reads the raw wire format because the SDK hides these nulls. It fails on main and passes with this fix, as does a new unit test.
• Compared every JSON path from a live server before and after, across plain, named-tool and auto-tool requests, streaming and non-streaming. No path that ever held a value was removed, and none was added.
• Real-server suites with Qwen/Qwen3-1.7B: test_function_call.py and test_namespace_tool_separator.py, 18 passed, including namespaced tools keeping namespace.
• Responses unit suites: 298 passed, 1 xfailed on openai 3.42.0. On openai 2.44.0, which CI pins, the regression tests pass plus test_responses_utils.py, test_streaming_events.py and test_serving_responses.py: 155 passed, 1 xfailed. pre-commit passed.
• Local testing was on macOS CPU. CI covers GPU, the Harmony path end to end (test_harmony.py) and test_simple.py, which needs Qwen/Qwen3-8B. I read SMG's parser but did not run SMG against this change.
• No model eval: this only changes how responses are serialized, not generated tokens.

Commands and output

Regression tests:

PATH="$PWD/.venv/bin:$PATH" VLLM_TARGET_DEVICE=cpu VLLM_CPU_KVCACHE_SPACE=2 python -m pytest \
  tests/entrypoints/openai/responses/test_function_call.py::test_wire_format_omits_unset_item_and_tool_fields \
  tests/entrypoints/openai/responses/test_protocol.py::test_response_json_omits_unset_item_and_tool_fields -v

main (openai 2.44.0):

E   assert not ({'async', 'caller', 'namespace'} & dict_keys(['arguments', 'call_id', 'name', 'type', 'id', 'namespace', 'status']))
FAILED test_function_call.py::test_wire_format_omits_unset_item_and_tool_fields[False]
FAILED test_function_call.py::test_wire_format_omits_unset_item_and_tool_fields[True]
FAILED test_protocol.py::test_response_json_omits_unset_item_and_tool_fields
3 failed in 44.84s

This PR (openai 3.42.0; on 2.44.0 these pass within the 155 above):

test_function_call.py::test_wire_format_omits_unset_item_and_tool_fields[False] PASSED
test_function_call.py::test_wire_format_omits_unset_item_and_tool_fields[True] PASSED
test_protocol.py::test_response_json_omits_unset_item_and_tool_fields PASSED
3 passed in 27.68s

Live server: vllm serve Qwen/Qwen3-0.6B --enable-auto-tool-choice --tool-call-parser hermes --reasoning-parser qwen3 --enforce-eager, openai 3.42.0, keys of a non-streaming response:

	main	This PR
function_call item	arguments, async, call_id, caller, id, name, namespace, status, type	arguments, call_id, id, name, status, type
reasoning item	content, encrypted_content, id, status, summary, type	content, id, summary, type
message item	content, id, phase, role, status, type	content, id, role, status, type
Function tool keys SMG rejects	allowed_callers, async, defer_loading, output_schema	none

On openai 2.44.0, FunctionTool only adds defer_loading of these four.

Details

The SDK types default every optional field to None, and the response was dumped with those defaults. The report in #47200 has two parts on current main:

• The status: null on function_call items that caused invalid type: null, expected a string was fixed by [Frontend] Make finish_reason and Responses item status truthful #59324.
• Echoed tools can still fail. SMG's FunctionTool rejects any key outside name, description, parameters and strict, so defer_loading: null should fail with an unknown-field error. New SDK versions add more such fields.

This uses exclude_unset rather than exclude_none because some nulls are meaningful: ResponseCodeInterpreterToolCall.outputs is set to None on purpose, and a client's explicit null should round-trip. So call sites that passed namespace=None or status=None now leave those fields unset. ResponsesResponse keeps its defaults, such as object, because only its output and tools are filtered, and model_dump() in Python mode is unchanged.

Not changed, to keep this small:

• Top-level user, text and usage can still be null. The OpenAI spec's own response examples return "user": null and "usage": null, and SMG accepts all three.
• content[*].logprobs can still be null, because call sites set it explicitly and SMG accepts it. Sending [] as streaming does at the start of a part would change a value, so it should be a separate change.
• vLLM-only fields such as kv_transfer_params stay as they are, since SDK users may read them as attributes.

Tradeoff: a non-SDK client that reads item["namespace"] without a default would now see a missing key. The schema marks these fields optional, and SDK clients still get None. The Rust frontend has no Responses API on main yet, so nothing needs to change there.

Not a duplicate: no open PR references #47200 or changes this serialization. This replaces #47184, which the stale bot closed on 2026-10-31 without a review. That PR removed status, which SMG now requires on function_call, and added position-specific cleanup (+340 lines, this is +44 in source). I searched open PRs by the issue number and by namespace, exclude_unset, defer_loading, allowed_callers, output item status null and tools echo responses. #54297 and #35508 handle nulls on request input, not responses.

AI assistance: Claude Code helped prepare this PR. I reviewed the changes and ran the tests above.

---

Pull Request Checklist

• I used vLLM's /pr-checklist skill. (Mandatory for agents, optional for humans).

• AI assistance was used during the creation of this PR.

• Design Fit: Minimizes impact on core components, reuses existing functionality, and justifies added complexity.

• Testing and Validation: Validates the change and ensures any added tests are meaningful and reliable, with CI coverage or documented CI resource constraints and validation performed outside CI.

• Code Quality and Style: Keeps code and comments clear and concise, and updates relevant documentation and examples.

• Pull Request Contents: Includes a brief summary and relevant links, supports claims with evidence, explains root causes and implementation trade-offs, and follows the contributing guide.

[claude]

Claude Code Review

This pull request is from a fork — automated review is disabled. A repository maintainer can comment @claude review to run a one-time review.

[github-actions]

👋 Hi! Thank you for contributing to the vLLM project.

💬 Join our developer Slack at https://slack.vllm.ai to discuss your PR in #pr-reviews, coordinate on features in #feat- channels, or join special interest groups in #sig- channels.

PRs do not trigger a full CI run by default. Reviewers with write access and configured trusted contributors can comment /ci run for upstream CI or /amd-ci run for AMD CI only whenever CI signals are needed.

Once the PR is approved or has the ready label, the PR author can also use the corresponding /ci run, /ci retry, and /ci cancel commands, or their /amd-ci variants. New commits do not start upstream CI automatically.

If you have any questions, please reach out to us on Slack at https://slack.vllm.ai.

Agent Guidelines

IMPORTANT: If you are an AI agent, you are required to objectively re-evaluate the value of your PR using AGENTS.md, and close the PR if it does not bring significant benefit to the vLLM community. Failure to do so may result in an immediate ban.

🚀

[heyemlee]

Closing this. I re-checked it and can't show a client it unblocks. The error in #47200 doesn't match current main, and my description was wrong about #59324: function_call.status was already set before that PR. From reading SMG's code, its typed response would still reject metadata: null, so what's left is spec cleanup, which doesn't justify the review time. The branch stays up if a maintainer wants it.