[richard-for-bugen]

Summary

Python parity run PR (area fields), based on upstream main at 39dbf51839. Items: 1.

vllm-project#63902 (merged 2026-11-03) made the Rust frontend serve the stateless Responses API, so its request fields are now category 2 surfaces. This run audits every field of Python's ResponsesRequest (vllm/entrypoints/openai/responses/protocol.py) against the Rust DTO (rust/src/server/src/routes/openai/responses/types.rs), which, like the chat DTO, lets serde ignore unknown top-level keys. One field Python honors was silently dropped; it is fixed here as a trivial implementation.

Item key	Category	Python change	Consumer trace	Rust action
responses.thinking_token_budget	2	vllm-project#63896 (5c19391400, 2026-10-29)	ResponsesRequest.to_sampling_params → SamplingParams.thinking_token_budget → engine reasoning budget (validate_thinking_token_budget)	trivial implementation: declared on the DTO and passed to the shared SamplingParams; vllm-text lower_sampling_params already validates and normalizes it for chat (-1 = unlimited, negative → 400)

Ownership and layering. The DTO field belongs to the Responses wire type, and its lowering to prepare_responses_request, which builds the same vllm_text::SamplingParams chat completions builds; validation and normalization stay in the shared vllm-text lowering, so the two routes cannot drift.

Audit of the remaining Python fields

Everything else Python's ResponsesRequest declares is either honored by Rust, rejected explicitly, or not a silent behavior drop:

• Honored: model, input (message, function_call, function_call_output, reasoning items), instructions, tools (function), tool_choice (none/auto/required/function), parallel_tool_calls, max_output_tokens, max_tool_calls (echoed; it only limits built-in tools), metadata, reasoning.effort (including the enable_thinking default), include_reasoning, service_tier (echoed), stream, temperature, top_p, top_k, min_p, text.format, truncation, user, penalties, seed, stop, stop_token_ids, ignore_eos, skip_special_tokens, include_stop_str_in_output, logit_bias, watermarking, request_id, session_id, priority, cache_salt, chat_template_kwargs, structured_outputs, kv_transfer_params, ec_transfer_params, vllm_xargs.
• Accepted no-ops, documented by [Rust Frontend] Add the stateless Responses API (streaming and non-streaming) vllm-project/vllm#63902: reasoning.summary, text.verbosity, include=["reasoning.encrypted_content"].
• Explicit 400 or 404 today:
  • background=true, previous_response_id, item_reference items, previous_input_messages (Harmony), and prompt;
  • other include values, including message.output_text.logprobs, and non-zero top_logprobs;
  • enable_response_messages, spaces_between_special_tokens=false, media_io_kwargs, and mm_processor_kwargs;
  • allowed_tools tool choices, input_file parts, input_image.file_id, and encrypted reasoning items;
  • built-in and namespace tools, and other item types, rejected at deserialization.
• store: not declared. A store=true request runs unstored, which is Python's implicit downgrade when VLLM_ENABLE_RESPONSES_API_STORE is unset. With the variable set, Python keeps a store; that case gets an env placeholder in the env run PR ([Rust Frontend] Python parity run 2026-10-14 (env) #10).
• Echo-only fields (prompt_cache_key, prompt_cache_options, prompt_cache_retention, safety_identifier): Python accepts them, does not act on them, and echoes them on the response object, where its ResponsesResponse also carries input_messages/output_messages as null. Rust drops the request fields and its response object lacks those keys. That is response-shape drift (category 4), queued as its own item rather than placed here.
• Render-only (mm_payload, return_mm_kwargs): ignored on /v1/responses in Python too.

Nested shapes match: function tools reuse the chat Function type (name, description, parameters, strict, defer_loading), and the input item and content part types are closed enums, so an unknown variant fails deserialization rather than being dropped.

Not in this PR (queued separately)

• Output logprobs (include=["message.output_text.logprobs"], top_logprobs): explicit 400; Python tests responses/test_basic.py::test_logprobs, responses/test_simple.py::test_streaming_logprobs.
• enable_response_messages: explicit 400; responses/test_simple.py::test_enable_response_messages.
• tools[].type="namespace" ([Frontend] Support OpenAI Responses API namespace tools vllm-project/vllm#47024, [Bugfix] Handle malformed namespace tools vllm-project/vllm#53763, [Bugfix][Frontend] Fix Responses namespace aliases in named tool calling vllm-project/vllm#59723): rejected by serde with a generic message; responses/test_namespace_tool_separator.py.
• The echo-only response fields above.

Needs Bugen

None.

Validation

New route test responses_thinking_token_budget_reaches_engine_core checks the engine request for absent, 256 and -1 (→ no budget). It fails with the field dropped (left: None, right: Some(256)).

Rust CI set from .buildkite/scripts/run-rust-frontend-cargo-ci.sh, on the tip:

cargo fmt --manifest-path rust/Cargo.toml --all -- --check
cargo sort --workspace --check rust
cargo deny --manifest-path rust/Cargo.toml --config rust/deny.toml check bans
cargo clippy --manifest-path rust/Cargo.toml --workspace --all-targets --all-features --locked -- -D warnings
cargo nextest run --manifest-path rust/Cargo.toml --workspace --all-features --locked --no-fail-fast

All green on the tip 4cc1c46a87: fmt, sort, deny (bans ok) and workspace clippy clean; nextest 2511 passed, 13 skipped.

The e2e run PR (#47) keeps test_thinking_token_budget.py::test_thinking_token_budget_responses_api skipped under the Rust frontend until this merges; it then gets enabled there.

Codex review (gpt-6.1-sol, high): approve with nits. It traced the field through prepare_responses_request → ChatRequestProcessor → lower_sampling_params → EngineCoreSamplingParams and confirmed the Python target is deliberate (vllm-project#63896 added Responses support, docs and a chat/Responses comparison test). One P3, fixed: the regression assertion ran inside the mock engine, so a mismatch stalled until nextest's timeout; the test now captures the budget through a channel and asserts after the response drains (red run fails in 0.07 s). Test-only follow-up, so no re-review.

Placeholders turn previously ignored inputs into explicit errors. This run has none: its one item makes a previously ignored field work.

[BugenZhao]

The e2e run PR (#47) keeps test_thinking_token_budget.py::test_thinking_token_budget_responses_api skipped under the Rust frontend until this merges; it then gets enabled there.

Let's fold this into e2e run PR

[richard-for-bugen]

Folded into #47 as e39538d (this PR's commit, cherry-picked), with f577830 dropping the Rust skip on test_thinking_token_budget_responses_api, which now passes under the Rust frontend.