Pull requests / #571

#571 Feat/responses api 451

closed · draft · @CC-David-CC · 0 评论 · 在 GitHub 查看

Server & APINVIDIA / CUDAModels & quantsSecurityDocumentation

描述

Adds the standard `/v1/responses` interface for #451, using the existing Strata service directly. Enable it once at startup with `--experimental-responses` or top-level `"experimental_responses": true`; it is disabled by default.

A small controller owns response records, typed output items and lifecycle transitions. JSON and SSE are views of the same execution, retaining the existing engine serialization and cancellation/draining. The monitor observes those transitions.

- Create, retrieve, paginate input items and delete; retained context supports independent `previous_response_id` continuations.
- Client-owned function-call/result loops, stable item/call IDs, and typed text/argument deltas.
- Responses `text.format` validation, existing authentication/CORS, atomic retained records, deletion coordination and restart recovery.

Unsupported parameters fail explicitly. Background execution/cancel endpoint, image input, reasoning representations, hosted tools, compaction and WebSockets are outside this patch. Strict schemas require `jsonschema`; completion is validated, without claiming grammar-constrained decoding.

## Validation

Merged upstream 0.1.38 (`99f3dbd`), preserving the Host checks on all methods and the browser-Origin check before Responses POST dispatch. Two new HTTP regressions cover those protections on Responses routes.

All eight serving/security modules passed: **231 passed, 5 tokenizer-fixture skips**, including all 41 Responses, 118 server, and 21 security tests. The OpenAI SDK contract check passed.

[Upstream sync and test outcomes](https://github.com/CC-David-CC/Strata-a5500/blob/feat/responses-api-451/docs/benchmarks/2026-10-03-responses-upstream-0138.json)

The published native CUDA check passed all 10 integration groups on an RTX 4090 with Coder IQ1_M, 4,096-token context and the original 0.1.34 base. That result is not a native-GPU test of 0.1.38.

[Enable and use Responses](https://github.com/CC-David-CC/Strata-a5500/blob/feat/responses-api-451/docs/RESPONSES.md) ? [Native evidence and source hashes](https://github.com/CC-David-CC/Strata-a5500/blob/feat/responses-api-451/docs/benchmarks/2026-10-02-responses-cuda.json)

站内延伸阅读

链到安装、模型与版本说明,便于 SEO/GEO,非官方 issue 正文。