Skip to content

fix(gateway): forward Responses requests intact and replay streams after release #2399

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
hidb4ai merged 3 commits into from
Nov 2, 2026
Merged

fix(gateway): forward Responses requests intact and replay streams after release #2399

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
3 commits
Select commit Hold shift + click to select a range
ea7c4d0
fix(gateway): forward Responses requests intact and replay streams af…
jasper-yl Nov 2, 2026
bbe6548
fix(gateway): replay Responses text events, failure status and degrad…
jasper-yl Nov 2, 2026
83a1557
fix(gateway): preserve logprobs in Responses replay
hidb4ai Nov 2, 2026
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
1 change: 1 addition & 0 deletions gateway/README.md
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -25,6 +25,7 @@ An optional Rust executor for PowerContext, defined by the gateway design in #21
- **Two deployment forms**: *combined* (Rust owns the public TCP port; Core listens only on a private Unix socket; all non-`/gateway` routes are reverse-proxied to Core over UDS) or *standalone* (TCP to a remote Core).
- **UDS client**: hand-rolled HTTP/1.1 over the Unix socket with a pooled, keep-alive connection pool (bounded idle set, connect/read timeouts, stale-connection single retry, content-length truncation and chunked-trailer consumption — covered by a real-socket regression test).
- **OpenAI-compatible ingress** at `/gateway/v1/chat/completions` and `/gateway/v1/responses` (same format in and out, no cross-format translation), authenticated by a gateway key that replaces provider keys.
- **Responses fidelity**: on `/gateway/v1/responses`, every input item other than a plain user message (developer messages, tool calls and outputs, reasoning, client-specific items such as Codex's `additional_tools`) is forwarded unchanged, and request fields the ingress does not model (for example `reasoning`, `include`, `prompt_cache_key`) are passed through for Core's request allowlist to decide; only the client-local `client_metadata` is dropped. The upstream request is always non-streaming; a streaming client receives the released (or permitted degraded-passthrough, still marked `x-pc-degraded`) 200 response replayed as Responses SSE events: `response.created`, `response.output_item.added`/`done` per item with each message content part as `response.content_part.added`/`done` around one buffered `response.output_text.delta`/`done` (or `response.refusal.delta`/`done`), and a terminal event chosen from the response status (`response.completed`, `response.incomplete`, or `response.failed` when the status is `failed`/`cancelled` or an `error` is present).
- **Middleware chain**: composable `GatewayMiddleware` trait; ships with resource limits (request-body precheck + concurrency semaphore, health probe exempt) and request-id propagation (`x-powercontext-request-id`, client-supplied or generated, echoed on responses).
- **Body-free evidence**: usage completeness (known/estimated/unknown), cache tokens, TTFT vs first-release latency, attempt counters — never request/response bodies.

Expand Down
Loading
Loading