Summary

• add native POST /responses and POST /v1/responses endpoints
• forward the original request text and upstream response body without JSON or SSE reserialization
• preserve structured and streaming upstream errors, retry metadata, and request cancellation
• detect nested vision input, classify agent-initiated input, and document the new endpoints
• add focused coverage for request forwarding, streaming, errors, headers, CORS, vision, and initiator behavior

Why this approach

This overlaps in endpoint scope with #219 and #230, but deliberately uses a raw pass-through instead of parsing and rebuilding Responses API output. That preserves typed SSE event lines and forward-compatible payload fields, and avoids the async-generator serialization issue reported on #230.

It also adds behavior not covered by those proposals: forwarding request text without JSON reserialization, propagating the inbound AbortSignal, stripping stale compression/length headers after fetch decoding, preserving non-2xx JSON and SSE bodies, and exposing retry/request headers through CORS.

#263 solves a different problem by adapting Responses-only models back into Chat Completions shape; this PR exposes the native OpenAI Responses API shape unchanged.

Validation

• bun test — 61 passing
• bun run typecheck
• bun run lint
• bun run build