feat(api-rs): OpenAI Responses-compatible /v1/responses ingress

[tarrencev]

Summary

Expose Centaur's durable, thread-keyed session control plane behind an OpenAI Responses-compatible POST /v1/responses — the wire API the Codex CLI speaks — so codex (via a custom model_provider base_url) or any Responses client can drive a fully-configured Centaur thread (tools, sandbox, durability).

The Codex analog of the Anthropic /v1/messages ingress. Stacked on the Anthropic ingress PR (it reuses the session-runtime helpers and ExecuteSessionInput fields that PR exposes) — review the /v1/responses commit here; rebase to main once the Anthropic PR merges.

What it does

• Endpoint: POST /v1/responses (OpenAI Responses shape); streaming SSE + non-streaming.
• Defaults to HarnessType::Codex — the Codex wrapper speaks the same wire format the client expects and uses the deployment's configured Codex model backend.
• Thread continuity: keyed on Codex's session id — the session-id header (equal to prompt_cache_key, stable across codex resume) → thread api:codex:<session-id>, so a resumed CLI session continues the same durable thread and warm sandbox. Absent → a fresh api:<uuid> thread.
• A ResponsesTranslator maps harness output to Responses streaming events (response.created → output_item.added/content_part.added → output_text.delta → output_text.done/output_item.done → response.completed), reusing the same harness-output parsing as the Anthropic translator (claude content blocks; codex dotted and slash-method events). The terminal usage carries total_tokens, which the Codex CLI requires.
• Centaur owns the harness/model/persona/tools, so request model, instructions and tools are accepted and model is echoed, but not threaded into the harness. v1 drives the turn off the last user message in input.

Validation

End-to-end against a live deployment with the codex CLI (custom model_provider → the ingress):

• basic reply — codex exec "Reply PONG" → PONG;
• tool use — codex exec "run echo …" → the inner agent runs the command and returns its output;
• streaming + non-streaming verified directly against /v1/responses.

cargo fmt --check · clippy -D warnings · cargo test green (Responses translator unit cases for both harness shapes).

Follow-ups

• Honor a client-selected model/instructions via a Codex-harness mapping.
• Reasoning items, client-side function_call round-trips, real usage tokens, x-api-key auth, full-history replay.