[bot] OpenAI Realtime API (WebSocket) not instrumented

Start in the openai_2_15_0 module by inspecting how TracingHttpClient wraps HttpClient and how client.beta().realtime() establishes its WebSocket connection. Identify a stable WebSocket or RealtimeClient hook, then determine how session events and deltas can be captured. Done means Realtime sessions emit a span with input/output JSON and available token metrics, or the limitation is documented if no stable hook exists.

Gap description

The openai-java SDK v2.15.0+ exposes a WebSocket-based Realtime API under client.beta().realtime(). The current instrumentation in openai_2_15_0 works by wrapping the SDK's HttpClient interface (execute / executeAsync), which only intercepts HTTP/HTTPS requests. WebSocket connections are established through a separate code path and are never passed through TracingHttpClient, so no span is created for any Realtime API session.

Affected surface

Why it matters

Realtime API sessions carry voice and text generation work that is billable and latency-sensitive — exactly the kind of LLM execution Braintrust is designed to observe. Without instrumentation, users see a complete blind spot for this API even when all other OpenAI calls are traced.

Analogous gap already tracked

Issue #120 tracks the same problem for the Google GenAI Live (streaming) API, which also uses a persistent bidirectional connection that bypasses the HTTP-layer hook.

Suggested approach

1. Identify the WebSocket-abstraction interface in openai-java (likely WebSocket or a RealtimeClient).
2. Use ByteBuddy to intercept session open/message/close events, or wrap the WebSocket factory the same way TracingHttpClient wraps HttpClient.
3. Accumulate audio/text deltas and emit a span on session close with braintrust.input_json / braintrust.output_json and token metrics where available.

Notes

• This gap exists only in the openai_2_15_0 module; no other provider currently wraps a WebSocket transport.
• If the upstream SDK does not yet expose a stable hook point, a note in the docs would at minimum set user expectations.