[BOT ISSUE] OpenAI Realtime API not instrumented — zero visibility for real-time voice/text sessions

[braintrust-bot]

Summary

The OpenAI Node SDK provides a GA Realtime API (client.realtime) for low-latency, bidirectional voice and text interaction with GPT-4o models. This repo has zero instrumentation for any Realtime API surface — no channels, no plugin handlers, no wrapper proxies, and no auto-instrumentation configs. Users building voice-enabled AI applications with the OpenAI Realtime API get no Braintrust spans at all.

What instrumentation is missing

No coverage in any layer:

• Wrapper (js/src/wrappers/oai.ts): No proxy for the realtime resource. Only chat.completions, embeddings, moderations, and responses are wrapped.
• Auto-instrumentation config (js/src/auto-instrumentations/configs/openai.ts): No config entry for any realtime.* methods.
• Channels (js/src/instrumentation/plugins/openai-channels.ts): No channel definitions for Realtime API.
• Plugin (js/src/instrumentation/plugins/openai-plugin.ts): No handler for Realtime calls.
• Vendor types (js/src/vendor-sdk-types/openai-common.ts): No type definitions for Realtime resources.

A grep for RealtimeSession, clientSecrets, and realtime.connect across js/src/ returns zero matches.

Key upstream API surfaces with no Braintrust tracing:

SDK Method	Description
client.realtime.clientSecrets.create()	HTTP POST — creates an ephemeral client secret/token for WebSocket connection. Accepts model, voice, tools, instructions, and other session config.
RealtimeSession class	WebSocket session that emits events including response.done (with full usage metrics: input/output audio tokens, text tokens), conversation.item.created, tool call events, and error events.

The clientSecrets.create() call is a standard HTTP request that could be instrumented like any other method to give session-level visibility. The RealtimeSession event stream is architecturally different (WebSocket-based) but its response.done events contain usage metrics that are directly analogous to what's captured for chat.completions.create() and responses.create().

Braintrust docs status

not_found — The Braintrust docs at https://www.braintrust.dev/docs/instrument/wrap-providers document OpenAI as a supported provider for standard LLM calls and streaming, but make no mention of the Realtime API, real-time voice, or WebSocket-based interactions.

Upstream reference

• OpenAI Realtime API guide: https://platform.openai.com/docs/guides/realtime
• OpenAI Node SDK Realtime resource: client.realtime (GA — not under client.beta)
• SDK source: https://github.com/openai/openai-node/tree/master/src/resources/realtime
• Supported models: GPT-4o, GPT-4o-mini (real-time audio and text)
• Usage metrics in response.done events: input_tokens, output_tokens, input_audio_tokens, output_audio_tokens
• Session configuration: model, voice, tools, instructions, temperature, modalities, turn detection

Local files inspected

• js/src/wrappers/oai.ts — no realtime property interception
• js/src/auto-instrumentations/configs/openai.ts — no Realtime configs
• js/src/instrumentation/plugins/openai-channels.ts — no Realtime channels
• js/src/instrumentation/plugins/openai-plugin.ts — no Realtime handlers
• js/src/vendor-sdk-types/openai-common.ts — no Realtime types
• e2e/scenarios/openai-instrumentation/ — no Realtime test scenarios