<!-- library-gap-audit: openai-agents-realtime -->
Summary
@openai/agents-realtime is a separate npm package (part of the OpenAI Agents SDK monorepo, distinct from @openai/agents-core/@openai/agents) that provides RealtimeAgent and RealtimeSession classes for building low-latency, audio-to-audio voice agents on top of OpenAI's Realtime API. This repo instruments the standard (text) Agents SDK tracing lifecycle but has zero instrumentation for this realtime package — no channels, no wrapper, no auto-instrumentation config reference it anywhere in source.
This is distinct from existing issue #1731, which covers the raw openai package's client.realtime resource (the base OpenAI Node SDK's Realtime WebSocket API). @openai/agents-realtime is a different npm package, under a different SDK family (the Agents SDK), with its own RealtimeAgent/RealtimeSession class implementations, its own tool-calling integration with the Agents SDK's tool system, and its own transport layer (WebRTC/WebSocket via @openai/agents-realtime's own connection handling) — not the same code path as client.realtime.
What's missing
Current instrumentation (js/src/auto-instrumentations/configs/openai-agents.ts) only hooks the MultiTracingProcessor.onTraceStart/onTraceEnd/onSpanStart/onSpanEnd lifecycle methods inside @openai/agents-core's dist/tracing/processor.{mjs,js}. This captures spans for standard (text) Agent/Runner execution.
Per OpenAI's own documentation, RealtimeSession/RealtimeAgent execution does not flow through this TracingProcessor pipeline at all — tracing for realtime/voice sessions happens server-side on OpenAI's Realtime API, not via the local MultiTracingProcessor that Braintrust's openai-agents config hooks. This is confirmed by an open upstream issue (openai/openai-agents-js#594, "Custom Tracing Processors for Voice Agents (Realtime API)", opened 2025-10-21, still open) which states that custom tracing processors are not currently triggered for RealtimeSession/RealtimeAgent workflows.
As a result, hooking the existing MultiTracingProcessor channel gives zero visibility into @openai/agents-realtime usage — a separate instrumentation approach is needed, e.g. wrapping RealtimeSession's connection/event-emitter surface directly (analogous to what #1731 proposes for the base SDK's RealtimeSession), to capture session config, tool-call events, and usage/audio-token metrics from events like history_updated, response.done-equivalent events, and tool-call events.
Weekly downloads
Weekly downloads: 1,534,984 (as of 2026-08-24; https://api.npmjs.org/downloads/point/last-week/@openai/agents-realtime — reported window 2026-08-17 to 2026-08-23)
Braintrust docs status
not_found for this specific package/surface. Checked:
- https://www.braintrust.dev/docs/integrations/agent-frameworks/openai-agents-sdk — documents
OpenAIAgentsTraceProcessor (TS) / BraintrustTracingProcessor (Python) for standard Agent/Runner.run execution, and separately mentions tracing "Speech and transcription spans... for voice agents" — but this refers to the classic STT→Agent→TTS voice-pipeline flow inside @openai/agents-core, which still runs through the standard TracingProcessor. It does not mention RealtimeAgent, RealtimeSession, or @openai/agents-realtime anywhere, and does not use the word "realtime" at all.
- https://www.braintrust.dev/docs/instrument/wrap-providers — lists OpenAI Agents SDK (
@openai/agents, min version 0.0.14) as a supported integration, with no mention of a realtime/voice-session variant.
Upstream references
Local files inspected
js/src/auto-instrumentations/configs/openai-agents.ts — only hooks @openai/agents-core's MultiTracingProcessor lifecycle methods; no reference to @openai/agents-realtime, RealtimeAgent, or RealtimeSession.js/src/instrumentation/plugins/openai-agents-channels.ts — channels only cover the trace/span lifecycle methods above.- Repo-wide grep for
agents-realtime, RealtimeAgent, RealtimeSession — only matches in pnpm-lock.yaml and e2e/scenarios/openai-agents-instrumentation/pnpm-lock.yaml (transitive lockfile entries), zero matches in any js/src/ source file.
e2e/scenarios/openai-agents-instrumentation/ — scenario covers standard Agent/Runner execution only, no realtime scenario.
<!-- library-gap-audit: openai-agents-realtime -->
Summary
@openai/agents-realtimeis a separate npm package (part of the OpenAI Agents SDK monorepo, distinct from@openai/agents-core/@openai/agents) that providesRealtimeAgentandRealtimeSessionclasses for building low-latency, audio-to-audio voice agents on top of OpenAI's Realtime API. This repo instruments the standard (text) Agents SDK tracing lifecycle but has zero instrumentation for this realtime package — no channels, no wrapper, no auto-instrumentation config reference it anywhere in source.This is distinct from existing issue #1731, which covers the raw
openaipackage'sclient.realtimeresource (the base OpenAI Node SDK's Realtime WebSocket API).@openai/agents-realtimeis a different npm package, under a different SDK family (the Agents SDK), with its ownRealtimeAgent/RealtimeSessionclass implementations, its own tool-calling integration with the Agents SDK's tool system, and its own transport layer (WebRTC/WebSocket via@openai/agents-realtime's own connection handling) — not the same code path asclient.realtime.What's missing
Current instrumentation (
js/src/auto-instrumentations/configs/openai-agents.ts) only hooks theMultiTracingProcessor.onTraceStart/onTraceEnd/onSpanStart/onSpanEndlifecycle methods inside@openai/agents-core'sdist/tracing/processor.{mjs,js}. This captures spans for standard (text)Agent/Runnerexecution.Per OpenAI's own documentation,
RealtimeSession/RealtimeAgentexecution does not flow through thisTracingProcessorpipeline at all — tracing for realtime/voice sessions happens server-side on OpenAI's Realtime API, not via the localMultiTracingProcessorthat Braintrust'sopenai-agentsconfig hooks. This is confirmed by an open upstream issue (openai/openai-agents-js#594, "Custom Tracing Processors for Voice Agents (Realtime API)", opened 2025-10-21, still open) which states that custom tracing processors are not currently triggered forRealtimeSession/RealtimeAgentworkflows.As a result, hooking the existing
MultiTracingProcessorchannel gives zero visibility into@openai/agents-realtimeusage — a separate instrumentation approach is needed, e.g. wrappingRealtimeSession's connection/event-emitter surface directly (analogous to what #1731 proposes for the base SDK'sRealtimeSession), to capture session config, tool-call events, and usage/audio-token metrics from events likehistory_updated,response.done-equivalent events, and tool-call events.Weekly downloads
Weekly downloads: 1,534,984 (as of 2026-08-24; https://api.npmjs.org/downloads/point/last-week/@openai/agents-realtime — reported window 2026-08-17 to 2026-08-23)
Braintrust docs status
not_foundfor this specific package/surface. Checked:OpenAIAgentsTraceProcessor(TS) /BraintrustTracingProcessor(Python) for standardAgent/Runner.runexecution, and separately mentions tracing "Speech and transcription spans... for voice agents" — but this refers to the classic STT→Agent→TTS voice-pipeline flow inside@openai/agents-core, which still runs through the standardTracingProcessor. It does not mentionRealtimeAgent,RealtimeSession, or@openai/agents-realtimeanywhere, and does not use the word "realtime" at all.@openai/agents, min version 0.0.14) as a supported integration, with no mention of a realtime/voice-session variant.Upstream references
Local files inspected
js/src/auto-instrumentations/configs/openai-agents.ts— only hooks@openai/agents-core'sMultiTracingProcessorlifecycle methods; no reference to@openai/agents-realtime,RealtimeAgent, orRealtimeSession.js/src/instrumentation/plugins/openai-agents-channels.ts— channels only cover the trace/span lifecycle methods above.agents-realtime,RealtimeAgent,RealtimeSession— only matches inpnpm-lock.yamlande2e/scenarios/openai-agents-instrumentation/pnpm-lock.yaml(transitive lockfile entries), zero matches in anyjs/src/source file.e2e/scenarios/openai-agents-instrumentation/— scenario covers standard Agent/Runner execution only, no realtime scenario.