[David Elner (delner)]

• Trace LiveKit voice sessions in Braintrust, connecting user and agent turns with speech recognition, model responses, speech synthesis, and tool calls.

• Support automatic instrumentation and a new wrapLiveKit wrapper:

  import { wrapLiveKit } from "braintrust";
  import * as agents from "@livekit/agents";
  
  const { voice } = wrapLiveKit(agents); // Adds Braintrust tracing.
  const session = new voice.AgentSession({ stt, llm, tts });

• With BRAINTRUST_CAPTURE_ATTACHMENTS=true, capture incoming audio and agent audio actually played, linking recordings to speech spans and accounting for interruptions.

• Add streamingAttachmentBufferBytes to configure audio attachment segment size, defaulting to 4 MiB.

session                              # Entire voice conversation
|
+-- user_turn
|   +-- user_speech                   # What the user said
|   `-- stt                          # Speech-to-text transcription
|
+-- agent_turn
|   +-- llm_response                 # Model requests a tool call
|   +-- lookup_order [tool]           # Tool execution (named after the tool)
|   +-- llm_response                 # Model answers using the tool result
|   +-- tts                          # Text-to-speech generation
|   `-- agent_speech                  # Speech actually played to the user
|
`-- ...                              # Further user and agent turns

With audio capture enabled:
  session                            -> Audio recording attachments
  turn and speech spans              -> Relevant ranges in those recordings

[Luca Forstner (lforst)]

I find all the code to process WAVs a bit overkill and I think it could be solved more cleanly. I left a counter-proposal in slack: https://braintrustdata.slack.com/archives/C083GCUTVDZ/p1792414129468479?thread_ts=1792378384.092499&cid=C083GCUTVDZ