feat: Add pipecat integration - #543

resolves https://linear.app/braintrustdata/issue/SDK-7/braintrust-pipecat-integration
resolves #550

AI Summary

Adds observer-based Pipecat instrumentation through setup_pipecat() and auto_instrument(). The integration injects a BraintrustPipecatObserver into PipelineWorker construction and turns pipeline, model, tool, transcription, and speech events into Braintrust spans.

The integration supports Pipecat AI 1.3.0 and the pinned latest version, with versioned VCR coverage and subprocess coverage for auto-instrumentation.

Usage

import braintrust

braintrust.init_logger(project="pipecat-demo")
braintrust.auto_instrument()

Manual setup and audio options are also available:

from braintrust.integrations.pipecat import setup_pipecat

setup_pipecat(
    capture_user_audio_attachments=True,
    capture_agent_audio_attachments=False,
)

The existing aggregate option remains supported:

setup_pipecat(capture_audio_attachments=True)

Audio attachment controls

Audio attachments can be controlled across integrations with generic environment variables:

BRAINTRUST_CAPTURE_USER_AUDIO_ATTACHMENTS=true
BRAINTRUST_CAPTURE_AGENT_AUDIO_ATTACHMENTS=false

For Pipecat, precedence is:

1. capture_user_audio_attachments / capture_agent_audio_attachments
2. capture_audio_attachments
3. The corresponding environment variable
4. The integration default

Setting BRAINTRUST_CAPTURE_AGENT_AUDIO_ATTACHMENTS=false disables LiveKit playback attachments while preserving agent_speaking spans and transcripts.

Captured PCM chunks are accumulated and logged as WAV/PCM Braintrust attachments with sample-rate, channel-count, frame-count, and byte-size metadata when available.

Validation

Coverage includes:

• Pipecat latest and 1.3.0 integration sessions
• LiveKit Agents latest and 1.3.1 audio-control coverage
• shared environment and integration utility tests
• pre-commit formatting, lint, and cassette-hygiene checks

Co-authored by StarfolkAI (@starfolkai)[bot]