Deepgram Flux detected user finished speaking.

Agent finished speaking (playback complete).

Deepgram Flux detected user started speaking (barge-in).

Audio data received from Twilio.

Logger for individual services (Flux, LLM, TTS, Player, Agent).

Application state -- just routing information. Conversation history is owned by Agent, not tracked here.

Twilio stream ended.

Twilio stream started.

Cancel agent response and clear Twilio buffer.

ElevenLabs streaming TTS service. Sends text chunks, receives audio chunks via callback. Audio is returned as base64-encoded mulaw a

Self-contained agent response pipeline. LLM is persistent (keeps conversation history across turns). TTS connections come from TTSPool (

Streams audio to Twilio at the correct rate. Features: - Independent playback loop (not affected by incoming messages) - Can be

Custom formatter with colors and clean timestamp.

Send audio to Deepgram Flux.

Deepgram Flux streaming service. Audio format: mulaw 8kHz (direct from Twilio, no conversion needed). Turn events: StartOfTurn (barge-in

OpenAI streaming LLM service. Manages conversation history and streams tokens via callback.

Unified logger for shuo. Class methods -- lifecycle events (server, call, websocket, stream) Instance methods -- event/action/transitio

A named point-in-time within a turn.

A named time range within a turn.

Start agent response pipeline.

Connection pool for ElevenLabs TTS WebSockets. - Pre-connects `pool_size` connections at startup - Dispenses warm connections via get()

Records spans and markers for each agent turn. All timestamps are stored as milliseconds relative to the turn's t0.

All trace data for a single agent turn.

A pooled TTS connection with its creation timestamp.

ANSI color codes.

Current phase of the conversation.