Skip to main content
All notable changes to this project are documented here. The format is based on Keep a Changelog, and this project adheres to Semantic Versioning.

​
Version 0.1.0

Released July 15, 2025

​
Added

​
Conversation History Limits

New history option with maxMessages (default 100) and maxTotalChars (default unlimited) to prevent unbounded memory growth. Oldest messages are trimmed in pairs to preserve user/assistant turn structure. 
const agent = new VoiceAgent({
  history: {
    maxMessages: 50,
    maxTotalChars: 100_000,
  },
});
Emits history_trimmed event when messages are evicted.

​
Audio Input Size Validation

New maxAudioInputSize option (default 10 MB). Oversized or empty audio payloads are rejected early with an error / warning event instead of being forwarded to the transcription model. 
const agent = new VoiceAgent({
  maxAudioInputSize: 5 * 1024 * 1024, // 5 MB limit
});

​
Serial Input Queue

sendText(), WebSocket transcript messages, and transcribed audio are now queued and processed one at a time. This prevents race conditions where concurrent calls could corrupt conversationHistory or interleave streaming output.

​
LLM Stream Cancellation

An AbortController is now threaded into streamText() via abortSignal. Barge-in, disconnect, and explicit interrupts abort the LLM stream immediately (saving tokens) instead of only cancelling TTS.

​
interruptCurrentResponse(reason)

New public method that aborts both the LLM stream and ongoing speech in a single call. WebSocket barge-in (transcript / audio / interrupt messages) now uses this instead of interruptSpeech() alone. 
// Interrupt both LLM generation and speech
agent.interruptCurrentResponse("user_speaking");

// Old method: only interrupts speech, LLM keeps running
agent.interruptSpeech("user_speaking");

​
destroy() Method

Permanently tears down the agent, releasing the socket, clearing history and tools, and removing all event listeners. A destroyed getter is also exposed. Any subsequent method call throws. 
agent.destroy();
console.log(agent.destroyed); // true

​
history_trimmed Event

Emitted with { removedCount, reason } when the sliding-window trims old messages. 
agent.on("history_trimmed", ({ removedCount, reason }) => {
  console.log(`Trimmed ${removedCount} messages: ${reason}`);
});

​
Input Validation

sendText("") now throws, and incoming WebSocket transcript / audio messages are validated before processing.

​
Changed

​
disconnect() is Now Full Cleanup

Aborts in-flight LLM and TTS streams, clears the speech queue, rejects pending queued inputs, and removes socket listeners before closing. Previously it only called socket.close().

​
connect() and handleSocket() are Idempotent

Calling either when a socket is already attached will cleanly tear down the old connection first instead of leaking it.

​
sendWebSocketMessage() is Resilient

Checks socket.readyState and wraps send() in a try/catch so a socket that closes mid-send does not throw an unhandled exception.

​
Speech Queue Completion Uses Promise

processUserInput now awaits a speechQueueDonePromise instead of busy-wait polling (while (queue.length) { await sleep(100) }), reducing CPU waste and eliminating a race window.

​
interruptSpeech() Resolves Speech-Done Promise

So processUserInput can proceed immediately after a barge-in instead of potentially hanging.

​
WebSocket Message Handler Uses if/else if

Prevents a single message from accidentally matching multiple type branches.

​
Chunk ID Wraps at Number.MAX_SAFE_INTEGER

Avoids unbounded counter growth in very long-running sessions.

​
processUserInput Catch Block Cleans Up Speech State

On stream error, the pending text buffer is cleared and any in-progress speech is interrupted, so the agent does not get stuck in a broken state.

​
WebSocket Close Handler Calls cleanupOnDisconnect()

Aborts LLM + TTS, clears queues, and rejects pending input promises.

​
Fixed

  • Typo in JSDoc: "Process text deltra" → "Process text delta"

​
Version 0.0.1

Released July 14, 2025

​
Added

  • Initial release of Voice Agent AI SDK
  • Streaming text generation via AI SDK streamText
  • Multi-step tool calling with stopWhen
  • Chunked streaming TTS with parallel generation and barge-in support
  • Audio transcription via AI SDK experimental_transcribe
  • WebSocket transport with full stream/tool/speech lifecycle events
  • Browser voice client example (example/voice-client.html)
This was the initial release of the SDK, providing the foundation for voice-enabled AI agents with streaming capabilities.

Build docs developers (and LLMs) love

Get started for freeTalk to us