Skip to main content

Changelog

This page tracks released versions of the SDK. Per-package changelogs live alongside each package (packages/*/CHANGELOG.md) and on npm.

1.1.1 (2026-07-11)​

Patch release for @llmrtc/llmrtc-provider-openai and @llmrtc/llmrtc-backend:

  • openai provider: fix streaming TTS failing with response.body?.getReader is not a function when the OpenAI SDK returns a Node Readable body (Node < 18 or apps loading the SDK's Node shims). The error was previously swallowed and every turn silently fell back to non-streaming TTS, adding seconds of latency.
  • backend: the CLI now exits at startup with a clear message on Node < 20 (override with LLMRTC_SKIP_NODE_CHECK=1) instead of failing in confusing ways at runtime.

1.1.0 (2026-07-11)​

Correctness release: makes the advertised behavior work end to end.

  • Core: system prompt is always sent and pinned outside the history window; runTurnStream accepts an AbortSignal (barge-in) and commits the partial response to history; onLLMEnd has real guardrail semantics (throwing cancels the response); concurrent turns serialize correctly; clearHistory transitions clear conversation history; PlaybookHooks are wired; tool executor abort/ordering fixes; deep tool-argument validation.
  • Providers: multi-turn tool calling now works across OpenAI, Anthropic, Gemini, Bedrock, OpenRouter, LMStudio, and Ollama; malformed tool arguments are flagged instead of silently becoming {}; Bedrock defaults to a current, invocable inference profile; Ollama NDJSON streaming is buffered correctly; ElevenLabs format mapping fixed; Whisper uploads are named by their actual container.
  • Backend: turns serialize per connection and barge-in aborts the in-flight turn in any phase; session reconnection binds the recovered history; TURN credentials refresh on a TTL; start()/stop() lifecycle fixes; real-time audio pacing.
  • Web client: vision frames are delivered in time to reach the turn; session resume follows the protocol and restores history (new reconnected event); close()/failure releases the microphone and camera; frame capture works outside Chromium; the non-WebRTC fallback (sendAudio/ttsChunk) is usable.
  • Packaging: packages resolve under require(), clean checkouts build first try, npm tarballs ship NOTICE and CHANGELOG files.

1.0.0 (2025-12-19)​

Initial public release.