Skip to content

v0.3.0

Latest

Choose a tag to compare

@luixaviles luixaviles released this 30 Apr 04:41
73d6237

Coordinated 0.3.0 release across all @ariontalk packages. Headline change: Gemini Live engine migrated from gemini-2.5-flash-native-audio-preview-12-2025 to gemini-3.1-flash-live-preview.

Engine-Gemini 0.3.0

  • Migrate to gemini-3.1-flash-live-preview. Default model updated.
  • Synchronous tool calls. The model now pauses output until a FunctionResponse arrives. NON_BLOCKING and SILENT scheduling are removed.
  • Greeting trigger uses sendRealtimeInput({ text }) instead of sendClientContent text turns (the latter is rejected by 3.1).
  • sessionResumption only attached on reconnect when a saved handle exists.
  • Listening-state transitions on either generationComplete or turnComplete (3.1 emits these as separate signals; previously stuck on "speaking" if only turnComplete was watched).
  • Bumps @google/genai to ^1.50.1.

Token-Server 0.3.0

  • Migrate to gemini-3.1-flash-live-preview.
  • Drop Behavior.NON_BLOCKING from the highlight tool config.
  • Add thinkingConfig.thinkingLevel = LOW for low-latency voice responses.
  • Bumps @google/genai to ^1.50.1.
  • Private package, not published to npm.

Widget 0.3.0

Version aligned with the rest of the @ariontalk package set. Picks up the new engine via the workspace lockfile. No widget-level API changes.

Core 0.3.0

Version aligned. No functional changes.

Plugin-Silero-VAD 0.3.0

Version aligned. No functional changes.

Upgrading

  • If you self-host the token-server, redeploy it (or pull this release) so it stops sending NON_BLOCKING in tool declarations. 3.1 Flash Live closes the WebSocket with code 1008 ("Operation is not implemented") if NON_BLOCKING is present.
  • If you pinned gemini-model="gemini-2.5-flash-native-audio-preview-12-2025" explicitly, switch to gemini-3.1-flash-live-preview. The default has changed.
  • Custom tool integrations: synchronous tool calls now require sending a FunctionResponse for every tool call, including unhandled ones, otherwise the model stays paused. The bundled engine-gemini already handles this for highlight_and_scroll.

Install

pnpm add @ariontalk/widget
# or via CDN:
# https://cdn.jsdelivr.net/npm/@ariontalk/widget@0.3.0/dist/ariontalk.js