Transcription

• Fix bugs where meta-models (e.g. cloud) were incorrectly validated against a specific provider — provider routing is now resolved server-side based on language support
• Improved language-aware provider selection: the best STT backend (Deepgram vs. Soniox) is now chosen automatically per language

AI

• AI model selectors now filter out deprecated models (GPT-3.5, Claude 2, etc.), date-snapshot versions, and non-chat models (codex, moderation, realtime)
• Local provider connection status (Ollama, LM Studio) is now shown with visual indicators — model selection is disabled when the provider is not running

Calendar

• Improve participant email resolution for Apple Calendar events with a better fallback chain (email → contact → URL → name)

Other fixes and improvements

• Fix Cmd+Q quit behavior — now triggers on keydown instead of keyup for more predictable behavior
• Status banner position adjusts dynamically based on timeline visibility
• Profile avatars now use Facehash-generated faces
• Fix hover jump on interactive buttons caused by gradient backgrounds
• Remove contacts permission request from onboarding