Electron 44 under the hood, markdown in AI Voice, and the old Open WebUI chat app retired.
AI Voice
- Replies render as markdown. Headings, lists, tables, links, images, and inline code show properly instead of as raw symbols, using the same parser and sanitizer the old Open WebUI chat app used. Fenced code blocks keep the one-tap Copy button that hands back the exact text.
- Your own turns stay raw text. They are mostly speech-to-text, so a stray
*or#never restyles them.
Open WebUI chat app removed
The original per-page Open WebUI chat app is gone; AI Voice on the Open WebUI backend replaces it with streaming replies, conversation memory, a model picker, profiles, and knob ring feedback.
- Existing pages migrate automatically on first launch: each becomes an AI Voice page on the Open WebUI backend. The first page's URL, API key, and model seed Settings → Auth if that was still blank; later pages keep their model as the page's pick.
- Two things do not carry over: voice now runs through the Wyoming STT/TTS servers on Settings → TTS/STT rather than Open WebUI's own Whisper, and the knob gesture is tap-to-toggle rather than hold-to-talk.
Under the hood
- Electron 44 (Chromium 152). The installer is about 10 MB larger because the graphics layer is now built into Electron rather than shipped as separate DLLs.
- Builds on Node 26. The toolchain now supports Node 20 through 26 (
yargs17.7.3 fixes the Node 25.7/26 load failure in electron-builder's CLI).