The chat showed 'Server offline' until the user opened Settings, and the selected model was forgotten on every restart. useAutoConnect fetches the models once after the persisted settings are restored; activeModelId is now persisted.
regenerateResponse removed the previous answer before checking for a model and then returned silently, leaving the conversation without a reply. It now checks first and explains that a model is required, like sendMessage. The duplicated streaming code in sendMessage and regenerateResponse is now a single streamAnswer helper.
- 17 files unreferenced by the app (knip): hello-wave, parallax-scroll-view, themed-*, icon-symbol, Button, GlassCard, ScreenWrapper, Models catalog, duplicated theme/color-scheme hooks
- scripts/reset-project.js and its npm script: it moves app/ to app-example and would wipe the project
- dependencies only used by those files: expo-image, expo-symbols, expo-web-browser, react-native-sse, @babel/plugin-transform-runtime
- GitHub Actions workflow running typecheck, lint, tests and the web export
The edit path duplicated the streaming flow with inline require() calls; it now reuses regenerateResponse, which also fixes:
- edited conversations were never saved to History
- the answer generated after an edit could not be stopped (no abort handle kept)
Stopping a stream now saves the partial conversation, like finishing or failing does.
Adds characterization tests for ChatScreen and useLMStudio.
evandro
Evandro Fonseca Junior