Roles Witch

1. New voice option — replies can now be spoken by an external OpenAI-compatible voice server, with its own address, key and voice per chat, plus a preview before saving. 2. Cleaner sound settings — audio settings split into "Speech input" and "Speech output," and a few clutter items (voice speed, one-time setup prompt) removed. 3. More reliable calls — sound handling under the hood was fixed, so calls hang less and audio cuts out less. 4. Smoother chatting — message menus (copy, edit, delete…) now work from a long-press anywhere on a bubble or photo, and can't be opened while the assistant is still typing its reply. 5. Better attachments — send PDFs and several photos at once, share message text, and tapping photos no longer slows down long conversations. 6. On-device mode improvements — photo understanding works properly, long chats stay within the model's memory, and switching between cloud/on-device is more predictable. 7. Nicer screens — new floating "new chat" button, per-chat streaming toggle, swipe to delete chats and tools, and a shorter acknowledgements list. 8. Bug fixes — avatar cropping saves what you actually selected, a nicer launch screen, and faster model list loading.