- Restructure floating input: dominant textarea with compact bottom
toolbar (model badge, thinking toggle, attach, send/stop).
- Model badge sizes to the current selection (not widest option) via
a layered transparent select, with truncate-on-overflow fallback.
- Auto-expand the conversation sidebar on desktop and slide chat
content right when open instead of overlaying.
- Add per-request thinking toggle (brain icon, default on, persisted
in localStorage) sending chat_template_kwargs.enable_thinking.
- Always disable thinking for title summarization.
- Generate chat titles before the main response to keep the SSE
stream from staying open past visible completion and to avoid
busting the KV cache between turns.
evan
6307a64c9c