## What / Why
This PR finishes the desktop refactor: the Tauri app is now a thin shell that launches the web server as a sidecar and loads the UI from `http://127.0.0.1:<port>`. All real backend logic lives in `packages/web/server/index.js`; desktop Rust keeps only native integrations (menu/dialog/notifications/updater/deep-link + window chrome).
This unblocks:
- consistent behavior across web/desktop/vscode (single backend)
- simpler desktop maintenance (no duplicated Rust backend)
- host switching between Local + remote instances in desktop
- reliable cold-start behavior on slow machines (VSCode + desktop)
## Key changes
- Desktop sidecar runtime
- build pipeline to bundle web dist + `openchamber-server` sidecar (`packages/desktop/scripts/build-sidecar.mjs`)
- robust local port selection (prefer saved/default, fallback to random; persisted in `~/.config/openchamber/settings.json`)
- improved PATH handling so the sidecar can locate `opencode` CLI (incl `~/.opencode/bin`, overrides, common bins)
- disable native right-click context menu in production builds (dev keeps it)
- Desktop instance switcher (Tauri-only)
- header button + modal to add/edit/delete remote hosts, set default, probe status/ping, switch back to Local escape hatch
- auth gate includes host switcher so you can recover when a remote host is broken/auth-required
- host list stored desktop-locally (not tied to the currently selected remote server)
- Notifications
- decision logic moved server-side; desktop notifications emitted via sidecar stdout and shown natively by Tauri
- prevent double-notifications on desktop Local origin (UI ignores SSE notification when native path is active)
- restore macOS notification sound
- Updates
- Tauri updater used only when viewing Local instance in desktop shell (avoid “remote web update” triggering desktop restart)
- Settings persistence & UX polish
- persist model favorites/recents via `/api/config/settings` (works for web + desktop; not origin-dependent)
- persist per-project sidebar collapse state in `projects[].sidebarCollapsed` via `/api/config/settings` (with debounce on toggles)
- macOS header sizing/traffic-lights offsets fixed (marketing macOS major injected from desktop; MultiRun header aligned)
- VSCode cold-start: keep retrying provider/agent loads after connection to avoid empty UI on slow machines
- misc lint/type fixes + bun.lock sync
- Desktop bootstrap / resiliency
- show onboarding screen when OpenCode CLI is missing (desktop Local origin), with retry hook to restart OpenCode after install
## Testing notes
- Desktop (macOS): switch Local <-> remote, set default host, verify auth gate recovery, native notifications (with sound), updater gated to Local
- Web: favorites/recents + per-project collapsed state persist across reload/restart
- VSCode: slow startup no longer results in missing providers/agents/models
- Compute default expanded count for stacked view to reduce DOM load
- Move diff processing to worker pool to improve UI responsiveness
- Add content cache key to avoid recomputing large diffs on render
* feat: added themes system
* feat: smart sidebar auto-hide for files/diff tabs + lower files sidebar threshold
* feat: added Checkbox component and update theming
- Add reusable Checkbox component for toggles across UI
- Replace several inputs with Checkbox in settings and commands panels
- Add DiffIcon and apply surface/border theming to key UI areas
* feat: Add convert-vscode-theme.cjs to convert VS Code themes to OpenChamber format
* refactor: remove unused permission logic from ChatInput
- Remove unused permission rules parsing logic from ChatInput
- Memoize renderTheme in DiffWorkerProvider to avoid unnecessary recalculations
- Remove forceOpaque helper in vscode theme adapter
* fix: guard VSCode theme loading in MarkdownRenderer
* feat: add custom user themes loading and reload
- Load user themes from ~/.config/openchamber/themes at runtime
- Expose /api/config/themes to fetch custom themes
- Allow theme reloading from Settings → Theme → Reload themes in the UI
- Add LRU eviction to diffCache (max 30 entries / 20MB)
- Single-file mode: render only selected diff, not all hidden ones
- Stacked mode: collapse files beyond first 10 by default
- Limit DiffWorkerProvider warmup to 10 files
- Reduce worker pool size (4 to 2) and AST cache (200 to 50)
* feat: add chat virtualization and turn grouping infrastructure
Implement VirtualMessageList with virtualization and overscan for smooth scrolling
Refactor MessageList to render only visible messages and pass scroll refs
Introduce TurnGroupingContext and provider to stabilize per-turn rendering and reduce re-renders
* feat: add web server session activity endpoint for visibility restore
Add /api/session-activity endpoint to expose tracked session activity
Use web server activity first when restoring visibility, with fallback to global status
Update server SSE to set session phase on activity events
* feat(chat): split TurnGroupingContext into UI/Streaming contexts
Add separate UI state and streaming contexts to reduce re-renders.
Skip animations for items visible in collapsed view when expanding.
Track shown parts in collapsed mode to fade in progressively
* feat(ui): cache turn groups with expand state and guard web activity
Add isExpanded to the turn cache key to trigger updates correctly
Expose isGroupExpanded from UI state so groups reflect expand/collapse
Limit web server session activity fetch to web runtime only