Baixar wmux 3.24.0 per a Windows

Added
- You can turn off "start wmux when Windows starts." wmux registers itself to launch at login on install, but that was never optional — if you didn't want it running every boot, there was no switch. Settings → General now has a Startup toggle that flips it on or off live (it reads and writes the actual Windows startup entry, so it always reflects the real state). Turning it off sticks across app updates — an update no longer silently re-enables autostart behind your back. Windows only; the toggle is hidden on other platforms. (#460)
- The orchestrator can pause and ask you a decision — and it waits, even across a restart. When the orchestrator hits a fork it shouldn't settle on its own — an ambiguous requirement, a risky or irreversible step, a real choice between approaches — it can stop and put the question to you instead of guessing. Its working loop parks (it stops auto-advancing on agent events and scheduled ticks) and a "Decision needed" card appears in the Orchestrator thread, with your options or a free-text answer; the loop stays paused until you respond. The pending decision is saved to disk, so it survives closing and reopening wmux — come back later and the question is still waiting, and answering it resumes the orchestrator from exactly where it paused, with your answer handed to it. It's the orchestrator's own judgment to ask, not a new setting to configure.
- The orchestrator now tells you when it hits a Claude rate limit. When the orchestrator's own Claude session hits (or approaches) a subscription rate limit, an amber notice now appears in its conversation — naming the window (5-hour / 7-day), the account it's running on (if you've bound one), and roughly when the window resets. A hard limit reads "limit reached … new turns keep using this account until you switch"; approaching a limit is a quieter heads-up. It's read from a first-class signal the Claude SDK emits (not guessed from an error), attributed to the account the session actually launched on, and de-duplicated so a burst of retries doesn't spam the thread. This is the detection groundwork for account switching; the orchestrator doesn't switch accounts on its own yet.
- Settings → Accounts now shows live usage per Claude account. Each registered Claude account can display its 5-hour and 7-day quota utilization (e.g. 5h 42% · 7d 71%) right next to its login badge, turning amber as a window crosses 80%. The number is read the moment a Claude turn ends in a pane bound to that account — not by polling on a clock — so idle accounts cost nothing and the reading is always the freshest one. It shares the existing opt-in usage toggle (off by default, since each refresh spends one small request against that account's quota), and a per-row ↻ button forces an on-demand refresh whenever you want the current number. Windows and Linux only (macOS stores Claude credentials in a shared keychain that can't be read per account); Codex accounts show login status only.
- Every pane header now has quick action buttons. The pane tab strip gained a small right-aligned cluster of four icon buttons — split right, split down, new browser, and maximize/zoom — so the actions that were previously keyboard-only (Ctrl+D, Ctrl+Shift+D, Ctrl+B Z) are now discoverable with the mouse, right next to the tabs and the close button. Each button drives the same store action the keyboard already does (no new behavior), carries a tooltip with its shortcut, and stays pinned to the right while the tabs scroll on narrow panes. The maximize button (divider-separated at the end of the cluster) shows a pressed/restore state while the pane is zoomed; it replaces the old hover-revealed corner maximize control, which used to overlap the cluster. The cluster can be hidden in Settings → Appearance → Layout for a minimal, keyboard-only chrome (default on) — hiding it restores the corner maximize control. (There is no header button for adding a second terminal to the same pane — one pane holds one terminal by design; Ctrl+T still adds a surface for power users.)
- The OpenCode bridge now also flags approval prompts and ignores sub-agent chatter. Building on the turn-completion signal, the OpenCode plugin now forwards a permission.updated (OpenCode asking to run something) as an "awaiting input" signal so the orchestrator can notice a pane blocked on an approval — debounced so an auto-approved permission ("permission": "allow") that resolves instantly never raises a false alarm. It also only signals for the root session now: a sub-agent going idle no longer wakes the orchestrator, only the top-level turn does. Re-copy integrations/opencode/plugins/wmux.js to pick this up.
Changed
- The agent & channels panel now has an obvious toggle at the foot of the sidebar. The control that reopens the right-side dock used to be a bare # glyph tucked into the far corner of the status bar — easy to miss once you'd collapsed the dock. It now lives at the bottom of the workspace list as a labeled Agent button (with a little robot icon), so getting the panel back is one clear click. Same toggle, same dock — just somewhere you can actually find it.
- The orchestrator reads terminals on completion, not by polling. Its guidance now says to rely on the automatic wake it gets when an agent finishes or pauses (rather than repeatedly reading a pane to check "is it done yet?"), to read the finished pane once, and to widen a read deliberately (larger tail_lines, then full_scrollback) only when the recent tail isn't enough to judge what happened. Fewer, cheaper reads — which is also what keeps the UI responsive while it works.
- The interface now speaks two colors: warm means "alive", cool means "where you are". Until now one amber did every job — running dots, links, focus, buttons — so nothing stood out. Each theme now splits its palette into a warm accent (things that are alive or need you: running dots, spinners, the terminal cursor, notification rings, unread badges — and the primary action button) and a quiet cool accent (things you navigate: links, jump arrows, the active tab underline, the focused pane edge, focus rings). Amber, Nightowl, Stars & Stripes and Taegeuk gained dedicated cool/warm counterparts; Catppuccin, Red Dynasty and Hinomaru already had two tones; Monochrome and Void stay deliberately colorless. A handful of latent mispaints came out in the wash — the browser loading dot, notification dots, the deck loop's running dot and the Git tab's current-worktree dot are all warm now, and running indicators no longer share a hue with warnings.
- Buttons, inputs, menus and dialogs got a machined, modern finish. Buttons now have a faint surface fill with a hairline edge and a subtle top highlight (and physically sink half a pixel when pressed); the important button on each surface is the one solid warm-accent fill; destructive actions are red-tinted until the final confirm. Inputs and the search bar are gently recessed and light up with a cool focus ring. Right-click menus float with soft shadows and rounded hover highlights, and paired view switches (like the editor's View/Edit) are proper segmented controls. This lands app-wide — toolbar, search, pane tabs, fan-out/rich-input/snippets dialogs, project trust and workspace profile dialogs, fleet cards, approvals, deck panels, first-run wizard.
- The theme picker shows real miniature previews. Instead of four abstract dots, every theme card is now a tiny mock of the app rendered in that theme's actual palette — background layers, accent glow, status dots — so you can see a theme before switching. The Custom card finally reflects your real custom colors as you edit them (it used to show a frozen sample palette forever).
- The orchestrator's control bar is just the controls now. The chip row above the orchestrator composer dropped the two canned-prompt buttons ("Agent status" and "PR status") — those were prewritten questions you can just type yourself — leaving the three controls that actually belong there: the agent mode, the loop, and schedules. Mode (the master off/manual/assist/orchestrate switch) now anchors the left, set off by a hairline from the two automations it governs, and the one-click fleet-recovery chip still appears on the right after a reboot when there are agents to bring back.
Fixed
- The wmux command now actually installs, so you can call it from any shell after installing. The installer was supposed to drop a wmux shim into a folder on your PATH so wmux … works in any terminal, but on packaged builds that step silently never ran: the code that installs (and uninstalls) the shim was loaded through a dynamic require that the production bundler leaves unresolved, so it threw behind a best-effort catch and did nothing — no shim, no PATH entry, no error you'd ever see. The loader is now bundled correctly, so install/update creates the shim and uninstall removes it as intended. (Same root cause as the Windows-startup registration; a leftover shutdown-time flush was quietly failing the same way and is fixed too.) (#463)