Baixar wmux 4.10.0 per a Windows

### Added

• **See when an agent is driving a browser tab, and stop it.** A browser tab
now shows a mark in its toolbar and on its tab while an agent is acting on
it. Stop ends the agent's current action and pauses it on that tab until
you resume; Take over hands the tab to you until you hand it back. Tabs of
your own Chrome that you lent to an agent can be returned from the
workspace's Chrome profile menu with Return to me. Before, the only sign
was the help bar, and there was no way to stop or take back a tab.

• **Office shows every workspace at once.** The Office plugin (by @bluemas) now seats every local workspace's agents in one office, so everyone who needs you queues at the same boss desk. Clicking an employee takes you to that workspace's pane. The office is scaled to fit the page, so it never scrolls. This needs a new plugin permission, `agent.state.read.all`, which the approval prompt lists on its own line; without it, Office shows only the current workspace, as before. An installed Office updates itself and asks for approval again (#2613).

• **Plugins shows what each catalog plugin looks like.** The detail for Office and Agent status now has a real screenshot of that plugin's page, between the credit line and the description. Before, you only saw a description until you installed and opened the plugin. A bundle you installed by hand shows no preview (#2618).

• **Run now can end a finished run's leftover session and start.** A finished
schedule run keeps its agent session open for up to 10 minutes, and until
now a Run now in that window was only refused with "Not started：previous
run's session still open", so you had to wait. Now the schedule offers to
end that session and start：an inline confirm names the session, Cancel is
focused, and confirming ends only that one session before the run starts. A
run that is still going never gets the offer, and scheduled runs are
unchanged. (#2625)

• **Bots in a group chat can react instead of replying, and a bot that stays quiet still shows it saw the message.** In a group chat, a bot woken by your message, or by another bot mentioning it, can answer with a single emoji. The emoji appears under that message as a small chip, the way a messenger shows a reaction. A bot that has nothing to add leaves a quiet acknowledgement there (👍, 👀, ✅ or 🙌, always the same one for that bot and message). Hover over a chip to see who reacted or who saw it. Before this, a bot with nothing to say left no trace, and a quick "agreed" took up a whole message. A reaction is only something you see：it never wakes a bot or starts any work.

• **A pane that hits its usage limit now offers what you can do next.** Click the limit chip in the pane's tab bar to open a small card. It also opens once by itself when the pane you are looking at hits the limit. From the card you can:
• Resume automatically when the limit reopens.
• Move the conversation to another account, when you have two or more. You can also tick "Always switch automatically".
• Continue in another agent that is installed and signed in. It opens in a new tab of the same pane, with a note about the work typed in for you to send.
• When every account is out, see the earliest time one reopens.

Before, the chip only offered a resume toggle and an account menu, and resuming was off unless you turned it on.

• **Keep the computer awake while agents work.** A new setting in Settings ›
General › Power stops the computer from going to sleep：Off, While working
(the default：only while an agent on this computer is working, as Fleet's
Running section counts it) or Always. The screen can still turn off.

• **Runs missed while asleep.** If the computer slept through scheduled runs,
the Schedules page says how many were missed and offers Run now, which
starts each of those schedules once, or Skip.

• **The Remote page checks Tailscale for you.** A short checklist shows
whether Tailscale is installed, signed in, has this PC on your network and
can see your other PC or phone, with one button next to the first thing
that's wrong (download it, open it, or how to add the other device).
Connecting to another PC over Tailscale now says straight away when
Tailscale is off instead of waiting (an address that another network in
the same range may own is still tried), and a connection you're using that
drops shows a notice within a minute：a PC link after 16 s down, or about
56 s when it went silent, on either PC; a workspace share after two failed
checks in a row. A brief blip that reconnects on its own stays quiet. wmux
never turns Tailscale on or installs it by itself.

• **Voice input：dictate into the Chat and Bots composers, or into a terminal pane.** The composers get a microphone button: click to start, click again to stop, and the text lands at the cursor for you to edit before sending. In a terminal pane, "Dictate into this pane" (⌘⇧S / Ctrl+Shift+S, rebindable, also in the command palette) types the text at the prompt without pressing Enter. A small indicator shows while wmux listens, and Esc cancels. Speech is turned into text on your computer by a local Whisper model. Your audio stays in memory and is never stored or sent. The first time, wmux asks before it downloads the model (about 80 MB, from huggingface.co); after that, dictation works offline. Korean and English are both understood whatever language wmux is set to.

• **A key for each plugin page.** Settings > Shortcuts and the command
palette list every page a plugin adds to the rail; give one a key and it
opens that page from anywhere, and closes it again. (#2646)

• **A terminal waiting for an answer shows as Needs you in Fleet.** When a command in a plain shell stops to ask something (`git push`'s `(y/N)`, "Press Enter to continue", ssh asking whether to trust a new host, a password), Fleet moves the pane to Needs you and shows the question. A password prompt shows only that it needs you at the computer; its text is never passed on. Before, such a pane looked like an idle shell. (#2649)

• **Let a paired PC start an agent on this one.** Settings › LAN has a new
section, off by default, where you pick which paired PCs may ask, which
folders (each the top of a git repository) and which agents, and the
permission mode the agent runs in. Each request shows a card with the PC,
the agent, the folder, the mode and the full prompt：Allow once, Always
allow this PC, folder and agent, or Deny. An allowed agent starts in a new
"From <PC>" workspace, on a fresh worktree of that repository, and its
header says who started it. By default every later instruction from that
PC asks again.

• **Ask a paired PC to start an agent.** On the Remote page, a PC's menu has
**Start an agent…**：pick an agent and a folder that PC allows, write the
instruction, and choose the pane here that gets the replies. Agents can do
the same with the `remote_spawn` tool, and scripts with
`wmux remote spawn`. That PC's person approves it on their card. If the PC
is off, the request waits and starts once it is on (up to 72 hours). The
agent's replies arrive here as a task.

• **The first run ends with your first bot's first task.** Once an agent is ready, the setup wizard picks a folder and offers "Give the bot its first job"：explain the folder, or draft a README in the chat. It first asks you to confirm that the task uses your Claude usage. Then it makes one bot for the folder and shows that bot's first answer, with a way to open the conversation. The bot only reads: nothing is changed, pushed or opened as a pull request until you allow its workers. Running the setup again never makes a second bot. Before, the wizard ended with a terminal demo and left the Bots page empty. (#2656)

• **Set up Grok and Antigravity from the first-run wizard.** The "Agent ready" step now has cards for Grok and the Antigravity CLI, next to Claude Code and Codex. Install shows the vendor's own install line and runs it in a terminal tab only when you click Run. Where the vendor documents more than one way on your system (Grok's WinGet package on Windows), you pick which one. Sign in then opens the agent, which signs in on its own; wmux can't check that sign-in, and the card says so. Before, these two had to be installed by hand. (#2657)

• **Qwen Code and Kimi Code CLI are recognised.** A pane running `qwen` is
pinned to its conversation when it starts, and after a restart its resume
reopens exactly that conversation (`qwen --resume <id>`), or Qwen's own
picker when there is none. A pane running Kimi Code CLI (2.x) is offered
Kimi's own session picker (`kimi --session`). Neither is in quick launch
yet：both are listed as not verified for fan-out.

• **Set up Qwen Code and Kimi Code CLI from the first-run wizard.** The "Agent ready" step now has cards for both. Install shows the vendor's own lines in the vendor's order：the install script first, then npm, then Homebrew for Qwen or pnpm for Kimi on macOS and Linux. Each line runs in a terminal tab only when you click Run. Sign in then opens the agent, which signs in inside its own session (`/auth` in Qwen, `/login` in Kimi). wmux can't check that sign-in, and the card says so. (#2659)

### Changed

• **The morning report carries no 1:1 thread titles.** Your 1:1 threads with bots still count as open threads. Before, the report listed each open 1:1's title, which is your first message there.

• **A bot's wake no longer picks up replies from another bot's 1:1 with you.** Before, a room-wide message could be matched to a card in another bot's 1:1, and the replies after it there could reach the woken bot.

• **Skipping a browser safety check now needs your approval.** On browser
panes without site protection, an agent that asks to skip a safety check
(running a blocked script, or reading sign-in data of email, banking or
login sites) or to save a download now needs your app