Baixar wmux 3.61.0 per a Windows

### Added

• **Every built-in shortcut can be moved or switched off.** Settings →
Shortcuts now lists all of wmux's keyboard shortcuts — workspace switching,
tab and pane cycling, focus movement, zoom, panels — each with its current
key. Click the key to record a new one, untick the row to hand the key to
the terminal, or reset it. Alt+Up / Alt+Down (previous / next workspace)
were the trigger：TUI apps such as Codex and Crush bind them, and wmux took
them with no way to give them back. A new key is refused when it would
take a key another shortcut uses, copy / paste, the prefix key, or a key
with no Ctrl / ⌘ / Alt. Sessions that switched shortcuts off under the old
toggle keep those settings (#1456).

• **A fan-out owner can read and nudge its workers' panes.** The agent or
orchestrator that started a fan-out can now read the terminal of each open
task it delegated, send it plain text, and stop it with Ctrl+C or Escape,
without anyone switching workspaces. Before, those panes were out of reach
because they live in other workspaces. Access ends when the task is
finished. Text read from a worker comes back marked as untrusted, since the
worker printed it. Text sent to a worker's pane may contain only letters,
tabs and newlines：control characters and raw writes are refused, so text
can't do what the key tool withholds. Note the limit：a pane agent's
identity is the pane id its MCP server reports. wmux checks which workspace
owns that pane, but it can't prove the caller is that pane, so another
program running as you could claim to be the owner.

• **Worker permission mode for fan-out.** Settings → Agents → Fan-out workers
picks the Claude Code permission mode workers launch with (Auto by default;
Accept edits, Bypass permissions or Manual). In every mode a worker may
record its task ledger row and read its mission channel without a prompt,
and is denied the wmux tools that reach beyond its own task：fan-out, typing
into other panes, messaging other agents and the browser.

• **"Allow wmux worker tools in Claude Code" button.** Adds the same minimal
allow list to `permissions.allow` in `~/.claude/settings.json`, keeping
everything else in the file. That file applies to every Claude Code session
on the machine, and the button says so.

• **Recent unattended fan-outs in Fleet → Approvals.** Shows what ran without a
prompt, when, from which workspace and in which repository.

• **The sidebar shows who each agent is.** Every agent row leads with a small
drawn monogram for its kind (C Claude Code, X Codex, G Gemini, A Aider,
O OpenCode, P Copilot, OC OpenClaude, K Kiro, R Grok; a terminal mark for a
shell), and a collapsed workspace shows up to three of them grouped by
status instead of a bare agent count. Agent rows also show live activity
and how long ago the agent last did something.

• **Fan-out tasks nest under the workspace that launched them.** Task
workspaces sit indented under their owner with a fold chevron and a
`N tasks · M need you` line, and a menu on that line closes every finished
task at once — worktree and workspace — while keeping any task with
uncommitted or unpushed work and saying why. A task counts as finished only
when every agent in it has finished. Tasks whose owner was closed
collect under "From closed workspace"; detached tasks stay on their own.
The `wtask：` prefix no longer shows.

• **Where a task came from.** A task row's fan-out mark says which workspace
fanned it out, who asked (you from the Multi Task dialog, the
orchestrator, or the pane that called it) and when, and a task workspace's
titlebar has a link back to its owner.

• **Recent activity order.** Settings › Appearance › Sidebar › Workspace
order now offers Recent activity next to Manual and Needs you first.

• **Resizable sidebar.** Drag the sidebar's inner edge (220–400px, default
264px); double-click to reset. Terminals resize once, when you let go.

• **"Changed since you last looked" dot.** A small dot on a workspace row
and its agent row says an agent finished or started waiting for you while
that workspace was out of view. It clears as soon as you look.

• **Finished fan-out tasks wait in one review queue.** Fleet has a new
"Ready to review" section：each fan-out task whose agents have all finished
and that is still open gets one row with its title, the workspace that fanned
it out, its branch, how many files and lines changed, its PR and state if
there is one, and how long ago it finished. From the row you can open the
diff, open or create the PR, jump to the task, or close it (a task with
uncommitted or unpushed work stays open and says why). The fan-out summary
under the owner workspace in the sidebar now reads `N to review` and opens
that section. Before, each finished task had to be opened one by one.

• Experimental Chat view can show and continue the same terminal conversation for Codex and OpenCode, alongside Claude Code, with provider-specific session identity and input guards. OpenCode TUI integration installs through `wmux setup-hooks` on supported versions.

• Shared agent-neutral conversation/activity UI and optional private managed Codex, OpenCode and ACP adapters. ACP is supplementary; opening Chat does not create a separate agent session.

• Start Claude or Codex from Chat with a first message in the same terminal, guarded against existing shell drafts and running processes (zsh/bash/sh with shell integration).

• Use one bottom composer for initial and follow-up messages, with explicit Claude Bypass and Codex YOLO startup mode selection.

• Add bottom-composer skill search for native Claude/Codex with keyboard navigation, agent-correct invocation, source labels, draft-preserving selection and honest discovery states. Keep idle connection text and session placeholders agent-neutral.

• Include native command destinations in the slash menu, add live Codex model/effort controls, and handle app/daemon version skew without restarting existing terminal sessions.

• Reduce Chat re-entry waits by overlapping history reads and subscription setup, preloading the view, and reusing bounded history only after fresh conversation and file identity checks.

• Preserve native Korean/IME composition in the Chat composer with synchronous input state, and prevent composition-confirming Enter from submitting a message.

• Right-align content-sized user bubbles inside transcript wrappers; keep agent replies left-aligned.

• Include the external WebSocket runtime in packaged apps so native chat support does not prevent startup.

### Changed

• **The README leads with what wmux does, not a wall of text.** A 16-second
clip at the top shows one prompt fanning out into three agents in separate
git worktrees, and one agent's question answered from an iPhone. A desktop +
phone image, shorter feature sections, and an install block near the top
replace the long pitch paragraph and the "Why wmux?" table. Stale figures were
corrected along the way (78 MCP tools in the full profile, scrollback up to
100K lines, 11 terminal palettes) (#1454).

• **Every README feature now has a clip.** Short recordings replace the still
images under Features：phone approvals, worktree fan-out and hunk adoption,
the orchestrator, channels, Fleet View, and surviving a quit or a crash. New
sections show the two ways agents browse (built-in panes, or a dedicated
Chrome over CDP) and Claude Code and Codex exchanging a review over wmux. The
Resume description now matches how it works：a second click adds the exact
session's `--resume` for Claude Code, and you press Enter (#1457).

• **Messages between agents keep their line breaks.** A message sent to a pane
running a detected agent (Claude Code, Codex and the other known agents) now
arrives with its real newlines, so code blocks, lists and diffs stay
readable. Before, every newline was folded into `␤` and the whole body
arrived as one line. Each body line is shown with a `│ ` prefix, so a message
cannot fake the end of its envelope and start a second one from someone
else. Shells and panes without a detected agent still get the folded form.

• **Fan-out tasks start from origin's default branch.** Each task's worktree
used to branch from whatever you had checked out, so a worker could start on
a stale local `main` or on an unrelated feature branch. A fan-out now fetches
origin's default branch once and creates every task branch from that same
commit, with no upstream set. The task's diff compares against that commit,
so upstream changes your checkout lacks no longer show up as the worker's.
If the fetch fails, tasks start from the last fetched origin branch, or from
your HEAD when there is none. The fan-out dialog, the fan-out result and each
task's mission channel then carry a warning.

• **Typing or pressing a selecting key at a pane that is waiting on an
approval is now refused for every agent, not only orchestrators.** A pane
agent that typed "1" or pressed Enter at a permission prompt used to answer
it blindly. It now gets a refusal saying the human answers the prompt in the
pane. Ctrl+C and Escape still work. This also applies to `wmux send` from a
shell. When the operator's policy refuses an orchestrator's approval press,
only that orchestrator gets the fallback of typing an answer.

• **Agents can fan out without an approval prompt.** A fan-out started from a
pipe or MCP caller (an orchestrator brain, an agent in a pane) now runs
without waiting for a click, with one toast so it is never invisible.
Before, every such fan-out raised a dialog that auto-denied after 30 seconds
when nobody was at the keyboard. Turn the prompt back on in
Settings → Agents → "Ask before an agent fans out".

• **Fan-out has runaway brakes.** A fan-out task cannot fan out again (one
level deep, including workspaces it creates), at most 8 f