Baixar wmux 3.38.2 per a Windows

Added
- You can hand a workspace a fresh orchestrator without losing anything
else. A New session control now sits with the other orchestrator
controls at the top of the deck. It retires the current orchestrator and
starts a new one that reads your project files again and takes a fresh look
at your panes, which is what you want once a long conversation has drifted or
a turn has wedged. There was no way to do this before: restarting wmux
deliberately resumes the same conversation, quitting the orchestrator's own
terminal just reopens it on that conversation, and deleting the workspace
takes every pane with it. Your panes, worktrees and conversation history all
stay, and loops and schedules keep running. It asks for a second click before
it commits, and stays available even mid-turn — a stuck turn is usually
exactly when you need it. (#711)
- You can set a dependent child task loose without stopping it. A Detach
from parent action now sits in the context menu of a child workspace that a
fan-out spun up. Detaching releases that task from its parent — the parent
stops tracking it and its mission channel folds away into Archived — while the
workspace itself, its worktree, branch, terminals and running agent all keep
going, now on their own. Before this, a child task was tied to its parent for
its whole life: the only way to end the relationship was to close it, which
tears down the worktree and the work along with it. Detach keeps everything
and simply cuts the leash. The cleanup scanner knows a detached worktree is
still live, so it is never offered up for deletion, and detaching something
twice is harmless. (#715)
Changed
- The orchestrator now runs on the terminal brain by default. The Command
Deck drives your own claude binary in an embedded terminal instead of the
Agent SDK, so it needs no API key and picks up its per-workspace CLAUDE.md
from disk every time it starts. The SDK brain has not gone anywhere: it stays
selectable under Settings › Orchestrator, and wmux still falls back to it
automatically when daemon mode is unavailable. Existing installs move over on
the next launch. The move is reversible and costs no history — each brain
keeps its own conversation, so switching back in Settings picks your SDK
conversation up exactly where it was. (#710)
- The workspace mode now says how Claude is started, not how chatty the
orchestrator is. The three modes never reached the Claude process at all,
so "Off" quietly meant "wakes less often" while the orchestrator kept
running. Now: Off does not start it (and stops one that is already
running), Assist starts it with edits accepted automatically while every
other permission prompt still stops and asks, and Danger — the mode
previously called Auto, renamed for what it does — starts it in bypass mode.
How often the orchestrator wakes on its own is now a separate setting, and
every existing workspace keeps the wake behaviour it already had. (#735)
Fixed
- A pane running a TUI that asks the terminal about its own modes no longer
freezes. opencode probes six terminal modes the moment it starts, and each
probe hit a miscompiled xterm handler that threw. The throw killed that pane's
parser for good: the process kept running and its title kept updating, but the
screen never painted again and every agent read of it came back
RPC timeout: input.readScreen (5000ms) — with nothing to explain why. The
renderer bundle is now built at a target that stops mangling the handler.
(#708)
- Reading a pane can no longer hang. The read path waited on the pane's
parser without a deadline, so a stalled pane took the whole read down with it
and the caller got nothing. Reads now wait a bounded moment and return what
the pane has, which is worth far more than a timeout. (#708)
- The orchestrator model picker now applies to the terminal brain. The
model only ever reached the orchestrator when you typed into the deck
composer, and the terminal brain has no composer — you type into its terminal.
Choosing a model therefore did nothing for it, and scheduled or event-driven
turns ran on whatever model your last typed message happened to leave behind.
The picker now applies on every path. (#710)
- A workspace that started without a running daemon no longer loses its
terminal conversation. With no daemon the deck quietly serves the SDK brain
instead, but it filed that conversation under the terminal brain's name. The
terminal conversation could not be resumed afterwards, the substitute brain
was told it had no memory to write to, and the workspace stayed on the
substitute even after the daemon came back. (#710)
- Full-power mode no longer looks available when it cannot do anything. It
only applies to the SDK brain, so under the terminal and Hermes brains the
toggle now explains that instead of silently ignoring the click. (#710)
- A dev build no longer reads the production config to decide the CDP gate.
The CDP decision in the main process ran before the instance-isolation
(WMUX_DATA_SUFFIX) block, so a -dev build consulted ~/.wmux instead of
~/.wmux-dev. Moved after the suffix block so each instance reads its own
config. (#613)
- A pane that finished its command no longer reports running forever.
Clearing a pane back to idle sat behind a 30-second window built to suppress
a "Task may have finished" notification that had since been removed. A
workspace switch (any pty:resize) inside that window made the clear vanish,
and nothing retried it — the activity monitor consumes its state transition
before invoking callbacks, and a quiet pane never produces another burst. For
a plain shell, where no agent is detected, that clear is the only path there
is, so the pane sat at an idle prompt reporting running until it was closed.
The window also blocked the correction for the false running a resize
redraw raises in the first place. Agent panes are unaffected: their precise
hook and detector statuses still outrank byte silence. (#733)
- The orchestrator can no longer close a pane to clear its status. When a
wedged pane held the Stop gate open, the terminal brain read "resolve these
panes" as "end these panes" and ran exit, then Ctrl+D, in a live shell the
user owned. The refusal text now names the prohibition, and session-terminating
input is refused outright when it targets a pane the caller's own turn is
currently blocked on. The scope is exactly that intersection — an orchestrator
that is not gate-held closes shells as before, and even a gate-held one may
close panes it is not blocked on. (#733)
- A request the orchestrator was working on no longer comes back to life
after you quit. wmux remembers what it was asked to do so a restart does
not lose your request — but that memory used to be permission to act, with no
expiry. Reopening wmux could hand the orchestrator an hours-old objective and
set it loose on your panes seconds after launch, even in a workspace set to
Off. A request that outlives the app is now parked: it is still recorded,
still shown to you, and still stops the orchestrator from claiming it
finished — but it drives nothing on its own. On the next launch you are asked
whether to resume it or drop it. Only your own message can wake it up again.
(#733, #735)
- The orchestrator is told not to close your panes in the one case where it
used not to be. The refusal that holds a turn open already warned it never
to close or kill a pane to clear a status. That warning was missing from the
refusals raised by a durable request — which is exactly the case a leftover
request lands on. (#735)
- What you ask for now is what the orchestrator works on. If a request was
already on the books, anything you typed afterwards was recorded as a note
underneath it rather than as the thing to do — so a message sent after
reopening wmux could end up filed under an objective from hours earlier, with
the old one still leading. Now, speaking to an orchestrator whose request
predates this session replaces that request with yours. A request it is
actively working on still collects your message as a follow-up, so nothing
running is abandoned, and answering "resume it" on the startup prompt still
resumes the original untouched. (#733, #736)
- Dropping a request no longer abandons the agents working on it in
silence. Starting a new session, or replacing a stale request, can discard
one that still has work delegated to other agents. Those agents kept running
with nobody waiting on them. wmux now tells you, and asks what to do about
them, whenever a discarded request still had something outstanding. (#736)
- The New session button no longer promises to keep the conversation it
discards. Its tooltip listed the conversation history among the things that
stay, when starting a fresh session is precisely what clears it — the reverse
of what happens, on the one control you reach for when the orchestrator is
stuck on stale context. It now says the orchestrator forgets the conversation
and the #commander transcript remains as the record. (#737)
- The unread count no longer vanishes the moment you minimise wmux. On
Windows and Linux the tray tooltip is the only place an unread count shows
once the window is hidden — and hiding the window was itself what erased it,
because the "background sessions running" note and the unread badge each
overwrote the whole tooltip. They now share it, so a hidden wmux shows both.
A count that arrives while wmux is still starting up is no longer dropped
either. (#729, #738)
Security
- The CDP remote-debugging port is now closeable from config. Browser
automation runs over Electron's CDP port on loopback, and unlike the other
same-user attack surfaces it only needs a loopback socket — no filesystem or
process access — to reach full renderer control. Set