Baixar wmux 3.50.0 per a Windows

### Added

• **Built-in shortcuts can be switched off.** Every advertised row in Settings → Shortcuts now has a checkbox; unchecking it fully unbinds the combo and the key reaches the terminal instead — Ctrl+T then opens Codex's own transcript rather than a new wmux tab. Custom keybindings rebound onto a disabled combo still fire, and the setting survives restarts.

• **The `core` MCP tool profile can now be selected：`wmux mcp register --profile core`.**
The profile itself shipped earlier, but nothing could actually choose it — every
registration path wrote the full entry regardless. `core` drops the `browser_*`
family, cutting roughly 27 KB of schema off the `tools/list` an agent reads
before it does any work; it is an optimization, not a permission boundary, so a
core-mode agent keeps exactly the authority an ordinary one has. Use it for
hosts doing terminal, pane, workspace, channel, and delegation work that never
touch the browser. `--profile full` moves a host back.

The default is unchanged and stays `full`：the browser tools are first-class,
and an agent driving the workspace browser works with nothing to wire up
precisely because the default registration carries them.

• **`browser_repl`：drive the browser through many steps in one tool call.**
A snapshot, a `refs.find`, a click, a wait and a text extraction used to cost an
agent five round trips, each one re-reading a page of tool output. The new
full-profile tool runs a JavaScript snippet instead, where every allowed
`browser_X` tool is `await browser.X(args)` with its usual arguments. Snapshot
results come back with the refs already parsed, a failed step throws a
catchable error, `console.log` is captured, and the run's call ledger (tool,
redacted arguments, timing, browser events) is printed with the result so the
agent still sees what happened. State survives between calls; a snippet that
overruns its timeout has its runtime terminated and the next call starts fresh.

The snippet runs in a worker thread and every browser step goes through the
existing tool handler, so automation leases, password redaction, ref maps and
the action trace behind `browser_replay` all behave exactly as they do for
direct calls. One thing does change：approving a single `browser_repl` call
covers the 21 tools it exposes (navigate, navigate_back, tabs, click, type,
fill, press_key, hover, drag, select, scroll, scroll_into_view, snapshot,
smart_snapshot, extract_text, extract_data, wait, console, highlight, resize,
dialog). Tools with a heavier footprint — `evaluate`, `screenshot`, cookies,
storage, downloads, uploads, PDF, tracing, replay and surface lifecycle — stay
separate calls. (#1177)

### Changed

• **`browser_smart_snapshot` returns a diff, and its refs hold still.** The
tool re-sent its whole indexed listing on every call, which on an
act-then-verify loop is the same few hundred lines over and over. A repeat
call now returns only what moved, against the same baseline machinery
`browser_snapshot` uses — with the same fallback to the full listing when the
diff stops being a saving, and a first line that always says which of the two
you got. Pass `full：true` for the complete listing.

What made that possible is a fix worth having on its own：smart refs were a
running count over the page, so one element appearing above another
renumbered it and everything after it. Replaying a ref you were given a
moment ago clicked the neighbouring element instead, and said nothing. A ref
is now tied to the element itself, so it means the same thing until that
element goes away — and it never comes back naming a different one. Clicking
by `smartRef` against a live page was broken outright and now works, and a
ref that no longer names one element on the page you are looking at — the
page navigated or reloaded under it, the element went away, several elements
now answer to the same description, or the click is aimed at a different tab —
is refused with an explanation instead of clicking whatever moved into its
place. Saved flows record these clicks in a form that can actually be
replayed; before, the step always failed on replay while the live click
looked fine. In packaged builds, where the underlying identity is not
available, the listing stays complete rather than diffed. (#1166)

• **The sidebar spends one less line per workspace.** The "Agents N" disclosure row is gone; its chevron and count now sit on the workspace row itself, so an expanded workspace starts listing its agents immediately. Measured at 21px saved per collapsed workspace — with eleven open, the difference between seeing the list and scrolling it. The idle label now yields to the agent count (its minutes moved to the workspace name's tooltip, which also means a truncated name is finally readable on hover).

• **Browser tools now run their own page scripts in an isolated world.** A
website can no longer see the scripts the agent injects to read the page, nor
hand them doctored results by redefining the DOM methods they call — the
snapshot, extraction, wait and scroll scripts share the page's DOM but not
its JavaScript, so they are invisible to it. `browser_evaluate` runs there
too by default and gained a `mainWorld` option for the expressions that
genuinely need the page's own globals — as does the backend that has no
isolated world to offer, which now says so in its answer. (#1172)

• **The agent types with a human rhythm instead of a metronome.** Inter-keystroke
delays used to be drawn flat from a 50–150 ms range, which no typist produces.
They now cluster around a median with a long tail, run longer after full stops,
commas and word breaks, and occasionally pause the way someone does mid-sentence.
(#1173)

• **Clicks move the pointer there first, and land off-centre.** A click used to
teleport the pointer onto the exact centre of the target and press. It now
travels from wherever the pointer was left to a point somewhere inside the
element, which is what the page sees when a person clicks it. Hovering and
dragging in the builtin browser now go through the same real pointer input,
rather than events synthesised inside the page. Two new pipe RPC methods back
this — `browser.hover.cdp` and `browser.drag.cdp` — and they are gated on the
`browser.click` capability rather than `browser.evaluate`. A caller granted
hover and drag is being granted pointer input, which is what `browser.click`
means; routing them through `browser.evaluate` would have required handing out
arbitrary script execution to move a pointer. (#1173)

• **Navigations carry the page they came from.** Following a link from one page
to another now sends a `Referer`, as a real click-through does. A first
navigation, a reload, or a jump out of a blank or browser-internal page still
sends none. (#1173)

• **Emulated devices are consistent about what they claim to be.** `browser_emulate`
used to change the User-Agent string alone, leaving the browser's Client Hints
answering with the real machine — so an emulated phone described itself as a
phone in one place and as your desktop in another. The user-agent data,
platform and Accept-Language now match the preset, and resetting the preset
restores all of them together. (#1173)

• **A remote terminal now looks remote.** Its tab carries a marker, its tooltip reads "Remote terminal — …", and copying or dragging out pane info reports it as a remote terminal with its host and session — instead of calling it a plain terminal and handing you a working directory that only exists on the other machine.

### Removed

• **The six `company_a2a_*` MCP tools are gone.** `company_a2a_whoami`, `_send`,
`_broadcast`, `_inbox`, `_ack`, and `_status` no longer appear in any tool
profile. Nothing drove them — no prompt, skill, or documented flow called one —
while every session paid for their schemas in the `tools/list` it reads before
doing any work. The full profile drops from 93 tools to 87 and from 76,410 to
73,610 bytes on the wire; the `core` (47) and `commander` (41) profiles are
unchanged, because neither ever carried them.

Two of the six have a direct workspace-level equivalent：use `a2a_whoami` in
place of whoami, and `a2a_broadcast` in place of broadcast. The other four do
not. Sending by department, lead, member name or `"CEO"`, the member inbox,
acknowledging inbox messages by id, and reading the department tree were all
Company-mode-only features, and they are removed from MCP rather than
replaced — `send_message` resolves workspaces, not company members, and
`channel_ack` acknowledges a channel `seq`, which is a different store from an
inbox message id.

**Company mode's UI is unaffected** — the panel, spawn, and the provisioning
RPCs behind them are untouched. Company A2A messaging, though, currently has
no caller at all：these tools were its only entry point. The pipe handlers are
kept rather than deleted, pending the Company mode re-evaluation.

### Fixed

• Settings keeps the "Check for updates" button next to "Install now" once an update is staged. It used to be replaced, so a downloaded update hid any newer release published afterwards until the 30-minute background poll caught up.

• **An isolated instance no longer hijacks the production MCP registration.** A `WMUX_DATA_SUFFIX` boot — the dev build's automatic `-dev` included — used to rewrite the `wmux` entry in `~/.claude.json`, `~/.codex/config.toml`, and `~/.gemini/settings.json` to its own (often disposable) bundle path, silently retargeting your daily agents. Isolated instances now skip those writes (`WMUX_MCP_REGISTER_EXTERNAL=1` opts back in), the `wmux mcp` CLI warns before touching production configs, and Settings → MCP explains the skip instead of silently doing nothing.

• `browser_replay` works in a packaged build. Every one of its actions — `list` i