Baixar wmux 3.32.0 per a Windows

Added
- Per-pane agent memory, live in Fleet View. Each pane's card now shows the resident memory of the agent process running in it (the CLI plus any child it spawned), attributed from the process tree and refreshed on the existing Fleet-View liveness batch — no extra polling while the panel is closed. Makes it obvious at a glance which session is the heavy one when the machine gets tight.
- Groundwork for installing the per-account usage statusline from within the app. The statusline setup logic (wmux setup-statusline) is now exposed to the renderer through IPC handlers (statusline:bridge:status / statusline:bridge:install), mirroring the hooks bridge pattern, so an upcoming Settings UI can let an app-only user (winget/Setup.exe) who never opens a terminal discover and enable it. No user-facing UI ships yet with this change. Same explicit user-click constraint as hooks: never auto-run at boot. (#555)
- The orchestrator now resolves questions it already knows the answer to, instead of freezing on them. In auto mode the Command Deck brain previously escalated every fork to a human decision and halted its loop — even when an operator rule it had already quoted answered the question (the real incident: it recommended the correct option, cited the rule, and still froze for a day). Three changes restore actual judgment. First, the brain's turn context now states its decision authority per mode: an auto workspace is explicitly told to resolve forks from standing rules and escalate only genuine residuals, while assist stays report-and-recommend and off gets no ambient instructions at all — previously all three modes received byte-identical guidance. Second, a new operator-editable deck-policy.md (seeded on first run in the wmux data dir) carries binding standing rules the brain may act on directly — unlike orchestrator memory, which deliberately remains non-executable background context; policy rules still can't grant tools or override safety, and risky/irreversible actions still require a human. Third, the escalation prompt itself now demands a resolve-first check (policy → conventions → memory) before raising a decision, and when the operator's answer corrects the brain's judgment, it persists that correction to memory so the same class of question isn't raised again.
- A decision left unanswered no longer blocks the orchestrator forever. A brain-raised decision used to freeze the workspace's autonomous loop indefinitely until a human clicked resolve. Pending decisions now have a TTL (default 30 minutes, configurable): once it elapses, the next heartbeat wakes the brain for a re-examine turn — re-state the question with better context, or, in auto mode only, resolve it itself via the new deck_resolve_decision tool by citing the standing rule that settles it, then proceed. The self-resolve path is server-enforced (auto mode + TTL actually elapsed + a substance floor on the resolution text — all three checked in the daemon-side RPC, not just the tool description; the citing of the rule is demanded by the prompt, while the server guarantees the resolution isn't a bare token), re-examines are debounced to once per TTL window, a decision whose timestamp was lost can never be self-resolved (human-only), the brain may replace its own stale question with a sharper one (a fresh pending decision still refuses a second raise), and ordinary wakes remain blocked while a decision is pending, exactly as before.
- The orchestrator now runs a periodic level review, and caps how many brain turns run at once. Two additions harden autonomous (Command Deck) orchestration. A level-review heartbeat re-reads each armed workspace's current per-pane state every few minutes and, if a pane needs attention that no live event ever surfaced (a dropped hook, an event lost during a busy stretch), wakes the brain to catch it — routed through the exact same mode/decision/budget/rate gates as an ordinary event wake, so it can never wake more than the live path would, only recover what it missed. Separately, a global concurrency cap limits the whole app to two autonomous turns in flight at once across all workspaces: a burst of activity spread over a fleet no longer spins up a pile of brain subprocesses simultaneously — the extra turns are deferred and retried, exactly like a workspace that's already mid-turn. Neither affects a turn you type yourself, which is never throttled.
- The usage statusline now shows the model's reasoning effort. The model label reads Opus 4.8 (high) instead of just Opus 4.8, so the effort level a pane is actually running at is visible without opening /model. Like the usage percentages, it costs nothing: Claude Code already pipes effort.level on the statusline's stdin, and it is simply omitted on models that don't expose one. The same parenthetical also absorbs the verbose " (1M context)" that Claude Code bakes into the display name of [1m] model variants — Opus 4.7 (1M context) (high) would have been two parentheticals and 20 extra columns, so it renders as Opus 4.7 (1M, high).
Changed
- Hidden workspaces stop holding terminal memory. A workspace you haven't looked at in a while (5 minutes) now has its terminals unmounted to reclaim renderer RAM, so memory no longer grows with the number of workspaces you keep open — only with the ones you're actually using. The daemon keeps every session running untouched; revealing a parked workspace remounts it and replays its screen from the daemon (fast, via the new attach snapshot below), and cross-pane search and agent screen-reads still reach parked panes by reading their content from the daemon, so nothing is silently skipped. On by default; a Settings toggle turns it off for anyone who prefers every workspace to reveal instantly.
- Large sessions reveal instantly instead of freezing. Revealing a pane with a big scrollback used to ship its whole raw history (up to 8 MB of terminal bytes) for the renderer to re-parse in one synchronous burst — the multi-second freeze on long-running agent panes. The daemon now parses that history once on its own side and sends back the compact final screen state instead, so a flooded pane paints in a fraction of the time; full-screen TUIs and anything the snapshot can't reproduce faithfully fall back to the exact old behavior. The parse is chunked so it never stalls output on other panes.
- Lighter idle footprint and smoother streaming. Several always-on costs were trimmed: four rarely-open overlays (Settings, command palette, and two others) now load only when first opened rather than at startup; the unread-notification indicator updates in constant time regardless of how many panes and notifications exist; streaming output repaints at a calmer cadence; and hidden panes release their GPU rendering contexts sooner. Under the hood, the optional shared-MCP-broker path (still off by default) no longer pays a per-connection memory cost.
- Ambient auto-wakes are now rate-limited, and the orchestrator serves fleet state locally. Under heavy fleet activity the orchestrator's view of the workspace tree used to require a round-trip to the renderer that a storm of hook signals could starve — main now keeps a local mirror of the last renderer-pushed snapshot and reads it directly on the hot path, so hook and terminal responsiveness no longer degrade as the fleet gets busy. On top of that, each workspace's ambient auto-wakes are capped to a sliding-window rate, so a runaway hook or detector loop can't turn the brain into a busy-loop (a running loop still iterates on its own budget). Edge wakes can now also carry a one-line fleet summary so the brain sees the wider picture without a poll.
- Install instructions now lead with the package manager. The README's Windows install section puts winget install openwong2kim.wmux front and center with a clear note that it avoids the SmartScreen warning — the direct Setup.exe download is demoted to a secondary "offline install" path with an explicit note about why the warning appears (the installer isn't Authenticode-signed yet).
Fixed
- The daemon reaper no longer keeps sessions with a corrupted timestamp forever. A session record whose lastActivity had become malformed made its idle-time calculation NaN, so every TTL comparison silently failed and the record was kept indefinitely — quietly defeating the idle-session reaper for that record. Such a record now has its lastActivity restamped to the current time on load, so its TTL clock restarts and it can age out normally, without risking a still-live session on a single bad timestamp. (#557)
- Browser read tools can no longer return another workspace's page. With two workspaces each holding a live browser surface, an agent's read-side tools (browser_snapshot, browser_evaluate, browser_extract_text / _data, browser_wait) could return the other workspace's page: an agent in workspace A calling browser_snapshot might get workspace B's URL and content. The write path (browser_open / browser_navigate) already routed by the caller's resolved workspace, but the read path did not — when a tool was called without an explicit surfaceId (the common case) the page-selection heuristic just grabbed the first non-shell page globally, blind to which workspace it belonged to. Browser surfaces are now tagged with their owning workspace at CDP registration time, and read-tool page selection resolves the calling session's own workspace and scopes to its surface — never falling back to a different workspace's page (it opens the caller's own surface instead when theirs is absent). Setups with a single browser, and callers that pass an explicit surfaceId, are unaffected. (#554)