Baixar wmux 3.25.0 per a Windows

Added
- The notification panel now remembers where you were. Closing and reopening the bell panel restores the exact scroll position you left off at, instead of yanking you back to the top of the list mid-triage. If new notifications arrived while the panel was closed, it snaps to the top instead so the newest entries are immediately visible (the list is newest-first). The memory is per-session and in-memory only.
- Codex approval prompts now raise a "needs your input" alarm, mid-turn. When Codex CLI pauses to ask "Would you like to run the following command?" (or the edit-approval and first-boot directory-trust variants), wmux now detects it and fires an awaiting_input notification — previously a Codex pane waiting on approval sat silent until you happened to look, because Codex's own notify hook only fires when the whole turn completes and says nothing about mid-turn pauses. The prompt strings were transcribed from a live Codex CLI 0.145.0 session and are matched whole-line, so an agent merely talking about an approval prompt doesn't trigger a false alarm. Clicking the alarm jumps to the waiting pane, like every other notification.
- Connect a different brain to the Command Deck — Hermes Agent is the first non-Claude orchestrator. A new Settings → Orchestrator → Orchestrator brain picker chooses which agent runtime drives the deck: Claude Code (the default, unchanged) or Hermes Agent over ACP — the open Agent Client Protocol, so further ACP-speaking agents are a configuration entry away, not a new integration. The connected brain gets the same fleet hands (wmux tools over MCP), the same server-side guardrails (it cannot close panes or leave its workspace), streams into the same deck chat with tool-call chips, and keeps its own conversation history across restarts — each brain's thread is separate, and switching back resumes where that brain left off. wmux never touches the brain's model credentials: Hermes authenticates through its own setup on your machine, exactly like your Claude login. Requires the vendor's CLI installed; if it isn't, the first command tells you plainly instead of hanging.
- Orchestrator full-power mode (opt-in): your Claude Code skills, CLAUDE.md and hooks, inside brain turns. By default the Command Deck orchestrator runs "raw" — a fully explicit contract that deliberately loads nothing from your ~/.claude setup, because personal hooks firing inside every brain tool call cost real latency and can feed events back into wmux itself. A new toggle in Settings → Orchestrator opts a workspace's brain into the full Claude Code ecosystem: skills become invocable, your CLAUDE.md applies, and your hooks run. The brain's tool guardrails do not widen with it — it still cannot close panes, spawn subagents, run shell commands, or edit files, a skill's embedded inline-shell syntax is replaced with a placeholder instead of executing, and while the toggle is ON the brain also cannot write its memory notes (Write is hard-blocked in this mode so that no personal permission rule can widen it — a conservative first cut; relaxing that is a separate decision). One thing to understand before opting in: your hooks are your own code — they run inside brain turns outside any wmux sandbox, exactly as they do in your own Claude Code sessions. The toggle applies from the next brain turn on any path — typed, scheduled, or event-woken — the conversation carries over, and a restart restores it.
Changed
- The orchestrator's guardrails moved into wmux itself — groundwork for connecting non-Claude brains. Until now, "the orchestrator can't close panes or tear down workspaces" was enforced by options passed to the Claude SDK — a future non-Claude brain wouldn't have honored any of it. Three server-side layers now enforce the same contract for ANY brain runtime: the brain's tool catalog simply doesn't contain teardown tools (they're never registered, so no brain can call them), the pipe refuses teardown methods from a brain even if a future bug re-exposed one, and a brain whose session credential has gone stale has every request refused outright instead of being quietly demoted to a caller with wider powers. A brain is also confined to its own workspace when focusing panes, matching the existing terminal-routing confinement. No behavior change for today's Claude orchestrator or for regular agent panes — this is the security floor the upcoming "connect an agent" feature stands on.
- Agent panes get ~50 MB lighter each — the browser engine now loads only when used. Every agent pane (Claude Code, Codex, Gemini) runs a small wmux helper that exposes wmux's tools over MCP; it used to initialize the entire browser-automation library at startup, costing ~80 MB of private memory per pane whether or not the pane ever touched a browser tool. The library now ships as a separate lazy chunk that loads on the first browser_* call: an idle helper measures ~32 MB, and only panes that actually drive the browser pay the full cost. Nothing changes functionally — same tools, same behavior, pay for what you use.
- Hidden panes stop rendering in the background — the CPU your invisible agents were burning comes back. Until now, every pane in every workspace kept parsing its output byte-for-byte even while hidden behind another workspace: one busy hidden agent cost roughly a fifth of the render thread, so the app felt heavier the more workspaces you kept around. The optimization that pauses this (previously an experimental opt-in in Settings) is now on by default: while a pane is hidden, the daemon keeps capturing its output but the pane stops painting it — the process runs at full speed and no output is ever lost. What you may notice, once per long-hidden busy pane: the first switch to it after updating may briefly catch up — large backlogs restore from the daemon's snapshot, and the pane says so while it happens. That's the pane painting what it skipped, not missing output and not a regression. Daemon-backed sessions only; local-mode panes are untouched. If you'd rather pay the background CPU for always-instant reveals, one toggle in Settings → Terminal turns it back off — and that choice sticks permanently; no later update will flip it back on you.
- The daemon goes quiet when you do. At idle, wmux's background daemon used to burn a steady ~3% CPU and rewrite every session's recovery buffer to disk every 30 seconds even when nothing had changed. Four diets landed: the process-liveness sweep runs every 15 s instead of 5 s (one tasklist spawn per tick on Windows was the single biggest idle cost — the trade-off is that a supervised job's death is now detected within ~28 s worst case instead of ~13 s, and wmux.json restarts follow the same bound); recovery snapshots are dirty-only (a session with no new output since its last dump is skipped, with a forced dump every 10th tick as a freshness backstop — crash recovery guarantees are unchanged and the on-exit dump still always runs); per-pane metadata (cwd/branch/ports) is only re-broadcast when it actually changed, instead of every 5 s per pane; and the shell's working-directory signal is de-duplicated at the source, so an idle prompt no longer re-announces the same cwd on every redraw. Both cadences are now knobs in ~/.wmux/config.json (daemon.livenessIntervalSec, daemon.snapshotIntervalSec) — see docs/performance.md. Note for daemon changes generally: they apply on release-to-release upgrades; a manual revert needs a daemon restart (tray → Shut down wmux).
- Agent torrents now reach the screen in batches, matching local mode. Daemon-backed panes forwarded every pipe chunk to the renderer as its own IPC message; they now coalesce for 8 ms exactly like local-mode panes always have, cutting IPC wakeups under heavy output with no change to what you see — output still lands before any exit or restore marker, in order.
- PID-reuse ghosts can't pin the daemon anymore. If Windows recycled a watched process id onto an unrelated program, the daemon would count that ghost as a live session forever — blocking its own idle shutdown and holding the session's buffer in memory. The liveness sweep now remembers each process's image name and treats a PID that answers under a different name as the death it really is (confirmed by an independent re-probe before acting, and a probe failure still never counts as death).
- A pane now tells you while it's catching up, and slow switches leave evidence. A revealed pane with a large backlog shows a per-pane catching-up state instead of silently presenting stale content — what's on screen is either current or visibly refreshing. Every reveal also writes one [wmux:reveal] line with a mechanism code (live / retained-catchup / dirty-snapshot / dirty-raw-fallback / resync-degraded) to the on-disk logs, and wmux doctor --performance summarizes retention state, pane counts, and recent resync counters — so a "my pane looks frozen" report can arrive with its own diagnosis attached. The new docs/performance.md covers what keeps running while a pane is hidden, expected reveal behavior, the ~/.wmux/config.json daemon knobs, and how to read those log lines.
Fixed