Baixar wmux 3.31.0 per a Windows

Added
- Per-account Claude usage in Claude Code's statusline (wmux setup-statusline). The global StatusBar widget shows one account's 5h/7d usage, but panes in a single workspace can run different Claude accounts — each pane needs its own number. A new wmux setup-statusline command sets Claude Code's statusLine to a wmux script that renders model · account · 5h N% · 7d N% on the line under the input box; because the statusline process inherits CLAUDE_CONFIG_DIR from its own claude process, every pane shows the usage of the account it actually runs on — including accounts selected by hand with $env:CLAUDE_CONFIG_DIR. The numbers are zero-cost: Claude Code ≥2.1 pipes the session's live rate_limits (5h/7d used percentage) to the statusline on stdin, so there is no extra API traffic and no token spend — the script reads stdin plus wmux's accounts.json (for the account name) and nothing else. Installs into the default ~/.claude profile and every registered claude account; a user's own custom statusLine is never overwritten. --remove / --status supported.
Changed
- Statusline reset indicators are easier to read, and the 7d window now shows one too. The ↺ reset marker sat flush against the time (↺00:30) and double-width terminal fonts could swallow the first digit — there's now a space (↺ 00:30). The 7d window previously showed only the percentage; since "how long until it frees up" is exactly what you want to know when it runs high, it now shows the remaining time (7d 84% ↺ 52h, or ↺ 2d4h beyond 48h). Re-run wmux setup-statusline to update the installed script.
Fixed
- The per-pane "Resume" chip no longer appears while the agent session is still running. The chip (the ↩ affordance that reveals the conversation UUID and types the exact --resume command) is meant to surface only once the agent has exited — typing into a live Claude/Codex TUI would land in the agent's input box. On panes without OSC 133 shell integration, though, its only "is the agent busy?" signal was a decaying activity heuristic: an agent that stayed quiet for 2 minutes (a long thinking stretch, or a finished turn waiting for your reply) read as gone, and the chip popped up mid-session. The gate is now edge-triggered on process truth: when a hook or the live banner proves the agent is running, the daemon attributes the actual agent process in the pane's process tree (one snapshot per launch, then a ride on the existing liveness batch — no new polling) and reports alive/dead alongside the session. The chip stays hidden the whole time the process lives — however quiet — and appears exactly on its exit, whatever the exit path (double Ctrl+C, /exit, Ctrl+D, a crash). Shell-integration panes keep their existing authoritative OSC 133 gate; panes where no process could be attributed keep the old heuristic as a last resort.
- macOS zsh panes now percent-encode the directory they report, matching Windows. The zsh integration hook has reported the pane's working directory (OSC 7) since v3.30.0, but it sent the raw path while wmux's parser percent-decodes what it receives — so a directory whose real name contains a percent sequence (say build%20cache) was silently tracked as build cache, and a directory name containing a raw ESC or BEL byte could cut the report short and smuggle terminal escape sequences into the pane. The zsh hook now percent-encodes the path exactly like the PowerShell/bash hooks shipped in the previous fix — spaces, Unicode, literal % and control bytes all round-trip byte-exact — and the encoder shields itself from user rc options (a KSH_ARRAYS setopt from oh-my-zsh-style configs no longer corrupts the encoding). Shell integration scripts refresh automatically (v9) on the next daemon start. (#541 follow-up)
Added
- Both resume affordances gained a --dangerously-skip-permissions toggle (Claude only, on by default). The persistent Resume chip and the post-reboot recovery pill now each show a checkbox that forces --dangerously-skip-permissions onto the resume command they type — the common case for anyone who runs Claude in bypass mode, and previously a flag you had to retype by hand every time. It's checked by default; on the chip the command preview updates live as you toggle it, and on the recovery pill a checked toggle types the whole claude --dangerously-skip-permissions --resume <id> line in one click (no more two-step assembly). Unchecking falls back to restoring the conversation's captured permission mode (e.g. --permission-mode acceptEdits) on an exact resume. The flag rides both the exact --resume <id> and the cwd-relative --continue fallback, since it's a launch preference rather than something tied to one conversation. Codex, which has no equivalent flag, doesn't show the toggle. Neither affordance ever presses Enter for you — the typed command still waits for your keystroke, so bypass is never re-granted automatically.
- OpenCode can now use the wmux MCP tools in packaged builds. The daemon authorizes the bundled MCP server by recognizing the connecting agent host's clientInfo.name (a curated first-party allowlist — the server intentionally never self-declares capabilities, since part of its tool surface maps to reserved wmux.internal methods that no declaration can grant). OpenCode wasn't on that list, so under production enforce mode every capability-gated tool it called (a2a_*, workspace_list, pane_*, …) was rejected with capability "…" was not declared by this plugin. opencode (clientInfo name captured live from OpenCode 1.17.11) is now recognized alongside claude-code and codex-mcp-client, granting it the same scoped method allowlist — not a blanket bypass; an explicit user denied still wins. (#536)
- Running many Claude sessions at once no longer degrades into hook storms, daemon false-restarts, and a frozen recovery dialog. A fleet of concurrent sessions (measured at 13, designed headroom now 30+) hit several compounding bottlenecks, each fixed at its source:
- Hook connection storms: every tool call in every session (and every parallel subagent) fired a PostToolUse hook that opened a fresh pipe connection — even though wmux keeps only ~1 activity signal per pane per 3s and discards the rest. The hook bridge now suppresses the discarded calls before connecting (per-pane stamp file, fail-open), so pipe traffic stays flat no matter how many agents run. The main pipe's per-second admission cap was also raised 30 → 120 (it was below legitimate hook traffic and its pre-auth rejections made the bridge retry, amplifying the storm), and bridge.log now rotates at 5 MB.
- Daemon false-restarts under load: two hot paths wrote the growing sessions.json synchronously on the daemon's event loop (every cd across the fleet, plus the 30s snapshot tick's re-parse + re-write). The stalls starved the health ping exactly when the fleet was busiest, and the supervisor would force-respawn a perfectly alive daemon. Both paths now persist immediately but asynchronously through the existing coalescing write queue.
- Frozen "daemon recovery" dialog: when anti-virus blocked the process probes that verify a slow-to-ping daemon, wmux popped a synchronous modal that froze the entire app — and offered to spawn a fresh daemon over the live one holding every session. The launcher now re-pings the daemon before ever asking (a daemon that answers is alive, whatever the blocked probes say), and the dialog, when still needed, no longer blocks the app.
- Per-chunk listener fan-out: every daemon output chunk woke one listener per session (O(N²) on a busy fleet) and spammed MaxListenersExceededWarning from the 11th session. Output is now dispatched through a single shared listener keyed by session.
- Fleet-wide re-renders: a 2s activity clock re-rendered every pane while any agent was running; the subscription now lives in the one small element that displays it, so idle panes do zero work per tick.
- Boot/switch storms: the daemon's per-socket RPC cap (50/s) throttled the app against itself during a mass reconnect — all panes share one authenticated socket — and could leave panes unattached; raised to match the global cap (200/s).
- wmux setup-hooks is now plugin-aware: with the wmux Claude Code plugin installed it removes (and stops writing) the duplicate settings.json hook entries that made every turn-end spawn two bridge processes.
- Session restore after an upgrade no longer comes back with blank panes and duplicate agent sessions. Upgrading (say 3.27 → 3.30) replaces the running daemon: the old one suspends every session to disk and a fresh daemon re-spawns them. But that fresh daemon only starts accepting connections after it finishes re-spawning them all, and cold-recovering a large fleet of agent panes is slow (measured ~19–23 s for 30–35 sessions). The launcher gave up waiting at a flat 15 s, decided the daemon had failed, and fell back to local mode — so every pane re-ran its workspace startup command from scratch, producing a fresh duplicate agent session next to a blank, never-reattached pane, and you had to recover each one by hand. The launcher now recognizes that a just-spawned daemon which is still alive but hasn't opened its pipe yet is recovering, not wedged, and keeps waiting (up to a 90 s ceiling) so the panes reattach to their real sessions instead. A daemon that actually crashes on startup still fails fast. (#537)