Эта страница ещё не переведена — показан английский оригинал.

Изменения

v0.23.0 #

• Audio pack. w-pack install audio sets up music production on W's PipeWire stack: Ardour 9, qpwgraph for routing, Audacity, and a plugin starter set (Surge XT, the LSP suite, Dragonfly Reverb). PipeWire takes over the JACK API — pipewire-jack replaces jack2, and stays if the pack is removed, because other apps link against it — and your account gets realtime audio priority from your next login. Ardour, Audacity and qpwgraph open in your W colours. The W assistant can work inside an open Ardour session — tracks, plugins, mixer, MIDI — once you turn on MCP HTTP Server (Experimental) in Ardour's Preferences → Control Surfaces; it is off by default because Ardour's server listens on every network interface without authentication. About 660 MiB. The installer offers the pack as well.
• W Help. SUPER+F1, Hub → Documentation and the new Open in Help button on documentation cards now open the W documentation as a site in its own window: pages by category, search, English and Russian, painted in your current theme. It is the copy on disk in /usr/share/doc/w/html, so it works offline. From a terminal: w-help [page[#section]].
• Assistant, Calendar, Power menu and Volume control start from any app. Their launcher entries carried an escape that GLib rejects, so Nemo, gio launch and the desktop portals could not start them — only W's own launcher could. Fixed.
• Arrow keys in W dialogs moved focus the wrong way in rows of buttons. Fixed.
• Assistant: remote MCP servers. A drop-in in mcp.d can name its server with URL= (streamable HTTP) instead of COMMAND/ARGS; every assistant host receives it in its own native form.
• Signed install images. From the next nightly build the ISO carries a Sigstore build-provenance attestation, made by the build itself. Check it with gh attestation verify w-*-x86_64.iso --repo tarkh/w; the edge release notes show the steps.
• A broken link to the FAQ in Keeping W up to date (English and Russian) is fixed.

v0.22.0 #

• 3D pack. w-pack install 3d sets up Blender 5 with the Cycles GPU runtime for your card (AMD HIP, Intel Arc oneAPI; NVIDIA renders through its driver). The W assistant can build and edit scenes in a running Blender — inspect them, run code, take viewport screenshots, export — with Blender's safe mode on and its telemetry off. Blender follows your W theme live; a theme preset you pick in Blender yourself is left alone. Every Open/Save dialog in Blender opens at the same size as the rest of W. The pack is heavy: about 2.5 GiB with its dependencies. On an AMD machine whose processor has integrated graphics, Cycles lists the iGPU as a second HIP device named after the CPU — untick it in Preferences → System, or renders split onto the slower chip.
• Engine upgrades for ComfyUI and InvokeAI. Both install their engine from upstream, so system updates never moved it. w-pack outdated now shows the installed and the newest stable version, and sudo w-pack upgrade moves it: the service is stopped (refused while it is busy), a rollback point is recorded, and if the new version fails to start and answer, the previous one is put back. To hold a specific version, set it with w-conf set comfyui VERSION <x.y.z> (or invokeai). Available engine upgrades show up in w-update status and in Hub → Packs with an Update button; they are never counted in the bar's update badge and w-update upgrade never runs them. The assistant can plan and apply the upgrade too.
• Ollama models live in the shared AI model store. With the ai-extra pack, Ollama now keeps its models in ollama/ inside the machine-wide store (/var/lib/w/ai-models by default) and follows w-conf set ai-models DIR like ComfyUI and InvokeAI. Models already in /var/lib/ollama are moved there once, during this update (instantly on Btrfs; otherwise copied, verified and only then removed). If both places already hold models, the store wins and /var/lib/ollama is left for you to merge by hand.
• The assistant answers everyday requests lighter. In the default goose host, tools of installed packs (Blender, GIMP, Inkscape) are no longer sent with every request; the assistant turns a pack's tools on when a task needs them, within the same turn. A plain question now carries 58 KB instead of up to 128 KB — faster and cheaper on paid models.
• Screen Recording in the launcher, next to Screenshot: it opens the region selector and starts recording; launching it again stops the take.
• Every Open/Save dialog opens at one size — floating and centred at 1000×640, in every app. Some GTK dialogs used to open tiled, and GTK saved that full-height tile as your size for every app afterwards. If your dialogs still open stretched after the update, that saved size is yours and W leaves it alone; reset it once with gsettings reset org.gtk.Settings.FileChooser window-size. LibreOffice dialogs opened from the Start Center no longer tile either.
• Terminal launches follow your terminal. Nemo's "Open in Terminal" did nothing (it looked for gnome-terminal, which W does not ship); it now opens your selected terminal in that folder. Apps marked to run in a terminal and gio launch also use it, through the standard xdg-terminal-exec package, which this update installs.
• foot gets blur. With foot as your terminal, background blur now follows the blur setting, including the Hub override; before, foot was never blurred.

v0.21.0 #

• InvokeAI pack. w-pack install invokeai sets up InvokeAI, an image generation app with a canvas editor, as a service on http://localhost:9090, under its own account. It shares the machine-wide model store with ComfyUI (/var/lib/w/ai-models): every model file already there is registered in place — nothing is copied — and deleting a model from InvokeAI's UI can never remove weights ComfyUI uses. The GPU backend (NVIDIA, AMD/ROCm, Intel, or CPU) is picked for your machine; the pack adds a launcher entry and an AI skill. It also installs opencv (four small packages) for the canvas's content-aware infill. On a machine with two GPUs (for example an AMD card plus the processor's integrated graphics), set device: cuda:0 in invokeai.yaml, or half the jobs run on the slower one. Removing either AI pack keeps the shared w-ai group for the other.
• Clicking a notification opens its source. The window of the app that sent it comes to the front, on its own workspace — Telegram shows the chat, Firefox the tab — instead of staying behind. "Saved" notifications for screenshots and screen recordings now show a preview of the capture, and clicking one opens the file in its default app.
• Screenshot editor opens at the size of your screen. Annotation (satty) used to squeeze every capture into a shorter canvas and open it shrunk; it now opens at its on-screen size, and a region capture opens right over the spot it was taken from.
• Super+Enter no longer fails at random. On systemd 262 the terminal occasionally opened no window at all (about one launch in five in testing). Ghostty now starts in a mode that avoids the race.
• Logging out no longer loses your saved session. If a save dialog stayed open during logout, the background session saver could overwrite the exit snapshot with a half-closed desktop, so windows came back missing or stretched. The saver is now paused for the exit and resumed if you cancel it.
• Gaming pack: a rootless w-pack setup gaming no longer misreports its own library subvolumes as plain folders in home snapshots, and the MangoHud config exists right after install instead of at the next login.

v0.20.0 #

• Screen recording. Super+Ctrl+R records the current screen, Super+Alt+R a region you drag out, Super+Ctrl+Alt+R the focused window. Each chord is a toggle: the same keys that start a take end it. While recording, the bar shows a pulsing dot and the elapsed time; click it to stop, right-click it to open the capture settings. Recordings land in your Videos folder. Encoding uses the GPU when the machine has one and falls back to the CPU with a notification, so a take never silently fails. Sound comes from the system output by default (AUDIO=mic or none to change it); MP4, MKV and WebM are supported. Recording all monitors at once is not possible — pick one. The update installs wf-recorder and imagemagick for you.

  One step by hand: your own bar config does not pick up the new indicator block on its own. To make the bar show it, reset that file to the shipped default:

  w-reset quickshell .config/quickshell/w/config/bar.json

  This replaces your bar.json (the bar layout) with the shipped one — the current copy is backed up first — so if you have customised the bar, you may prefer to add the recording block by hand instead.

• Screen capture settings in the Hub. Hub → Displays has a new "Screen capture" tab: encoder, sound, format, frame rate and screenshot annotator, all changeable from the keyboard. The same settings are available as w-record get|set and w-screenshot get|set, and, as everywhere in W, as w-conf set --user record KEY VALUE.

• Window screenshots and recordings now have rounded corners and a shadow. With no annotator (--copy, --save) or in Satty, a window screenshot gets transparent rounded corners matching your theme and a soft macOS-style shadow. Turn the shadow off with w-conf set --user screenshot WINDOW_SHADOW no. Window recordings can paint the corners black too (w-conf set --user record WINDOW_MASK yes); that is off by default because it forces CPU encoding for window takes. Flameshot captures are unchanged. Themes with square windows get no mask.

• Active-window screenshots include the window border. The capture used to cut the frame off along the straight edges, because the border is drawn outside the window's own rectangle. This has been so since window screenshots arrived in v0.19.0; it also applies to window recordings and to the Flameshot preselection.

• The selection overlay looks the same for screenshots and recordings: no fill, a 1 px frame and a neutral 65 % dim (Flameshot's dim is now the same depth). The overlay fades out instantly, so it no longer ends up in the capture.

• Satty opens at the size of your capture instead of a fixed 70 % window. It is capped at 80 % of the screen: a small region no longer opens as a huge window, and a full-screen capture still fits with margin.

v0.19.0 #

• Screenshots are now annotated in the same window you select in. Super+Shift+I (region), Super+I (screen) and Super+Ctrl+I (window) freeze the display and let you drag the area and draw on it in one surface, with the toolbar next to the selection — no second window opens. Flameshot does the annotating; it captures through the same Wayland path as the rest of W and wears your active theme. For region and window captures the bar steps below the overlay instead of covering the strip you are dragging through. The clipboard and plain-save captures are unchanged and still open nothing.

  Satty, the annotator this replaces, stays installed. If the overlay misbehaves for you — wrong monitor, slow to appear — switch back without reinstalling:

  w-conf set --user screenshot ANNOTATOR satty

  w-conf unset --user screenshot ANNOTATOR returns to Flameshot. One case always uses Satty and ignores the setting: a single capture spanning every monitor, which has no Flameshot equivalent.

• Two new keybinds save the screenshot straight to a file. Super+Alt+I (screen) and Super+Ctrl+Alt+I (window) write a PNG into your Screenshots folder with no annotator and no dialog. Rebindable like any other key, e.g. w-hotkeys set screenshot_screen_save "SUPER + F12".

• New W-Pack: gaming. Steam, Lutris and Heroic on one shared Proton/Wine/MangoHud/gamemode foundation, plus ProtonPlus, protontricks, gamescope, ntsync autoloading, and the game-tuned scx_lavd scheduler (installed, not enabled). Off by default, ~4–5 GiB:

  w-pack install gaming

  The pack enables the multilib repository and installs the 32-bit graphics stack your GPU actually needs before Steam resolves its Vulkan dependency — left to the resolver it picks the wrong provider and breaks Vulkan on AMD/Intel. Your game library is kept out of @home snapshots (it runs to hundreds of gigabytes and re-downloads from the storefronts), and the MangoHud overlay (Shift_F12) and the Heroic window are dressed in your active W colours. Log out and back in once after installing — the pack adds your account to the gamemode group, and group membership only applies to new sessions.

• w-pack bundles can declare a pre.sh — a root step that runs before the package list, for what a package list cannot express: enabling a repository, syncing a fresh database, pinning a hardware-specific package ahead of a virtual dependency. Only bundle authors are affected; w-pack refresh still never runs it.

v0.18.1 #

• Fixed: locking at the exact moment of a suspend no longer freezes half-way and replays the animation on resume. When suspend arrives via the lid, a bare systemctl suspend, or critical battery, the lock screen used to cut off mid-fade-in and the rest of the animation played over the desktop after wake. The lock now skips its fade-in when a suspend is already in flight — the screen is going off anyway — while normal locks (Super+L, idle, w-power sleep) keep it.

v0.18.0 #

• New W-Pack: ComfyUI. Local image/video generation as a system service on :8188, with the GPU-matched engine build (NVIDIA/AMD/Intel Arc/CPU) chosen automatically, a shared model store under /var/lib/w/ai-models, and ComfyUI-Manager enabled by default so custom nodes install from the UI. Extra launch flags go through w-conf set comfyui ARGS "..."; extra environment variables go in an admin-owned 20-local.conf drop-in. Off by default — install with w-pack install comfyui.

• New W-Pack: LocalSend. AirDrop-alike file transfer, integrated into Nemo's right-click menu ("Send with LocalSend") and scoped to the home firewall zone. Its accent colour follows your W theme automatically, the same way every other GTK4/libadwaita app does.

• w-conf now accepts -- to end option parsing, for setting values that start with a dash:

  w-conf set comfyui ARGS -- --lowvram

• Dev pack: added cmake, clang, and ninja to the common toolchain.

• Fixed: ai-extra could silently skip installing the GPU-accelerated Ollama backend if the plain ollama package was already present on the machine (e.g. pulled in earlier as a dependency) — it was mistaken for a deliberate choice and left alone. It now installs the right backend regardless.

• New w-power sleep command. Suspending now waits for the lock screen to finish drawing first, avoiding a black or half-drawn locker on resume; the Power menu and idle-timeout suspend now use it.

• Fixed: some RGB mice/keyboards could still come back at their factory colour after resume, even with the resume hook introduced in v0.17.0 — systemd could kill it before it finished detecting the device. It now runs in a way that survives to completion.

v0.17.0 #

• Home folders now follow the system language. If your interface is in Russian, ~/Downloads becomes ~/Загрузки, ~/Pictures → ~/Изображения, and so on for all six standard folders — including ~/Screenshots, which the screenshot tool (Super+Shift+S) now resolves dynamically instead of hardcoding. The rename happens at the next login after the update and is one-time: folders you moved are left alone. To opt out:

  w-conf userdirs LOCALIZE=no

  If you run a non-Russian locale this changes nothing for you.

• OpenRGB devices re-colour after suspend. Keyboards, mice and other OpenRGB-supported lighting used to go back to their hardware default after waking the machine and never caught up until the next theme switch. A new system-sleep hook (w-rgb) re-applies the theme's RGB palette on resume.

• w-fingerprint now controls the fingerprint scanner on the lock screen. Two modes: native (the default, unchanged behaviour — fprintd authenticates through PAM) and wake (the scanner stays armed during hyprlock, so touching it unlocks the screen without reaching the keyboard). Switch with:

  w-fingerprint lock-sensor wake

• Toolchain state hygiene: Go builds no longer litter ~/go. GOPATH, GOMODCACHE, GOCACHE and GOBIN are now routed into ~/.local/ and ~/.cache/ — outside @home snapshots — at every login, across TTY, SSH and the graphical session. The subvolume is carved automatically; a pre-existing ~/go is left untouched (safe to delete by hand). Future tools (cargo, gradle, npm…) will follow the same pattern.

• w-pack deploy_managed now preserves file modes. A managed file with an executable bit (e.g. the graphics pack's mcpinkscape hook) kept it in the bundle tree but arrived on the machine as 0644. It now keeps the mode the file has in the pack — matching how apply_rootfs and the directory branch (rsync -a) have always worked.

• w-sync re-applies packs when the pack tooling itself changes. Previously only changes under scripts/packs/ triggered w-pack refresh; a fix to w-pack or its deploy SDK now does too.

• Hyprlock: fingerprint-aware auth config shipped. A new ~/.config/hypr/hyprlock-auth.conf template enables w-fingerprint lock-sensor wake to work out of the box on fresh installs. If you use w-reset hyprlock (e.g. after the v0.16.0 lock-screen rework), the auth fragment is included automatically.

v0.16.1 #

• Session restore: tabbed windows and Firefox windows come back correctly. Two longstanding defects, both present up to and including v0.16.0:
  • Window groups (tabbed windows) came back scattered over the workspace instead of as groups. Every member of a group reports the same content rectangle, so the restore could not read the split tree and fell back to tiling the windows in arrival order. A group is now one tile of the tree — its tab bar accounted for — and is reassembled after the tree is in place: tabs in the saved order, the tab that was on top on top. Dwindle only; a floating group returns as separate floating windows and a group's lock state is not restored.
  • Firefox windows came back as one window of tabs plus one blank (or swapped). W launched firefox once per saved window, but a second launch is a remote command to the instance already running and opens a blank window. Firefox keeps its own session, so it is now launched once and each window it brings back is placed where a window with the same title was. At logout W also finishes Firefox's own "closed in series" memory in its session file — Firefox's own pass stops at any older closed pop-up and drops the rest, which on a real profile is why a second window went missing.
  • Firefox restores the tabs inside a window only if its own Settings → Home and startup → "Open previous windows and tabs" is on; W does not switch that on for you. Hub → System → Session now says so when you turn restore on.
  • If you keep personal relaunch overrides in ~/.config/w/session-apps.tsv, the file gains an optional flag column; two-column rows keep working. Flags: own-session (the Firefox case above) and = in the command column to keep the recorded command line.
• RGB lighting is more reliable. The theme's colour was sent to the openrgb daemon by guessing when it had come up; on a slow login the send happened before the daemon knew its devices and was lost. W now waits until openrgb reports devices (up to 20 s) before colouring, so RGB matches the theme on the first try. This affects the RGB feature introduced in v0.16.0.
• The desktop guide (w-guide, Hub → Documentation) and the on-box assistant explain that window groups come back as groups and where Firefox's tab-restore setting lives — in English and Russian.

v0.16.0 #

• Lock screen reworked (hyprlock + hypridle). Two defects, both present in every release so far:

  • the blurred workspace was replaced by a flat colour — at random on a manual lock, almost always on wake from suspend. The cause is a race in hyprlock 0.9.6 that W's immediate_render = true opened; it is now off, so the screenshot is captured before the session is locked (about 30 ms later, not noticeable). When a fallback colour is ever needed it is the theme's surface tint, not pure black.
  • the machine could suspend unlocked: hypridle released its sleep inhibitor before hyprlock had a frame. inhibit_sleep = 3 is now pinned, and the lock command switches the display on first, so a blanked screen still yields a valid screenshot. The background is also darker and blurrier (three blur passes, brightness 0.65) so the clock and the password field read cleanly over bright windows.

  hypridle.conf and the lock colour file are rendered by W and arrive with w-sync. hyprlock.conf is yours and an update never overwrites it, so after the update run, as your user:

  w-reset hyprlock

  It backs up your current ~/.config/hypr/hyprlock.conf first. If you have customised that file and want to keep it, make the two changes by hand instead: in general set immediate_render = false; in background set color = $w_lock_bg_fallback, blur_passes = 3, brightness = 0.65, contrast = 0.85, vibrancy = 0.2, noise = 0.015. Takes effect at the next lock.

• RGB lighting follows the theme. Keyboards, mice, mainboards and other OpenRGB-supported devices are recoloured on every theme switch and login, using three pigments calibrated for LEDs rather than for a screen. Hub → Appearance → Settings gains an RGB on/off toggle and a saturation row (soft / medium / crisp); the same from the terminal: w-appearance rgb on|off, w-appearance rgb-level soft|medium|crisp. On by default at medium; set rgb off to leave your devices alone. The update installs openrgb and i2c-tools; verify detection with openrgb --list-devices.

• New pack: virt. Desktop virtualization — libvirt + QEMU/KVM, virt-manager for full control and GNOME Boxes for one-click VMs. Install with w-pack install virt; it defines the default network and storage pool, so virt-install and virt-manager work on first launch.

• OpenCode is a W AI host. w-ai host install opencode, then pick it in Hub → AI or run w-ai --host opencode. It gets W's tools, approvals and identity like Claude Code and Codex; besides its own Zen subscription it accepts any models.dev provider (opencode auth login -p <provider>).

• AMD GPUs: the GPU step no longer fails. It asked pacman for mesa-vdpau, which Arch no longer ships, so the transaction — and with it a fresh install on AMD hardware — stopped there. Removed; amdgpu_top (a GPU monitor) is installed instead.

• usbutils (lsusb) is now part of the base system.

• w-style status labels ANSI slot 5 "ansi magenta" instead of "accent (magenta)" — the accent is the primary row above it.

• For people building W itself: w-pack list --all shows maintainer-only bundles (new w-dev); scripts/apply.sh refuses to reconfigure bare metal from an unmanaged checkout unless W_APPLY_ON_HOST=1; W_SUDO=run0 lets build-iso.sh elevate through polkit instead of a sudo password.

v0.15.0 #

• New pack: graphics. GIMP 3.2, Inkscape 1.4, Krita 6, and G'MIC filters, themed to match W out of the box. Install with w-pack install graphics.
• The on-box AI assistant can drive GIMP and Inkscape directly, through a new per-pack MCP-server channel: install graphics and the assistant gets working tools for both editors, visible in w-ai status as MCP extra:.
• AI assistants no longer ask twice for W's own tools. Goose, Codex, and Claude Code each used to raise their own approval prompt before running a W-curated tool — installing a knowledge pack is already consent (the polkit prompt for anything privileged is unchanged). Goose's first-run telemetry prompt is now also answered off by default.
• The tool list assistants receive is markedly smaller — cut by roughly a fifth to nearly half, depending on host — so more of a small local model's context stays free. Same tools, leaner descriptions.
• The assistant's skill catalog is generated once and read the same way on every host (Goose, Codex, Claude Code); the old flat llms.txt file is gone (removed automatically on update).
• The assistant now knows how to work with LibreOffice documents headlessly (convert/replace text via UNO) and how to find, pull, and run containers with quadlet/Podman — ask it and it can carry these out itself.
• Obsidian vaults created after your first login now get W's theme immediately, not just at the next login.
• Bitwarden's floating window is slightly bigger (900×800 instead of 800×790).

v0.14.0 #

• Display scale: only values Hyprland actually accepts. Hyprland takes a scale factor only when both the width and the height of the mode divide by it exactly (on its 1/120 grid); anything else it rejects on reload, substitutes its own nearest value with a red notification, and from then on your config no longer matches what is on screen. W checked only the width, so the Hub's Displays panel offered scales such as 1.5 on 2560x1600 or 1.75 (may blur) on 2880x1800 that the compositor would never apply. Now:
  • the Hub's Scale list is Auto plus the scales the compositor accepts for the selected resolution, in roughly quarter steps (a panel that cannot divide by 1.5 offers 1.6 instead). The fixed preset list and the may blur tag are gone; a value you set by hand stays in the list even when it is not one of the picks;
  • new w-monitor scales <output> [WxH] prints the full accepted set for an output (or for any resolution you name), with the recommended picks marked;
  • w-monitor scale <output> <value> refuses a value Hyprland would reject and names the nearest accepted one, instead of writing a config the compositor then overrides. Arrives with w-sync and takes effect at once — the shell is reloaded as part of the update. A scale already saved in your config is left alone; if Hyprland has been substituting it, pick one from the new list. The displays guide, w-monitor --help and the on-box assistant describe the change.

v0.13.2 #

• Nothing changes on an installed machine. Second half of v0.13.1: the nightly build's check now also has ImageMagick's SVG decoder (librsvg), without which the boot-logo verification failed instead of skipping. w-sync sees a version bump and nothing else.

v0.13.1 #

• Nothing changes on an installed machine. This release only fixes the project's own nightly build: the check that runs before the ISO is assembled needed ImageMagick (v0.13.0's theme-tinted boot logo is verified with it) and the build container did not have it, so the first nightly after v0.13.0 stopped before building anything. The tool is now installed there. w-sync will pick this up as a version bump and nothing else.

v0.13.0 #

• Fixed: DNS lookups timing out for 10+ seconds on some networks — since v0.1.0, and until now only on certain Wi-Fi/router combinations (seen on a MacBook's Broadcom Wi-Fi behind a home router). W's resolver reaches its DNS-over-TLS servers with TCP Fast Open, and on such paths those streams complete the handshake and then silently drop every byte; about one in five still worked, so the resolver never fell back to plain DNS and every lookup simply hung — pacman reported Resolving timed out on every mirror, and app bundles failed to install during first boot. TCP Fast Open is now off system-wide (/etc/sysctl.d/80-w-dns.conf; nothing else on a W desktop uses it client-side, and w-kernel harden off leaves it alone). Arrives with w-sync and takes effect immediately, no reboot. The DNS step now also runs a resolve probe after applying the policy and logs a warning if it fails. The network guide and the on-box assistant describe the change. New installs additionally: a failing app bundle at first boot no longer cancels the ones after it — the failed ones are listed in the log and retried on the next boot as before.

• The boot logo follows the theme's colour. A theme without a logo/ of its own inherits W's brand mark and now gets it painted in the theme's accent on both the Plymouth splash and the GRUB menu (the primary container colour on dark themes, the primary colour itself on light ones, where the container is too pale to read on the canvas). The built-in w theme renders byte-for-byte as before. Themes generated with w-theme new from now on carry the two new keys (W_PLYMOUTH_LOGO, W_GRUB_LOGO); a theme you generated earlier keeps the plain mark until you rebuild its palette with w-theme edit <name>. A theme that ships its own logo/ is untouched. The theming guide and the on-box assistant know about it.

v0.12.0 #

• The scratchpad has a button in the bar. The workspaces block now shows Hyprland's special workspace as a glyph button (nf-md-layers, 󰌨) after the numbered ones: it lights up while the scratchpad is shown on the focused monitor, and a click toggles it — the same as Super+S. Until now the scratchpad, once anything was sent to it (Super+Shift+S), appeared in the bar as a raw special:scratchpad text button sorted before workspace 1, and clicking it did nothing useful — since v0.1.0. The glyph is overridable like every other block icon: "icon" in the workspaces block of ~/.config/quickshell/w/config/bar.json (an empty string or a missing key keeps the built-in one, so a bar.json you already edited needs no change). w-session's own parking workspace never shows up there. Arrives with w-sync; the bar restarts on its own.

v0.11.0 #

• Codex is a full assistant host. OpenAI's Codex CLI now gets W's layers the way Claude Code does — actions (w-mcp), identity, and W's knowledge through w-mcp — as launch flags, so nothing is written into ~/.codex and a codex you start yourself in a project of your own stays a plain Codex. Until now w-ai host install codex only pointed at a README and the W layers had to be pasted into ~/.codex/config.toml by hand. Now:

  w-ai host install codex     # OpenAI's standalone installer, per-user (~/.local/bin)
  w-ai host login codex       # browser sign-in with your ChatGPT account
  w-ai profile use codex      # or: Hub -> AI -> the codex profile

  From then on w-ai, w-ai ask and the Super+W palette open Codex. The shipped codex profile runs on the ChatGPT login and W exports no key; set PROVIDER=openai in the profile to pay per token against an API key instead (w-ai key set openai). Codex picks its own model in the session (/model); a MODEL= in the profile pins one. If you had pasted the old [mcp_servers.w-mcp] block into ~/.codex/config.toml, it is redundant now and harmless either way. The AI guide and the on-box assistant know the new commands; /usr/share/w/ai/hosts/codex/README.md has the details.

• Fixed: the installer broke on disks whose model name contains spaces — since v0.1.0. The disk step split its list on whitespace, so a Samsung SSD 970 EVO Plus either left the wizard spinning behind the previous message (it looked like a hang at Cloning the repository…) or produced shifted, nonsensical entries. The list is built properly now, and empty card readers and the live USB you booted from are no longer offered as install targets. New ISO only — installed systems are not affected by this line.

v0.10.0 #

• New bundle: telegram. Telegram Desktop from Arch's own repository, with the active W theme rendered into Telegram's palette — colours, and a plain chat background in the theme's tone — and kept in step when you switch themes. On an account that has never launched Telegram the theme is in place from the very first start; an account with history applies it once from Chat settings and Telegram re-reads it on its own after that. A checkbox in the installer, or sudo w-pack install telegram.

• Per-app window rules are drop-in files. W's own live in /usr/share/w/hypr/rules.d/<app>.lua (the screenshot annotator's rule moved there from hyprland.lua); yours go in ~/.config/hypr/rules.d/<app>.lua and win over W's for the same app. A broken drop-in shows up in hyprctl configerrors without taking the others down. Bundles ship rules this way and w-pack reloads running sessions on install and remove, so a rule works at once, not after the next login. Machines installed before this release need one step: the loader lives in your own hyprland.lua, which updates do not touch. Once, after w-sync update:

  w-reset hyprland .config/hypr/hyprland.lua && hyprctl reload

  (a backup is taken automatically). Without it the Bitwarden and screenshot-annotator rules below stay inert.

• Bitwarden pops out as a dialog. The window the SSH agent raises from the tray for every signing request — unlock, approve — now floats centred at dialog size instead of wedging itself into the tiling layout. Turn on Close to tray in its preferences: with the upstream default, Super+Q quits the app and the SSH agent inside it.

• Fixed: ~/Pictures belonged to root. Since v0.1.0 the install created it on the way to Pictures/Screenshots as root, so anything else wanting to write there — a browser download, a file manager — was refused. The next update hands the directory back to you.

• Fixed: installing or removing bitwarden over a root SSH session ended in «could not switch the SSH agent» (it worked from the Hub and sudo). The bundle's per-account steps now run with the account's own runtime.

• New ISO: the installer checklist offers the telegram bundle. Installed systems are not affected by this line.

v0.9.0 #

• W speaks Russian. Choose Russian as the system language and everything W draws itself follows: the desktop shell and the login screen, the Hub, every notification the W tools send, the W entries in the launcher (searchable in Cyrillic), the built-in help of all 36 w-* commands, the w-info catalog and the offline reference in Hub → System → Documentation — 36 reference pages now exist in Russian next to the guides translated in v0.8.0. What the commands print to the terminal stays English, and so do third-party programs that ship no translation. Any other language, or LANG=C, falls back to English.

• A language is more than LANG: new w-langpack. It knows what a locale needs beyond the setting itself — Firefox in that language, a spell-check dictionary, translated man pages, a console font that can draw it — and installs what is missing (w-langpack plan, sudo w-langpack apply). Switching language with w-locale set or in Hub → Input sets the font at once and offers the packages. The profile also runs on every update, so the next w-sync update fills in what your current locale is missing.

• No more empty boxes. The Noto font family (about 415 MiB: all scripts, CJK, colour emoji) is now part of the base system, so a Japanese web page, a Telegram emoji or a Hindi filename renders instead of showing tofu — and Japanese renders with Japanese glyphs, not Korean ones. Installed machines get the fonts on the next update.

• Fixed: an update could reset the text-console keymap to US. The keymap chosen at install lives in /etc/vconsole.conf, and since v0.1.0 every update that touched the system files overwrote it with the stock us. The file is now yours; updates leave it alone. Check with cat /etc/vconsole.conf and, if it says us and you chose something else, set it back with localectl set-keymap <map>.

• Bundles can be removed. sudo w-pack remove <bundle> is the declared reverse of install: it takes W's own wiring out — services, managed config (backed up first), theming, AI skill — and leaves the packages unless you add --packages. Your data is never deleted, and paths that hold any are listed. w-pack unsetup <bundle> does the same for your own account. Until now removing a bundle's packages by hand did not stick: the next w-sync update put the bundle back — for ai-extra that meant the service and a 635 MB model. The Packs panel in the Hub has a Remove button for both; the irreversible package half is deliberately terminal-only.

• New bundle: office. LibreOffice (with the metric-compatible fonts that keep Windows-authored documents laid out as intended), pdfarranger and xournalpp for cutting up and annotating PDFs, and Obsidian for notes — the first freeware pick in a bundle, fetched by your own pacman from Arch's servers, never shipped in the ISO. Obsidian gets the W colour theme as a snippet in each of your vaults, switched on unless you have chosen your own snippets. A checkbox in the installer, or sudo w-pack install office.

• The Hub asks before it deletes anything. One confirmation dialog for every destructive action — a theme, a layout, a hotkey or profile, an AI profile, a bundle — that tells you what actually happens (which theme takes over, that keys stay in the keyring) and starts on Cancel, so a stray Enter cannot confirm. The notification history is grouped by day.

• w-term --hold -e <command> keeps the terminal window open after the command finishes and asks before closing — for commands whose output is the point. w-locale list --names prints each language's own name and country; the Hub language picker shows them.

• Fixed: reference pages. Every page's subtitle ended in a double full stop since the reference appeared; the w-session summary was cut off at a comma in the catalog and the docs; the w-style page was missing the second half of its help, examples included.

• ll and la no longer print a header row. In eza that -h is --header, not human-readable sizes, and the header was the one English line left in a listing on a translated system. Existing ~/.zshrc files are untouched; the change reaches you on a new install or with w-reset shell.

• New ISO: the language menu shows names in their own script («Русский», not «Russkiy»). Installed systems are not affected.

v0.8.0 #

• The documentation now covers every Hub panel — and speaks Russian. Six new guides join the four already shipped: displays, input, network, power, AI and packs, ten in total. The ? in a Hub tab header opens the matching page from twelve places now instead of two, including tabs that had no way in at all — System → Security, and System → Rollback, which opens the recovery page. The whole tree — start page, FAQ and all ten guides — is translated into Russian and appears on its own when the system language is Russian. The per-command reference stays English: it is generated from the commands' own help text. The project README links into the tree too, so the docs are readable on GitHub without installing W.

• Fixed: on a Russian system the documentation would not open at all. Every page is looked up in your language first and falls back to English. Since v0.7.0, where that fallback arrived, a race in the file loader killed the retry, so a non-English system got an empty viewer and nothing else. It falls back correctly now — and there is a Russian tree for it to find.

• The documentation reads better. Rewritten renderer: real spacing between paragraphs and lists, themed links, commands set off from the prose, and code shown on its own block with shell highlighting. You can select the text with the mouse and copy it with Ctrl+C, and scroll a page from the keyboard. Documentation in Hub → System is no longer stepped over when you move through the Hub with the arrow keys.

• Flatpak is no longer installed with the base system. It is a second software ecosystem with its own runtimes and store, and a machine that takes its software from the repositories never needs it — so it is now an optional bundle: Flatpak + Flathub, the Bazaar store, Flatseal, and the W theme carried into the sandbox. New installs get it as an unchecked box in the installer, or later with sudo w-pack install flatpak.

  On machines already running W nothing is removed — Flatpak stays installed and keeps working. If you use it, run sudo w-pack install flatpak once to put it under bundle management; after that sudo w-pack refresh flatpak tops up the sandbox theming that a system update used to do on its own.

• Fixed: four reference pages showed escaping the commands never print. The pages for w-ai, w-mirrors, w-pack and w-theme carried stray backslashes in front of quoted command names.

• Fixed: the FAQ named a command that does not exist. Recovering from a black screen is w-monitor reset --all, not w-monitor reset all.

• Fixed: a selective update could run its steps out of order. The module list w-sync used had drifted from the one the system applies: eight modules were missing from it and got appended at the end, so time settings, power, mirrors and five others ran after package installation rather than in their place. It never produced a failure. Both lists are now generated from one registry, and a check proves they agree.

v0.7.0 #

• W's documentation now ships with the system and opens on the desktop. Press SUPER + F1, or pick Documentation in Hub → System, for a start page, an FAQ, four guides (desktop, updates, theming, security) and a reference page for every w-* command. It lives on disk at /usr/share/doc/w/, so it works with no network — including the times you need it most. Some Hub tabs now carry a ? in their header that opens straight to the matching page. The reference pages are generated from each command's own help text, so they cannot drift from what the commands actually do.

• Rollback has a place in the Hub. Hub → System → Rollback lists your snapshots — which one you are running, when each was taken and what it was taken for — with a Restore button on each. It drives the same w-rollback the terminal does, so nothing can happen here that could not happen there. Opening the tab asks for your password every time: reading the snapshot list is a privileged call, and W does not hold that grant open between uses.

• Fixed: a rollback could fail halfway through on an encrypted install. snapper always creates its .snapshots as a btrfs subvolume; W keeps its snapshots elsewhere and mounts them over that path, so the empty leftover sat nested inside the system root. On a rollback, Limine's restore tool moves every subvolume nested in the outgoing root — and that one is a live mount point, which the kernel refuses to move. The restore stopped at "Moving child subvolumes", left /var/lib/machines and /var/lib/portables in the old system, and the system it kept could not be removed afterwards with snapper delete. Every encrypted install made so far carries the leftover; this update clears it, on existing machines as well as new ones. If W reports that it left a non-empty .snapshots alone, nothing is wrong — it found files there and will not touch them. Unencrypted (GRUB) installs were never affected.

• A rollback no longer finishes on a red error line. "kernel.default: Read-only file system" appeared right after "Restore complete". A rollback registers the system it replaces as a snapshot, which makes the running system read-only on the spot, so a small bookkeeping write failed — by design, harmlessly, and it was never meant to be shown.

• RECOVERY explains the read-only system after a rollback. The recovery page that ships to disk now says why the machine you are on becomes read-only the moment a rollback runs, and that the answer is simply to reboot. Its cleanup instructions also no longer list /.snapshots among the nested subvolumes — after the fix above, it is not one.

v0.6.1 #

• Nothing on your machine changes in this one. It repairs W's own nightly install test, not the system that test installs. The check used to start asking questions the moment the freshly installed machine answered on the network — which is a second or two before the machine has finished starting up. So it called the lock screen "not running" a fraction of a second before it started, and turned a perfectly good build red; on another night, with the same code, it would have reported a clean system without ever having looked at the end of the boot. It now waits for startup to finish before it asserts anything. No file that runs on an installed machine was touched, so this update costs you nothing but the version number.

v0.6.0 #

• Fingerprints can be enrolled from the desktop. Until now the authentication card and the lock screen could only use a finger enrolled by typing fprintd-enroll into a terminal. Hub → Input → Fingerprint now lists all ten slots and adds or removes any of them. There is a CLI behind it — w-fingerprint (status, enroll <finger>, delete <finger>, fingers) — and it is the only supported way to talk to the reader; the Hub calls nothing else. Neither needs root or a password prompt: fprintd lets you manage your own prints. The tab appears only on a machine that has a reader. One thing that surprises people: fprintd answers an active desktop session only, so w-fingerprint status over SSH reports the reader as unavailable (reason=unauthorized) while it sits right there — run it from your session. The on-box assistant knows the command too.

• Two 12-hour clock formats in the bar. The clock cycles through its formats when clicked; hh:mm A and hh:mm:ss A now sit beside HH:mm and HH:mm:ss. This is a default, and bar.json is yours once it exists, so the new presets reach new installs and newly created accounts only. To take them on a machine you already run — this discards your own bar settings:

  w-reset quickshell config/bar.json

• Russian documentation. README, SECURITY, CONTRIBUTING and RECOVERY now have Russian versions under docs/ru/, with a language line under the heading of every page. English stays the default. RECOVERY is the page that also ships to disk — whoever reads it is offline by definition — so an installed machine now carries both /usr/share/doc/w/RECOVERY.md and /usr/share/doc/w/RECOVERY.ru.md, and the on-box assistant points a Russian-speaking user at the second one.

v0.5.1 #

• Cancelling an authentication prompt no longer counts against you. Dismissing the auth card — or pressing "Use password" — was recorded by pam_faillock as a failed login. Three cancels and the password prompt refused you for ten minutes ("Authentication token manipulation error"), while a fingerprint still let you straight in. On current polkit the PAM helper is a socket-activated system unit, so cancelling only closes the socket: the helper lives on, reaches pam_unix with a dead conversation, and every abandoned window was tallied. /etc/pam.d/polkit-1 now carries its own copy of the password chain, in which a dead conversation fails on the spot without touching the counter. Wrong passwords are still counted and the lockout still works — only cancelling stopped being an attempt. If a machine is locked out right now, it clears by itself after ten minutes, or at once with:

  sudo faillock --user "$USER" --reset

• "Use password" during a fingerprint prompt now actually switches to the password. v0.5.0 announced this, but it never worked on any machine: the button dropped its flag into $XDG_RUNTIME_DIR, while the stack that reads it runs inside a unit with ProtectHome=yes, where /run/user does not exist. The flag was invisible, so the password you typed simply sat in the queue until the reader's 30-second timeout — exactly the behaviour that release said was gone. The channel moved to /run/w/fp/, and the button takes effect immediately.

• Cancelling a fingerprint prompt frees the reader instead of holding it for half a minute. The rough edge named in v0.5.0 — the reader staying busy for the rest of its 30 seconds after you switch to the password — is gone, and on readers with a lamp it no longer stays lit. W now ends an abandoned verification the only way pam_fprintd accepts, by restarting fprintd through a rule scoped to that one unit, that one verb and your own active session: 83 ms to the password prompt, against 27 seconds before. The card also stops offering the finger while the reader is still busy, rather than asking you to touch a device that cannot answer.

• The card no longer says whose password it wants when there is nobody else it could be. "Password for user X" is shown only when the account is not the one you are logged in as, or when there is a choice to make.

  Updating deploys all of the above, but the running session keeps the auth agent it started with. Restart it — or just log out and back in:

  systemctl --user restart w-authd

• Under the hood. The nightly image build now initialises pacman's keyring before upgrading it, so a build can no longer finish green while the image quietly distrusts the new master keys. The issues the nightly canary opens are prefixed [canary] and name the stage that actually failed, instead of reading "Edge image fails to install" over a run in which the image installed, booted and passed every check.

v0.5.0 #

• Choosing a kernel and switching hardening on now work on encrypted installs, where both were silently doing nothing. w-kernel and the hardening step wrote /etc/default/grub and ran grub-mkconfig unconditionally. An encrypted install boots through Limine, which reads /etc/kernel/cmdline and never looks at GRUB's config — but GRUB is installed there all the same, so the write succeeded, the regeneration reported success, and nothing anywhere said otherwise. On every encrypted machine the result was that w-kernel set did not change which kernel booted, and the hardening boot parameters never reached the kernel: w-kernel harden status answered mixed — sysctl drop-ins on, not one cmdline token. w-kernel now detects the bootloader and writes what that bootloader actually reads.

  Updating fixes this for you — the hardening profile is applied again through the corrected path — but boot parameters only take effect at the next reboot. To see where a machine stands:

  w-kernel harden status
  cat /proc/cmdline

• The kernel is now selectable from the Hub, and linux-lts is offered as a third choice. W Hub → System → General lists zen (W's default), vanilla and lts, marking which is installed, which is the default and which is running; picking one opens a terminal, because installing a kernel is a long download plus a DKMS rebuild. Hardening became a switch in W Hub → System → Security. That Security tab used to be hidden on machines without Limine — that is, on every unencrypted install, which is exactly where bootloader-independent hardening was least reachable. It is now always shown, and the Secure Boot rows inside it are what depends on Limine.

  From a terminal:

  w-kernel list
  sudo w-kernel set lts
  sudo w-kernel remove vanilla
  sudo w-kernel harden off

  set never uninstalls anything, so switching back to a kernel you already have is offline and instant, and remove refuses to take away the running or the default kernel. A kernel change now also raises the "reboot needed" badge in the bar.

• Your SSH agent is now something you choose: w-ssh. W still ships gcr-ssh-agent and still defaults to it, so nothing changes unless you ask. What changed is that the session no longer hardcodes that one agent's socket: SSH_AUTH_SOCK is a fixed path — $XDG_RUNTIME_DIR/w/ssh-agent.sock — with a symlink behind it pointing at whichever agent is active. A password manager that carries SSH keys takes the slot in one command:

  w-ssh list
  w-ssh use bitwarden
  w-ssh status

  Bitwarden, 1Password, a plain ssh-agent on a fixed path and gcr are known out of the box; any other agent is one SOCKET_<name>= line in ~/.config/w/ssh.conf (or /etc/w/ssh.conf for the whole machine) and it appears in w-ssh list with no code change — w-conf cat ssh shows the catalogue. w-ssh use none leaves the session with no agent at all. The stable path is exported by the session, so it is in place after your next login.

  One thing to know before switching: an agent holding a full vault offers every key it has, and sshd gives up after five (MaxAuthTries), so a large vault can fail to log in anywhere. w-ssh sync fixes that by generating per-host IdentityFile/IdentitiesOnly selectors from the keys in the agent, reading each key's name in the vault as the list of hosts it belongs to. They live in a generated file included from ~/.ssh/config; w-ssh include adds that include line, which has to sit at the very top of the file to apply to more than one host — w-ssh status says so if it does not.

• A Bitwarden bundle, installed only if you ask for it. w-pack install bitwarden — it is not offered during installation and nothing is pulled in by default. W ships no password manager and takes no position on which one you should use; what it ships are the slots one plugs into — the polkit prompt that biometric unlock speaks to, the Secret Service that stores the unlock key, the clipboard filter that keeps copied passwords out of history, and now the SSH agent above. The bundle is just the wiring: the app from the Arch repository (that package ships the polkit action, so biometric unlock works without the app's own setup step), its SSH agent socket agreed with w-ssh, and tray autostart.

• A fingerprint prompt now looks like one. While the reader is being waited on, the authentication card shows a fingerprint glyph and no password field — there was nothing you could usefully type into it — and turns itself into the password card the moment fprintd gives up. "Use password" is now immediate, where previously pam_fprintd held the conversation for its full 30-second timeout and a password typed in the meantime simply sat in a queue. This needs a reader with an enrolled finger (fprintd-enroll); a machine without one sees exactly what it saw before. One rough edge remains: after you switch to the password, the reader itself stays busy for the rest of its 30 seconds. It no longer holds anything up.

• The W mark is the same size in the boot splash, on the login screen and on the desktop. On a HiDPI panel it was three different sizes: Plymouth drew it at double size and blurry (it applied GNOME's device-scale heuristic and then bilinearly upscaled every sprite), the greeter at a fractional scale such as 1.5 picked a 4K wallpaper master for a 2K screen, and only the desktop had it right. All three now derive it from the panel they are actually drawing on, and the splash logo is rasterised from the theme's vector at the exact pixel size (this adds librsvg). Ordinary 1x monitors were already consistent and look unchanged. A side effect worth having: on a multi-monitor setup, a head with a different DPI is no longer forced to share another head's scale.

• sudo w-reset no longer breaks the thing it restores. Two defects, both present in released versions, both invisible to any static check because they only happen at runtime. Restoring a single file lost its executable bit — the real damage being w-reset updatesys, which left /usr/bin/w-sync unable to run, so the update client could no longer deliver its own fix. And the vendor copy of /etc/w/update.conf had CHANNEL=stable hardcoded even though the channel is detected from the checkout, so restoring it quietly took the machine off edge. The mode now comes from the pristine copy, and the vendor copy is built from the same detection a fresh install uses.

  If you ran sudo w-reset updatesys on a machine before this release, it needs one command by hand before it can update at all, and a look at the channel afterwards:

  sudo chmod 755 /usr/bin/w-sync
  w-sync status

• The assistant's desktop knowledge is split into three skills. w-desktop had grown to carry five subsystems and fourteen tools and sat one byte under the size limit skills are held to, which meant new knowledge about one subsystem was being paid for by dropping knowledge about another. It now covers the compositor, windows, key bindings, the shell UI and screenshots; w-displays covers monitors, the greeter's screen and night light; w-session covers session memory and named layouts. Every tool is still there under the same name. Asking what your resolution is no longer drags layouts and compositor configuration into the answer.

• Smaller things. A first boot no longer logs Failed to start Ghostty: the vendor unit was the only one under the graphical session target without a "needs a display" condition, and W has a path on which that target comes up without one — the terminal warm-up itself was never affected. Limine's snapshot manager now says when it evicts old snapshot boot entries to stay inside the EFI partition, instead of doing it silently — which matters more now that a machine may carry three kernels.

v0.4.0 #

• W now refuses an update it cannot verify. Until this release, trusting an update meant trusting the transport: HTTPS, plus the assumption that the address in /var/lib/w/src/.git/config still pointed at the real W. Everything w-sync does after a pull runs as root, so that assumption was worth root on your machine. Every release is now published as an annotated tag signed with the W release key, and w-sync update verifies that signature against the key shipped on the machine (/usr/share/w/update/w-release.allowed_signers) before it takes a snapshot and before it pulls. A tip that carries no release tag, or one not signed by a trusted key, stops the update dead and leaves the machine bit-for-bit as it was. w-sync status gained a Signed: line saying which way this is set.

  Verification is on by default, including on machines installed before this release. If your checkout tracks something other than the official repository — a fork, or a branch of your own — there are no signed release tags to find, so this update records VERIFY_SIGNATURE=no in /etc/w/update.conf for you. A value already in that file is never overwritten.

• If w-sync update has been dying on what looked like a credentials problem, this release is the fix — and the fix has to arrive by hand once. The install image is tagged edge, and the nightly build moves that tag. git fetch --tags refuses to move a tag it already has and fails the entire fetch, so on a machine that had once seen that tag the next w-sync update died before it could do anything, and reported it as though the repository could not be reached. The fetch now passes --force --prune-tags. Since the broken code is the code that would have to run to deliver its own fix, a machine already in this state needs one command first, run as the user who owns the checkout:

  git -C /var/lib/w/src fetch --tags --force --prune-tags origin
  w-sync update

• The status bar is now composed by you, screen by screen. W Hub -> Appearance -> Bar is a new tab listing every block the bar can show, once per connected monitor, with the whole bar switchable off per screen as well. The same thing from a terminal — or through the assistant, which knows these commands:

  w-bar status
  w-bar block HDMI-A-1 tray off
  w-bar monitor eDP-1 off

  Choices live alongside the rest of your bar settings in ~/.config/quickshell/w/config/bar.json. Switching a bar off keeps them, so switching it back on restores the same composition. Bar position moved out of Appearance -> Settings into this new tab; w-appearance bar-position is unchanged.

  Two things follow for a machine that already exists. A new install now ships four blocks switched off — system monitors (CPU/RAM/temperature/disk), network, brightness and keyboard backlight: they are readouts rather than everyday controls, and each is two clicks away in the Hub. Your own bar is left exactly as it is — that file is yours and updates do not touch it. But blocks are addressed by an id that your bar.json predates, so the new tab and w-bar will list nothing on your machine until the file has them. Taking the new default — which replaces your bar customizations, after backing the current file up — is:

  w-reset quickshell config/bar.json

  Adding "id": "<name>" by hand to the blocks you want addressable works just as well; the bar reloads as you save.

• The Hub menu was reorganised. It had grown into a list in the order things were added: twenty tiles in which a settings section and a one-shot command looked alike. It is now an even 4x4 — ten sections (appearance -> devices -> connectivity -> power -> system -> extensions) followed by six quick actions. Security and Date & time were each a section holding two or three controls; both are now tabs of System, where the boot-options tab appears only on a machine that has them. Calendar and Screenshot are no longer tiles: they are launcher entries now, the calendar still opens from the clock, and the screenshot entry takes a region (the other modes capture instantly and would photograph the launcher closing). The Power tile is called Power menu, so it no longer reads as a synonym of the Power section next to it. Separately: a highlighted row is no longer clipped at the edge of a scrolling list, and a dropdown in a bottom row now opens upwards instead of drawing past the card — which also fixes the NumLock dropdown under Input, wrong since it shipped.

• Three launcher entries never worked, and a fourth is new. Power menu, Volume control and Assistant have been in the launcher since they shipped and did nothing when clicked: their Exec= line was quoted in a way the desktop-entry specification does not recognise, so what reached the shell was a syntax error. Each had another door — a hotkey, the logo's right-click menu, a bar block — which is why it went unnoticed until Calendar, which now has an entry of its own and no other door. All four work.

v0.3.0 #

• W has moved to its permanent home — https://github.com/tarkh/w. Every release up to and including v0.2.0 was published from a scratch repository used to shake the release machinery out, and that repository is being deleted. The address is baked into the install image and into /etc/os-release, so this release is the move: a new install points at the new repository on its own. A machine installed from an older image keeps pulling from the old address and stops updating the moment it disappears — point it at the new one once and it resumes:

  git -C /var/lib/w/src remote set-url origin https://github.com/tarkh/w
  sudo w-sync update

  The entries below this one describe releases of that earlier repository, so their tags and release pages no longer exist. Everything they describe is present in the system you are running; the changelog moved across intact on purpose.

• Switching off your last display no longer strands the desktop. With every connected output disabled, the machine came up with a session that runs, answers on the network and draws nothing — a black screen with no way back that does not involve a second computer. The Hub never offered the switch on a single-screen machine, but two ways around that existed, and both are now closed:

  • w-monitor disable run from a text console skipped the "not the last active output" check completely, because the check asked the running compositor and there is none on a console. It has been skippable that way in every release so far. The check now answers from the kernel's list of connected outputs when there is no compositor to ask, so it holds on a console too.
  • The dangerous case is not one command, it is time passing. Switching off the laptop panel while an external monitor is plugged in is a perfectly good thing to want, and the machine only breaks later, when the external one is unplugged — a moment at which nothing was checking anything. The display layout is now judged once more immediately before the compositor reads it, separately for your session and for the login screen, and if it would leave every screen off, exactly one is switched back on. Every other choice you made is left alone, and on a healthy configuration this does nothing whatsoever.

  Should you still land on a black screen, RECOVERY.md now has a section written for that symptom — Ctrl+Alt+F2, w-monitor reset --all, sudo w-monitor greeter reset, reboot — and it ships on the machine too, at /usr/share/doc/w/RECOVERY.md, which is where you will need it.

• The install image is rebuilt nightly, and no longer right after each release. The ISO the README links to is one rolling image built from the current tree. Publishing a version no longer forces a rebuild on the spot, so for a day or so after a release the download can still be the previous night's image. It makes no difference to what you end up with: an image here is a bootstrap medium, and the first update pulls the machine up to the current tree regardless.

v0.2.0 #

• w-rollback — one command that puts an older system back. W has taken a filesystem snapshot before every package transaction and every w-sync update since the beginning, but nothing turned one back into the running system. sudo w-rollback run now does: it lists the snapshots, asks which one to go back to, and asks you to confirm before it changes anything. The system it replaces is kept, not deleted, so a rollback can itself be rolled back. It works on both kinds of install — encrypted (Limine) and plain (GRUB) — and both paths were walked on real machines rather than reasoned about. w-rollback status tells you where you stand without changing anything.
• The rollback W did have never worked. The assistant's w_snapshot_rollback action — present in every release up to and including v0.1.7 — called snapper rollback, which repoints the btrfs default subvolume, while W boots an explicit subvol=@. It reported success and changed nothing, every single time. If you ever asked the assistant to roll back and it told you it had, it had not — and your system was not damaged either, because nothing happened at all. The assistant can now only start a rollback: it opens w-rollback, and the password and the confirmation are yours.
• RECOVERY.md — what to do when the machine will not boot. A page written by symptom: won't boot, boots wrong, an update made things worse. It ships on the machine as well, at /usr/share/doc/w/RECOVERY.md, which is where you will need it — a machine that will not boot has no browser.
• You are now told when you are running from a snapshot. Booting a snapshot from the boot menu is the recovery path, its root is read-only, and the desktop looked completely normal — so anything saved outside /home vanished at the next reboot without a word ever being said. A notice at login now says so, on both bootloaders.
• The installer no longer asks about Secure Boot. The checkbox was cosmetic: the answer was read three times and reached nothing — no keys, no signing, whichever way you answered it. Turning Secure Boot on for real needs the firmware in Setup Mode and a UEFI toggle, neither of which an installer can do for you, so the option is gone and the encrypted install's final screen now points at Hub → Security: it enables Secure Boot in two steps and then seals TPM2, after which the passphrase stops being asked at every boot. This affects new installs only; nothing changes on an installed system.
• The on-box assistant now knows the upgrade order for laptops with Broadcom Wi-Fi: the driver goes in before the system upgrade. The other way round, DKMS rebuilds the old driver against the new kernel, the build fails, and the machine comes back with no Wi-Fi — which on a laptop with no Ethernet port is also the channel you would have fixed it through.

v0.1.7 #

• The install image builds again. Arch removed the broadcom-wl package from its repositories on 1 September, and the image's package list still asked for it, so every build since — including the one for v0.1.6 — failed before producing anything. The list now follows upstream, which dropped the same package the same day. If you are updating an installed system, nothing here changes it; this restores the download.
• Broadcom Wi-Fi on old chips, and what changed. That package was what let the live session drive a handful of older Broadcom chips — BCM43142, BCM4360, the ones in machines like the 2013–2015 MacBook Pro. Arch ships only a source version now, so the default image no longer covers them in the installer, exactly as the official Arch image no longer does. Installed systems are unaffected: the installer still detects such a chip and installs the driver for it. What is affected is installing on a laptop that has one of those chips and no Ethernet port — there the installer can no longer bring the network up. For that case you can build yourself an image that does: ./scripts/build-iso.sh --broadcom-wl, see "Building your own ISO" in the README.
• The README now has a section on building your own image, instead of a note buried in the install steps.

v0.1.6 #

• Folders in the file manager are the theme's colour again. W recolours the Papirus folder icons to the tone of the active theme, and that had quietly stopped working in v0.1.2 — the release that moved W's internals out of /usr/local. The recolouring step looked for its helper on PATH, where it no longer is, found nothing, and skipped without reporting anything, so nothing in the logs said so. If you installed or updated since v0.1.2 your folders have been stock Papirus blue; updating fixes them, and already-open windows pick up the new icons when you restart them.
• A new default wallpaper for the w theme, in all four resolutions. The colours of the theme are unchanged — the picture was drawn to fit the palette, not the other way round — so nothing else about your desktop moves. If you are using your own theme or your own wallpaper, this does not touch it.
• Nothing else changes on an installed system.

v0.1.5 #

• The unencrypted install path is now tested as well. Every published image was already installed onto an encrypted disk and checked before this; now the plain btrfs + GRUB layout — what you get by answering "no" to encryption — is installed and checked alongside it, on the same image. The two use different boot loaders, so passing one said nothing about the other.
• Your assistant now answers update questions correctly. It knows that this repository publishes one commit per release (so "one behind" means one version, not one commit), that pinning REF in /etc/w/update.conf holds a machine in place, and that w-update and w-sync are two different updates — packages from Arch, and W's own configuration.
• Nothing else changes on an installed system.

v0.1.4 #

• Every published image is now installed, not just built. After each build a runner boots the very file on the download page, installs it unattended onto an encrypted LUKS2 + Limine + TPM2 disk — the stack a default install produces — reboots into the result and checks it: no failed services, no crashes, the login manager up, the boot loader correctly staged. If an image would not install, that is now known the same day instead of on your machine.
• Nothing on an installed system changes in this release; updating one only restamps its version.

v0.1.3 #

• The install image on the releases page is now rebuilt every night from this repository, instead of whenever one happened to be built by hand. W is a thin layer over a distribution that moves daily, so an image that sits still only gets further from Arch — and an upstream package that stops building now turns into a failed build the same night, rather than into a failed install on your machine weeks later. Same page, same link, still exactly one image.
• Nothing else changes. No tool, no default and no configuration is different in this release; updating an installed system only restamps its version.

v0.1.2 #

• W's own commands now live in /usr/bin and its internal helpers in /usr/lib/w/, instead of /usr/local. /usr/local is yours: nothing W installs will appear there any more, so what you put in it is once again the only thing in it.

• If you are upgrading an existing install, remove the old copies by hand and reboot — w-sync does not delete files, and /usr/local/bin comes before /usr/bin in PATH, so anything left behind keeps shadowing the updated tool, w-sync itself included:

  sudo rm -f  /usr/local/bin/w-* /usr/local/bin/papirus-folders
  sudo rm -rf /usr/local/lib/w
  sudo rm -f  /usr/local/src/w-windowblind.c
  sudo reboot

  Leave /usr/local/bin/sudo alone — that is W's sudo-rs shim and it works by being found first. A fresh install needs none of this.

v0.1.1 #

• The README now says where the first-boot logs live, so a failed install is one ls /var/log/w/ away from an explanation instead of a guess. Thanks to @tarkh (#1).

v0.1.0 #

First public release.

• Beta of the W desktop: Hyprland under uwsm with the Quickshell UI (bar, launcher, notification centre, calendar, clipboard, volume and brightness, power menu, network and Bluetooth tray, lock screen with fingerprint support), a session that remembers open windows and named layouts, and a unified authentication dialog.
• Theming across GTK, Qt/Kvantum, icons, terminal, shell prompt, Firefox chrome, bootloader and boot splash from one palette. w-theme new <image> builds a full theme from any wallpaper.
• TUI installer in English and Russian — disk, LUKS encryption with Secure Boot, locale, keyboard, timezone, users, update channel, optional software bundles — plus an unattended --preset mode.
• btrfs layout with snapper timelines for root and home, zram in place of swap.
• Updates over the edge channel: w-sync takes a home snapshot, pulls, and applies only the modules a change actually touches. w-reset restores any module to its W default.
• The w-* tool set: updates, themes, wallpaper, power, monitors, keyboard, pointer, night light, keyboard backlight, DNS, firewall, logs, time, screenshots, package bundles. w-info lists them all.
• An on-box AI layer, opt-in and off by default, with every state-changing action behind polkit.