| Age | Commit message (Collapse) | Author | Files | Lines |
|---|
|
The reverse image search never worked in practice: the clipboard-and-paste
route to TinEye was the only one left after scripted uploads hit bot and
consent walls, and it still left finding and replacing the file by hand.
wallhaven-dl, a separate app, does both the search for a larger copy and
general wallpaper search and download, so the button now launches it and
the TinEye code is removed rather than hidden.
The button is always shown, not only for a low-res staged image, since the
app is useful beyond upscaling. It launches through gtk-launch and the
app's desktop entry, which needs no ~/bin on quickshell's PATH and keeps the
app's taskbar icon, and closes the panel so the window does not open under
the overlay.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
|
|
When the image staged for the targeted screen gets a yellow or red
quality note, a button under the mock screens copies it to the clipboard
and opens TinEye in the browser, where Ctrl+V runs the reverse search and
the matches can be sorted by size. It acts on the staged image, not the
hovered one, because the pointer leaves the thumbnail to reach it.
The image goes through the clipboard because a search engine takes no
local file in a URL, and scripted uploads did not get through: TinEye
answered 403 to curl and 405 with a browser user agent, and Google Lens
redirected to its consent page, dropping the image. The paste in the
browser is the upload, and only on the user's click.
The copy is converted to PNG with magick. Copied as itself, a JPEG put
only image/jpeg on the clipboard and Firefox pasted nothing into the
page; it hands web content image/png.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
|
|
The wallpaper tab showed every image under ~/Pictures/wallpapers in one
grid. It now groups them by top-level folder, each behind a header that
opens and closes it, and adds the distribution's /usr/share/wallpapers:
its loose images as "system", its Plasma packages as "Plasma". A package
carries one picture at several resolutions under contents/images/, some
with a dark version under contents/images_dark/, so it shows once, the
dark version where there is one and the largest of those, labelled with
the package name rather than "5120x2880.png". Groups are keyed apart from
their names, so a local folder called "system" does not share open state
with the system group. Open state lives in the Wallpapers singleton,
because the panel is destroyed on close.
The list is one flat ListView of rows, a header or a row of thumbnails,
so only rows in view (plus cacheBuffer) are instantiated and decode. A
first version made each folder one delegate, which loaded every image of
a folder as soon as it opened; the system group alone is 727. The rows
live in a ListModel and a toggle inserts or removes only that folder's
rows: assigning a fresh JS array instead resets the view, which loses the
scroll position and re-decodes everything visible. cacheBuffer is clamped
at 0 because the column's width reads negative before layout and the
ListView warns on each negative value.
The scan's find calls go through a small shell function rather than eval.
Each thumbnail's 2px border tells how it will look on the targeted
screen. `file -N -0` reads every image's size from its header after each
scan, ~0.6s for 1161; the NUL keeps names containing ": " intact, and the
size is the last WxH in the description because a JPEG's "density 72x72"
comes first. wallp runs swaybg -m fill, which scales to cover, so the
measure is the enlargement factor: <=1 green, <=1.33 yellow, else red, in
the udt palette colours. A note under the mock screens, in that colour,
gives the size, the factor and why it will look soft or pixelated, for
the hovered image or else the staged one. The staged image keeps its
accent border. Screen sizes are fixed, as the mock's already were.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
|
|
The balloons now wear the Hyprland window border: 2px, the udt gradient
read from ~/.cache/udt/udt-border.lua (the file look_and_feel.lua reads,
watched so a wallpaper change follows), rotating, with the windows'
corners.
A Rectangle border takes one colour and its radius is circular, while the
windows use rounding 15 at rounding_power 4, a superellipse. So the body
and the border are both drawn by a Shape from an SVG path whose corners
are the cubic fit to |x|^4 + |y|^4 = 1, and the border is a ring filled
odd-even, which keeps the translucent body from showing the gradient
through.
One angle drives every balloon, so they turn together, and the
animation runs only while the stack is mapped. Hyprland's borderangle
(speed 30) reads as 3s a turn, which looked slow on a balloon that size,
so the turn is 1.5s, tuned by eye.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
|
|
A slot query for an unloaded model loads it into VRAM, so any status
probe against the router needs autoload=false. Found when a probe during
the ai module's design reloaded Gemma.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
|
|
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
|
|
Engines are read-only, since stopping llama-server or sd-server takes
its web UI down too. Apps start and stop through their own controls,
detached so a start outlives a quickshell reload. Fanfictioner has no
start (a run needs a story) and stops by pid, children first, so its
sd-cli runs go with it and nothing else does.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
|
|
fanfictioner runs sd-cli one-shot, not through sd-server, so the sd row
read down while a CLI run held VRAM and generated. A running sd-cli now
reads up and busy, its model marked (sd-cli), and outranks the web UI.
The model is the first --diffusion-model/-m: sd-cli's prompt follows and
is free text that may contain "-m".
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
|
|
Six TSV lines per sweep: llama and sd engines, four apps. llama's slot
probe passes autoload=false, since a plain /slots?model= loads the model
into VRAM. sd-server has no progress endpoint, so busy is its log
having been written in the last 5s.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
|
|
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
|
|
Engines (llama.cpp, sd.cpp) read-only, apps get start/stop through their
existing controls. Records that /slots?model= autoloads an unloaded model
unless autoload=false is passed, found when a design probe reloaded one.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
|
|
Qt's atomic write builds the file as an unnamed O_TMPFILE. When the
target does not exist yet it links that straight into place, and the
only event naming the file is IN_CREATE: no close_write, no moved_to.
statusctl watch did not listen for create, so the first write of each
mode in a session went unseen. On a fresh session that is the write
that matters: a Teams call started an hour in left the waybar pill
reading off until some later rename happened to redraw it.
The test reproduces the event with a hard link, which is what the
linkat from Qt's temp file looks like to inotify.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
|
|
It was a hand-installed copy, which can drift from the tested source
silently. Replaced with a symlink after confirming the two were
identical.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
|
|
The auto-detector now sets presentation while Teams holds a mic stream,
alongside the game checks. Teams runs for the whole session, so its
process alone is no signal and would have kept lock, DND and the
breaktimer off all day. A call is Teams holding a source-output, which
pactl lists tagged with application.process.binary = "teams-for-linux";
verified matching during a live call.
It reuses status.gaming rather than status.presentation, so a call
ending clears only the automatic half and never a manual toggle. The
statusctl tooltip for that state said "game running", which is now
wrong, so it reads "auto: game or call". The file and waybar class keep
the gaming name: renaming them touches four places for no behaviour
change.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
|
|
An edit to shared/Notify.qml was never picked up by the running
shells, while an edit in the same pass to a component's own file did
reload, so the shell looked current and ran the old sanitize. Record
that shared/ edits need a forced reload of each linking component.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
|
|
Balloon and drawer bodies switch from RichText to Qt's MarkdownText
(CommonMark plus GitHub extensions). Raw HTML still goes through Qt's
HTML importer, so body-markup senders keep working.
Markdown images fetch their destination the same way an <img> does,
which the existing sanitize filter would miss since it only inspects
tags. sanitize now escapes every "![" so the syntax renders as literal
text, local or remote; parsing CommonMark destinations and reference
definitions to allow local ones is not worth the grammar. Verified in
the shell log: before the filter reached the running shell, a test
body fetched https://example.org/a.png; after, a second one did not.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
|
|
The crossfade only brought a loaded cover to the front when the image's
source compared equal to the wanted URL. It never did for a title with
spaces or brackets: the URL carries the title as a percent-encoded
fragment cache buster, and QUrl reads it back decoded, so "#Foo%20Bar"
came back as "#Foo Bar". The load finished, loaded() bailed, and the old
cover stayed. Confirmed with a standalone qml test.
The slot being loaded is now tracked by index instead. Stale status
changes cannot match: only the back image is ever reloaded, and a newer
want resets it before assigning.
Seen working on screen by the user.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
|
|
The OSD background held one Image, so a track change swapped covers in a
single frame. Two images now take turns: the new cover loads into the
hidden one and is brought to the front only once it reports Ready, then
both fade over 300ms. Waiting for Ready is what keeps a slow or remote
art server from fading through black.
The blur, scrim, sheen and grain are grouped under one item whose
opacity follows artOn. It goes false when the player has no art or the
cover for the current track errors, so the panel fades back to its plain
translucent look instead of snapping, or keeping a stale cover that
reads as a fault. It goes true again only when a new cover has loaded,
so returning from plain never flashes the previous one.
The back image's source is cleared before being set, so a URL it
already holds still reloads and reports a status. onWantChanged misses
the value at load, hence the Component.onCompleted call.
Not yet seen on screen: the shell reloaded it cleanly but the fade was
not watched through a track change.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
|
|
The art background is opaque, so Hyprland's blur behind the OSD cannot
show through it; making the art translucent to let it through would
also thin the scrim that keeps light covers legible. Fake the frosted
glass instead: a faint top-down sheen in Theme.text and a fixed grain,
painted once by a Canvas and masked to the panel's rounded corners.
The grain is a Canvas rather than a ShaderEffect because a shader needs
a compiled .qsb beside it; the Canvas repaints only on resize.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
|
|
When the current track has art, the volume OSD now draws it across the
whole panel, blurred (MultiEffect blur 0.5, default blurMax) and masked
to the panel's rounded corners, since clip only cuts to the bounding box.
The thumbnail in the track row stays sharp on top.
A dark scrim (Theme.base at 0.82) sits over the art. The palette is dark
and the text light, so a light cover made the panel unreadable, and more
blur did not fix that; only darkening did. The scrim is a sibling of the
MultiEffect, not a child: nested inside it, it did not paint at all, which
is why raising its alpha first appeared to change nothing.
Without art the panel keeps its translucent, compositor-frosted look.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
|
|
Both now follow the scheme but only read their theme at startup.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
|
|
install.sh's tab is the overview of a whole run, so it leads as "General
recap"; the consumer tabs follow, labelled GRUB, SDDM and Homepage. File
names stay the producers' short ids, and one without a label shows under
its own name, so a new producer still needs no edit here.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
|
|
install.sh now hands its output to the steps window on every run, so the
Theme tab stops piping it through udt-steps, which would write the tab a
second time.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
|
|
StepsWindow shows unified-desktop-theme's pending manual steps, one tab per
consumer, centred on DP-3 on the top layer with on-demand focus so it can
stay open while commands are pasted elsewhere. udt-steps writes the files
and calls the new `steps` IPC verb; indented lines become command blocks
with a copy button. The Theme tab pipes install.sh's output into it.
The icon switch now also reruns udt-accent (SDDM power icons) and
waybar-theme-udt's install.sh (taskbar theme, module icons), chained after
the gsettings write both read. Cursor changes write ~/.config/udt/cursor.lua
instead of editing hypr-theme's tracked environment.lua.
The theme notice and the Udt.qml comment claimed install.sh reloads
nothing; it reloads Hyprland, waybar, kitty and conky now.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
|
|
notify-actions.sh called rofi with no -theme, so clicking a notification
that carries actions opened rofi in its stock theme instead of the udt
one every other menu uses. Pass ~/.config/rofi/udt/menu.rasi, the small
fixed-options layout the screenshot and notes menus already use.
menu.rasi shows only the listview, so the "Notification" prompt and the
filter input are not drawn; a notification's actions are few enough that
picking from the list is enough.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
|
|
Quickshell 0.3.1 keeps the QLocalSocket of a failed connect, so setting
connected = true again never retries: an overlay started before the
assistant (at login, or by a restart) stayed dark for good. The Socket
now lives in a LazyLoader that the retry timer turns off and on.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
|
|
Reads desktop-assistant's event socket, JSON lines on a Unix socket the
assistant owns: state, context use, the running tool and audio levels.
A circle shows the name and used/total context, a wrench and the tool
name while one runs, and pulsing dots while it transcribes or thinks.
Around it three sine ripples follow the loudness, the user's voice and
the reply in two palette colours from Theme, so a udt change recolours
them live. It fades out a configurable delay after the reply.
The ripple is driven by FrameAnimation, not a Timer: a 16 ms Timer
drifted against the display refresh and visibly stuttered. Standalone
like the other components; it reconnects every 2 s while the assistant
is down, and the assistant runs the same without it.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
|
|
A Loader keeps the height of its last item after active goes false. When
a browser tab playing media was closed, its MPRIS player vanished,
Player.active dropped and the window shrank to 100px, but the Loader
still reserved 46px plus the Column spacing on top. The volume row was
pushed down and clipped at the bottom edge, which is the "stuck" OSD a
hot reload used to cure: a fresh Loader starts at zero height.
Hiding the Loader while inactive makes the Column skip it. Both the
retained height and the collapse were confirmed with headless qml6 runs.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
|
|
A label is the only place the reminder says what a bind does, so
cutting it short loses the part that matters, as with the column moves'
scrolling-layout note. A long one now takes a second line.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
|
|
Each key showed its plain action as text. It now carries the real
cap's legends, Nerd Font glyphs checked by name in Inconsolata, with
the shifted or Alt legend below and the cap in the palette colour
nearest the real one. Hovering still lists every action.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
|
|
For the keybind reminder's macropad keycaps. blue reads termblue, not
blue: udt-palette's blue is the active-border role, magenta under
candy-night, and termblue is the terminal's actual blue. The defaults
are Macchiato's, as for the rest.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
|
|
The pad glyph on each row was too small to read. The left fifth of the
overlay now draws the pad at size, each key labelled with its plain
action; hovering a key or an encoder lists everything it does, held
keys and turn direction spelled out. padActions() in Binds.js inverts
the grouped rows into that per-key list, and PAD's turn values became
words for it. The panel is pinned at a fixed height: centred, it moved
whenever the hovered key's list changed length.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
|
|
The macropad's firmware sends ordinary combos, so the overlay listed
them as keyboard shortcuts nobody presses. A PAD table in Binds.js maps
each combo to its key (1-12) or encoder (top or bottom, turned or
pressed) and what is held with it, and the row draws the pad small with
that key lit. The workspace row is built in a loop, as it is in
keybindings.lua. Every entry was checked against hyprctl binds -j.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
|
|
Hyprland's key names are its own (mouse:272, XF86Calculator), not what
is printed on the keyboard or how the key is pressed, which defeats a
reminder. A fixed map in Binds.js renames them for display only; an
unlisted key shows as before. SUPER reads as Win for the same reason.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
|
|
Groups flowed left to right in rows, so each row was as tall as its
longest group: short groups sat above empty space and the block ran
past the top and bottom of the screen. Columns filled to the screen's
height pack the groups and keep everything visible.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
|
|
Modelled on window-switcher: a Scope with a 1x1 keepalive PanelWindow
plus a LazyLoader'd overlay, opened via IpcHandler (SUPER+k is bound
in a different repo). hyprctl binds -j is asked fresh on every opening
so the list matches whatever Hyprland has loaded, with Binds.group()
from task 1 turning it into the grid's {name, rows} groups.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
|
|
A description like "toString: X" made byName["toString"] resolve to
the inherited Object.prototype.toString instead of undefined, so the
group was never created and .rows.push threw. Object.create(null)
gives byName no prototype to shadow.
Also note in a comment that CAPS/MOD2/MOD3/MOD5 modmask bits are not
decoded and silently drop from the combo string.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
|
|
Binds.js parses hyprctl binds -j output into display groups for the
upcoming keybind reminder overlay. Group name and row label come from
the "Group: Label" convention in each bind's description; a bind
without one still surfaces under "Undescribed" rather than vanishing,
so a gap in the Hyprland config stays visible.
Pure functions with no QML dependency, so test_binds.js checks them
directly with node, stripping the .pragma library line the QML
runtime requires but node does not understand.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
|
|
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
|
|
The list comes from hyprctl binds -j on every opening, so it is always
what Hyprland has loaded, loop-generated binds included. Labels come
from each bind's description option, since hyprctl reports Lua actions
as an opaque __lua handle.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
|
|
The volume OSD drew only its "Output 72%" label, with the icon, bar and
device line clipped below the panel edge. The code and the theme were
not at fault: a hot reload fixed it, and a debug line logged just after
showed the volume row laid out correctly at the top of the panel. The
stuck state lived in the running process, so the fix is a restart.
quickshell has no reload subcommand and no IPC target for it, and a
hot reload needs a real content change to a file. reload.sh kills and
restarts each named component, or every directory holding a shell.qml,
and prints the qs count afterwards, since the process is qs and not
quickshell.
It must run from the user's terminal. A detached qs started from an
agent's tool call dies with the call, which would leave the desktop
without its shells; AGENTS.md now says so under Reloading.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
|
|
libvirt reports "running" as soon as QEMU is up, long before the guest
has booted, so the panel had no way to say when a VM is actually usable.
After a successful start, a detached script polls the VM's sshd and
sends a notification once it accepts connections, or a warning after
five minutes.
Host and port come from `ssh -G <vm>`: a Host alias named after the VM
may put sshd on a non-default port (buildsystem listens on 2290, and a
probe of port 22 never succeeded). Without an alias `ssh -G` echoes the
name back, and the guest agent's IPv4 address on port 22 is used. The
probe is a plain `nc -z`, so no authentication or key is involved.
Detached rather than a Process so several VMs can wait at once and the
wait survives a shell reload. Only starts issued from the panel are
watched; a VM started from virt-manager or the CLI is not.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
|
|
status.gaming is written by the shell's game detector and presentation is
the OR of it and the manual file, but statusctl read only the manual half.
A game therefore asserted the presentation effects while the waybar pill
still read deactivated.
Report gaming as a third state, for presentation only. The watch follows
status.gaming as well as its own file, and emits only on a real change so
an unchanged rewrite draws nothing.
Toggle is refused while a game holds the mode: the detector reasserts the
file within its poll, so the write would not stick. set stays allowed, the
manual half is worth setting for when the game exits.
gaming stays out of MODES. The detector owns that file and nothing here
may write it.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
|
|
PCSX2 joins gamescope, Steam and DuckStation in the game detector, so
presentation is set automatically while a PS2 game runs and the idle
lock, DND and breaktimer stay out of the way.
The binary is /usr/bin/pcsx2-qt, the same -qt shape as duckstation-qt,
so it takes the same pgrep -x exact name match and cannot self-match the
way the bracketed steam pattern has to guard against.
Verified against the real emulator rather than a stand-in: with
pcsx2-qt running, status.gaming reads 1.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
|
|
Game detection probed only gamescope and Steam, so a standalone
DuckStation session never turned presentation on. Add a third probe for
the duckstation-qt process.
pgrep -x rather than the bracketed pgrep -f the Steam path needs: this
is an exact match on the process name, so it cannot match the sh running
the check, and the binary name is under the 15 character comm limit so
it is not truncated.
|
|
A failed pairing showed the raw bluetoothctl output as a plain red line that
only cleared on the next attempt, and the desktop notification is suppressed
while the drawer is open, so the failure was easy to miss and impossible to
dismiss.
Show it in a red card with a close control that clears pairing.error. It has
no timeout, and clearing the source means the next identical failure shows
again rather than staying hidden.
|
|
BluetoothRow only read BluetoothDevice.battery, which is BlueZ's
org.bluez.Battery1. A gamepad never publishes that interface: a Sony pad
exposes its charge as a kernel power_supply instead, so the percentage never
appeared.
Read every power_supply whose name ends in a MAC on a 30s timer while
anything is connected, keyed by address. BlueZ still wins when it has a
value.
|
|
Presentation is worth having automatically while a game runs: it is a
fullscreen app that already holds the screen, and the idle lock, DND and
breaktimer are all unwanted during it.
Detection is a 5s poll for a gamescope process or a Steam game binary,
either one enough. The Steam pattern is bracketed so pgrep -f cannot match
the sh running the check and report a game forever.
The result goes to its own status.gaming file, and presentation becomes the
OR of it and status.presentation. That separation is the point: a game
ending clears only the gaming half, so a presentation set by hand survives
a game session instead of being clobbered on exit.
|
|
The drawer absorbed the indicator's device actions but not the settings
window, so there was no path to the full KDE Connect app. Add an Open app
button beside Refresh, launching kdeconnect-app through execDetached so the
window outlives a quickshell reload.
|
|
The hand-rolled buttons only reacted to hover, so a press gave no visual
confirmation until the action's outcome appeared. kdeconnect made this
obvious: its row buttons shell out to kdeconnect-cli, and nothing on
screen acknowledged the click.
Button.qml now fills solid accent while the mouse is held (red for
danger), with the label flipped to base for contrast. TransportButton
gets the same treatment, and the window-switcher close button lightens
on press. Toggles, tiles and tabs are left alone: their own state change
is already the feedback.
|
