- A workspace Hyprland makes for a switch (createworkspacev2) grows into the row
over the slide instead of appearing at full size; the window keeps the full
room meanwhile.
- Every config.toml save restarted the colour Animation, and Animation::new
runs its clock even from a colour to itself, so every window redrew for 0.4 s
per save. The colours now only fade when they change.
- The file watch could read config.toml halfway through the shell's own save
and reset settings to their defaults. Saves and rereads share one lock, the
file is written to a new file and renamed over, and the shell's own saves
are skipped by the watch.
- A song change or new player only pops the island when the focused window
isn't that player's app (parse::same_app, tested).
appearance.surface.design picks the look (chips in the config window); each
design's opacity, rim and light are settings. surface::shell now draws the
island, idle pill, workspace and tray pills, tray menu, control centre and
popup, so a design applies everywhere. On see-through glass faces fade with
.opacity(). Shell text settings can list their allowed words (chips).
Hyprland drops an empty workspace the moment it's left, so its number vanished
and the row jumped before the disc slid. Its slot now stays for the slide,
shrinking and fading, and the pill narrows with it.
All tunable values (sizes, timings, text sizes, colours, polls, equalizer
presets and bands, hotspot, extra output) are settings declared once in
config.rs, read from ~/.config/amane/config.toml, applied live on save.
shell.conf's four values move over on first start.
settings.rs: a normal window (control centre gear, or ipc config) with a
page per part, a page per Hyprland category, search and reset per row.
hypr.rs: Hyprland's options from hyprctl descriptions/getoption, applied
with hl.config and kept in ~/.config/hypr/amane.lua.
Switching from a workspace with the bar shown to a fullscreen one slid
the bar away while the workspace peek slid in from hidden, so the pill
dipped and came back. A peek now starts from the part's current
position, so a shown pill stays still for the peek and then leaves.
The tray, which peeks don't show, slides away on its own.
Tested on DP-2 (2 -> 3 and back), frame by frame.
edge::spacer: a full-width, empty, click-through Background layer as tall
as the pills reach (TOP_GAP + COMPACT_H = 44 px) with Zone::Reserve.
Hyprland keeps tiled and maximised windows below it and lets fullscreen
windows cover it, so space is reserved exactly where the bar stays.
Verified: reserved 0 44 0 0 on all monitors, a tiled window moved from
y=636 to 680, a fullscreen one still covers 1920x1080, idle 0 frames.
The map menus' width follows the monitor: from its right edge to the
island's compact pill, less a 16 px gap (the pill's pop is 15 px a side);
822 px on 1920. AREA_H 480, no alternating rings (simulated: no label
overlaps up to 8 nodes at this size). The password popup centres on it.
- control.rs: Menu::Ethernet; the panel morphs into the card's details
(gathered off the drawing thread first) instead of a separate popup;
prompt::copy now takes a callback so both use wl-copy the same way
- prompt.rs: the Details popup is gone
- map.rs: 920x460, at most 8 nodes, lighter nearness pull (simulated:
no label overlaps up to 8 nodes while drifting)
- workspaces.rs: a peek on a workspace change shows the workspaces alone
- tray.rs: the shell serves org.kde.StatusNotifierWatcher (only when no other
watcher runs) and is its host; icons from the theme or from IconPixmap
(png.rs, a stdlib png writer); left/middle/right click and scroll; dbusmenu
menus drawn in the island's look; tray IPC verbs for testing
- edge.rs: the bar hides only on monitors showing a fullscreen window, driven
by Hyprland's event socket; peeks of the workspaces (with the disc sliding)
and the island (song change, new player), with a pop on the island
- prompt.rs: Ethernet details popup (addresses, gateway, DNS, MAC copyable)
- map.rs: 1100x560 maps, every other node a ring further in
- tools/fake_tray.py: a fake tray app, by bus name or by path
- prompt.rs: popup for Wi-Fi passwords, Bluetooth PINs/passkeys/codes, saved
passwords (Copy through wl-copy on stdin) and failure reasons; joins through
iwctl, pairs through a bluetoothctl session; only popups with a field take
the keyboard (an exclusive grab let stray keystrokes close failure popups)
- menus: the panel is the only shape; the menu window fades its content in
once the panel has taken the menu's size
- maps: clicking the bubble searches again (iwctl scan, bluetoothctl scan)
with rings running out from the click
- Ethernet tile and pill icon from sysfs; link speed badges (Ethernet speed,
Wi-Fi RxBitrate from iwd diagnostics, at most once a minute)
- edge.rs: bar hidden until the pointer rests on the top edge; nothing is
read or drawn while hidden
- IPC: control wifi join|password|search, bluetooth pair|connect|search,
prompt answer|cancel, status prompt=
The shell registers as iwd's agent and as BlueZ's default agent (KeyboardDisplay),
each only while its daemon runs (GetNameOwner, never a call that would start it), and
again after a restart. RequestPassphrase, RequestPinCode, RequestPasskey,
RequestConfirmation, RequestAuthorization, DisplayPinCode and DisplayPasskey show a
small popup over the control centre that takes the keyboard (Enter answers, Escape
cancels); the answer is the reply to the daemon's call. Amane can only reply with
success, so Cancel leaves the call unanswered and cancels the pairing or the joining.
New Wi-Fi networks with a password are clickable on the map now; IPC control prompt
answer|cancel.
The menus are 820 px wide (the panel resizes to them). This desktop is a bubble in
the middle, networks and devices float around it, nearer the stronger (or the more
connected) they are, rings ripple out while scanning. Each is tied to the bubble by a
line: grey and wobbling while idle, running out from the bubble after a click until
it's connected, solid with a travelling pulse once connected.
The map is one Canvas: node discs and icons are shapes in it (glyph_shapes), the node
widgets only take the pointer. Seven icon Canvas widgets had cost 4.6 ms of gpu a
frame and halved the frame rate; now 1.8 ms at 60 fps. The menu window is sized to
its menu.
The pill drops the icon of a service that isn't running (iwd, bluetoothd, the openvpn
service) and is as wide as what it shows. A tile whose service isn't running is grey
with a grey strike through its icon. Leaving a menu with the pointer fades its content
where it is and shrinks it straight into the pill; the panel no longer grows back to
its full size on the way.
The tile or block grows into its menu; the panel fades its content and takes the
menu's size underneath. Wi-Fi lists networks (join saved/open, disconnect, stop
iwd), Bluetooth lists devices, VPN has a switch per openvpn service, Sound has
every output with its own slider and a default picker. A second Sound slider for
the output named by extra_output in shell.conf (NR515). Screen and keyboard light
are written to sysfs directly: logind refuses a shell without a seat.
Wi-Fi and hotspot (iwd, never D-Bus-activated by a read), VPN (openrc openvpn.*),
Bluetooth (bluez), battery, volume, screen and keyboard light, performance mode
(platform profile + governor + EPP through doas -n), and the external GPU with the
programs using it (drm fdinfo). Icons move to glyph.rs.
where everything lives (repo, library, skill, desktop and its Hyprland
setup), the user's ground rules, the current state (master, the three
open glass branches with benchmark numbers, what runs on the desktop),
architecture including hover, instant transport, monitors, theme and
idle pill, Hyprland 0.56 pitfalls, performance rules, develop and test
workflow, open threads
- README install: the desktop's amane CLI saves ~/.cache/amane/amane-shell,
so amane compile + amane run is the way to start the shell
- amane-remote.sh anim: with one window per monitor, gaps of 500 ms and
more are other islands resting, not stutter
the Theme service waits on amane::watch_file (inotify on the folder, so
matugen's rename-over saves count) for both ~/.config/amane/theme and
shell.conf, settles 50 ms for multi-write saves, and skips writes that
change nothing. measured on the desktop: rename-over, in-place and sed -i
writes picked up in 55-68 ms (was up to 500 ms), identical rewrites
ignored, no activity while idle (was 4 reads a second). deleting the
theme file takes effect at the next start
- matugen/amane.colors: a matugen template (primary, on_primary,
primary_container, primary_fixed, on_surface, on_surface_variant,
outline_variant, surface_container high/normal/lowest) that skwd-wall
renders into ~/.config/amane/theme on every new wallpaper
- theme.rs: a Theme service reads that file and shell.conf twice a
second (matugen replaces its output, which a file watch can lose) and
fades all ten colours together on one Animation; without a theme file
the shell keeps its pink on purple
- the island, the idle pill and the workspace pill draw with the theme;
white overlays (chips, tracks, hover tints) stay as they are
- shell.conf: cover_colors (true: the open media card and equalizer take
the cover's gradient as background, accents stay the wallpaper's) and
transition_ms (0 switches instantly), both applied without a restart
- parse: settings() and rgb(), with a test (6 pass)
- checked on the desktop with a hand-written theme file: pill, workspace
pill and card recolour, cover_colors switches the card background live
the island is always there now: with no player it shows a pill with a
drawn music note and 'Nothing playing', in the cover gradient with the
pink glow. click-through, static (draws once). checked on all 4 monitors
- workspace pill: clicks did nothing. amane's Workspaces::focus sends
'dispatch workspace N', which hyprland 0.56 with a lua config turns into
hl.dispatch(workspace N) and rejects. dispatch the lua form the user's
keybinds use, hl.dsp.focus({ workspace = N }), through hyprctl on its own
thread (named workspaces with negative ids are left out)
- island: 'magic-island has no child to draw' at startup without a player,
and the island then never appeared. hyprland 0.56 configures a layer
surface before its first commit (seen with WAYLAND_DEBUG), so a window
that starts visible(false) gets a size, amane runs the empty view, logs,
and waits for frame callbacks an undrawn surface never gets. with no
player the island is now an empty click-through window instead
- verified on the desktop: 0 'no child' lines, island draws at startup and
shows on all 4 monitors when a player starts
- one window per monitor (namespace magic-workspaces, Overlay like the
island), showing that monitor's hyprland workspaces in the island's
look: cover gradient, pink glow
- focused workspace pink, shown-elsewhere soft pink, empty ones dim;
hover eases in (one at a time), a click switches workspace
- special workspaces are skipped; name-only ones show their first letter
- event-driven through Workspaces: redraws only on hyprland events and
hover
- play/pause flips its icon (card and pill) the moment it's clicked; the
clicked state is shown until the player has answered
- after play/pause, next or previous, Media is read again at 150 and
600 ms instead of waiting for its once-a-second poll, so title, time
and play state follow within a fraction of a second
- the cover loader checks 4 times a second instead of once
- ipc: island play-pause, the same path as a click (keybinds, tests)
- the card opens on the monitor that was clicked (keybinds: the focused
monitor); the others keep their pill. moving the card to another monitor
restarts its open animation there
- only the focused monitor's window (and the one with the card) reads
Spectrum, so the other pills show still bars and rest
- hover lights only the pill under the pointer
- measured with music on 3 monitors: pill 10.7% of a core (was 27.5%),
open card 20.7% (was 58%), ~62 frames a second in total, 0-1 slow
frames per open/close (was 7-14), back to one island's cost
stopping the shell leaves its cava running (parent 1), reading on from a
closed pipe and ignoring SIGTERM; every restart while music played leaked
one. on start, pkill -KILL -f our cava.conf before starting our own
window_per_monitor with the same view: every monitor shows the island,
all sharing one state (open, hover, player). measured with music on 3
monitors: 180 frames a second in total, pill 27% of a core (was 9%),
open card 58% (was 21%), gpu 1.4 / 3.0 ms per frame each; open/close
drops 7-14 of ~175 frames (was 0-1)
one pink glow for the pill and the open card (opacity 0.3, blur 22),
0.7 while the pill is hovered; the dark drop shadow is gone. same single
shadow as before, so no extra cost (pill 1.3 ms gpu)
- the open card keeps a faint pink glow: the pill's shadow blends from a
dark drop shadow to a centred pink glow (opacity 0.3, blur 22, inside
the window margin); measured 2.95 ms gpu with the card open, as before
- eq bands: hovered, the track widens 4 px and brightens, the knob grows
25% with a full pink ring, the frequency label turns pink
- time bar: hovered, the track thickens 4 px and brightens, the knob grows
- previous/play-pause/next: no halo, circle or growth any more; they go
bold instead (an outline of the glyph's own colour, 0 to 2 px; pause bars
widen), and glowing() is gone
- each hovered spot eases a highlight in and out over 220 ms
- pill: widens 10 px and its drop shadow turns into a centred pink glow
(the same shadow the pill already had, so no extra cost)
- previous/next: swell 12%, icon turns pink, pink halo; play/pause swells
10%, halo, its pink warms toward white
- chips and presets brighten and grow 4-6%, dropdown rows slide right
- halos only exist while hovered (a shadow is two gpu layers per frame)
- one button lit at a time, and highlights reset when the card or the
dropdown changes: amane matches hover targets by position in the frame,
so a target that disappears never hears the pointer leave
- README: features, layout, requirements, install, IPC, runtime files, limits
- AGENTS.md: the user's ground rules (branch + approval, never touch real
players' sound), architecture, measured performance rules, type-check
and test loop, desktop test driver and its pitfalls, open threads
- tools/: amane-remote.sh and fake_mpris.py, so the test loop doesn't
depend on the Hermes skill
- .gitignore: the Cargo files amane generates
- the pill, the card and the fade veils share one 150-degree gradient from
the cover's palette (most common, second, background), each darkened to
a fixed brightness so white text stays readable; veils span the whole
shape so their gradient lines up with the backdrop's
- measured on the desktop: card open 3.1 ms gpu (solid fill was ~3), pill
1.7 ms (was 1.0), open/close animation unchanged (1 and 0 slow frames)
- pill levels: lows left, highs right, no mirroring; every bar still grows
up and down from the middle line
- one linear timeline: the pill fades out, the empty shape grows, then the
card fades in on the settled shape (and backwards on close)
- fades use a surface-coloured veil instead of .opacity(), and the shape
has no clip, so neither costs a gpu layer
- measured on the desktop: open went from 11 frames over 20 ms to 1,
close from 4 to 0
- the spectrum capture runs whenever the shown player plays, and the pill
shows it as 7 mirrored, glowing bars (about 9% of one core while playing)
server2