amane-config/AGENTS.md

AGENTS.md

Handoff for agents continuing work on the Magic Island shell (Hermes, Claude Code, Codex, a new session). Read README.md first for what the shell does and how to use it; this file is how it's built, how to work on it, and where things stand. Last updated 2026-10-13.

Where things are

• This repo: /root/amane-config on the dev host; remote git@server1:fabolous005/amane-config.git.
• Amane library and docs: /root/amane (docs/src/*.md, source in src/). The desktop has its own copy at ~/git/amane.
• Hermes skill amane (~/.hermes/skills/software-development/amane/): distilled Amane docs plus lessons. references/recipe-media-island.md and references/remote-visual-testing.md cover this project; scripts/ holds copies of tools/ (keep them in sync).
• Desktop: ssh -o BatchMode=yes fabian@192.168.2.169. Hyprland 0.56.2 with a Lua config (~/.config/hypr/*.lua; API stubs in /usr/share/hypr/stubs/hl.meta.lua). Intel Xe iGPU (Raptor Lake). Monitors: eDP-1 1920×1200 at 1920,620; HDMI-A-1 1920×1080 at 0,620; DP-2 1920×1080 at 3840,0; HDMI-A-4 at 5760,0 only when connected (on 2026-10-13 three were connected, all 1920 wide). ~/.config/amane on the desktop is a git checkout of this repo.
• Desktop daemons the control centre uses: OpenRC services iwd (Wi-Fi), bluetooth (bluez), openvpn.<name> (VPN); doas rule permit nopass fabian for rc-service, sh and cat; wl-copy, ip, lspci. iwd and bluetoothd are often stopped; the shell then greys their tiles.

Ground rules from the user

• Branch per big change, ask before merging. New branch for each big change; merge into master only after explicit approval. "In main branch" or "merge directly to main" means commit straight to master. Small follow-ups in the same request can share a branch. Push only when the user asks.
• Never touch the user's real players' sound (Cider, Firefox). Before any island preset|output call or volume change, check amane ipc call island status names a fake test player. Test audio goes to the inaudible null sinks.
• The user's own amane: they sometimes run it from a terminal. Don't kill it without asking; tools/amane-remote.sh start refuses when it finds a shell it didn't start.
• Desktop access is SSH only. Don't probe the network, SSH keys or other hosts. If SSH fails, report it and wait (the user once had to restart sshd).
• Leave the desktop clean after testing: fake players and null sinks stopped, shell running without the frame log (see "After testing").
• Edit files with the patch/write tools, not scripts that rewrite files. The user installs missing tools themselves. Scripts take instance values (hosts, names) as arguments.
• D-Bus only for reading status. Actions go through the CLIs: iwctl to join, scan and disconnect, bluetoothctl to pair and connect (the user had a D-Bus agent version reverted, afe963d). The one exception, chosen by the user: the shell serves org.kde.StatusNotifierWatcher and calls tray apps' dbusmenus for the tray. Ask before serving anything else on the bus.
• Never wake iwd by accident: any D-Bus call to net.connman.iwd starts it (it then joins Wi-Fi and adds a route). Check org.freedesktop.DBus.NameHasOwner first; only talk to it while it has an owner.
• Root only through doas -n with fixed command lines (rc-service <svc> start|stop, cat /var/lib/iwd/<profile>.psk). Strings from other programs (SSIDs, device names, D-Bus metadata) go to Command args, never into amane::spawn/output (sh -c). Passwords go to wl-copy on stdin, never as an argument; iwctl's --passphrase is visible in ps for about a second, which the user accepted.
• Hermes approval prompts are switched off; edits to AGENTS.md may still prompt (security.protected_instruction_files), and the user approves AGENTS.md rewrites first.
• Report honestly what was and wasn't checked on the real desktop.

Current state

master (pushed): the island (pill, card, CD ring, EQ, output and player switching, hover animations, instant transport buttons), an idle pill when no player is open, one island per monitor (card on one monitor, live bars only on the focused one), a workspace pill per monitor, and wallpaper colours through matugen. Merged from control-centre on 2026-10-13, then extended on master:

• Control centre (top right): a status pill (Wi-Fi or hotspot, VPN, Bluetooth, Ethernet, battery; icons of stopped services hidden) that opens a panel: tiles for Wi-Fi, Bluetooth, VPN, Hotspot, Ethernet; volume with the output, brightness, keyboard light, performance mode, external-GPU programs. The panel morphs into its menus as one shape: Wi-Fi and Bluetooth maps (devices floating round this computer, click to join/pair, right-click a saved network for its password, click the centre to search again), VPN list, sound outputs, Ethernet details with copyable addresses. The maps reach from the right edge to the island's pill, less 16 px. Popups (magic-prompt) ask for passwords, PINs and passkeys and say why something failed.
• System tray right of the workspace pill.
• Bar hiding: only on monitors showing a fullscreen window; the pointer resting on the top edge brings it back. Events bring one part in for a moment (a peek): the workspace pill alone on a workspace change, the island on a song change or a new player. A peek starts from where the part is, so switching from a shown bar to a fullscreen workspace keeps the workspace pill still (no dip) while the tray leaves on its own.
• Reserved space: windows that aren't fullscreen stay below the bar (44 px at the top of every monitor).

On 2026-10-13 the user was asked to try the mouse-only parts (tray clicks, top-edge hover reveal, Copy buttons, PIN pairing) and answered "All works"; the agent itself only drove them over IPC.

Open design branches (pushed, not merged; the user is reviewing them). They branched from 036b326, 13 commits before master's control-centre work; merging one will need a rebase (conflicts not checked yet):

Branch	What it is
glass-base	glass (0.05 to 1, default 1 = solid) in shell.conf; new src/surface.rs with shell(), which draws the island, idle pill and workspace pill; with glass below 1 the open/close fade uses .opacity() (a veil can't hide content on see-through glass). README "Glass" section with the Hyprland layer rule.
design-frosted	glass-base plus an Apple-style surface: default glass 0.62, thin tint, hairline white rim (accent when hovered), white sheen along the top third, soft low black shadow instead of the accent glow.
design-liquid	glass-base plus clear glass: default 0.4, denser toward the bottom right, specular gradient (bright top left, faint bottom right), 1.5 px light rim warming to flame when hovered, keeps the accent glow.

Each branch has review images in docs/design/. A blur value was added to shell.conf on all three and then reverted: the user wants the blur strength set only in Hyprland (~/.config/hypr/looks.lua has decoration.blur.size = 3). When the user picks a design, merge glass-base plus that design into master.

Measured 2026-10-09 with the silent "Synth Test" player playing, 4 monitors (Amane GPU time per frame, CPU of amane-shell):

	Pill	Open card	Open/close
master (solid)	1.6 ms, 12%	3.3 ms, 25%	2.5 ms
glass-base at 0.5	1.7 ms, 13.5%	3.0 ms, 25.5%	2.4 ms
design-liquid at 0.4	2.2 ms, 16%	3.3 ms, 23.5%	2.9 ms
design-frosted at 0.62	2.9 ms, 22%	3.3 ms, 24%	2.9 ms

Hyprland's CPU stayed at about 9.5% in every run, blur rule on or off. All builds, master included, had 25–40 stutter frames (20–500 ms gaps) per 3 opens and closes, against 0–1 measured earlier with fewer monitors; the cause wasn't found (Cider and Firefox were running).

On the desktop right now: the agent-started shell runs a master build (11646ca; pid in ~/.cache/amane-remote-test/shell.pid, no frame log). ~/.config/amane is a checkout at 036b326 with master's src/ deployed over it, so git status shows changed files. shell.conf has glass = 0.70 (set by the user). amane is not autostarted by Hyprland: after a reboot the user starts it.

Theme pipeline: Skwd-wall v2's matugen integration amane renders ~/.config/amane/matugen/amane.colors into ~/.config/amane/theme (Skwd scheme: scheme-fidelity). ~/.config/hypr/windowrules.lua has the glass layer rule (^magic-.*$, blur, ignore_alpha = 0.5).

Constraints of Amane itself

• Only amane and the Rust standard library: amane compile generates Cargo.toml, so no crates. External tools run as processes (std::process::Command, amane::output).
• Amane redraws the whole window every frame, so every effect is paid at the frame rate.
• Amane can only blur what the shell draws itself; frosting the desktop behind the shell needs a Hyprland layer rule.
• amane compile saves ~/.cache/amane/amane-shell; amane run starts it.

Architecture

Piece	Kind	Job
island.rs Island	input-only Service	open: Animation (0 pill, 1 card), expanded, selected player (bus name; None follows the playing one), open menu, time-bar scrub, hover (per-Spot animations), opened_on / pointer_on (monitors), playing_now (a clicked play/pause until the player answers). Also the view and the island IPC handler (command).
sound.rs Sound	Service, 1 s poll	pactl list sinks + sink-inputs → each player's streams (matched by process ancestry, then app name), volume, mute, output, equalizer. Curves saved by player identity to ~/.config/amane/island-eq.
spectrum.rs Spectrum	Service, 200 ms check	While the shown player plays: pw-record --target <its stream> | cava -p $XDG_RUNTIME_DIR/magic-island/cava.conf → 24 levels. On start it pkill -KILLs cava left by an earlier shell (cava ignores SIGTERM).
art.rs Art	Service, 250 ms poll	Cover URL → local file (curl for https), copied to palette-cover, which Palette (opened in main.rs, 8 colours) watches. Only used for cover_colors.
theme.rs Theme	Service, amane::watch_file	10 colours from ~/.config/amane/theme and shell.conf (cover_colors, transition_ms; glass on the glass branches), faded together on one Animation<Colors>. Every view starts with theme::begin() (subscribes, returns cover_colors) and draws with theme::now(). No theme file: the original pink on purple. watch_file watches the folder, so matugen's rename-over saves count (55–68 ms to redraw).
workspaces.rs	view + input-only Hover	Workspace pill per monitor (namespace magic-workspaces, top left): Workspaces filtered by output() == monitor.name, special workspaces skipped. Click runs hyprctl dispatch 'hl.dsp.focus({ workspace = N })' on a thread.
parse.rs	plain functions	pactl/cava/proc parsing, filter-chain config, pw-cli params, EQ store, middle ellipsis, key = value settings, #rrggbb; for the control centre also iwctl/bluetoothctl output (plain_lines, bluez_reason, pairing_event), iwd profile names and psk files, Wi-Fi bars, link speeds, hyprctl monitors/clients, ip addresses and routes, tray icon scoring and menu mnemonics. Unit-tested on its own.
control.rs Control	view + input-only Service	The control centre (magic-control, top right) and its menu window (magic-control-menu, no shape of its own: the panel morphs to the menu's size, then the menu's content fades in). Menu (Wifi, Bluetooth, Vpn, Sound, Ethernet), menu_content(menu, map_w, ..) for both windows, map_width(monitor_w), the control IPC handler.
system.rs Links, Power, Gpu	Services, threads	iwd (D-Bus reads, only while it has an owner), openvpn services, bluez service state, wired cards from sysfs, battery, brightness, keyboard light, power profile, eGPU programs from fdinfo. Actions: join_wifi, scan_wifi, disconnect_wifi (iwctl), scan_bluetooth, rc-service via doas, saved_password, wired_details.
prompt.rs Prompt	view + input-only Service	The popup (magic-prompt): password, PIN, passkey, confirm, show a code, saved password (Copy), failure reason. Runs the bluetoothctl pairing session. Only popups with a text field take the keyboard. prompt::copy(text, done) is the one wl-copy.
map.rs	widget functions	The Wi-Fi/Bluetooth maps: nodes on an ellipse round the centre bubble (nearer = stronger), floating, lines (wobbling, running out while connecting, solid with a pulse), search rings. AREA_H 480, NODES 8 (a 9th makes labels meet at 822 px wide; the sizes were picked by simulating the drift).
glyph.rs	shapes	Every icon as Shapes; maps draw all of theirs into ONE canvas (glyph_shapes), as each Canvas has a large fixed GPU cost.
edge.rs Edge	Service, Hyprland sockets	Reads .socket2.sock events (fullscreen, workspace, focus, windows) and asks .socket.sock for monitors/clients; no polling. Hides the bar on fullscreen monitors (edge.part(Part, monitor), hidden_by), the 1 px magic-edge strip that brings it back, peeks and the island's pop (edge.pop), the workspace highlight slide (switched), and spacer, the magic-reserve strip that reserves 44 px for the bar.
tray.rs Tray, TrayUi	Services, bus threads	Owns org.kde.StatusNotifierWatcher (only if no other watcher runs) and is its host; items by bus name or path (path-only owners found by asking all connections at once); icons from themes or IconPixmap; dbusmenu menus in magic-tray-menu; the tray IPC handler.
png.rs	plain functions	Stored-deflate PNG writer for pixmap tray icons (amane draws images from files).

The equalizer: one setsid -f pipewire -c eq-<key>.conf per player with libpipewire-module-filter-chain, 10 biquads (low shelf 31 Hz, peaking 63 Hz–8 kHz, high shelf 16 kHz). The player's streams move into sink magic_island_eq.<key>; gains change live with pw-cli set-param magic_island_eq.<key> Props '{ params = [ "b1:Gain" 6.0 ... ] }'. A player only gets one when its curve or output first changes, or it has a saved non-flat curve.

The view (island.rs view)

• One LayerWindow per monitor (window_per_monitor, namespace magic-island, Layer::Overlay because the user's windows are fullscreen), sized for the open card; an InputArea lets clicks outside the island through.
• No player: idle(), a static click-through pill with a drawn note and "Nothing playing".
• Monitors: tap(monitor) opens the card where it was clicked, expand() (IPC) on the focused monitor; open_on restarts the animation when the card moves. Other monitors render phases(0.0), their pill. Only the focused window and the one with the card read Spectrum (reading a Service is what subscribes a window), so the rest stay idle.
• phases(open) splits one linear 560 ms timeline: pill fades out over 0–0.18, the empty shape grows over 0–0.6 (cubic ease-out), the card fades in over 0.6–1 (smoothstep). At most one face is built per frame. veiled() fades a face with a backdrop-coloured veil instead of .opacity().
• Background: the wallpaper surfaces as one 150° three-stop gradient (backdrop()); with cover_colors the card's background blends to backdrop_colors(Palette) (dominant, colors()[1], background(), darkened by deepen() to 0.22 / 0.15 / 0.07 brightness) as it opens. White-alpha overlays (CHIP, TRACK, CHIP_HOVER, CLEAR) stay unthemed.
• Hover: hovering(spot) returns the on_hover closure easing that Spot to 0/1 over 220 ms; the view reads island.lit(spot). Amane matches hover targets by their index in the last frame, so a vanished button never hears the pointer leave: entering a spot fades all others, and collapse()/set_menu() drop stale highlights. Effects are colour blends and .scale()/.translate(); transport buttons go bold (bold_icon: the glyph's own outline, 0–2 px).
• Transport: play/pause flips at once through playing_now; after any transport click refresh_soon() re-reads Media at 150 and 600 ms (Media polls once a second). island play-pause over IPC takes the same path.
• pill_levels(): 7 bars, bar i = loudest of spectrum slice i, ×1.4, linear, drawn up and down from the middle. ring(): 40 candles, levels from mirrored(), 3 paths (glow, body, flame tips).
• Moving the pointer off the open card collapses it.

Hyprland 0.56 (Lua config) pitfalls

• Never return visible(false) from a layer view. Hyprland configures a layer surface before its first commit, so a hidden-at-start window gets a size, Amane logs "<name> has no child to draw" and the window stays blank forever. With nothing to show, return window.click_through().child(Rectangle::new().width(0.0).height(0.0)).
• Workspaces::focus does nothing: it sends dispatch workspace N, which the Lua config rejects. Use hl.dsp.focus({ workspace = N }). Test dispatches harmlessly on the already-focused workspace.
• hyprctl eval / hyprctl repl run Lua in one shared state; globals (e.g. a layer rule kept in a variable) vanish when Hyprland reloads its config.
• Layer-rule match.namespace is a full match: ^magic- matches nothing, ^magic-.*$ works. Layer rules can only switch blur on; blur strength (decoration.blur.size, passes) is global.
• $XDG_RUNTIME_DIR/hypr/ keeps folders of dead instances; the live one has .socket.sock (amane-remote.sh picks it). A shell started with a dead signature gets an empty Workspaces list.
• hyprctl layers lists hidden layer windows too (alpha 0); filter on alpha > 0.
• Reserved space and fullscreen: a layer with Zone::Reserve reserves its whole height (amane's measure_thickness), and only on a window anchored to one edge (or an edge plus both sides), so corner-anchored or card-sized windows can't reserve. Hyprland keeps tiled and maximised (fullscreen 1) windows out of the reserved strip and lets fullscreen 2/3 cover it. Bar windows must keep Zone::Ignore, otherwise they get pushed down below the strip. Check with hyprctl monitors (reserved: 0 44 0 0) and a tiled window's at:.
• Hyprland state comes from its sockets, not polling: .socket2.sock streams events, .socket.sock answers monitors/clients. Fullscreen 2 and 3 hide the bar; 1 (maximised) keeps it.
• A peek (workspace or island) only shows on the focused monitor, and only matters on a fullscreen one; test with a workspace hop between two fullscreen workspaces there, then restore the focus.

Performance rules

• Measure before and after every visual change, baseline first, with the same player and monitor. The GPU-busy counter (gt0/gtidle/idle_residency_ms) reads 100% even with the shell paused, so compare Amane's per-frame gpu time and top CPU instead.
• Don't resize a shape with the full card inside it; grow the empty shape, then fade the content in.
• .clip() and .opacity(<1) each add a GPU layer (a clip over the whole card once took the iGPU from 47 to 9 fps). A shadow is two layers per frame: halos only while hovered.
• Gradients cost: a gradient background plus 10 gradient fills halved the frame rate; one backdrop gradient is affordable. A veil over a gradient must be the same gradient at the same size (Amane places gradients by each rectangle's own bounds).
• Draw repeated marks (candles, bars) as one Path per look.
• Streaming data needs amane::request_frame() in the view while levels are live; a Service write alone got about 11 redraws a second from Hyprland.
• Rectangle::child holds one child; stack layers with Stack.

Develop

1. Type-check on the dev host without the desktop, in a scratch crate outside the repo (don't commit a Cargo.toml). Scratch at /root/.hermes/cache/scratch/amane-check/ (pruned after 24 h idle; recreate it):

   [package]
   name = "amane-shell"
   version = "0.1.0"
   edition = "2024"
   
   [[bin]]
   name = "amane-shell"
   path = '/root/amane-config/src/main.rs'
   
   [dependencies]
   amane = { path = '/root/amane' }
   
   [workspace]
   

   Then cargo clippy there. The first build takes minutes.

2. Unit tests: rustc --edition 2024 --test src/parse.rs -o /tmp/parse_test && /tmp/parse_test (10 tests).

Testing on the desktop

tools/amane-remote.sh HOST COMMAND; the full list is at the top of the script.

H=fabian@192.168.2.169
tools/amane-remote.sh $H deploy src        # back up the remote src once, then replace it
tools/amane-remote.sh $H build             # amane compile
tools/amane-remote.sh $H start             # stop the shell it started earlier, start one WITH frame logging
tools/amane-remote.sh $H fps 3             # frames, average gap, gpu wait
tools/amane-remote.sh $H anim island       # open/close 3x; stutter = 20-500 ms gaps
tools/amane-remote.sh $H shot magic-island out.png
tools/amane-remote.sh $H sinks start       # null sinks island_test_a/_b
tools/amane-remote.sh $H players start 4   # fake MPRIS players; "Synth Test" plays a synth into island_test_a
tools/amane-remote.sh $H ipc island status
tools/amane-remote.sh $H sh 'COMMAND'      # anything inside the graphical session's environment

• Pick a fake player with ipc island next until status names it, before anything that changes sound; stopping the fake players lets the island fall back to the user's player.
• ipc island open opens on the focused monitor (the user's, if they're at the desk). Read the geometry from hyprctl layers -j on that monitor right before each grim capture. The user's pointer can open or collapse the island mid-test.
• With one window per monitor the frame log mixes windows; resting ones show ~1000 ms gaps.
• Design screenshots: set glass in the remote shell.conf (applies live), wait ~1 s, capture pill, workspace pill and the card (ipc island open, wait ~2.5 s).
• Hover can't be tested with hyprctl dispatch movecursor (no motion events reach the surface); real motion needs a virtual pointer and moves the user's mouse, so ask first.
• Before trusting timings, check uptime: a background Gentoo build once made every build look broken.
• A fresh shell once exited with "lost the gpu (Device is lost)" (Mesa warns Xe is experimental); starting again worked.
• To tune a visualiser, sample real levels read-only (pw-record --target <stream> | cava, timeout -s KILL on each stage; a plain timeout on the pipeline orphans cava) first. Typical music sits at 0.15–0.35.
• In pkill -f patterns over SSH, use [p]ython3 ... so the pattern doesn't match the remote shell itself (a plain pgrep -f fake_tray.py once killed the test script).
• Control centre and tray without a mouse: ipc control open|close|menu <wifi|bluetooth|vpn|sound>|ethernet|wifi search|status, ipc tray list|menu|open N|activate N|choose ID|enter ID|back|close. control status ends with prompt=<what the popup shows>. A fake tray app: python3 tools/fake_tray.py name|path (by bus name, or path-only like libappindicator; it prints every call it gets). Kill it by pid.
• Maps with many nodes: with Wi-Fi off the maps are empty. Deploy a scratch copy of src/ whose map::map injects fake nodes (never commit it), then redeploy the real src. Pick sizes by simulating the drift in Python, not from single screenshots.
• Measure gaps and alignment from pixels, not by asking a vision model: ffmpeg -i shot.png -vf crop=1920:1:0:Y -f rawvideo -pix_fmt rgb24 row.raw and look for the border pixels.
• To see the island and the control centre together, open them on a monitor without a fullscreen window (focus one of its workspaces first, restore after).

After testing

tools/amane-remote.sh $H players stop
tools/amane-remote.sh $H sinks stop
tools/amane-remote.sh $H sh 'P=$(pgrep -f "^/usr/lib/python-exec/python3[.0-9]*/python3 fake_tray"); [ -n "$P" ] && kill $P'   # fake tray apps
tools/amane-remote.sh $H sh 'kill $(cat shell.pid); for i in $(seq 50); do pgrep -F shell.pid > /dev/null || break; sleep 0.1; done
  rm -f shell.log; setsid -f sh -c "echo \$\$ > shell.pid; exec ~/.cache/amane/amane-shell" > /dev/null 2>&1'

Open threads

• The user picks a glass design; then rebase glass-base plus that design onto master (they predate the control centre, whose windows draw their own surfaces), merge, and update README/AGENTS.md.
• The candle ring is still mirrored (mirrored()); the pill's bars aren't.
• The Amane library could fix the two Hyprland 0.56 issues upstream (handle a configure on a hidden layer window; use the Lua dispatch in Workspaces::focus). Not started.
• island-eq is tracked in git but rewritten live by the island, so the desktop checkout always shows it changed; untracking it was proposed, no answer yet.
• Leftovers offered for deletion, no answer yet: src.before-test/ on the desktop, seven merged local branches on the dev host. New since: the merged local control-centre branch, and old screenshots in the desktop's ~/.cache/amane-remote-test/ (dimtest.png, grab.png, shot.png, ws-after.png, shots/).
• On a monitor narrower than about 1500 px the maps stop at 600 px (MIN_MAP_W) and run under the island.
• The open/close stutter rise (25–40 slow frames per 3 opens and closes on every build) is unexplained.
• See README "Known limits".