The hands' drop-shadow outline sat on their own compositor layers, so the GPU
recomputed it every frame a bubble moved. It is now on an inner box, painted
once into the hand's layer. The bubbles rise beside the watch, never behind it.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
iOS starts an installed web app at the manifest's start_url ("/"), so the
query parameter was dropped. Opened with ?parentMode=1, the page now points at
a second manifest whose start_url keeps it, and names the app "Musik Eltern".
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The remote's hauppauge.rem names the Guide button KEY_GUIDE, not lircd's
KEY_EPG, so the remote never toggled the clock.
The overlay is now hidden by opacity rather than display: none, so it is
already laid out and painted when Pos1 or Guide shows it. The drop-shadow
filter on the whole watch is gone (it re-rendered and re-blurred the dial on
every second-hand tick), the hands sit on their own layers, and the bubbles
and countdown cards only exist while the clock is on screen.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Hiding the vertical scrollbar with overflow: hidden also stopped the wheel from
scrolling. Keep overflow: auto and hide only the bar, as the shelf rows already do.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Authentication (01), one input policy for skip/mute/volume steps (04), the
firmware's interleaved serial frames (05), podcast as an album kind (06), and eight
smaller cleanups (07). Each says why, what was decided, the design, the tests and
when it is done.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
- python-backend/README: track analysis is implemented (2e0e6ad); the cache layout,
the module table (lirc, screen, timers, podcasts, tippen) and the route table
(track analysis, HA proxy, typing game) were missing pieces.
- docs/REPO_OVERVIEW: Mein Zimmer and the clock are implemented; Home Assistant is
also reached over HTTP through the room page's proxy, not MQTT only; the IR remote,
scripts/, the optional config sections and figure kind were not mentioned.
- web/README: the key table was the mockup's (Ctrl transport keys, TAB over 'all');
?pi=1 is now a preset with per-browser overrides, not a read-once flag; the room
page and the typing game exist; eslint is part of the checks.
- Docstrings that described earlier designs: HaConfig (written as browser-direct in
the same commit that built the proxy, a8ed350), services/__init__, events,
playback, media, colors, tippen progress and AppHeader, lowPower.ts, and the
'Hoerbuecher' spelling.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
- claude-design/Dolphin Beats Music Player.zip (32 MB): the design tool's export of
the mockups already unpacked next to it. Still in history (30065bb) if needed.
- claude-design/pasted-*.png: byte-identical copies of dolphin-mascot.png and
dolphin-remote.png that nothing references.
- python-backend/notebooks: the chord-recognition course material with stale paths.
effect_debug.ipynb stays.
The hardware models are untouched, and the history keeps its size: removing files
from the tree does not shrink the pack.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
- wire.with_figure: the device uses dataclasses.replace directly.
- DeviceConnected/DeviceDisconnected: emitted, never subscribed to, and a duplicate
of ConnectionChanged(target="firmware").
- ScreenService.toggle(): never called.
- GET /api/lirc (LircConfigOut, api.lircConfig): no front-end reads it.
- POST /api/volume's delta_percent (api.nudgeVolume): no front-end sends it.
- LibraryCache.shrink_stored_covers: a one-off migration that re-read every cover
header on every scan. _INDEX_VERSION 7 rebuilds any cache that could still need it.
- Firmware: the MFRC522 library and its unused globals (lib/nfc/rc522.c is the
driver in use), and mfrc_version.h, a reader test sketch nothing included.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The hub sent every open tab a full snapshot on every StateEvent, LED writes
included - placing and removing a figure was 13 snapshots, 9 of them for LED
changes the snapshot does not even contain. List the events that matter, and make
a test fail when a new StateEvent is sorted into neither list.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
a7fb56c pinned mypy to Python 3.11 for the Pi. In a dev venv that also has the
analysis extra, numpy's stubs use 3.12 syntax, so mypy aborted on them before
checking a single file - and from the next day on four real errors went in
unnoticed (ebab03b, 0ef5a04, 9fd106b). Skip following numpy/librosa instead; only
librosa_analyzer.py touches them. Fix the four errors (variables reused with a
narrower type).
There is no CI, so scripts/check.sh is the gate: pytest, ruff, mypy, vitest, tsc
and eslint in one command.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
play_figure resumed only when the player's playlist *was* the figure's playlist
object. rescan_library rebuilds every Playlist, so after the refresh button - or
the podcast poller landing a new episode, every few hours on its own - putting a
figure back started it over. Playlist is a frozen dataclass: compare by value.
The album picker's "keep what is already playing" check had the same identity
comparison and restarted the album after a rescan; it compares by value too.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
It has been untouched since 2026-09-21 while the web front-end got every feature and
the Pi performance work since (thumbnails, ?pi=1, number keys, clock, volume steps).
Keeping it meant a second, already stale copy of the search, keymap, Home Assistant
client and API types. Nothing outside slint-frontend/ referenced it.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Each fix to the home-screen app's layout on iOS broke something else, and they
all traced back to black-translucent: it runs the page under the clock, so any
top-anchored control without its own inset sat behind it (the play view's back
button), and since iOS 26 it sizes the app one status bar short with a strip at
the bottom that CSS cannot paint (WebKit 301108). The status bar is opaque now
("black"), the page starts below it, and lib/fullHeight.ts - the JS workaround
for the short viewport - is gone with the CSS that depended on it.
viewport-fit=cover stays, so everything against a screen edge now adds that
edge's inset from shared --safe-* variables: the corner back button, the
remote-assign button, the assign toast, the tool buttons and the tab rail did
not before. The rules are written down under "Layout on phones" in the README.
On a phone held sideways the mini player and the tabs share one row instead of
taking half the screen.
Home-screen installs read the status-bar tag once: remove and re-add the app.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The unpinned platform had moved to Arduino core 3, which the firmware no longer
built against. Pinned now: pioarduino's platform 55.03.312-1 (Arduino core 3.3.12,
ESP-IDF 5.5.5) - PlatformIO's own espressif32 platform still ships core 2 - plus
MFRC522 1.4.12, AceButton 1.10.1 and native 1.2.1.
- LED strips: the legacy RMT driver is replaced by the new rmt_tx API with a bytes
encoder, at the same 50 ns tick, bit timings and byte order. ESP-IDF 5 aborts at
boot when both drivers are linked, and the driver now picks the channel itself,
so Esp32DriverRGBW::begin() takes only the pin.
- Button backlights: ledcAttach instead of ledcSetup/ledcAttachPin, and ledcWrite
is addressed by pin rather than channel.
- Rotary encoder: gpio_reset_pin instead of the removed gpio_pad_select_gpio.
The touch pads stay on the legacy touch driver: on the plain ESP32 it only warns,
its conflict check passes as long as the new driver is not linked (it is not), and
moving would change the raw counts the thresholds were tuned against.
src_filter is build_src_filter now. Not yet tried on the device.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
A figure folder that holds album folders instead of audio can play any of them,
up to six, in folder order (depth-first, alphabetical). Each is its own album,
titled "Fuchs - <folder>" and of the figure's kind. Putting the figure down plays
the album it played last, or the first, straight away; another figure placed in
between forgets the position in that album, not the album itself.
Meanwhile the ring offers all of them: one section per album from a fixed palette,
edges shifting towards the neighbours' hue, the selected one bright and wobbling.
The knob moves the selection one section per detent and wraps; pressing plays the
selected album, and left alone for 8 s the picker fades into the figure's colours
and the music carries on. While open, the device hands the knob to the picker
(`capture_knob`): turns do not change the volume and the press is not published to
Home Assistant. A tick per step and a confirm "pling" go with it.
The firmware only draws: one new ring effect, LED_WHEEL_EFFECT_SEGMENT_SELECT (23),
re-sent on every detent with the section to glide from. The look is tunable through
a temporary `general.figure_selection` config section until the right values
become constants.
Every UI sound - the screen's blubb, the parents' timer, the picker - now follows
the music volume with a floor, times a new parent-mode "Effekt-Lautstärke"
(`general.effects_volume`).
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Einfg on the kiosk keyboard, or Guide (KEY_EPG) on the IR remote, shows a big
analog watch over whatever page is open, built from the Delphin Uhr mockup
without its swimming dolphin. The overlay stays mounted and is only hidden, so it
appears at once. Beside the watch, only while they run:
- The bus: on school days, a silent countdown to `general.clock.bus_time` (07:25)
that starts `bus_lead_minutes` (60) before. School days skip weekends and the
school and public holidays of `general.clock.subdivision` (DE-BY), fetched daily
from openholidaysapi.org and cached under the library cache. A badge on the
clock says whether there is school today, and why not.
- The dolphin timer, set in parent mode (5/10/15/30 minutes, or until a time).
It lives in the backend, so a parent's phone can start it for the kiosk. When it
runs out the kiosk opens the clock at 00:00 and a soft bell chimes on the
device; it clears after a minute, or at the next key, tap or remote press.
Parent mode is now a way in rather than a page: `?parentMode=1` adds a 🔧 button
next to the ⌨ one, or an "Eltern" tab in the phone's dock, so the URL can stay
bookmarked. The timer comes first there, the device settings below it.
Unmapped IR buttons are logged at debug level, to find out what irserver calls a
new one.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Delphin Uhr.dc.html is the reference for the kiosk's clock overlay. The bundle
now also carries it, its bus and dolphin-with-watch images, and the dinosaur
skin's artwork.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Next to the five volume bars, a − and a + button move the volume by
`volume_increment` device units, the same step as the IR remote's volume keys and
the rotary encoder, via a new `{steps}` form of POST /api/volume. The bars still
jump to a percentage.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The three shelf rows fit the screen, and each scrolls sideways on its own, so the
page around them no longer scrolls vertically.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
With `general.screen.off_command` set, the monitor switches off once nothing has
played and nobody has done anything for `auto_off_minutes` (15 by default, 0 for
never, editable in parent mode). "Doing anything" is any intent a person could
have caused: a remote key, a web command, or a key press or tap in the kiosk,
which the page reports at most every 30s via POST /api/screen/activity. The first
such intent after an auto-off runs `on_command`, so the remote wakes the screen
too. The auto-off itself is silent; switches a person asks for still blubb.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The IR remote's digits and the kiosk keyboard's number keys now both emit
PlayRemoteSlotRequested, and one reaction resolves the slot for either (a podcast
show to its newest episode, an album to the album). POST /api/remote/press/{digit}
is the keyboard's way in. Mid-search a digit is still part of the query.
A remote slot can start one song instead of a whole album: `track_index` on an
album slot plays from that track, then the rest of the album. Shift+A on the play
screen arms "this song" where A arms "this album", and the assign popup offers
both. A series slot cannot name a song, and an index past the album's end is
rejected when saving.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
ScreenService plays a synthesised two-bubble "blubb" (musicmouse.blubb) after each
switch, through aplay on the music's ALSA device at the player's volume: rising for
on, falling for off, going by the last word the toggle command prints. Nothing plays
with simulated audio or at volume 0. Built from samples rather than shipped as a
file, like the front-end's WebAudio blips, so its loudness follows the volume.
Del maps to a new "screen" action on every page, which App.tsx sends to the new
POST /api/screen/toggle only when the page is served from localhost - the kiosk on
the device - so Del on a tablet does not switch off a screen in another room.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
KEY_POWER emits ToggleScreenRequested, which a new ScreenService turns into the
optional general.screen.toggle_command (musicdolphin: ansible's kiosk-screen). It is
a service, not a reaction, so the subprocess runs in its own task instead of holding
the bus; a press while one is still running is dropped, and failures, hangs and a
missing command are only logged.
irserver reports every line of a held button as repeat 00, so a slightly long press
on Next skipped several tracks and Power would switch the screen off and on again.
A line of the same button within 0.25s of the previous one now counts as a repeat.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
irserver (musicdolphin's IRTrans daemon) serves lircd's network protocol on port 8765,
but sends "key_play hauppauge" where lircd sends "KEY_PLAY Hauppauge", so every press
was dropped. Uppercase button names when parsing and compare the remote name
case-insensitively.
irserver also always reports repeat 00, so a held transport button is not debounced
there; volume ramps as intended.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Sizing the stage from the screen height was not enough on iOS 26: with the layout
viewport a status bar short, WebKit clips at the short viewport and draws nothing
below the document, so the dock still floated over a dark strip. In standalone mode
`<html>` now takes the larger of `100vh` and the screen height too, and the stage
follows it.
Coming back from the background, or a relaunch restoring the page, is when iOS is most
likely to hand back a short viewport and it does not always fire `resize`, so the
height is also re-applied on `pageshow` and `visibilitychange`.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
In the iOS home-screen app the layout viewport - `100dvh`, `innerHeight`, even a
`position: fixed; inset: 0` box - can come back one status bar short, which left the
dock floating that far above the bottom of the screen with an empty strip below it.
In standalone mode the stage now takes its height from `screen` instead (swapped for
landscape, since iOS reports it in portrait), via a `--full-height` variable that
lib/fullHeight.ts sets only there; the keyboard inset is measured against the same
height. A Safari tab, Android and the kiosk are unchanged.
The mascot and app title above the group switch only cost a row on a phone, so they
are gone; the connection/figure pill moves above the group switch, centred, and only
shows while it has something to say.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
With the ambient canvas off under ?pi=1, the backdrop blur no longer buys anything:
it paid for itself by giving each panel a compositing layer above the animating
canvas, and with nothing animating underneath it was just cost - enough to make the
browse rows visibly sluggish on the device. So the Pi profile now drops every backdrop
blur (`glassBlur`, which the music app and the typing game share), and turns the
drift and the moving mascot back on: plain CSS transform loops the compositor runs by
itself, which measured as noise.
The knob set is consolidated to match: one glass toggle instead of panel and card
blur, one shadows toggle for cards and tiles, the typing game's drift and view
transitions folded into the shared ones, and the backing-store cap fixed by profile
rather than a setting. The parent dialog groups them into Darstellung / Abspielen /
System and hides the canvas tuning while the canvas is off. Stored settings move to
musicmouse.perf.v2, since v1 drafts held the old preset in full.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Two kids, two looks: ?theme=dolphin or ?theme=dinosaur picks one, and the
choice is remembered in the browser so the installed home-screen app keeps it.
Everything that differs per child lives in one Skin object per file under
web/src/lib/skins/; layout and feature toggles in lib/theme.ts stay shared.
The dolphin renders pixel-identical to before.
The hardcoded sea hues (210 music, 175 typing game, 340 accent) are now CSS
variables written once at startup, and the Gecko-friendly sRGB gradient stops
come from the skin. The dinosaur is jungle green with a lava-orange accent,
its own mascot, icons, manifest and browser colour.
Nothing underwater is left in the dinosaur skin: leaves instead of bubbles
(play-view canvas, room page, typing game), a walking mascot, lava blobs in
the bubble game, a hatching egg instead of the treasure chest, dino pets that
walk or fly, a prehistoric speed ladder with the T-Rex as the hero rung, and
renamed games and light colours. Pets and ladder rungs are keyed by the
backend's existing ids, used as slots, so the backend and save format are
unchanged. Pets show as emoji until drawings exist.
python-backend/.gitignore: ignore tippen-curriculum-dino.yml, the dino
device's personal curriculum, like the original.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
On an iPhone the sticky group switch sat on a fixed teal fade, which was the
wrong colour as soon as the ambience took on another album's tint (brown for
Conni). It is now a neutral frosted blur that fades out, so it suits whatever
is playing, and it reaches up under the status bar so scrolled content never
shows behind the clock.
After the search keyboard had been up, 100dvh in the home-screen app came back
a status bar short and the white page showed as a strip along the bottom. On
phones the stage is pinned to the viewport edges instead, with the sea colour
behind it.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
- New SHOW_APP_TITLE switch (off): the mascot + title header goes, the status pill
stays on a slim row of its own.
- SHOW_ROW_TITLES off; with only icons, the shelf icon moves into a column left of
the tiles so an icon-only shelf is no taller than one without.
- Help sheet regrouped into sections in a compact two-column grid, and reachable
from a small keyboard button (not on phones or the typing page).
- Ctrl+d/Ctrl+u half-page jumps removed.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
A podcast's episode cover was only fetched right after the feed sync downloaded that
episode, so anything placed by hand (or downloaded before the og:image fallback
existed) never got one - in Wissen macht Ah only the three newest episodes had their
own art. The sync now fetches the cover of every on-disk episode in the feed window
that lacks a sidecar image. One that can't be found or fetched is recorded in
.missing-covers.json and left alone for the same 7-day backoff as failed downloads,
so a feed without per-episode art doesn't cost a page fetch per episode every poll.
The scanner's per-episode fingerprint now includes the sidecar cover, so one added
after the episode was first scanned invalidates its cached entry instead of being
ignored until the index is rebuilt. Episodes without a sidecar fingerprint exactly
as before, so existing caches stay valid.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Phones got the kiosk's keyboard-first layout squeezed down: shelves cut off
by the tab rail and a header taking a fifth of the screen. They now get
their own layout, chosen by a media query (narrow, or short landscape on a
touch screen); tablets and desktop are unchanged.
- Pick a group (Musik/Hörbücher/Podcasts), then a category, then scroll
through a grid of smaller albums.
- A floating glass dock: mini player, Musik/Zimmer tabs (no typing tab - a
phone has no keyboard for it) and a search button that turns into a real
search field, raised above the on-screen keyboard.
- Album sheet and room page get compact phone variants.
Browse state stays in the shared UiState, so resizing or rotating keeps
your place.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The report was an ssh command string, which the remote user's login shell runs - and
on the devices root's is fish, which rejects `if ...; then`. The copy itself finished,
but the script then exited 127 at the very end, so a successful sync looked failed.
Feed the report to `sh -s` on stdin instead.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The second of the web front-end's three pages. Lights get a toggle, five brightness
steps and six colours; shutters get four presets, up/stop/down and a window that
reflects the position Home Assistant reports. Its own violet hue rather than the
player's teal, so the two pages read as siblings rather than one bleeding into the
other - the same reason room.css has its own tokens.
The backend keeps Home Assistant's URL and token and relays the calls, so this client
learns only which entities exist, exactly as the browser does. A 404 from GET /api/ha
means Home Assistant is not configured, which is a normal answer: the tab simply does
not appear.
Polled, not subscribed, and only while the page is on screen. HA's own auth-and-
subscribe websocket is more machinery than a room panel needs, and a lamp's state is
not worth a request every 2.5 seconds when nobody is looking at it - this device has a
music player to stay out of the way of. A tap patches the entity locally and records
when; a poll that was already in flight when the patch landed is dropped for that
entity rather than putting the lamp back to "off" until the next one catches up. That
reconciliation is the one non-obvious thing in useHomeAssistant.ts and it is carried
over for the same reason.
Nothing on a shutter card animates locally. A real cover reports its own movement and
position, and a client-side guess at where it will end up is precisely what would
fight with that - the design mockup animated it because it had no backend to ask.
Two additions beyond a straight port, both because a kiosk has no pointer:
CTRL+TAB cycles the pages. The web's tab rail is pointer-only, which sits badly with a
front-end whose README claims every screen is reachable from the keyboard. Plain TAB
was already the group cycle, and CTRL+TAB is the idiom for the outer one everywhere
else.
--page opens directly on a page, which a kiosk may well want and which is the only way
in with neither keyboard nor pointer.
Verified against the real Home Assistant: the shutter reported open and the lamp off,
and both drew that way. Nothing was switched - those entities are hardware.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Slint cannot read woff2, which is the only form web/public/fonts/ carries, and Raspbian
has no fonts-nunito to install instead - so the device was going to fall back to
DejaVu Sans, which is not what any of this was drawn in. The upstream variable TTF
(SIL OFL, weights 200-1000) is small enough to vendor and covers every weight the
design uses, including the 800/900 the headings lean on.
SLINT_DEFAULT_FONT takes one file and makes it the primary font. shell.nix now points
at this copy rather than at nixpkgs' own, so development and the device render from the
identical file, and the binary falls back to finding it beside the executable when
neither the dev shell nor the device's launcher has set the variable.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
The React UI is sluggish on the kiosk and the reasons are browser-shaped: a compositor
deciding what gets its own layer, a requestAnimationFrame loop repainting the viewport,
and a JPEG decode per album card per cold start. The last three commits chased that
through the service worker and the cover pipeline. This is the other direction - the
same backend, the same layout, the same German copy, without a browser.
Browse and play only. The typing game, the room-lights page, parent mode and the
IR-remote assignment stay web-only, and nothing here touches python-backend.
Four things carry the performance claim, in rough order of how much they should matter:
Covers are downscaled once, ever. The backend serves one size - 640px on the long edge,
no ?size= - and a grid card is 180. src/covers.rs fetches each cover once on a worker
thread, resizes it, and writes a JPEG thumbnail to ~/.cache/musicmouse-slint/covers/.
Two tiers and only two, which is what makes the cache worth keeping: a per-widget pixel
size would give each album a dozen near-identical files and a fresh decode for each. The
cache is dropped when the websocket announces a rescan, because that is the one moment
the backend rewrites the art behind an unchanged cover URL - the trap sw.js had to learn
about the hard way.
The grid is virtualized. Rust hands the UI the album list pre-chunked into rows and the
view puts those in a ListView, which instantiates only what is on screen. A flat list of
660 cards gives it no rows to skip, hence the chunking. Same purpose as
content-visibility: auto on .grid > .card.
The progress bar animates between the 2 Hz pushes rather than running a clock, so
interpolation costs a property evaluation per frame on the render side and there is no
equivalent of usePlaybackClock. It is suppressed for the frame a track changes on, so a
new track jumps instead of sliding across two unrelated positions.
Nothing on screen animates by itself. The ambient canvas is not ported, for the reason
lib/lowPower.ts already gives: a loop repainting the viewport is a floor you cannot get
under while it runs at all.
The device must not use FemtoVG. On Mesa V3D it draws every runtime-loaded Image as
solid black (slint-ui/slint#11785, open), which here means every album cover; Skia and
the software renderer are unaffected. So `kiosk` is linuxkms + Skia and FemtoVG stays
the default only for desktop development. Both profiles are verified to build; the kiosk
binary links libinput/libgbm/libdrm and no X11 or Wayland at all.
lib/search.ts, lib/keyboard.ts and lib/format.ts were already pure functions with their
own tests, so they port across as pure Rust with theirs: 43 tests, no window required.
The .slint files are layout only - nothing in them formats a number or picks a word.
Verified against the real 660-album library rather than a fixture. tools/headless-shots.sh
renders the UI inside a nested headless compositor and grabs a frame per screen, which
is how that was checked on a machine whose session was locked; a Slint window needs a
real compositor, and under bare Xvfb nothing maps at all.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Search was still stalling for seconds a keystroke. A trace across five keystrokes said
the renderer main thread was blocked for 9.4 s in a single task with *zero* V8 samples
inside it - no JavaScript ran at all. What ran instead, on the worker pool during those
same 9.4 s: ImageDecodeTask 7.6 s, RasterTask 1.7 s, and only 229 ms of actual "Decode
Image". The main thread was waiting on cover bytes, not computing anything.
Two causes, and the first one was mine. The stale-while-revalidate handler added earlier
today made every cover do a Cache Storage read *plus* a network fetch *plus* a cache
write, all serialised through one worker thread. Measured with
Network.setBypassServiceWorker: worst keystroke 5580 ms through the worker against
461 ms without it. So covers no longer go through the worker at all. That costs nothing
here: this worker only registers on localhost or over HTTPS, which on this setup is the
kiosk on the device itself, where the backend is the same machine and a cache lookup is
strictly more work than asking for the file. The shell caching, which is what makes it
installable, stays.
The second is that decode cost goes with pixel count, and 640 px was headroom for a
tablet at devicePixelRatio 2 that nobody had asked for. A browse grid paints a card
132 px wide; the largest any screen asks for is 340. At 384 px, typing "conni" over 343
albums, keydown to painted, three runs:
before 3735, 270, 5580, 88, 57 ms
after 873, 75, 17, 24, 26 ms
`shrink_cover` never scales art up, so dropping the limit cannot be applied by
re-reading the cache - only by going back to the original art. _INDEX_VERSION 6 does
that: the index is discarded and every album is scanned again through store_cover.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
The downscaling landed and the device kept serving 1920px art, because _cover_for has
two branches and only one of them went through the cache. An album with a cover.jpg
already sitting beside its audio - which most of this library has - had its `cover`
point straight at that file, so FileResponse served whatever the internet had given it.
The cache directory was full of tidy 640px files that half the albums never used.
Both branches now store through the cache, in _cover_for and in _cover_for_episode
(sidecar and shared-folder art alike). The undownscaled bytes still come back alongside
the path, because colour extraction wants the real thing.
That re-points `cover` for every album, but only for albums that are actually rescanned,
and the scanner reuses anything whose fingerprint is unchanged - so _INDEX_VERSION goes
to 5. An index from 4 is discarded and rebuilt, which is what that mechanism is for and
is cheap by design.
Two tests asserted `cover == <the library file>`, which is precisely the behaviour being
changed. The episode one was also checking something real - that an episode's own
sidecar art beats the show's shared cover - and both covers now live under their own
album id, so it reads the colour back out of the stored file instead of comparing paths.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Shrinking the covers changed nothing on the device, twice over, because two layers were
independently serving the old bytes and each had the same wrong premise written into its
comment: that a cover is "content-addressed by album id". The id addresses which album
the art belongs to. The bytes behind the URL change whenever the art is reprocessed.
The backend sent `public, max-age=604800`, so every browser that had loaded the page in
the previous week kept decoding the 3000px original out of its own disk cache without
asking. It now sends `no-cache`, which does not mean "do not store" but "revalidate
before reusing" - the file stays cached and the usual answer is a 304.
The service worker cached covers cache-first on a URL that never changes, which means a
client could serve a stale cover for ever; bumping the cache name fixed today's covers
and would have had to be done again for the next batch. It now does
stale-while-revalidate: the cached copy is returned immediately, so a grid of album art
still never waits on the network, and a fresh copy is fetched behind the page for next
time. Neither layer needs a version bump when art is reprocessed again.
The revalidation is off the paint path entirely, which is what makes `no-cache`
affordable here: the service worker answers first, the network call happens behind it.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
A DevTools trace recorded on musicdolphin put the whole question beyond argument:
ImageDecodeTask 29,939 ms x20 (~1.5 s each)
Decode LazyPixelRef 3,924 ms x20
Decode Image 3,530 ms x22
Paint 99.7 ms
Layout 77.3 ms
all app JavaScript <400 ms across 22 seconds
Image decoding is not the largest cost on that page, it is very nearly the only one.
The covers should already have been small - the backend downscales to 640 px now - but
the page was still decoding them at 1920, 1600, 1400 px, with deliveryType
"cache-storage" on every one. The service worker caches cover art cache-first, keyed on
URL, on the premise that it is "immutable per album id". The id did not change; the
bytes did. So every client that had ever loaded a cover kept serving the 3000 px
original from its own disk and never asked the backend for the new one.
Renaming the cache is the retirement mechanism the worker already has - `activate`
deletes every cache that is not one of the two current names - so COVERS becomes v2,
with a note saying that changing how covers are produced means bumping it.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Covers. Art out of an ID3 APIC frame is sized for a record sleeve: this library
averaged 3000x3000 and 580 kB per cover, 140 MB across 284 albums. The browser was
decoding nine megapixels - around 36 MB of bitmap - for every cover it painted, to show
it in a 185 px card, on a Pi with 2 GB of RAM. The largest any screen in this app asks
for is 340 px (the play view), so cache.store_cover now downscales to a 640 px long
edge, which leaves room for a tablet at devicePixelRatio 2 and cuts the decode about
twentyfold. Pillow was already a hard dependency, for colour extraction.
scan_library reuses an album whose fingerprint is unchanged without re-reading its
tags, so covers already on disk would never be rewritten - hence shrink_stored_covers(),
a pass at the top of a scan. Reading a JPEG's dimensions only parses its header, so
after the first run it costs one small read per album. Art already small enough is
returned byte-identical rather than re-encoded, so repeated scans cannot slowly grind
it down, and anything Pillow cannot read is passed through untouched: a cover that is
too big is a performance problem, a cover that is missing is a visible one.
Animation. Halving the ambient canvas to a quarter of the pixels at 30fps took it from
53.5% of a core to 25%, and 25% was still not good enough to use. A requestAnimationFrame
loop repainting the viewport is a floor you cannot get under while it runs at all, so
?pi=1 now switches it off outright rather than thinning it, along with the decorative
CSS loops, the view transitions, the typing game's bubbles and its next-key pulse. The
pets stay on screen but hold still, through the same path prefers-reduced-motion already
took - taking the animation away is the point, taking away what she earned is not.
.stage keeps its own static gradient, so there is still a sea behind everything.
Both blurs on the panels stay on. Dropping those measured five times worse.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Searching a single letter over a real library puts 340 cards in one grid, each with a
cover image, a shadow and a backdrop blur, all of them live whether or not they are on
screen. On musicdolphin that saturated the renderer badly enough that a DevTools
Runtime.evaluate could not be scheduled on the main thread inside 30 seconds, which is
a fair description of what "sluggish" felt like.
content-visibility: auto on the grid's cards is the browser's own answer: off-screen
cards skip layout, paint and compositing and come back as they scroll near the
viewport. The grid's tracks are sized by minmax(180px, 1fr) rather than by card
content, so skipping that content cannot move the columns; contain-intrinsic-size
supplies the block-axis guess and `auto` remembers each card's real size after its
first render, so scroll height and offsetTop - which BrowseView's keep-the-selection-
on-screen effect reads - stay honest. Not behind ?pi=1: there is no visual difference
to trade away.
SHOW_CARD_BLUR separates the repeated frosted surfaces (every card, every list row)
from the handful of panels wrapped around them, because the two behave nothing alike
and lumping them together is what made the first Pi profile five times slower. A panel
is one live backdrop copy that buys its whole subtree a compositing layer; a card is
one of three hundred sitting directly over the animated canvas. ?pi=1 now drops the
card blur and keeps the panel blur.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
The first cut of ?pi=1 switched off the two things that look most expensive - the
backdrop-filter glass blur and the decorative CSS animation loops - and made the device
slower, not faster: 53.5% of a core idle became 118%. Measured on musicdolphin, sum of
the Firefox process tree, idle on the browse screen, 15s average:
full app 53.5%
+ ambient canvas at half resolution 25.0% <- the whole win
+ 30fps cap on top of that 25.0% (no idle change)
+ decorative CSS animation loops off 24.6% (noise; left on)
+ backdrop-filter glass blur off 118.1% <- 5x worse
The blur is what promotes each glass panel to its own compositing layer. Without it the
animated canvas and the whole album grid above it collapse into one layer and every
canvas frame repaints all of it. The most expensive-looking CSS in the app is what was
keeping the rest of it cheap. SHOW_GLASS_BLUR and SHOW_DECORATIVE_ANIMATIONS therefore
go back to unconditionally on, in both the music app and the typing game, with the
numbers written down next to them so the next person does not repeat this.
What is left is one real change - paint a quarter of the pixels - plus three caps that
only bite while something is playing and so are not in the table above: 30fps on the
canvas, 60 bubbles alive at once, and ten progress-bar re-renders a second. Those three
are unmeasured; measuring them means playing audio.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Two separate costs, both measured on musicdolphin (Pi 4, 1920x1080 kiosk), where the
app was burning ~70% of a core with nothing happening on screen.
Per-frame work. The ambient canvas repaints a full-screen gradient plus a particle
field every frame, and `usePlaybackClock` pushes a React setState per animation frame
into both PlayView and PlayerBar for the whole length of a track. `?pi=1` (lib/
lowPower.ts) makes those cheaper rather than switching them off: the canvas paints a
quarter of the pixels at 30fps with a bubble cap, and the clock renders ten times a
second - a progress bar advances one pixel every few hundred ms and its label has
one-second resolution, so nothing on screen can tell. Only the effects with no cheap
version actually go: the backdrop-filter glass blur and the decorative CSS loops.
Also drops a redundant full-canvas clearRect that the opaque gradient always covered.
Search. normalize() runs a Unicode NFD decomposition, and albumMatches/songMatches
called it on every album title and every track title on every keystroke - 4969 of them
for a track search, whose answer cannot change until the library does. buildSearchIndex
does it once per library payload; a keystroke is now String.includes over strings that
already exist. On the real library that is 33ms -> 3.4ms for an eight-letter track
query on a laptop, and this runs on a Pi. The same index partitions albums by shelf and
pre-sorts each shelf's categories, which App and BrowseView were deriving separately
from the same data.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Requiring 3.13 meant the device needed an interpreter the distribution does not
have, which is what dragged uv in, and uv then had to be matched to the Pi's
32-bit userland by hand and to build Pillow from source because no armv7 wheel
exists for a 3.13 ABI. Dropping to 3.11 removes all of that: apt provides the
interpreter and piwheels has prebuilt armhf wheels for the native dependencies.
The 3.13-only syntax was shallow - PEP 695 throughout, which converts back
mechanically:
type X = Y -> X: TypeAlias = Y
type Handler[E: Event] -> E = TypeVar("E", bound=Event) plus a plain alias,
which is generic anyway because it carries a TypeVar
def f[T: Bound](...) -> a module-level TypeVar
Also drop the one @override (3.12, and static-only), and stop the lirc test
harness calling Server.close_clients(), which is 3.13: the scripted handler now
releases its connection when asked, which is what that call was there to force.
Verified on 3.11.14 - 485 passed, mypy strict clean - and still 496 passed on
the 3.14 dev venv, which additionally has the analysis extra.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
