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>
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>
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>
A show that has published for years is unbounded: GEOlino Spezial alone is 358
episodes and 5.8 GB, and the device it syncs onto is a 30 GB SD card that also
holds the rest of the library. Nothing stopped the 6-hourly poll from eventually
filling it.
Cap each show's folder at general.podcast_episode_limit (default 50, null to
keep everything), pruning the oldest past that after each sync pass.
The same limit caps what is downloaded, and it has to be one number for both.
Prune to the newest N but keep fetching everything the feed offers, and every
poll would re-download exactly the episodes the previous one deleted - forever,
at full size, since missing_episodes() decides purely from what is on disk.
There is a test for that specific loop.
Pruning only touches files named the way this module names them
(YYYYMMDD - Title.ext), so feed.txt, folder.jpg, the failed-download record and
anything placed by hand are all left alone; a parse that fails means "not ours",
not "delete it". An episode's sidecar cover goes with it. A pass that only
deleted still reports a change, because the library needs the rescan just as
much as it does after a download.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
python-backend/.gitignore lists /.musicmouse-cache, but index.json and seven
cover JPGs were committed before that ignore existed, so the ignore never
applied to them. index.json is a 2 MB machine-specific blob holding 6606
absolute /home/martin/... paths, rewritten on every app start - it showed up as
modified in git status permanently, and a deploy checkout would have shipped a
dev machine's index to the device.
The files stay on disk; only the tracking goes.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
A lesson's `keys` was capped at two outright, which made it impossible to write
"K, with the four keys before it still in play". The cap now counts only what is
*new*; keys already taught can be listed freely, and the spotlight falls on the
new ones alone, so a lesson titled "K" drills K instead of spreading itself
evenly over all five keys it names. A round that introduces nothing new still
spotlights its whole list - that is the "mixed" replay, unchanged.
Also documents `unlocks:` paths as relative to the library root. Absolute and
"~" paths still resolve, so an existing file keeps working, but a relative path
survives moving the library.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Pearls were earned every run and spent on nothing: the aquarium fills up by
finishing worlds, not by paying for it. A counter that only ever goes up is one
more stat competing for attention on the result sheet and the home screen, and
one more field to carry through the run payload, the progress file and both test
suites.
Stars and the animal ladder already say how a run went, so nothing is lost.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
A first pass over an unanalysed library is hours of librosa, and there was no
reason for a desktop to spend them one core at a time. Analysis now runs in a
process pool sized by `general.library.analysis_workers`, defaulting to one per
core bar one (capped at 8) when the key is absent.
Two consequences worth knowing before touching this: whatever an `Analyzer`
returns has to be picklable, and an analyzer that keeps state on its own
instance - a test double counting calls - only behaves as written with
`analysis_workers=1`.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Letters-round lessons only had bubbles and jellyfish as eligible arcade
modes; with jellyfish gone, bubbles is now always used instead of the
two alternating - no lessons need to be dropped.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
The three tab buttons now share one frosted glass pill instead of each
floating separately, and BrowseView/RoomView reserve TAB_RAIL_CLEARANCE
of right-side padding so a full-width row or grid never renders under
it - verified against the real library, where dense rows (e.g. Conni's
82 audiobooks) previously butted right up against the rail.
Also gitignore the real (non-.example) tippen-curriculum.yml and
tippen-progress.json, matching config.yml's own privacy treatment,
since the curriculum now contains real reward paths into a personal
library.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Moves tippen from a standalone app into web/ as a third tab (audio player /
smarthome / typing), replacing the old single room-toggle corner button with a
vertical icon tab rail. Curriculum and progress now come from the backend
(musicmouse/tippen/*) instead of a build-time YAML import and localStorage.
Adds reward-driven lock rendering: Cover/BrowseView/AlbumModal show a question
mark for locked albums/tracks with a hint on what unlocks them, and
ResultSheet gets a new unlock-animation block alongside the existing
lesson-unlock and aquarium-creature celebrations.
CSS from the two apps is merged carefully: identical rules (bubble/card/
key-cap/view-enter/backdrop-enter and their keyframes) are shared as-is,
while rules that bake in each app's own hue are kept separate under a
`tp-` prefix and scoped to the typing tab's own .tp-stage wrapper, so
neither app's look bleeds into the other's.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Moves the typing app's lesson plan and progress from client-side YAML/localStorage
into the backend, and adds a reward system that ties passing a lesson to unlocking
part of the music library. Lock state is always recomputed live from curriculum x
progress x the live library, never persisted separately.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
- Rename all German identifiers, types, mode ids, file names, CSS classes
and data-attributes to English throughout tippen/src; only user-facing
text (lesson titles, word lists, labels, spoken praise) stays German.
- Add a word/nonsense-word list to the "Übung: die Grundstellung" home-row
lesson in the curriculum.
- Remove the unused HandHint component and speech.ts, and carry forward
the in-progress App.tsx/component/generator/progress edits from other
sessions.
- Refresh the regenerated music-library cache index.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
_extract_audio wrote transcoded audio to a .tmp temp path, so ffmpeg
couldn't guess a muxer from the filename and aborted on every video
episode whose source was still reachable. Pass -f mp3 explicitly instead
of relying on the extension.
Separately, a permanently 404'd episode was retried on every startup and
every 6-hour poll forever, since nothing remembered past failures. Track
failed downloads per show folder with a 7-day backoff before retrying.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
- podcast_feeds.py: fetch per-episode cover art from a feed's itunes:image
(when it's genuinely distinct from the channel image) or, failing that,
from the og:image on the episode's own linked page; extract audio from
video-only enclosures via ffmpeg; match podcast-dl's filename convention
(illegal characters become "_" instead of being dropped) so enabling
feed.txt on an already-downloaded show doesn't re-download its back catalog
- scanner.py: _cover_for_episode now checks for a same-stem sidecar cover
image before falling back to embedded ID3 art and the shared folder cover
- scanner.py/__init__.py: fixed a bug where an album folder nested one level
deeper than usual (an age-range grouping folder, say) was mistaken for an
empty album and skipped; added periodic progress logging for long scans
and analysis passes
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Adds a TCP client for lircd's classic protocol: play/pause/next/prev/
volume/mute map to the same intents every other front-end already
emits, and number keys 0-9 play an assigned album/audiobook from the
start or a podcast show's newest episode, resolved fresh on every
press. The mapping is configured in config.yml and editable from the
frontend: a small "Taste zuweisen" button on the play screen (or the
A+digit keyboard shortcut) opens a 10-key picker to assign whatever is
currently playing.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
A show folder under Kinderpodcasts/ opts in by containing a feed.txt marker
naming its RSS feed. A new PodcastFeedService polls every such feed every 6
hours (and once at startup), downloads any episode not already on disk using
the existing YYYYMMDD - Title.ext convention, and triggers the same
rescan-and-broadcast sequence "Bibliothek neu einlesen" already uses - now
shared via App.rescan_library() instead of duplicated. A show with no
feed.txt is untouched, so there is no new config section for this.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
The background worker now runs a real librosa analyzer (tempo, beat grid,
per-second energy/valence curves) instead of only the null baseline, kept
behind build_analyzer() so a plain checkout without the analysis extra still
runs fine. The web player reads that per-track analysis and drives a new
animated "Ambience" background (bubbles, colour, current) that reacts to the
beat and the mood curve as the track plays, plus a debug overlay for tuning
it. Also adds a one-off script to backfill podcast cover art from iTunes.
Implements the room-control page from the design mockup: a scenes row above cards
for shutters, color lamps, and brightness-only lamps, all driven by a new
`general.ha` config section (server URL, token, ordered device/scene lists).
The backend proxies every Home Assistant call server-side (GET/POST /api/ha/...)
rather than the browser calling Home Assistant directly, so the long-lived token
never leaves the LAN device and Home Assistant's own CORS settings don't need to
know about musicmouse at all. Card kind (shutter/color/brightness-only) is
inferred at runtime from what Home Assistant reports about each entity, not
configured explicitly.
Also stops tracking python-backend/config.yml, which had drifted into the repo
despite its own header saying it shouldn't be - it now carries real credentials
locally and needs to stay untracked.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
