"The tts dispatcher thing and the text highlight is completely out of sync on
desktop." Firefox + speech-dispatcher (like iOS) reports no onboundary events,
so the highlight rode a pure estimate - and a 130-word section drifts seconds
away from an engine that speaks at its own pace.
- On platforms without boundaries the system engine now speaks one sentence
at a time, and every finished utterance re-anchors the clock: measured
duration vs estimate, smoothed (half old, half new, clamped to 0.25-4x).
Drift cannot accumulate past a sentence.
- The per-word estimate is punctuation-aware (full stop ~260ms, comma ~130ms).
- Once a boundary does arrive (Chrome), the rest of the section goes back to
a single utterance with the exact clock - the accurate path is unchanged.
- When a boundary lands mid-section, the remainder is spoken as one utterance
(it used to skip to the next section).
podcast-test's engine stub finishes utterances now and pins the whole thing:
one sentence first, the next following on its own, the highlight inside the
spoken text, a boundary collapsing the remainder, the highlight advancing.
53 checks. Version 0.7.10, SW v102.
"The music thing button with the Frutiger Aero doesn't work on the podcasts."
The podcast page is a full-screen overlay at z-45 and the control sat at z-40:
visible through the transparent page, but every tap landed on the overlay.
The control now takes a data-raised="player" position on /podcast/[id] -
z-index 46, docked to the header's corner (top under the header pill, right
mirroring the max-w-3xl geometry) where neither the transport bar nor the
voice panel lives - and the transcript gained pt-32 so the episode card clears
it. Study/review/exam feeds keep their "steps aside" behaviour.
aero-music-test: a podcast section asserts the control exists, that
elementFromPoint at its centre hits the button, and that a tap expands the
credit bar. 31 checks. Version 0.7.9, SW v101.
"The podcasts don't even work on my Linux setup on Firefox." Firefox on
Linux speaks through speech-dispatcher; without it speechSynthesis is present
but empty and silent - getVoices() is empty, speak() does nothing, no events
- so the player looked broken.
- pressing play with zero installed voices (or a start that never arrives
within 1.7s) raises a banner naming the cause, with the Linux fix
(sudo apt install speech-dispatcher) and the iOS one where it applies;
- the banner offers "Use the natural voice" when a provider key exists (it
switches the source, keeps playing and remembers it), "Add a provider key"
when not (opens Settings) and a link into the picker;
- natural-voice errors (refused autoplay, provider refusal) show in that same
banner instead of hiding inside the picker.
podcast-test fakes a present-but-empty speechSynthesis and pins the banner,
the Linux line and the one-tap switch. 48 checks. Version 0.7.8, SW v100.
"It only plays once, and when you pause it you can't unpause."
WebKit leaves speechSynthesis paused after a cancel(), and every later
speak() is queued into that paused state - silently. Fixes:
- resume() whenever the synth is paused, before speaking and after our own
cancel;
- a ~90ms beat before a speak() that follows a cancel (several engines drop
it otherwise);
- no cancel when nothing is pending.
The speech engine also no longer rebuilds on rate or voice changes: both are
read live through getters, so a voiceschanged arriving after the first
utterance cannot leave a second engine speaking over the first. A rate or
voice change restarts the current part from the same word instead (the Web
Speech API cannot change either mid-utterance), making the change audible at
once; the natural engine applies speed live and resumes across a voice swap.
The suite's speech stub now models the quirk (cancel -> paused, speak while
paused is dropped and counted) and pins that nothing is ever spoken into a
paused engine. 44 checks. Version 0.7.7, SW v99.
Two reports, one root: the platform's speech engine.
"Loud cracking at the end of every sentence" is the browser/OS TTS popping
between its internal sentence chunks (an idle cancel makes it worse). Hygiene
for that path: never cancel when nothing is pending, normalise the spoken text
(control characters, ellipses, dashes), and a 140ms beat between sections so
speech never runs into the engine's tail.
The real fix is a second voice source that does not use the platform engine:
- lib/providers carries TTS profiles (OpenAI gpt-4o-mini-tts, Groq
playai-tts, any OpenAI-compatible /audio/speech behind the custom provider),
male voices first.
- /api/tts proxies them with the same guards and the key still in headers.
- the picker gains System / Natural, with the provider's voice list.
- lib/podcast-audio plays one clip per section through a hidden <audio>
element, measures the clip's real duration and gives each word a slice of it
proportional to its spoken length - the highlight follows the audio exactly,
and pause/seek/chapters become real audio operations. Clips cache per
session, the next section generates while the current plays, and a voice
change disposes the element and its blobs. Speed applies live (setRate)
without dropping the cache.
podcast-test stubs /api/tts with a real silent WAV and pins the path: the
voice list, the request body, exactly one playing element, and the highlight
riding currentTime. 43 checks. tests/lib.mjs launches Chromium with the
autoplay policy relaxed (a real device has the listener's gesture).
Version 0.7.6, SW v98.
Safari hands web pages one voice per language, whatever the iPhone has
installed, so the picker cannot fix that on iOS. It now carries the way out
instead: Accessibility -> Spoken Content -> Voices -> your language, download
an Enhanced/Premium voice and make it preferred. The speaker button wears an
amber dot while the note is waiting, and the voice-less banner appends the
same guidance. Detection covers iPadOS too (maxTouchPoints, since it reports
as a Mac).
The suite fakes an iPhone user agent, a single Daniel voice and pins the dot
and the note's actual steps (37 checks). Version 0.7.5, SW v97.
"The scrolling works kinda weirdly" + "maybe a less robotic voice, preferably
male".
Scroll: the transcript now follows by LINE, not by word - words on the same
line never move the page (that was the jitter), a line only scrolls once it
leaves a comfortable band, small corrections are instant rather than animated,
and a wheel/touch/drag on the transcript pauses the following for four seconds
instead of fighting the reader. Chapter jumps and word taps re-centre at once.
Voice: lib/podcast-voices ranks what the platform installed (natural / neural
/ online names up, espeak / SAM / compact down, a male name preferred) and the
player picks the best one automatically; a new picker in the chrome lists
everything in the episode's language, applies it live and remembers the choice
per language. A note points at the system settings when only robotic voices
exist. The engine takes the chosen voice instead of guessing.
podcast-test: 34 checks now (voice ranking, filtered picker, persisted pick,
same-line silence, line following, manual-scroll suspension). Version 0.7.4,
SW v96.
The third segment of the note-powered wizards (Study deck / Homework /
Podcast) takes the same intake - photos, documents, pasted text, the share
target - and turns it into a spoken episode.
- prompts: PODCAST_SYSTEM_PROMPT writes for the ear (plain prose, no markdown,
no visual references), chaptered into 3-8 sections around a length target.
- /api/podcast mirrors /api/generate's plumbing (rate limit, model gate,
provider keys, heartbeats, the 2x deadline, the optional vision pass).
- storage: frezmo.podcasts.v1, zod-validated, included in the backup and the
encrypted sync payload; a new library tab lists episodes with listen /
rename / delete-with-undo, and the wizard switch deep-links in.
- player: the transcript is word-wrapped spans and the word under the voice
rides a lyric-style highlight that keeps itself centred. Two clocks: exact
onboundary events where the browser reports them, a per-word length timer
where it does not (iOS Safari); the first boundary retires the timer.
Pause = cancel + re-speak from the word, which makes resume, chapter jumps
and tap-any-word-to-seek the same operation; the position saves on pause,
page-hide and unmount, so reopening resumes. Speed rebuilds the engine and
keeps playing; the Aero soundtrack pauses itself when the voice starts; a
browser with no voice gets an honest note and a readable transcript.
- the script sits on a glass pane (plain paragraphs over Aero's grass failed
contrast; the scan now covers the player too).
tests/podcast-test.mjs (17th suite) stubs the model and the speech engine and
drives the highlight by hand: 26 checks. All suites green, contrast clean.
Version 0.7.3, SW v95.
"Trying to use a different provider on the web app glitches out completely" -
an iPhone photo of the first-run gate with iOS's saved-password sheet laid
over it ("password for proxmox...") and the provider field stuck under the
keyboard.
- The key fields (gate, Settings, provider rows) were type=password with
autocomplete=off, which iOS/Safari and 1Password/LastPass read as a login.
They now carry autoComplete="new-password", data-form-type="other",
data-lpignore="true", autoCorrect/autoCapitalize=off and inputMode=text:
the mask stays, the password overlay goes. The custom base URL gets
inputMode=url and the same hints.
- The gate's keyboard handling scrolled to the Continue button instead of the
field being edited, so with the provider rows open the keyboard covered the
input and the locked wrapper could not be scrolled to it. The wrapper is
user-scrollable again, alignment anchors on the focused element (Continue
is the fallback), and the corner controls hide while the keyboard is up
rather than landing on the card.
providers-test drives a fake visualViewport (340px, installed before mount)
and pins the three layout contracts plus the hardening attributes. All 16
suites green, contrast clean. Version 0.7.2, SW v94.
The deploy gate (all 16 suites back to back) had the timed-run check read 23
where the suite demanded >= 25: under the fake clock the suite's own waits
advance the app's interval, so a loaded machine can tick a few extra seconds
between the card switch and the read. The assertion keeps its meaning - a
fresh budget, not the exhausted card's leftover (~0) - with a looser low end
and a comment explaining why. No app behaviour changed.
"It should play by default when the Frutiger theme loads in." The embed URL
now carries auto_play=true as a second try alongside the immediate play()
call, and when the browser still refuses (a reload carries no gesture), the
retry rides the next pointerdown/keydown anywhere OUTSIDE the player - the
first tap on anything starts the music. A 700ms throttle keeps a flurry of
taps from skipping through the set, taps on the player's own controls are
left to their handlers (no blip on the x), and an explicit pause or dismiss
is respected: no later gesture undoes it. Looking at the player (expanding
the circle) starts a still-pending soundtrack too.
aero-music-test simulates the refusal (__scAllowPlay) and pins all three
paths: one attempt anyway, the next tap starts it, an explicit pause stays.
28 checks. Version 0.7.1, SW v93.
Selecting Aero starts a random track from the Frutiger Aero playlist
(Bigblamboii's set on SoundCloud); every following track is drawn at random
too, so the set shuffles itself. The player is the credit bar from the mock:
a 44px glass circle with a note by default, expanding into pause + the
track's title + "SoundCloud - artist" (both reported by the widget).
- Only Aero ever touches SoundCloud: the embed and the widget API are created
while the theme is on and the listener has not turned it off; light/dark
request nothing (pinned by a suite check).
- The off choice persists (freezmo.aero-music.v1); re-selecting Aero in
Settings is fresh consent (applyTheme clears it).
- Autoplay rides the theme-selection gesture; after a reload the circle comes
up paused (browsers require a gesture) and one tap starts it.
- Volume 45; the iframe is 1px and inert, only the audio engine; the credit
area collapses back to the circle without stopping the music.
- Geometry: clears the phone tab bar, lifts above the chat composer and the
wizard bar, steps aside under immersive overlays and while the cookie
notice shares the corner. 44px targets on phones.
- Cookie policy gained the third-party-embed paragraph it was owed.
tests/aero-music-test.mjs (16th suite, stubbed widget API, 21 checks). All
suites green, contrast scan clean. Version 0.7.0, SW v92.
Follow-up to the smart-review breakdown: "I only have 2 decks" turned out to
be three deck cards (one clipped at the window edge) and 15 + 13 + 6 = 34
due. The line now carries each deck's activity total next to its due count -
"Due by deck: Sky deck - 15 of 15 - Humanismus - 13 of 15 - ... - 6 of 13" -
so the arithmetic is checkable inline without opening a deck.
groupByDeck returns {label, count, total}; new i18n key review.dueOfTotal
(en + cs). review-test pins the format. Version 0.6.9, SW v91.
"It contains mistakes I either did more than once or from decks that I've
removed - there's no way there's 34 to repeat with 2 decks." The counts were
honest (dueItems/mistakeItems only resolve live deck+activity pairs; a seeded
repro with 37 SRS rows and 30 live shows "30 due"), but nothing in the UI said
where an item came from: a deck created twice from the same notes shows every
question twice, and a deck that survived a delete or a sync looks like a ghost.
Both read as a broken list.
- SessionFeed gains showDeck (review + mistakes): a deck chip above every
card, so two identical questions are visibly two decks.
- The Smart review card breaks its numbers down under the buttons:
"Due by deck: Sky deck - 17 - Sky deck (2) - 17", capped at three decks
plus "+n more" (groupByDeck in lib/learning).
- New i18n keys (en + cs, plural forms for "+n more").
review-test pins it: a twin-deck seed renders the breakdown and both deck
chips in the feed. Version 0.6.8, SW v90.
"Terribly high CPU usage, the MacBook gets really hot" - a fanless MacBook
Air on a 144Hz screen. Measured on that machine with headed engines
(process CPU plus the per-process GPU time macOS keeps): an idle Aero
library cost LibreWolf ~110% CPU and ~230ms/s of GPU, the study feed ~130%,
Chromium up to ~57% CPU and ~350ms/s of GPU, Safari ~220-250ms/s of
WindowServer GPU. Priced one by one:
- the gradient mask on .ice-floats re-rendered the animated layer offscreen
on every frame (~75 points of Gecko CPU). The floats now rise through a
plain rectangular band clear of the header, the bottom chrome and the
rail, and fade in and out inside it, so the clip never cuts a visible one;
- the overlay grain forced a backdrop readback per frame for a texture that
measured a mean <=0.11/255 difference: removed;
- the invisible blob drift made every frame a full-screen redraw: the blobs
are painted once;
- static glass (answer tiles, feedback panel, chips, inputs, dropzones,
lists, segmented controls, Aero hint chips) re-blurred the moving scene
every frame: only floating glass blurs now, and Aero's baked chat chrome
drops a live blur that never showed under the bake;
- the scene also parks while the window has no focus and under any dialog
or sheet scrim (.scene-pause); the loading icon is a compositor-run CSS
animation instead of a framer-motion main-thread loop.
After, focused and animating: LibreWolf library 109% -> 33% CPU, study
134% -> 34%, chat 130% -> 42%; Chromium GPU 2.5-4.4x lower with 7-22 points
less CPU; Safari's WindowServer GPU -20-26% in Aero, about half in dark.
Unfocused or under a sheet: nothing. bubbles-test sweeps every float through
41 phases against the band and pins the parking and the cost contract;
tests/heat-probe.mjs (macOS) repeats the measurement; the suites find
Playwright's Chromium on macOS.
All 15 suites pass, lint clean, the contrast scan lost a finding. Version
0.6.7, SW v89.
Co-authored-by: Cursor <cursoragent@cursor.com>
Scrolling through a deck without answering recorded a 0% attempt and listed
every card under "Missed items" with its answer revealed. Only answered cards
count now: the score is correct/answered, the attempt stores the answered
count, an empty run records nothing, and the cards scrolled past wait under
"Not answered (n)" - listed as prompts with "not counted as mistakes" - and
"Answer the rest (n)" starts a catch-up run of just those cards (same restart
path as retry-missed, so history and the SRS still see the answers). "Aced
all" requires nothing left waiting too.
phase2 pins the flow: all-skipped records no attempt and no mistake list;
one-of-two reads "You got 1 of 1 right" + "Not answered (1)" with attempts:1,
bestScore:100; the catch-up run records the remaining answer. Exam mode keeps
its own rule: in a timed exam an unanswered question is a miss.
Version 0.6.5, SW v87.
The uniform wash (background-glass) made the rail and the chat chrome nearly
opaque - "why is the sidebar so white and non-translucent... its supposed to
be translucent and blurry!". The rail, the chat header, the welcome card and
the composer now use the same recipe as every other surface in the app
(background-blur under the shared gloss + veil + --glass-bg, exactly what the
deck cards do), and the dedicated wash asset is deleted. Answer bubbles keep a
thin 0.3 veil over the same blurred wallpaper.
Measured: the rail runs 181,229,241 over the sky to 133,174,75 over the grass
(translucent, d166) and the chat/library screens still scan at 0 below-minimum
text-contrast findings in all three themes. The chat suite now pins both the
shared recipe and the translucency, so a future "make it uniform" pass cannot
quietly reintroduce an opaque panel. Phone touch keeps the flat 32% glass.
Version 0.6.3, SW v85.
The phone (touch) chat chrome carried a 55% flat white veil while every other
Aero surface uses the thin 32% Ask-pill tint, so the header, welcome card and
composer read as plain white on a bright sky. They now use the same flat glass
as the cards, the answer/user bubbles keep 42% for long reading, and the
composer's field is a lighter 42% tile. The shared Aero wash block moved above
the touch media query so its background-image can no longer resurrect the bake
on phones.
The timed run worked but was invisible: the only cue was a small chip among the
streak/XP counters. The chrome pill now draws the countdown along its lower
edge, the chip is bordered (a control, not a stat), turns rose with the bar in
the last ten seconds, and fires one soft haptic tick at ten seconds.
Contrast scan clean on the phone chat in all three themes; activities suite
pins the countdown bar (25 checks). Version 0.6.2, SW v84.
From your notes: the generator now stores a verbatim source quote with every
activity (schema + prompt + a server-side clamp so a chatty model cannot drop a
card), and the feedback panel shows the sentence that carries it with the quote
highlighted green. Decks made before the change, and quotes the model trimmed,
fall back to a local IDF-weighted match over the saved notes text
(lib/notes-source.ts): nothing is shown unless at least two distinct terms
land. The passage hides in exam mode with the other reveals.
Read aloud: a Web Speech API speaker control on the text a learner reads
(flashcard faces, questions, statements, sentences, instructions) via
lib/speech.ts and components/SpeakButton.tsx — one utterance app-wide, a voice
matched to the deck's language, and no button at all where the browser has no
speech synthesis. It used to sit inside the flashcard's flip button (nested
buttons, tapping it flipped the card); it now lives next to the hint row.
Timed runs: the feed chrome carries a timer chip (off by default) that gives
each card 30 seconds; a card that runs out is recorded as a miss through the
same path as a wrong answer (history + SRS agree), shows a toast, and the feed
moves straight on with a fresh budget. handleAnswered ignores a second result
per card, so a timeout plus a manual answer cannot record twice.
The generator's type spec now names what each type suits (language decks lean
on flashcards/matching/fill blank, ordering needs a real sequence, short answer
reasoning), so the deck-aware guidance covers all seven types.
New activities-test suite (24 checks: quote path, local path, the green mark,
the speech stub, the timed run on Playwright's fake clock), registered in
run-all; phase2 pins that the notes passage stays hidden in exam mode; en+cs
keys for every new string; no new dependencies. Version 0.6.1, SW v83.
The header and composer were full-bleed bands (flush to the top/right/bottom
viewport edges, composer corners off-screen) next to a rail that floats with a
12px margin. The new --layout-gutter (12px) spaces the rail, header, empty-state
card and composer identically on every side; --radius-panel/-control/-chip
(-pill) replace the ad-hoc radii across the shared classes, so every panel is
rounded on all four corners and shared controls match in every theme. The
desktop-only ::after hairlines and the left-corner-only radius are gone.
Aero: the rail and the chat chrome (answers included) share one even recipe -
the wallpaper extra-blurred into a hue-free wash (background-glass.webp) under a
uniform ~50% white veil and a whisper of gloss. Measured within ~10 RGB units
across the rail top->bottom and the card/composer left->right; the blue->green
smear and the teal composer are gone. The composer's field is a clean even tile
with no wallpaper gradient inside it.
Chat fixes: send is a filled primary tile shaped exactly like attach (dimmed
while the input is empty), suggestion/context chips use the shared .chip token
(40/44px target, visible rim, full-contrast text, raw model id in the tooltip),
the AI note reads at 12px --text-muted (.ai-disclaimer), the selected sidebar
row carries the accent border + accent icon, the theme toggle is back in Aero
(the moon flips to Dark; Settings brings Aero back), and the empty state is
centered exactly between header and composer.
Outside CSS: the wallpaper's top-left HUD-circle decoration (it peeked through
the 12px gutters on wide windows) was healed out of background.webp and
background-blur.webp; the orbs/bubbles ride one masked .ice-floats layer that
fades them under the chrome bands and the rail strip, so none is sliced by a
panel edge (still parked under reduced motion).
Version 0.6.0, SW -> freezmo-v82. Tests: chat suite pins the new geometry (12px
gutter, 24px corners, even wash on all four panels, send-vs-attach shape,
chips/input radii); mobile pins the phone chrome and the Aero toggle; bubbles
pins the floats layer + its reduced-motion opt-out.
The pushed copy pass renames the library heading to "Your freezer"
("Tvůj mrazák" in Czech) and reworks the gate subtitle plus several
Czech strings. Five suites pinned the old heading — regression, i18n,
native, wizard and providers — and now assert the new one; the README
prose and project tree follow. Key parity is unchanged.
Version 0.5.9, SW -> freezmo-v81.
1. Bugs fixed:
- Backup schema now accepts theme: "aero" so backups exported under
Frutiger Aero pass parseBackup on import, and applyImport in "replace"
mode replaces SRS schedules, answer history, and chats instead of
merging into existing SRS or skipping empty arrays.
- Typed FillBlank cards bind the input value to chosen ?? value instead
of chosen ?? "", and normalise() preserves Unicode letters (\p{L}\p{N})
when grading non-ASCII answers.
- OpenAI-compatible vision calls (contentFor in lib/openai-compat.ts)
prefix data:image/jpeg;base64, when routes pass stripped base64.
- Streaming calls (ollamaChatStream, providerChatStream) keep their
AbortController active across the entire body stream and API routes
forward request.signal so stopping a chat or cancelling a run aborts
the upstream request immediately.
- HomeworkCard uses homeworkModeLabelKey so "check" records display the
Checked label instead of Step-by-step guide.
- Desktop dropzones on /create and /homework accept dropped documents as
well as images.
- LibraryScreen filters reported questions out of its due and mistake
counts to match ReviewClient.
- Clearing a provider key resets generation if it pointed at that
provider; chat conversations persist and restore provider; ModelPicker
keys items by provider:name.
- RouteTransition no longer mutates module-level lastPath during render.
- Service worker asset handler falls back to Response.error() when an
uncached asset fetch fails offline.
- Confetti and .ice-orb / .ice-bubble respect the in-app reduce-motion
toggle.
2. Performance & cleanup:
- Deduplicate window storage listeners across all six localStorage
stores (one listener while subscribed instead of N^2 notifications).
- Use todayAnswerCount in LibraryScreen and pass sorted folders to
DeckCard.
- Memoize MessageRow with stable callbacks in ChatScreen and compute
last message IDs in a single pass.
- Close ImageBitmap after drawing in lib/image.ts and destroy pdf.js
documents in finally in lib/documents.ts.
- Avoid re-parsing serialized sync payloads in POST /api/sync.
- Deduplicate copyText (lib/share.ts -> lib/clipboard.ts), chatMessageId,
describeExtraction, ALL_FOLDERS, useIsPhone, and remove unused helpers
and boilerplate public/*.svg files.
Version 0.5.8, SW -> freezmo-v80.
Docking the two chat bands into the rail card removed the floating cut
but the edge stayed a straight vertical line - still a seam ("this
still looks bad"). Both bands now keep the rail's own 12px gutter and
end in the rail card's 1.5rem radius on their left side, so the ending
reads as deliberate; the header divider spans the flat part of the
bottom edge, from the corner curve inward. Phones keep the plain
full-width borders.
The chat suite now pins the geometry: 12px gutter at the seam, 24px
left corners on both bands, zeroed band borders, inset ::after
hairlines. Chat 104/104, mobile 126/126, regression 50/50.
Version 0.5.7, SW -> freezmo-v79.
app/not-found.tsx renders a standalone screen: the root layout's ice
scene is the backdrop, the glass card/rim/bubbles come from the shared
recipes (full Aero with that theme on, frosted glass in light/dark),
and the mascot is the scene's glossy ice orb at full strength with a
snowflake and a "404" chip. Two ways out - back to the library and
into the chat - both 44px targets, en + cs.
The regression suite pins the 404 status, both links and the Czech
copy; the pixel-contrast scan covers the screen in all three themes.
The service worker no longer caches non-OK navigations as the offline
shell (a typo'd URL used to write the error page into the shell
cache). Version 0.5.6, SW -> freezmo-v78.
The chat header and composer are full-bleed bands inside the main
pane, which starts right at the sidebar on desktop - so both ended in
a hard vertical edge beside the floating rail card (the 12px gutter
showed the raw scene and the bands' fill + backdrop blur read as a
color step at the seam), with their hairlines starting in mid-air at
that edge. Both bands now reach the card's right edge (the rail's own
12px padding, expanded or collapsed) and their hairlines are inset to
the content box; phones keep the plain full-width borders. The chat
suite pins the dock. The "Chat is set to ignore notes text" banner
went away - the scope chip already says what the chat may read.
Version 0.5.5, SW -> freezmo-v77. Chat 104/104 (2 new), mobile
126/126, wizard 95/95, phase3 31/31, bubbles 34/34, i18n 13/13,
native 33/33, regression 44/44, contrast 0 findings, axe clean.
Full e2e: 14/14 suites.
Buttons: the recipe's thin veils let btn-secondary/btn-ghost/glass-pill
blend into the grass and sky - they now carry a stronger pale-glass veil
(the Ask-pill look) over the blurred scene, ghosts plus a rim and
full-contrast text. Homework make/check results gained the same red "!"
report control as the feeds and the guide (a synthetic short-answer
activity maps into the report flow; pinned in phase3).
Reduced motion: the manual toggle only fed framer's MotionConfig, which
strips transform animations - while components branched on the OS-only
useReducedMotion - so translucent sheets still spring-faded their opacity
in over the page (the "ugly visual glitch with the settings panel"). All
16 components now read useAppReducedMotion() (OS or toggle), and the
settings sheet/scrim and the library's phone sheet go instant. Verified
frame by frame: opacity 1.00 from the first frame.
Contrast 0 findings, axe clean; mobile 126/126, chat 102/102, providers
20/20, wizard 95/95, phase3 31/31, bubbles 34/34, i18n 13/13, native
33/33. Version 0.5.4, SW -> freezmo-v76.
The gate (user-key installs) now headlines Ollama with a "Free" badge and
adds an "or - Use a paid provider instead" option that opens the same
provider rows as Settings, framed as billed by the provider directly.
Saving a provider key there lists its models and switches study
generation to it automatically, so the gate steps aside with no Ollama
key at all.
Gemini joins the registry via its OpenAI-compatible endpoint
(generativelanguage.googleapis.com/v1beta/openai); the privacy policy
names it (en+cs).
Providers suite grows to 20 checks: the gate scenario runs against a
second user-key-mode app instance (free badge, paid option, provider
rows, pass-through, generation switch). All other gates green:
contrast 0 findings, mobile 126/126, chat 102/102, wizard 95/95,
bubbles 34/34, regression 44/44, i18n 13/13, native 33/33.
Version 0.4.9, SW -> freezmo-v73.
Settings -> Providers takes keys for OpenAI, Anthropic (their official
OpenAI-compatibility layer), DeepSeek, Groq, OpenRouter and any custom
OpenAI-compatible endpoint. One adapter (lib/openai-compat) covers
listing, streaming chat, JSON generation and vision, and every model
flow can run on a provider: the chat picker lists each enabled
provider's models (labelled, remembered), and a "Study generation"
choice moves decks, homework, grading and attachment transcription onto
a provider model.
Keys stay in this browser exactly like the Ollama one
(freezmo.providers.v1): never stored server-side, never in a backup or
the sync payload, forwarded only with the request that needs them. The
provider test asserts the key reaches the provider while the routing
header stays on the app side.
New suite tests/providers-test.mjs (13 checks) stubs both protocols:
model listing, chat streaming, study generation and the Settings flow.
Privacy policy updated (en+cs). Contrast 0 findings, axe clean, mobile
126/126, chat 102/102, wizard 95/95, bubbles 34/34, regression 44/44,
i18n 13/13, native 33/33. Version 0.4.7, SW -> freezmo-v72.
- Aero: blue gradients back on the active segment and every primary
(Review due, the card's Study button, the composer's send button); a
frosted backing behind the page heading and the All decks label; a
bottom-weighted rail tint; chat header/composer/welcome panels; calmer
orbs; stronger card veil + a stronger blur bake.
- Dark/Light: a dedicated --track token makes progress bars read (the
first attempt shifted --surface-sunken and the pixel scan caught the
sheet-contrast regression); brighter dark secondary text; darker light
segmented labels and difficulty tints; stronger light card border and
shadow.
- Chat UI: friendly model chip ("Gemma 4 31B") with the raw name in a
tooltip, larger chips, plain centered disclaimer, vertically centered
welcome panel, distinct chat-plus icon for new chat.
- Chat models: every chat-capable model from the server's /api/tags is
selectable (favourites first, search field when long, live-list
validation in the route); default CHAT_MODEL_DEFAULT or gemma4:31b.
- Reduce motion: Settings toggle (default off) that parks the scene,
collapses CSS animations and drives framer-motion's calm mode.
- Folders hint, "New" badge on fresh decks.
Contrast: 0 findings, axe clean, mobile 126/126, chat 102/102, wizard
95/95, regression 44/44, i18n 13/13, native 33/33.
Version 0.4.6, SW -> freezmo-v71.
A user's run failed with the wizard's generic "check your Ollama
settings" card and could not be replicated with the same key - because
the key and model were never the problem: a raw error (a mid-run socket
drop, a wifi/LTE switch, the app backgrounded by the phone OS) leaked
out of the NDJSON reader as a DOMException, and the wizard can only
show the generic card for unknown errors.
The reader now classifies everything: abort -> "cancelled", stall ->
"the model stopped responding", anything else on the wire -> "the
connection dropped mid-run, try again". Pinned with a fetch stub that
kills the stream mid-run in the wizard suite (95 checks).
Contrast 0 findings, mobile 124/124, chat 101/101, regression 44/44,
i18n 13/13, native 33/33. Version 0.4.1, SW -> freezmo-v70.
A user got stuck on the wizard's loading screen: a hung model meant
silence, and silence was indistinguishable from work - the server's own
hard deadline is 2x OLLAMA_TIMEOUT_MS (ten minutes by default), and the
client waited on a dead stream indefinitely.
Both generate routes (deck, homework) now send an NDJSON heartbeat
every 15s, and the client guards the two quiet phases: no response at
all for 45s, and no bytes for 90s mid-stream, both surfacing the
regular error state with a Retry. The caller's abort still maps to the
friendly "cancelled" state.
Verified with a stalled route and Playwright's fake clock in the wizard
suite (93 checks), plus contrast 0 findings, mobile 124/124, chat
101/101, regression 44/44, i18n 13/13, native 33/33.
Version 0.4.0, SW -> freezmo-v69.
The touch fallback from 0.3.8 (a real backdrop-filter over the live
scene) was the wrong trade: every pane re-blurs on every frame while
the bubbles animate, and phones choke - input lag, "buttons take ages
to respond". Touch devices now drop both the bake (which mispaints
there) and the live blur: the panes keep the thin Ask-pill tint over
the sharp scene, and the hand-rolled activity fills keep their own
tints - transparent glass at zero per-frame cost. Desktop keeps the
baked blur.
tests/mobile-test.mjs pins the flat contract (no baked layer, no
backdrop-filter on a card) with a touch context: 124/124.
Contrast 0 findings, axe clean, chat 101/101, regression 44/44, i18n
13/13, native 33/33. Version 0.3.9, SW -> freezmo-v68.
Mobile engines repaint background-attachment: fixed late (iOS treats it
as scroll), so the baked glass showed a stale crop of the wallpaper
until a tap forced a repaint - "some of the UI looks bugged until I
click it". On touch devices (max-width 639px or pointer: coarse) the
bake is dropped for a real backdrop-filter over the live scene, so the
bubbles blur through the glass as they pass and the position is correct
by construction. Hand-rolled frost fills keep their own tints there;
desktop keeps the baked layer (measured cheaper and verified correct).
Pinned with a touch-context check in the mobile suite (124 checks).
Contrast 0 findings, axe clean, chat 101/101, regression 44/44, i18n
13/13, native 33/33. Version 0.3.8, SW -> freezmo-v67.
Bubbles: ten -> eighteen (interleaved, 18px to 96px, 10-23s runs), with a
tighter specular hotspot, a reflected-light glint low on the sphere, a 1px
glass edge and a brighter inner rim on both the bubbles and the drifting
orbs. Bodies stay translucent with capped opacity, so wallpaper text stays
readable; bubble-perf shows no added frame cost.
Chat: the conversation scrolls up to the glass header, and a line clipped
exactly at that edge read as a rendering bug over the busy wallpaper - the
scroller now carries a 14px top mask in Aero, shorter than the column's
16px top padding so a fresh conversation shows no dimming.
Contrast 0 findings, axe clean, chat 101/101, mobile 123/123 (bubble count
pinned to 18), regression 44/44, i18n 13/13, native 33/33.
Version 0.3.7, SW -> freezmo-v66.
The white surfaces were reading as a milky white panel (a card top was
~71% white). The frosted recipes now use a 14% base veil plus the gloss
over the pre-blurred wallpaper, small chrome (rows, chips, segments,
buttons) keeps a touch more veil, and nothing nested inside a sheet or
dialog uses the viewport-aligned layer (it showed the raw dark water
under the settings theme labels at 1.5:1 — nested chrome falls back to
the flat fills over the genuinely blurred sheet).
The first-load text bug was the boot script: its dark test was
`s && !a ? s === 'dark' : matchMedia(...)`, which for a stored aero fell
through to the OS preference, so a dark-preferring machine loaded with
both `dark` and `aero` classes and every `.dark` rule painted over the
wallpaper until a theme toggle re-applied the classes. The stored value
now decides alone; a dark-colorScheme check pins it in the mobile suite
(123 checks).
Background: the user's latest Gemini scene (1280x714) re-encoded to
public/aero/background.webp (206 kB) with a fresh pre-blurred sibling.
Contrast: 0 below-minimum findings, 0 axe violations, all three themes.
Version 0.3.1, SW -> freezmo-v62.
The scene is the user's own Frutiger Aero image now: green hills, a river,
an eco-city, bubbles and daisies. Re-encoded to public/aero/background.webp
(152 kB) with a fresh 3 kB pre-blurred variant for the glass, and the
WallpaperSafari credit line, its i18n keys and the README credit are gone
(the image is made by the user with Gemini).
Contrast: 0 below-minimum findings and 0 axe violations, all three themes.
Version 0.2.8, SW -> freezmo-v60.
UI bug from a phone screenshot (short answer, Czech):
- The textarea was flex-shrunk inside the slide's column (58px instead of
~82px), clipping the typed answer mid-line. Every direct child of a
study/review/guide slide is now shrink-0 and the slide scrolls instead.
- Two AI disclaimers were visible at once: the activity's own after the
grading box plus the feed's row. The activity-level copy is gone; the
feed/guide row (which also carries the report control) is the single one.
The phase-2 check now asserts the note sits below the graded feedback.
- The mobile suite's backdrop-tap check is deterministic too (the spring
animation made the measured box mid-flight).
Ordering activities: the deck prompt now allows them only when the notes
hold a real sequence (chronological events, process stages, procedure steps,
a lifecycle, a ranked scale) and points at matching / multiple choice /
true-false when they do not - "sort these" was landing on decks with no
order at all.
Version 0.2.4, SW -> freezmo-v57.
The sheet's spring animation meant the box measured right after the dialog
appeared could be mid-flight; tapping 24px above it sometimes landed inside
the sheet once it settled (the deploy gate caught it twice — production was
never touched). Tap a fixed point near the top instead: the sheet is
max-h-[90dvh], so its top edge cannot reach y=12.
Three consecutive mobile runs: 119/119 each.
The wallpaper is the one from the mockup: "The S Aesthetic That You
Definitely Recognize Frutiger Aero" via WallpaperSafari
(https://wallpapersafari.com/w/IlWuz7) — the URL in the request pointed at a
different (teal abstract) wallpaper, this is the grass/sky/water one that
matches the screenshot. Re-encoded to a 237 kB WebP in public/aero/, used as
the scene in aero only (light/dark never download it), and credited in
Settings next to the theme picker.
The drawn CSS scene (rays, reeds, skyline, caustics) is off over the photo;
the ten rising bubbles stay, and the orbs became extra bubbles. Buttons and
segments are liquid glass now — translucent, glossed top half, bright rim,
backdrop blur — instead of the solid aqua fill. Helper lines that sit
straight on the wallpaper (the keyboard tip, the AI note) get a small glass
chip: dark text on grass measured 2.5:1.
Also carries: the underwater scene + bubbles work, the floating cookie
notice, the Czech title (Tvůj mrazák) and the review-suite fix that made the
last deploy fail its gate.
Contrast: 0 below-minimum findings, 0 axe violations across all three
themes. Version 0.2.2, SW -> freezmo-v56.
Frutiger Aero, closer to the mock:
- The scene is water now: sun shafts through a bright surface, caustic
patches, reeds on the left, a city silhouette on the right, and ten soap
bubbles that actually rise (own keyframes, varied sizes/speeds/sway; the
spans are display:none outside aero so they cost nothing elsewhere, and
they sit still under reduced motion).
- Pale blue glass at 0.8: still reads as glass, but dark navy text clears
4.5:1 over the water. The phone tab bar mixes the sheet glass instead of
the page — in aero the page *is* the water, which left its labels at 3.9:1.
- The aero page-level text (muted/subtle) is much darker for the same reason.
- The PWA status bar follows the new water blue (#2fa7e6).
The cookie notice is floating glass now (blurred backdrop, rim, shadow): as a
plain card it read as part of the page background instead of a notice.
Czech: "Tvoje mraznička" -> "Tvůj mrazák" (library title), test updated.
Contrast: 0 below-minimum findings and 0 axe violations across every screen,
all three themes. Version 0.2.1, SW -> freezmo-v55.
Frutiger Aero (Settings -> Theme): a third theme, not a dark/light variant.
A sky gradient with a strip of grass, the glacier blobs become clouds, the
orbs soap bubbles, glass gets the classic horizon gloss and the buttons a
glossy aqua. The light/dark toggles (header, rail, chat header) hide while it
is on and the picker says why; the pre-paint script applies the class before
first paint and the PWA status bar follows (#3fa3e0). Its muted/frost text is
darker than Light's because those labels sit directly on the sky.
Contrast, measured on real pixels (tests/contrast.mjs, now covering aero):
- The light settings sheet was translucent over a dark scrim, so its muted
text measured ~2.8:1. Sheets and dialogs now use --glass-sheet-bg (a more
opaque surface), and Light's --text-muted is a touch darker.
- Result: 0 below-minimum findings and 0 axe violations across every screen,
all three themes, phone and desktop.
Also: the suites now seed the cookie notice as already dismissed (they run as
a returning user); the regression suite opts out to test the notice itself —
the previous deploy failed its gate because the notice blocked a chat chip
click. The notice also hides during immersive sessions (overlay-open), so it
can never cover a study feed.
Version 0.1.8, SW -> freezmo-v54.
Three plain-language documents in both app languages (lib/legal.ts, rendered
by components/LegalPage.tsx): /privacy describes the local-first storage, the
key never leaving the browser and the three opt-in server copies (sync,
reminders, reports); /terms covers the GPL-3.0-only software, your own Ollama
key, fair use and liability; /cookies states what is true — no cookies at all
(no Set-Cookie, no document.cookie), strictly-necessary local storage instead,
no tracking.
A cookie notice tells every new browser the same thing once and remembers the
dismissal (freezmo.consent.v1). It sits at z-44: below the immersive study
feeds so it can never cover a session, above the phone tab bar. Settings links
all three documents. The regression suite checks the pages render with real
sections and that dismissing the notice sticks.
Version 0.1.7, SW -> freezmo-v53.
- axe is clean across every screen in both themes, phone and desktop. Fixed:
the hidden file inputs (aria-hidden + tabIndex -1 — they are opened by
labelled buttons, so they should not be in the accessibility tree), the
heading order (deck/homework/smart-review cards are h2 under the page h1;
LearningSection takes a level so it stays h2 on the library and h3 inside
the settings dialog) and landmarks around the study and review feeds.
- tests/contrast.mjs (new diagnostic): samples rendered pixels, because axe
cannot judge text on blurred glass — it reports those nodes as incomplete.
It skips disabled controls (1.4.3 exempts them) and content behind an open
sheet. Result: no below-minimum text on any scanned screen.
- The report control exists in the homework guide too: under each "Try it
yourself" exercise, labelled "Report exercise", without a skip (there is no
queue to skip). lib/question-reports now takes deckId/deckTitle directly.
- The rail's two creation actions are buttons now (.nav-action) rather than
quiet nav rows — once the library is full they were the only way to start
something on desktop and read as just more navigation. "Make my homework"
is "New homework" wherever it is an action (rail, empty state, share
chooser, manifest shortcut); the page title and mode switch keep their
wording.
Version 0.1.6 (three task bumps), SW -> freezmo-v52.
Audit pass with the threats that actually apply to a self-hosted, account-less
app: getting the Ollama key throttled or banned, and someone grinding the app
down. Per-route windows already existed; these are the missing layers.
- lib/model-guard.ts: every route that spends a model call now takes a slot —
60 calls per window across all clients (RATE_LIMIT_MODEL_TOTAL), 3 in flight
per IP, 6 in total — and the slot is released when the call finishes, on
errors and on client disconnects. Over the concurrency cap answers 429
kind=busy, not a punishment. Slots are taken after every cheap rejection so
a malformed request cannot hold one (found by the new suite: early 400s
were leaking slots).
- Images must be base64 data URLs (dataUrlSchema in lib/schemas): a plain URL
in an images field made Ollama fetch an arbitrary address on our behalf.
- Push endpoints must be a real push service (PUSH_ALLOWED_HOSTS, defaulting
to FCM/Mozilla/Apple/WNS): the server POSTs to whatever is stored.
- The rate-limit identity is the LAST x-forwarded-for hop. Reading the first
let a client spoof its own IP and dodge every per-IP limit.
- Disk guards: reports.jsonl stops at REPORTS_MAX_BYTES (8 MB), reminder
devices cap at PUSH_MAX_SUBSCRIPTIONS (50).
- Security headers: frame-ancestors 'none' + X-Frame-Options DENY, nosniff, a
referrer policy and a Permissions-Policy without camera/mic/geo/payment.
New tests/security-test.mjs (13 checks, registered in run-all) proves each
guard against a slow stub Ollama with deliberately tiny limits. README:
security notes + rate-limit tables updated. Version 0.1.3, SW -> v51.
Versioning: package.json carries the app version (0.1.2 here; one bump per
task with `npm run version:bump`, patch 0-9 rolling into the minor at 10),
baked into the build via lib/version.ts, reported by /api/healthz and shown
in Settings next to the running commit. AGENTS.md documents the rule; the
native suite checks both the API and the Settings row.
Fix: on desktop the app header is hidden, so every page owns its top padding —
/create was the only one without `lg:pt-10`, which slammed the Study deck /
Make my homework switch against the top of the window while /homework had a
comfortable gap. Both now sit at 40px; the wizard suite asserts the clearance
and that the two pages stay in step (90 checks).
SW -> freezmo-v50.
The MacBook ran hot browsing the app: the ice scene animates constantly, so
every backdrop-filter surface re-ran its blur on every frame. Measured on
desktop (2x CPU throttle, software renderer): p50 frame 66.6ms baseline,
33.3ms with the static surfaces not filtering, 16.7ms with the scene's
animation also off; frames over 34ms: 128 -> 14.
The glass recipe is split: floating layers (sheets, dropdowns, dialogs, the
sticky header/chrome/tab bar) keep the blur (backdrop-blur-lg, what phones
already used); static layers (cards, the sidebar rail, the gate, the study
prompt card) keep the identical tint, rim, highlight and shadow without the
filter. The study prompt card moves from glass-panel to card + glass-solid,
which is the same surface recipe.
Quality: a deterministic pixel diff (scene motion frozen, both themes,
desktop + phone) shows a ~1% colour shift in the card backdrops — invisible.
The mobile axe contrast audit and the bubbles suite stay green.
tests/bubble-perf.mjs now also reports the glass cost. README: decision 61.
SW -> freezmo-v49.
The full-screen loading screen shown the instant a Link starts navigating
was rendered inside the Smart Review card. That card has backdrop-filter,
which makes it a containing block for position: fixed — so the overlay was
confined to a 44px strip of the card and the loading block spilled below the
fold instead of covering the viewport. It is now portalled to <body> with a
soft page scrim behind it; the review suite measures that it covers the
viewport (390x844).
SW -> freezmo-v48.
Sync was push-on-demand only: a deck frozen on the laptop stayed in dirty
until someone opened Settings and pressed Sync now. The two stores whose
writes mean "the user just made something" now schedule one debounced
background syncNow("auto") through lib/sync-auto: debounced 2.5s so a burst
of writes is one push, one run at a time, silent on failure (the data stays
dirty and Settings still offers the manual sync), retried on the browser's
online event, and imported dynamically so the sync code stays out of the
first load with no import cycle.
markSyncDirty() now reports whether it actually marked, which keeps pull and
backup-import paths quiet while a real import pushes like any other local
change. Answers and chat keep the existing manual behaviour on purpose —
they change every few seconds while studying.
Tests: phase-1 (a deck write pushes itself, then Sync now is a no-op — 50
checks) and phase-3 (a check run pushes the new homework record — 30 checks).
README: decision 59. SW -> freezmo-v46.
- "Check my work" now takes photos and documents of the attempt: documents
are read in the browser into text, photos go to /api/homework as
solutionImages and are transcribed by the vision model before the check
("Reading your solution..."). The finished attempt comes back in
meta.solutionText and is stored as the record's attempt, so "Redo"
re-checks the same work without keeping the files.
- Deleting a deck now drops its due schedules and mistakes from the library's
smart review immediately (the counts resolve against the live deck list);
Undo brings them back, because the learning data itself is never erased.
- The phone override for the glacier blobs was still at 0.1/0.2: it is now a
quarter of the original value there too (0.025/0.05), matching the desktop.
- License: GPL-3.0-only (LICENSE, package.json, README section).
Tests: phase-3 photos-only API + desktop attachment checks (28/28), review
suite deck-delete checks (18/18), i18n/mobile/phase-2 green. README:
decisions 57-58. SW -> freezmo-v45.
