A month on the tile, and music without opening the player

1.4.8 is the opposite of the hotfix before it. The calendar tile gets a second face showing the whole month, there is a new music tile, a new 3×3 size, and a filter for which calendars feed any of it.

Both new faces were designed in full before a line of code: 26 decisions, each with the alternative that was rejected. The starting point was working code from a sibling project of mine, read, taken apart into decisions, and rewritten against this project's rules rather than copied. What came across and what deliberately did not is most of what follows.

A whole month on the calendar tile

It is a second face of the existing calendar tile, held in the same field that already holds the clock's dial and the weather's face — the third time that pattern has earned its place. Switch it with the badge in the corner in edit mode, exactly like the clock. A tile from before the update reads as the old face, so every existing calendar tile looks identical after upgrading and there is nothing to migrate.

A new tile type was the obvious alternative and it is worse. The picker already has 13 entries; a fourteenth called "calendar (month)" next to "calendar" makes the user guess what the difference is before either one is on screen. Switching face by size was the other option — small means teaser, large means month — and it breaks the promise the size ladder makes, which is that it is a menu of shapes. It would also deny a small month grid to someone who wants exactly that.

The geometry is always the same: seven columns, one row per week. Only the contents of a cell change, on a single threshold computed from the cell as actually measured — not from the tile's width in grid units, which says nothing about available dp across different screens.

Below the threshold: a dot. No day number, no weekday header.
Above it: the day number, a dot beneath it, and the weekday header.

Three signs that never change meaning: the plate says the day exists, the dot says something is on it, and a brighter plate says it is today. The dot is binary and never a count — at 2×2 there is no room for one, and at 8×4 a number squeezed under the day number is a mark nobody reads. The query is identical either way, so being binary buys one rendering rule rather than two.

Today used to carry an outline around its digit, and that was wrong: an outline draws around the number and so competes with it for the same few dp. On the phone it overlapped its own digits at ordinary font sizes and could only get worse as the system font grew. A fill sits behind the text and cannot collide with it at any size.

Why a threshold is needed at all: the desktop is eight columns, so a grid unit is about 40 dp on a 360 dp screen and a 2×2 tile is about 84 dp. Seven day columns leave about 12 dp per day, and a two-digit number needs about 14. That is not a font-size problem — the numbers simply do not fit, and an unreadable number is worse than a dot, because it pretends to say something.

Tapping a day opens the system calendar on that day, with the turnstile, because that is still opening an app, just with an argument. It degrades by itself: at 2×2 a day cell is around 12 dp, far below the touch minimum, so there the whole tile simply opens today. No extra rule required.

There is no month navigation. Horizontal is the page pager plus the turnstile's reserved meaning, vertical is desktop scrolling and the tile's own page rotation. Browsing months is the calendar app's job, and the tap already goes there.

The first day of the week comes from the region, not the language

The locale is taken from composition, not from the process default, because Android 13's per-app languages make those two different answers.

The trap underneath that: the week-fields lookup reads the first day of the week from the region, not the language. A locale carrying only a language answers Sunday for every language — Polish included. Verified on the JVM: pl gives Sunday, pl-PL gives Monday. And per-app languages can hand composition a bare pl, which would have started a Polish user's week on Sunday.

So the region is taken from the locale when it has one and borrowed from the system locale when it does not. Borrowing the phone's region is a better answer in its own right, not just a patch: the start of the week is a regional convention, so someone in Poland running the launcher in English still starts on Monday. The language stays with the composition locale, because that is what draws the day letters.

It draws the month before it knows what is on it

The grid, the numbers and today's brighter plate render immediately. The dots wait for the first real emission from the query.

That is not a nicety. An empty grid of dots at start-up is exactly the bug that already shipped once on this project — a "0" and "set a city" flashing at people with a full desktop. A month grid drawn without dots is indistinguishable from a month that is genuinely empty.

It is now a standing entry in the regression checklist rather than a one-off fix, because this class of bug is invisible to whoever writes it: with a warm process you never see it, the value is correct two frames later, and the only people who meet it are users cold-starting with real data.

Wide sizes stopped being a stretched grid

The first device pass was blunt: 4×4 and 2×2 good, 8×4 and 4×2 weak. There is one cause and it can be named. The month grid stops carrying more information as it grows wider. Cells get wider but not taller, so 8×4 spends 50 dp on a day to say precisely what a 24 dp cell said at 4×4. The rest is whitespace.

Three layouts now, chosen from the measured tile rather than from a list of known sizes:

shape	layout
width − height ≥ 120 dp	split: a square month on the left, an agenda on the right
that, and width ≥ 1.8 × height	agenda only
anything else	grid only

The first condition is phrased as "how much is left beside a square month", and that is not an accident: a month is square by nature, so the question is whether the remainder is enough for an agenda. At 8×4 that leaves 184 dp — a real list. At 4×2 it leaves 92 dp, too little for a time and a title, so that shape does not split.

The 120 dp threshold is computed, not eyeballed. The requirement was that width 7 must look like width 8, because a tile trimmed by one unit changing its layout is a distinction nobody asked for. Of the shapes that must split, 7×4 leaves the least: 123 dp on a 320 dp screen, 138 at 360, 158 at 412. One hundred and twenty fits every one of those, and width 6 stays out of the split on all ordinary screens at 82, 92 and 105 dp. The first version used 140, which is exactly why 7×4 fell back to a bare grid.

The left half of the split is literally the same month grid used at 4×4, not a second rendering of it. The agenda comes from the flow the dashboard already lives on, filtered with the same grouping the dashboard's calendar section uses, so the two surfaces cannot disagree about which day an event belongs to.

The agenda shows two days, not one. A wide tile usually has room to spare, and the honest thing to put there is the next thing that will matter, not whitespace. Tomorrow is added only when rows actually remain, so a packed day pushes it out entirely instead of stealing rows from today. Both days carry a dated heading, including in the split layout — a reversal of the first version, which dropped the heading because the grid beside it already highlights today. That argument stopped being true the moment a second day appeared: the headings are what tell the two lists apart.

Rows truncate rather than compress. A list longer than the space loses the far end of the horizon; it does not shrink the type. A calendar that fits one more row at the cost of every row's legibility has stopped answering "what now".

The largest system font broke both halves at once

Reported from the phone at the largest font step, with screen zoom untouched. Three symptoms, one cause: the thresholds were constants in dp, and text grows in sp.

The regression checklist already says it outright — nothing about the tiles scales with that setting, they are a fixed grid of millimetres, while every sp inside them does. The tile stays the same size; the letters in it do not.

Day numbers cut to one digit. The fixed comparison still said "print numbers" to a cell that had stopped fitting them.
Dots vanished. At double scale the number took the cell's whole height, so the dot under it was pushed out.
The agenda ran outside the plate, painting onto the tiles beneath.

Both thresholds now multiply by the font scale, floored at 1 — a smaller-than- default font does not buy the grid the right to draw in a tighter cell than the layout was designed for. At the largest font the whole month drops to dots, and that is the intended outcome rather than a capitulation: the checklist accepts text that is not there and rejects text that is cut.

Only the vertical budget scales. Whether the agenda fits beside the grid at all does not, deliberately: running out of width ends in an ellipsis, which is proper degradation, while running out of height ends in cutting.

Then I broke the weekday header, and that one was a design error

The fix worked — the month dropped to dots at the largest font, as intended — and took the M T W T F S S header with it, which nobody had asked for. Reported from the phone immediately.

The cause is not in that fix. It is in a rule I had written earlier: the header shares a threshold with the numbers, because a grid without dates has nothing to label. That sentence is true in exactly one case — a one-cell tile, where the column is 12 dp and a letter does not fit either. Everywhere else it glues two different questions together:

A day letter is one glyph on one line. It needs its own line, and it needs the weeks to still have something left.
A date is two digits with a dot beneath, inside a week row. It needs the row to be tall, not merely wide.

At double scale the second failed and the header went with it, despite sitting in 25 dp columns where the letter had fitted all along.

Two separate tests now. The header appears wherever the weeks can live with it, and gives way only on a one-cell tile, where taking a line would leave the month with nothing. The cell height is computed from what the weeks actually get, so the header still cannot starve the numbers it labels. At the largest font the result is a header plus dots. A week row's height deliberately does not scale with the font: a plate and a dot are not text.

Music, without opening the player

A new tile type. Album art as the plate, title, artist, skip and pause, and a tap on the body opens whatever the sound is coming from.

The session is picked live rather than pinned to the tile: the first one playing, or the first in the list when nothing is. The system already orders that list by activity, so "the first one playing" is the one being listened to. Pinning a player to a tile was the alternative, and it demands an answer to "what does a tile pinned to Spotify show while YouTube is playing" — a separate feature, not a variant of this one.

The reference implementation I read has a real bug here that this one does not: there, the controller is fetched afresh on every button press, so "next" can land on a different session than the one the tile is currently drawing. Here the controller is held, and the action goes to the session that is on screen.

Three states, not two

This distinction is easy to miss and it changes the design.

Playing — art, title, artist, transport with a pause icon.
Paused, session alive — the same, with a play icon. The metadata here is real, not stale. A media session usually survives a pause, the controller is still returned, and play resumes that exact player. The difference from state 1 is the icon and nothing else — no dimming, no greying. Dimming would suggest "out of date", and it is up to date; the only thing that changed is that the sound stopped, which is precisely what the icon says.
No session at all — the player was killed, or the phone restarted. There is no metadata and nothing to resume, so this is the empty face, without a + affordance: + promises that the action happens on this tile, and opening a player is not that. Tapping opens the device's default music app, which turns a dead tile into a useful one without making a promise.

Album art on the plate is a correctness requirement

The art fills the whole plate with a gradient scrim from the bottom; text and transport sit on the scrim. That is not decoration.

The function that picks a readable content colour takes one Color and judges its luminance. An image has no single luminance, so white text on light album art simply disappears. The scrim is a known colour — and it is the scrim we compute contrast against, not the picture.

The scrim carries the tile's colour, not black. The first version was flat black and read like a system media notification rather than like this tile. The colour is mixed into black rather than used at full chroma, and that is not taste: the scrim exists so that contrast can be computed, and darkening whatever colour the user picked guarantees the bottom of the gradient stays a known, legible ground even with a bright accent.

Two collisions had to be settled. The desktop draws one photo through the tile plates, so album art would be a second image in the same plate: while something is playing, the art wins — the tile stops being a window and becomes its own image, and returns to being a window when the session goes. Plate transparency applies to the art exactly as it applies to a colour.

The sizes go 2×2, 4×2, 4×4, 8×4. The floor is 2, not 4, and that took an argument. Three transport buttons need about 40 dp each, so about 120 dp, so three units — which suggests a minimum width of 4. But that minimum also bounds the size ladder, so it would exclude the 2×2 rung, where this tile is a designed variant (art plus one centred play button) rather than a layout failure. The definition in the code is "the narrowest width at which this tile still says something", and a play button on album art says something. The touch problem is solved by removing two buttons, not by squeezing in three. The artist drops at 4×2 — the only place text disappears on shrinking, because 84 dp of height is spent by three 40 dp buttons and one line of title.

The icon lost a second, and only in one direction

Pause switched instantly. Start took about a second. The asymmetry was the clue: the playing check tested for exactly the "playing" state, but playback begins by passing through buffering. The icon was not waiting for our event — it was waiting for the player's buffer to fill. A pause has no such state on the way.

Fixed in two layers: buffering and seeking stopped counting as paused, which also removed a latent bug where skipping to the next track flashed the icon back to "play"; and an optimistic switch, set before the command is sent, so the icon turns in the same frame as the touch. It only holds until the session disagrees, so the cost of being wrong is a pause pressed in the player right after the tap.

And then buffering had to get its own sign

Folding buffering into playing hid it under the pause icon. On a slow stream the tile then shows a pause, makes no sound and does not move — which is exactly what a hung tile looks like. Reported from the phone right after the previous fix.

So the boolean gave way to three values — paused, buffering, playing — and the work divides cleanly. Drawing looks at all three and puts a spinner in the button's place while buffering. Pressing the button looks at "engaged", because pressing during buffering means pause and that is what the user wants. Session selection looks at "engaged", because a session that is buffering is the one being listened to. And the art's drift looks at "engaged", because a tile freezing the moment a stream stalls would read as broken all over again.

Seeking goes to playing, not buffering, deliberately: a skip usually takes a frame or two, and a spinner flashing on every "next" would be worse than no spinner. A skip that genuinely has to wait passes through buffering anyway.

The optimistic switch changed meaning for the better along the way. Pressing play now shows a spinner rather than a pause icon — "heard you, working" instead of "playing", which was not yet true.

One deliberate exception: the spinner does not obey the animation-speed multiplier. Everywhere else zero means "finished", but that rule is about decoration — the turnstile, arrivals, the art's drift. This is feedback, and a progress indicator frozen mid-arc says "hung", which is the very thing the spinner exists to prevent.

The drift, and a bug that was already there

The album art drifts the same way the gallery tile's photo does, from one shared implementation. Those are the only two places where an image fills a plate, and a second, subtly different movement would read as a fault rather than as variety.

Two rules make it more than an ornament. The drift runs only while something is playing, which turns it into a state signal: a tile that moves is a tile that is playing, readable from across the room without squinting at an icon. And zero on the speed multiplier means finished — no transition is composed at all, the image stands.

Which fixed a bug that was already shipped. The gallery tile did not respect the speed multiplier — not one reference to it in that file; it stopped only in edit mode. The multiplier is documented as applying to every animation in the app, so anyone who had turned all animations off still had a drifting photo on their desktop. Copying that pattern would have spread the bug to a second tile instead of fixing it, so both moved onto the shared path at once.

And one that arrived with the change: the drift scales the image to 1.15 and the music face clipped nothing, so the art ran past its plate and lay over neighbouring tiles — visible immediately in a screenshot from the phone. The gallery tile has carried that clip since it got its own drift. The lesson is worth writing down because it is not a one-off: anything given this movement must be in a clipped container. Scaling beyond its own bounds is built into it, not a side effect.

Why a track title does not break the rule from 1.4.6

1.4.6 settled that the unread count on the flip's back face is metadata the tile already prints, and that a sender or a message snippet is excluded, because those change a row in Data Safety, the lock-screen decision and the privacy weight of the whole app. That post ended by saying a sender on that face would be a separate feature with a separate answer.

The music tile shows a title and an artist. This is that separate answer.

It does not break the rule, and it is worth being exact about why. What was ruled out is the content of a notification. A track title does not come from a notification — it comes from the media session's metadata, which the playing app publishes so that the system can display it. The same words already stand on the lock screen and in the shade, by that app's decision rather than ours. The badge data still carries not one character of notification text; listener access here is the key to the session API, not to content.

The counter-argument, which I am not hiding: the tile hangs on a desktop that anyone glancing at the phone sees, and a podcast episode title can carry exactly the class of information the earlier rule protects against. The difference is that the user adds this tile themselves and removes it themselves, whereas a badge arrives on every tile at once. Visibility here is a choice, not a side effect.

Operationally: no new row in Data Safety, because nothing leaves the device, and no new permission. Music goes through the media session manager, which needs notification-listener access — and this launcher already has it for badges. Someone who turns that access on for music will also get counts on their tiles, because badges have no separate switch. That is deliberate and it needed zero lines of code: it is how the permission already behaves.

Coming back from system settings has to revive the tile without a restart, and that needed its own mechanism. There is no result to await and no dialog to hang a callback on — the launcher simply returns to the foreground holding a permission it did not have when the flow started subscribing. So the subscription is rebuilt from scratch, gated on the access state having actually changed, so an ordinary return to the desktop does not re-register system listeners.

There is no marquee on the title. It ellipsises like every other piece of text on every other tile. A long title is not a special case deserving its own rule.

Choose which calendars feed the tile

Until now every calendar was pulled without asking. With one noisy work calendar a dot appears under every day of the month and the grid stops saying anything. The old teaser face never had that problem because it showed one event at a time; the month grid exposes it.

The interesting part is what did not turn out to be needed. The project's rule is that a new non-nullable field in the backup format needs a version gate, because the JSON decoder will materialise it as false/0 in an old file rather than as the default. The real test is sharper: does what the decoder puts in place of the missing field differ from the truth about the launcher that wrote that file? Here it does not. A missing field gives an empty set, an empty set means "all calendars", and a launcher from before the picker showed exactly all of them. That is a concrete payoff of encoding "empty means all", not a matter of taste.

The selection enters the repository as a flow rather than being handed in by a screen, so the tile, the dashboard and the picker all see one answer. It is composed with the ordinary refresh trigger, so changing the selection updates the tile immediately rather than at the next tick — with one caveat found in review: that flow has to be de-duplicated, because the settings flow emits on every settings write and every one of those emissions would otherwise be a query to the content provider.

The summary row has to name the "all" case in words. Stored, that case is an empty set, and "0 selected" would read as a tile showing nothing — the exact opposite of what it means.

A known limitation, accepted: the IDs belong to the provider, so an account removed and added again comes back with a different ID and stops matching a saved selection. That reads as a calendar quietly returning to the tile, which is a better failure than the reverse — a filter hiding a calendar the user can see in every other app.

Also in 1.4.8

A new 3×3 size. Three grid units by three, so a row works out as 3 + 2 + 3 = 8: two large squares at the ends and a small tile between them. Every type gets it, including folders and the gallery, whose floor drops from 4×4. Weather fits three days with glyphs at that width, and the seven-day face trims to three instead of falling back to the current temperature; music shows title, artist and the full transport. The todo and clock tiles still compute their layout in whole cells and can look like 2×2 there — that is open work, not a hidden fallback.
Icon size from 60% to 250%, and Light / Normal / Bold type. One typographic scale now covers every tile face; large readings keep their sizes but take the chosen weight. The app name no longer moves the icon: its strip is a measured, invisible copy above the icon, so both ends of the face take the same space.
The badge dot became a pin. In dot mode every tile — 1×1, larger, folder — draws it in the top-right corner, 2 dp in, with no plate under a full-colour icon. The counter stays in the bottom strip beside the name. A notification arriving must not move or resize the icon.
All-day events were showing a day early, west of UTC. They begin at UTC midnight, and three places converted that with the local zone — so at a negative offset an all-day event landed on the wrong day on the tile, in the grid's dots and on the dashboard. A pre-existing bug, found while building the agenda, now fixed through one shared helper and covered by a test that had been passing falsely.
Edit controls hide while you move a tile. Delete, size, colour and the type control get out of the way during a drag and come back when you put it down.

What's next

1.4.9, and it is two things at once: Chrome's "Add to Home screen" finally has somewhere to go, and the settings screen stops being seven collapsible groups.

sqTile is on Google Play, in English, Polish, German, Italian and Ukrainian. Free, with a single one-time Pro unlock and no subscription. No ads, no analytics SDK, no account.