OpenSign

Each screen is a browser on a cheap box, and each box keeps its own time; those clocks wander seconds apart. At first every display ran its own slideshow counter from whenever it loaded, so screens showing the same folder fell out of step, and a display that reconnected started again at image one.

I wanted lockstep without WebSockets or a push service. The fix landed in two steps, twelve minutes apart. The slide index became a pure function of time, floor(now / period) mod N, and then “now” became the server’s clock. Each display estimates its offset with a round-trip-corrected sample, re-measures every five minutes, and re-arms its timer for the next period boundary. Screens flip together, to within the accuracy of that sample.

The same idea runs through the rest. The clocks can show the server’s wall time by shifting a date by its UTC offset, with no timezone database, and the verse of the day is the day number mod 710, so every screen agrees without coordinating. Changes arrive through a 5-second config poll, and a reload token in that response lets one Reload displays button, or a server restart after an update, refresh every screen.

The frosted-glass widgets blurred on Safari-based displays and showed no blur at all in Chrome, Edge and Opera. A test page never reproduced it, because files in public/ are copied unminified. The cause showed up only in the CSS the browser actually received. Vite 8’s minifier, Lightning CSS, collapses the prefixed and standard backdrop-filter declarations into one, keeping whichever is written last, and Chromium ignores the -webkit- spelling. It is a known upstream issue.

Putting the standard property last fixed the symptom. Then I closed the class of bug: a small Vite plugin inspects every emitted stylesheet and fails the build if it holds fewer standard backdrop-filter declarations than prefixed ones. It checks the output, not the source, because the output is what the screen gets.

The verse-of-the-day widget has to show exact text: the public-domain Berean Standard Bible asks only that text carrying its name match the official wording. The first cut of 30 verses was recalled in an AI-assisted session rather than extracted, and flagged at the time as unchecked. Ten minutes later the check came back: 9 of 30 were wrong, 5 with changed words (one missing the verse’s closing clause) and 4 in punctuation only.

Since then a script does the work. It pulls a ranked reference list from OpenBible.info and the verbatim BSB text, and asserts that the download carries the public-domain dedication. It strips psalm headings only by exact match: 21 are listed, 7 psalms are declared to have none, and any other psalm opening fails the run. A trailing “Selah” and one closing musical note are trimmed the same narrow way. Forty references are excluded by hand because they don’t stand alone without context.

Licensing lives in code too. Other translations use each install’s own API.Bible key, kept out of the config every display reads and never returned by the status route (the file endpoint is another matter; see the honest notes). Fetched verses are cached so screens share one call, purged at 30 days, and each view is reported to API.Bible, as its terms require. One trap surfaced on the way: red-letter text arrived without verse ids, so filtering by id silently deleted it. The parser filters by paragraph style instead.

The same layout has to read on a narrow strip display, a portrait 27-inch monitor and a 55-inch TV, and the people arranging it are not designers, so there is no free positioning. Five widgets (clock, calendar, countdown, giving QR code, verse of the day) sit over any of three modes (Text, Images, Stand By) and drop onto an orientation-aware 15-point grid: 5×3 or 3×5, odd counts so there is a true centre. A move onto an occupied point is refused. Each of four fixed sizes grows away from its anchor, so it never spills off its edge, and widget internals are sized in vmin, so a card keeps its proportions on any screen.

The verse card measures itself against a live 120-character sample. Longer text shrinks in 2% steps to a 70% floor; anything over 243 characters is skipped for the next verse, and a licensed translation falls back to the built-in text after 10 fetches without a fit.

Even spacing measures the widgets that share a column (portrait) or a row (landscape), holds the first and last against the edges, shares the leftover space equally, and re-measures whenever a size changes.

Signage usually fails for operational reasons, not code: a box loses the server after days of uptime, a screen needs a person after every update, or the dev setup differs from what ships. One FastAPI process on one port serves both the built frontend and the API. A split into separate dev and prod servers was tried in July and removed the same day in favour of parity.

In the field, one display stopped reaching the server after several days up, even though a speed test on that box passed, and a reboot cleared it every time. The cause was never pinned down, so the fix went into operations: a scheduled nightly reboot became a Required item in the display-box checklist, beside auto-starting the browser at the kiosk URL and turning off OS sleep; the server itself gets a stable address (a static IP or reserved lease) because every display points at it. The browser’s wake lock needs a secure context, so over plain LAN HTTP it does nothing; the OS setting is what keeps those screens on.

Images are served in place by path, so nothing is uploaded or copied. Picking a file opens the real OS dialog on the server machine: through osascript on macOS, because tkinter crashes off the main thread there and Homebrew’s Python ships without Tk, and through tkinter on Windows and Linux.