Story Reel Auto Pager

A self-advancing story reel

One HorizontalPager holds every card; a single LaunchedEffect keyed on the current page runs the timer for whichever card is showing. Nothing about the reel is per-card except the cards themselves — the progress bar, the timer, the tap zones and the close button all live in the shell, so ten cards can be ten different layouts without ten copies of that chrome drifting apart. The pager's own currentPage is the only place "which card" is recorded; a second counter anywhere else is a second answer waiting to disagree with it.

Traps

The timer counts frame deltas, not elapsed wall time, so a long hold cannot bank progress. Sampling a clock across a pause and using the raw difference lets whatever accumulated while paused land in one jump the instant it resumes. Take the delta from the previous frame every frame, paused or not, and only add it to the running total when nothing is holding the card back:

while (elapsed < holdMs) {
    withFrameMillis { frame ->
        val delta = if (lastFrame == 0L) 0L else frame - lastFrame
        lastFrame = frame
        if (!pressed && !busy && !pagerState.isScrollInProgress) elapsed += delta
    }
    progress = (elapsed.toFloat() / holdMs).coerceIn(0f, 1f)
}

All three gates matter for different reasons: pressed is the read-longer hold, busy is a save or share in flight (advancing under a share sheet would hand the next card's image to whatever the user picked), and isScrollInProgress covers a manual swipe already carrying the reel to the next card on its own. Missing any one of them is a pause that only sometimes pauses.

An empty onLongPress is not a no-op — it is what makes releasing a hold resume instead of navigate. detectTapGestures behaves differently depending on whether onLongPress is null: with it null, every release — however long the finger was down — resolves as an ordinary tap and fires onTap. Passed a non-null callback, even one that does nothing, a hold that outlasts the long-press timeout consumes the gesture before release and onTap never fires for it:

onLongPress = { },   // deliberately empty, deliberately not omitted offset ->
    val from = pagerState.targetPage
    goTo(if (offset.x > size.width / 2f) from + 1 else from - 1)
},

Delete the seemingly-dead lambda and a press held to read a line carefully still fires onTap the moment the finger lifts — the exact skip the hold was supposed to prevent. Verified against androidx.compose.foundation:foundation's TapGestureDetector.kt: the plain up-event wait (waitForUpOrCancellation()) only runs when onLongPress is null; passed a callback, the detector waits on waitForLongPress() instead, and a successful long press consumes to up and returns before onTap is ever reached.

A tap handler that can fire again before the last animation lands must read the target page, not the current one. pagerState.currentPage is the page actually on screen; pagerState.targetPage is where an in-flight animateScrollToPage is headed. A second tap arriving mid-animation, read off currentPage, re-issues a scroll to the page that is already being left — read off targetPage, it steps one further from wherever the first tap is already going. This isn't in tension with nowplaying-pager-no-feedback-loop reading currentPage on purpose — that skill syncs external state once a scroll has settled; this one decides where a tap goes mid-animation, a different question.

The segmented bar's count is the filtered card list's size, never the full set of card kinds. A year with no biggest day or too few dated plays for a decade breakdown drops those cards entirely (WrappedCard.entries.filter { ... }), and every index downstream — the pager's page count, the close-vs-advance decision, the segment row — reads that filtered list's size, not the enum's. Sizing the segment row off the enum instead prints empty trailing segments for cards this particular year was never going to show.

Resolve every card's data before the first frame mounts, not card-by-card as the user pages through. The whole year is computed in one function that returns a single immutable result, gated behind a three-state holder — loading, not-enough-data, or ready — and the pager only exists inside the ready state. A spinner appearing between two cards mid-story, instead of only once before the first one, means some card's data is being fetched lazily on arrival; move it into the same upfront computation everything else went through.

A card meant to be captured cannot take its colour from the viewer's current theme. The reel builds its whole scheme from the year's own artwork with isDark fixed to true, regardless of the device's light/dark setting, because the same card is captured as a share image — a picture that came out light-mode on one phone and dark-mode on another is not the same object, even though both are "correct" renders of the same screen. Anything captured for export needs its inputs pinned the same way; the device's ambient settings are exactly the kind of input that must not leak in.

The capture region is the poster inside the card, not the card's own slot. Every other card hands its whole slot to capturable() (see compose-capture-to-share-image), but the one card with its own Save and Share buttons passes the capture modifier to its inner panel only — the eyebrow above it and the two buttons below it are how you ask for the image, not part of it. The same slot composable also applies its bottom chrome padding before backgrounding and capturing, so the capture sees the same bounds a plain card gets; reversing that order bakes the padding into the exported picture as empty ground along one edge.

Verifying it

Run from the Kotlin source root holding your ui/ and viewModel/ packages — here, ui/screen/home/wrapped/WrappedScreen.kt is the reel screen and viewModel/WrappedUiState.kt is the state it reads; adjust both paths to your own layout.

1. All three pause gates are on the timer's own condition, not just one of them:

   grep -n "!pressed && !busy && !pagerState.isScrollInProgress" ui/screen/home/wrapped/WrappedScreen.kt
   

   Pass condition: found. Missing any name here is a case that no longer pauses.

2. The empty long-press callback exists and is not simply absent:

   grep -n "onLongPress = { }" ui/screen/home/wrapped/WrappedScreen.kt
   

   Pass condition: found. Absent, a hold-then-release will fire onTap on release regardless of how long the hold was.

3. The tap handler reads targetPage, the header reads currentPage:

   grep -n "pagerState.targetPage\|pagerState.currentPage" ui/screen/home/wrapped/WrappedScreen.kt
   

   Pass condition: targetPage appears inside the tap-gesture block; currentPage appears where the currently-shown card is read for display.

4. The segment count and the pager's page count both trace back to the filtered list:

   grep -n "cards.size" ui/screen/home/wrapped/WrappedScreen.kt
   grep -n "WrappedCard.entries.filter" viewModel/WrappedUiState.kt
   

   Pass condition: every on-screen count uses cards.size; the filter exists exactly once, upstream of all of them.

5. Export the same card twice: once with the device in light mode, once in dark. The two exported files must be pixel-identical — two exports taken in the same device theme prove nothing here. Any difference means the device's own theme setting reached a value inside the capture.