Platform capabilities as composable expect functions

Three shapes cover almost everything, and they are worth telling apart because each fails differently:

@Composable expect fun getScreenSizeInfo(): ScreenSizeInfo   // a measurement
@Composable expect fun KeepScreenOn()                        // an effect with an undo
@Composable expect fun rememberIsInPipMode(): Boolean        // a subscription, read as state

Every declared type is a common type: the measurement returns a small data class declared in the shared module, not a platform window object. That is the whole point — callers in shared code stay free of platform types, and the platform detail lives in exactly one file per target.

A capability one platform lacks still gets a full actual. It returns the neutral value:

@Composable actual fun rememberIsInPipMode(): Boolean = false   // this platform has no such mode
@Composable actual fun KeepScreenOn() { /* no display timeout to hold off here */ }

Traps

A stub must be correct, not absent. TODO() in an actual compiles, ships, and stops the app the first time that platform paints the screen — and because it sits behind a capability nobody tests on the secondary platform, it reaches users. The neutral value is almost always obvious: false for "is this mode active", an empty body for an effect, an empty list for a query. Write it, with the reason on the same line.

An empty actual is a decision, and it is not the same decision as "the platform cannot". Both look identical in the source. In the file inspected here the desktop target implements the measurement fully — from ordinary shared Compose APIs — while leaving the keep-awake effect empty under a note to implement it later. The full measurement proves nothing about the empty effect: they are unrelated capabilities, and an empty body in that company says nobody wrote anything, not that the platform cannot. Before concluding a platform lacks a capability, check that someone actually looked — see noop-actual-not-platform-limit.

The neutral return is a real answer, and callers act on it. A stub returning false for a windowing-mode query does not mean "unknown" — every if in shared code reads it as "not in that mode" and lays out accordingly. Correct for a platform with no such mode; wrong for one that has it and simply is not wired yet. If that difference matters, make the capability able to say "not supported here", and make the callers handle it.

@Composable belongs on the expect and on every actual. It is part of the declaration, not an implementation detail, and a mismatch is rejected — but the confusing case is the reverse: an actual that does not need composition still carries the annotation, so do not "clean it off" the stub.

Effect-shaped capabilities must undo themselves. The keep-awake actual sets the flag inside a DisposableEffect and clears it in onDispose; the subscription actual registers its listener the same way and removes it. A capability that sets a device-wide flag and never clears it outlives the screen that asked for it, and the symptom appears somewhere else entirely. If an actual has no onDispose, it is either genuinely stateless or a leak.

A measurement needs a cache key, and the right key differs per platform. The Android actual wraps its computation in remember(configuration), so it recomputes on configuration change and not on every recomposition. The desktop actual reads a Compose-provided window value that is already observable and needs no key. Copying the Android shape over as remember(Unit) freezes the value at first composition and never updates on resize — a bug only the platform with resizable windows shows.

A measurement reports the container, and your own chrome may be inside it. A title bar drawn by the app, above the content, is still part of the container the measurement reads — so the height it returns includes a strip no layout can use, and every consumer computing "how much room have I got" overflows by exactly that strip. It reads as a styling difference rather than a wrong number, and anything gating on the measurement inherits it (responsive-gate-size-not-platform). Two things make the correction survive: one constant, read by both the bar that draws the strip and the actual that subtracts it, and a conditional subtraction:

// adapted — the flag is published once, at window creation, before the first frame
object WindowChrome { const val TITLE_BAR_HEIGHT_DP = 40; @Volatile var inWindowBar = false }
val chromeTopPx = if (WindowChrome.inWindowBar)
    with(density) { WindowChrome.TITLE_BAR_HEIGHT_DP.dp.roundToPx() } else 0
val contentHeightPx = (window.containerSize.height - chromeTopPx).coerceAtLeast(0)

The condition is not "which platform": the same platform draws its own bar in some configurations and lets the system decorate the window in others, and a system decoration lives outside the container size — subtracting there under-reports by the same amount. Check which windows call the measurement, too: a secondary window with different chrome needs its own answer.

The platform actual is where the activity hunt lives, and how it asks decides what happens when there is none. The Android file inspected here walks the context chain twice, in two functions: one returns null when it runs out, the other throws. A composable built on the throwing walk takes the screen down when hosted outside an activity — a preview, an embedded host, a test. Pick per capability: a measurement can fall back to zero, a mode subscription probably cannot subscribe.

Version branches and their suppressions live in the actual, and that is correct. The Android measurement carries @Suppress("DEPRECATION") for its older-API branch. Keeping that inside the actual is what stops the suppression from applying to shared code, so resist hoisting it.

Verifying it

Confirm every expect has an actual in every source set you build, by name:

grep -rhoE "expect fun [a-zA-Z][A-Za-z0-9_]*" --include="*.kt" . | sed 's/expect fun //' | sort -u | while read -r f; do
  printf '%-30s actual in: %s\n' "$f" "$(grep -rlE "actual fun $f\b" --include="*.kt" . | sed -E 's|.*/src/([^/]+)/.*|\1|' | sort -u | tr '\n' ' ')"
done

Read the column, not the count: a declaration listing fewer source sets than its siblings is either a target that will not link or a hierarchy where one parent covers several — both worth knowing before a release build tells you.

Then find the stubs, so each one is a decision you have seen:

find . -path "*/src/*Main/*" -name "*.kt" | while read -r f; do
  sed -E 's://.*$::' "$f" | grep -vE "^[[:space:]]*$" \
  | grep -A1 "actual fun " | grep -B1 "^[[:space:]]*\}$" \
  | grep -oE "actual fun [A-Za-z_][A-Za-z0-9_]*" | sed "s|^|$f  |"
done

Stripping comments first is what finds the ones written as a bare comment inside an otherwise empty body — exactly how a stub is usually spelled. Each hit needs a reason next to it; the ones without are candidates for noop-actual-not-platform-limit.

And confirm nothing ships an unwritten branch:

grep -rn -A4 "actual fun " --include="*.kt" . | grep -E "TODO\(\)|NotImplementedError"

Finally, for a measurement actual, confirm the chrome strip has one definition and two readers — the composable that draws it and the actual that subtracts it:

grep -rn 'TITLE_BAR_HEIGHT\|_BAR_HEIGHT_DP' --include='*.kt' . | grep -v '/build/'

A literal height in the bar plus a separate literal in the actual is the drift this prevents; one reader and no other is a subtraction nothing keeps honest.