fix: cut the stranded-IME grace to 120ms and pin the upstream cause

The grace was 400ms, and once a window wedges its IME animation that full
400ms is paid on *every* dismissal, on every screen — which reads as the
padding lagging behind the keyboard rather than as a bug being corrected.

400ms was never protecting against anything real. `collectLatest` + `delay`
already means "no movement for X ms", because each animation frame emits a
new sample and cancels the pending wait. So the grace only has to outlast
the dead time between the target flipping and the animation's first
onProgress — not the animation. Measured over 12 real Gboard transitions on
a Pixel 8: that dead time is 17-36ms (closes 24-36, opens 17-23), and every
frame after it lands within 11ms across a ~264ms animation. 120ms clears the
worst case by ~3.3x. Set too low this degrades to a cosmetic snap, never to
wrong padding, since the target is always the truthful reading.

Also records what the workaround is working around. The defect is upstream:
a cancelled IME animation never delivers onEnd, so
`InsetsListener.runningAnimation` stays set, `onApplyWindowInsets` matches
neither branch, and `composeInsets.update()` is never called again —
`WindowInsets.ime` is dead for the life of the window. Bisected to
foundation-layout 1.4.0 (1.3.0 updated unconditionally and could not wedge),
still present in 1.12.0 and 1.13.0-alpha01. Compose's self-heal is scoped to
`SDK_INT == R`, and `WindowInsetsHolder.resetState()` only runs when the
holder's accessCount goes 0 -> 1 — which never happens in a single-Activity
app whose shell always reads insets. Filed as b/552500419.

Confirmed on-device that the bug is real and permanent underneath: with both
treatments disabled the inset pinned at 957px for 85s while the window
reported the keyboard gone, and `imeAnimationTarget` stayed correct
throughout — which is why reading it works.

ComposeImeInsetWedgeTest reproduces that upstream state deterministically in
~3s and is the repro attached to the bug. The failing half is @Ignore'd so
CI stays green; re-run it by hand after a Compose upgrade, and when it
passes, SafeImeInsets can be retired. The passing half is left enabled on
purpose: it guards the premise this fix depends on, so if a future Compose
release stopped keeping imeAnimationTarget current we would hear about it
instead of silently reading a second dead value.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Gs2gi3sZfQ7SHrVm2njLMw
This commit is contained in:
Vitor Pamplona
2026-08-25 18:19:23 -04:00
co-authored by Claude Opus 5
parent 416cd32eb3
commit 9eed37c5f4
3 changed files with 594 additions and 2 deletions
@@ -0,0 +1,379 @@
# Upstream issue draft — Compose `WindowInsets.ime` permanently wedges after a cancelled IME animation
Target: Google IssueTracker → **component 612128 (Jetpack Compose)**.
The library-specific component the docs link to (856989, from the "Create a new issue" button on
the Compose Foundation release notes) does not grant public Create Issues permission, so this is
filed one level up with a routing request at the top of the body.
Status: **FILED as https://issuetracker.google.com/issues/552500419 (b/552500419)** on 2026-08-25,
against component 612128 with a routing request. Remaining open item: the AOSP commit that introduced `runningAnimation`
between 1.3.0 and 1.4.0-alpha01 has not been identified (android.googlesource.com returned 403
to automated fetch). Adding the commit link before filing would help triage.
---
## Title
`WindowInsets.ime` stops updating permanently when an IME animation is cancelled without `onEnd` (regression in 1.4.0, still present in 1.13.0-alpha01)
## Routing
Please reassign to the owner of **`androidx.compose.foundation` / `foundation-layout`**
(WindowInsets). Filing here because component 856989 — the target of the "Create a new issue"
button on the [Compose Foundation release notes](https://developer.android.com/jetpack/androidx/releases/compose-foundation)
— does not grant Create Issues permission to external accounts. That documented path being
unusable by the public is arguably a separate docs bug worth fixing.
## Affected versions
* **Broken:** `androidx.compose.foundation:foundation-layout` **1.4.0 → 1.12.0 (current stable) and 1.13.0-alpha01**
* **Not broken:** 1.3.0 and earlier
* Verified by inspecting published `-sources.jar` for 1.2.0, 1.3.0, 1.4.0-alpha01…rc01, 1.4.0,
1.5.0, 1.6.0, 1.7.0, 1.8.0, 1.9.0, 1.10.0, 1.11.0, 1.12.0, 1.13.0-alpha01.
`runningAnimation` and its guard are absent in 1.3.0 and present from 1.4.0-alpha01 onward,
textually unchanged since.
* Reproduced on a Pixel 8, Android 17 (API 37). The API-30-only self-heal (below) means API 31+
has no recovery path at all.
## Summary
If a `WindowInsetsAnimation` is prepared and started but never ended — what a cancelled IME
animation looks like — `InsetsListener.runningAnimation` stays `true` forever. From that point
`onApplyWindowInsets` matches neither of its two branches, so `composeInsets.update()` is never
called again and **`WindowInsets.ime` is frozen for the remaining life of the window**.
Every `Modifier.imePadding()` in the app then holds a keyboard-height gap open with no keyboard
on screen, permanently. `WindowInsets.imeAnimationTarget` keeps reporting correctly, because
`updateImeAnimationTarget()` is called outside the guard — that asymmetry is the only reason a
workaround is possible at all.
## Reproduction
Deterministic instrumented test, ~3s, no gestures and no timing dependence. It drives Compose's
own listener through the cancelled-animation sequence using **public** interfaces
(`WindowInsetsAnimationCompat.Callback`, `OnApplyWindowInsetsListener`); reflection is used only
to obtain the listener instance for the view. Inside the androidx codebase `InsetsListener` is
directly accessible, so `listenerFor()` can be deleted and the rest of the test used verbatim.
```
FAIL aCancelledImeAnimationMustNotWedgeTheAnimatedInset
expected:<0> but was:<957>
PASS theAnimationTargetSurvivesTheWedge
```
The second test is expected to pass and is included on purpose: it pins the asymmetry between the
two readings, and would catch a "fix" that broke `imeAnimationTarget` instead.
The full test source is attached below.
## Root cause
`compose/foundation/foundation-layout/src/androidMain/kotlin/androidx/compose/foundation/layout/WindowInsets.android.kt`
```kotlin
override fun onPrepare(animation: WindowInsetsAnimationCompat) {
prepared = true
runningAnimation = true // set here…
}
override fun onStart(animation, bounds): BoundsCompat {
prepared = false // …prepared cleared, runningAnimation left set
return super.onStart(animation, bounds)
}
override fun onEnd(animation: WindowInsetsAnimationCompat) {
prepared = false
runningAnimation = false // …cleared ONLY here
…
}
override fun onApplyWindowInsets(view: View, insets: WindowInsetsCompat): WindowInsetsCompat {
savedInsets = insets
composeInsets.updateImeAnimationTarget(insets) // unconditional — stays correct
if (prepared) {
if (Build.VERSION.SDK_INT == Build.VERSION_CODES.R) {
view.post(this) // self-heal, API 30 ONLY
}
} else if (!runningAnimation) {
composeInsets.updateImeAnimationSource(insets)
composeInsets.update(insets) // the animated inset — never reached when wedged
}
…
}
```
After a cancelled animation: `prepared == false` (cleared by `onStart`) and
`runningAnimation == true` (never cleared, because `onEnd` never came). Neither branch runs.
`composeInsets.update()` is dead.
### Why the existing self-heal does not help
`run()` exists precisely to handle a cancelled animation, but:
1. it is gated to `Build.VERSION.SDK_INT == Build.VERSION_CODES.R` (API 30 only), and
2. it is posted only from the `if (prepared)` branch, and returns early unless `prepared` is still
`true` — which `onStart` has already cleared.
So it covers "cancelled between `onPrepare` and `onStart`, on API 30". It does not cover
"cancelled after `onStart`", on any API level.
### Why applications cannot recover
The only reset is `insetsListener.resetState()`, called from `WindowInsetsHolder.incrementAccessors()`
when `accessCount` transitions `0 → 1`. `accessCount` is driven by `WindowInsetsHolder.current()`'s
`DisposableEffect`, so it only reaches 0 when *every* insets consumer leaves composition
simultaneously.
In a single-Activity app whose shell (scaffold / bottom bar / drawer) always reads insets, that
never happens — the holder is created once and lives for the whole process. There is no public API
to force the reset. `WindowInsetsHolder` is `internal`.
Multi-Activity apps mask this: a new Activity means a new `View`, a new holder, and fresh state, so
the wedge dies with the Activity and reads as a transient glitch.
### Regression point
1.3.0's `onApplyWindowInsets` had no such gate and could not wedge:
```kotlin
override fun onApplyWindowInsets(view: View, insets: WindowInsetsCompat): WindowInsetsCompat {
if (prepared) {
savedInsets = insets
if (Build.VERSION.SDK_INT == Build.VERSION_CODES.R) view.post(this)
return insets
}
composeInsets.update(insets) // unconditional once onStart cleared `prepared`
return …
}
```
1.4.0 introduced `runningAnimation` and the `else if (!runningAnimation)` guard. Its own comment
states the intent:
> `// If an animation is running, rely on onProgress() to update the insets`
> `// On APIs less than 30 where the IME animation is backported, this avoids reporting`
> `// the final insets for a frame while the animation is running.`
i.e. a **one-frame** cosmetic flash on **API < 30** was fixed by making the update path conditional
on a flag that only `onEnd` clears — trading a single wrong frame on old devices for permanent
state corruption on all of them. The compensating recovery was never widened past `SDK_INT == R`.
## Real-world impact
Observed in a production Compose app (Amethyst, a Nostr client; single-Activity, `NavHost`,
77 `imePadding()` sites):
* On a Pixel 8 / Android 17, after ordinary manual use, `WindowInsets.ime` pinned at 957px while
the window reported `ime frame=[0,0][0,0]` — keyboard gone — and stayed pinned for 85+ seconds
until the process was restarted. Nothing in the app cleared it.
* Instrumented `WindowInsets.ime` vs `WindowInsets.imeAnimationTarget` across the failure:
```
17:12:58.803 animated=882 target=957 ← healthy open, 13 intermediate frames
17:12:58.902 animated=957 target=957
17:13:00.584 animated=957 target=0 ← dismissed; animated frozen
17:13:02.430 animated=0 target=957 ← reopened; snaps, no intermediate frames
17:13:03.479 animated=957 target=0 ← dismissed; frozen permanently
```
Note the loss of per-frame updates after the wedge: healthy transitions carry ~13 intermediate
values over ~264ms; post-wedge transitions carry none.
* Because the app never navigates away from its single Activity and its shell always reads insets,
`accessCount` never returns to 0, so the wedge is permanent for the session. Sessions in this app
routinely run for days.
The trigger for the underlying cancellation was not isolated — it is infrequent and required
extended manual use to hit. The defect being reported is not the cancellation itself but that
Compose enters a state it can never leave when one occurs. The attached test reproduces that state
directly and deterministically.
## Suggested fixes
Roughly in order of how targeted they are:
1. **Generalise the existing self-heal.** Post the `run()` reconciliation on all API levels, and
arm it after `onStart` as well as after `onPrepare`, so that an `onApplyWindowInsets` that
arrives with no intervening `onProgress` clears `runningAnimation` and applies `savedInsets`.
This preserves the API<30 one-frame behaviour the guard was added for, while bounding the
failure to a frame rather than forever.
2. **Reconcile on dispatch.** In `onApplyWindowInsets`, if `runningAnimation` is set but no
`onProgress` has been received since `onStart`, treat the animation as finished and update.
3. **Expose a reset.** A public way to reach `WindowInsetsHolder.resetState()` (or a documented
condition under which it runs) would at least let applications self-heal. Today they cannot,
short of reflection into an `internal` class — which R8 can rename or strip in exactly the
release builds where this occurs.
(1) or (2) is preferable: (3) only makes the bug survivable rather than fixing it.
## Environment
* `androidx.compose.foundation:foundation-layout` 1.12.0 (Compose BOM 2026.08.00)
* Pixel 8 (`shiba`), Android 17 / API 37, gesture navigation, Gboard, 120Hz
* Also inspected: 1.13.0-alpha01 — identical listener code
---
## Attachment — the failing test
```kotlin
package com.vitorpamplona.amethyst.ui.insets
import android.view.View
import android.view.animation.LinearInterpolator
import androidx.compose.foundation.layout.ExperimentalLayoutApi
import androidx.compose.foundation.layout.WindowInsets
import androidx.compose.foundation.layout.ime
import androidx.compose.foundation.layout.imeAnimationTarget
import androidx.compose.runtime.getValue
import androidx.compose.runtime.mutableIntStateOf
import androidx.compose.runtime.setValue
import androidx.compose.ui.platform.LocalDensity
import androidx.compose.ui.platform.LocalView
import androidx.compose.ui.test.junit4.createComposeRule
import androidx.core.graphics.Insets
import androidx.core.view.OnApplyWindowInsetsListener
import androidx.core.view.WindowInsetsAnimationCompat
import androidx.core.view.WindowInsetsCompat
import org.junit.Assert.assertEquals
import org.junit.Rule
import org.junit.Test
/**
* Upstream regression test for androidx.compose.foundation:foundation-layout.
*
* A `WindowInsetsAnimation` that is prepared and started but never ended — which is what a
* cancelled IME animation looks like — leaves `InsetsListener.runningAnimation` set forever.
* `onApplyWindowInsets` then matches neither of its two branches, so `composeInsets.update()`
* is never called again and `WindowInsets.ime` is dead for the life of the window.
*
* Introduced in 1.4.0 (absent in 1.3.0, where `onApplyWindowInsets` updated unconditionally
* once `onStart` had cleared `prepared`). Still present in 1.12.0 and 1.13.0-alpha01. The
* compensating self-heal (`view.post(this)` -> `run()`) is scoped to `SDK_INT == R`, so on
* API 31+ nothing clears the flag; `WindowInsetsHolder.resetState()` only runs when the
* holder's accessCount transitions 0 -> 1, which never happens in an app whose shell always
* reads insets.
*
* [aCancelledImeAnimationMustNotWedgeTheAnimatedInset] FAILS on every version from 1.4.0 on.
* [theAnimationTargetSurvivesTheWedge] documents the asymmetry that makes a workaround possible
* and is expected to PASS — `updateImeAnimationTarget` is called outside the guard.
*/
class ComposeImeInsetWedgeTest {
@get:Rule val rule = createComposeRule()
private val keyboardHeight = 957
private fun imeInsets(bottom: Int): WindowInsetsCompat =
WindowInsetsCompat
.Builder()
.setInsets(WindowInsetsCompat.Type.ime(), Insets.of(0, 0, 0, bottom))
.setVisible(WindowInsetsCompat.Type.ime(), bottom > 0)
.build()
/** Compose's own listener for this view. Private class, but both interfaces it exposes are public. */
private fun listenerFor(view: View): Any {
val holderClass = Class.forName("androidx.compose.foundation.layout.WindowInsetsHolder")
val companion =
holderClass.getDeclaredField("Companion").run {
isAccessible = true
get(null)
}
val holder =
companion.javaClass
.getDeclaredMethod("getOrCreateFor", View::class.java)
.run {
isAccessible = true
invoke(companion, view)
}
return holderClass.getDeclaredField("insetsListener").run {
isAccessible = true
get(holder)!!
}
}
private fun anim() = WindowInsetsAnimationCompat(WindowInsetsCompat.Type.ime(), LinearInterpolator(), 250L)
private fun bounds() =
WindowInsetsAnimationCompat.BoundsCompat(
Insets.NONE,
Insets.of(0, 0, 0, keyboardHeight),
)
@OptIn(ExperimentalLayoutApi::class)
@Test
fun aCancelledImeAnimationMustNotWedgeTheAnimatedInset() {
var animated by mutableIntStateOf(-1)
lateinit var view: View
rule.setContent {
view = LocalView.current
val density = LocalDensity.current
animated = WindowInsets.ime.getBottom(density)
}
rule.waitForIdle()
val listener = listenerFor(view)
val onApply = listener as OnApplyWindowInsetsListener
val callback = listener as WindowInsetsAnimationCompat.Callback
// Baseline: with no animation in flight the inset tracks normally.
rule.runOnUiThread { onApply.onApplyWindowInsets(view, imeInsets(keyboardHeight)) }
rule.waitForIdle()
assertEquals("baseline: the inset must follow a plain dispatch", keyboardHeight, animated)
// A cancelled animation: prepared and started, but onEnd never arrives.
rule.runOnUiThread {
callback.onPrepare(anim())
callback.onStart(anim(), bounds())
}
rule.waitForIdle()
// The keyboard is gone and the window says so. The animated inset must follow.
rule.runOnUiThread { onApply.onApplyWindowInsets(view, imeInsets(0)) }
rule.waitForIdle()
assertEquals(
"WindowInsets.ime must still track the window after an animation was cancelled " +
"without onEnd; it is instead frozen at the keyboard height forever",
0,
animated,
)
}
@OptIn(ExperimentalLayoutApi::class)
@Test
fun theAnimationTargetSurvivesTheWedge() {
var target by mutableIntStateOf(-1)
lateinit var view: View
rule.setContent {
view = LocalView.current
val density = LocalDensity.current
target = WindowInsets.imeAnimationTarget.getBottom(density)
}
rule.waitForIdle()
val listener = listenerFor(view)
val onApply = listener as OnApplyWindowInsetsListener
val callback = listener as WindowInsetsAnimationCompat.Callback
rule.runOnUiThread { onApply.onApplyWindowInsets(view, imeInsets(keyboardHeight)) }
rule.waitForIdle()
assertEquals(keyboardHeight, target)
rule.runOnUiThread {
callback.onPrepare(anim())
callback.onStart(anim(), bounds())
}
rule.waitForIdle()
rule.runOnUiThread { onApply.onApplyWindowInsets(view, imeInsets(0)) }
rule.waitForIdle()
assertEquals(
"updateImeAnimationTarget is called outside the guard, so this reading stays truthful",
0,
target,
)
}
}
```
@@ -0,0 +1,191 @@
/*
* Copyright (c) 2025 Vitor Pamplona
*
* Permission is hereby granted, free of charge, to any person obtaining a copy of
* this software and associated documentation files (the "Software"), to deal in
* the Software without restriction, including without limitation the rights to use,
* copy, modify, merge, publish, distribute, sublicense, and/or sell copies of the
* Software, and to permit persons to whom the Software is furnished to do so,
* subject to the following conditions:
*
* The above copyright notice and this permission notice shall be included in all
* copies or substantial portions of the Software.
*
* THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
* IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, FITNESS
* FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR
* COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN
* AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION
* WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.
*/
package com.vitorpamplona.amethyst.ui.insets
import android.view.View
import android.view.animation.LinearInterpolator
import androidx.compose.foundation.layout.ExperimentalLayoutApi
import androidx.compose.foundation.layout.WindowInsets
import androidx.compose.foundation.layout.ime
import androidx.compose.foundation.layout.imeAnimationTarget
import androidx.compose.runtime.getValue
import androidx.compose.runtime.mutableIntStateOf
import androidx.compose.runtime.setValue
import androidx.compose.ui.platform.LocalDensity
import androidx.compose.ui.platform.LocalView
import androidx.compose.ui.test.junit4.createComposeRule
import androidx.core.graphics.Insets
import androidx.core.view.OnApplyWindowInsetsListener
import androidx.core.view.WindowInsetsAnimationCompat
import androidx.core.view.WindowInsetsCompat
import org.junit.Assert.assertEquals
import org.junit.Ignore
import org.junit.Rule
import org.junit.Test
/**
* Upstream regression test for androidx.compose.foundation:foundation-layout.
*
* A `WindowInsetsAnimation` that is prepared and started but never ended — which is what a
* cancelled IME animation looks like — leaves `InsetsListener.runningAnimation` set forever.
* `onApplyWindowInsets` then matches neither of its two branches, so `composeInsets.update()`
* is never called again and `WindowInsets.ime` is dead for the life of the window.
*
* Introduced in 1.4.0 (absent in 1.3.0, where `onApplyWindowInsets` updated unconditionally
* once `onStart` had cleared `prepared`). Still present in 1.12.0 and 1.13.0-alpha01. The
* compensating self-heal (`view.post(this)` -> `run()`) is scoped to `SDK_INT == R`, so on
* API 31+ nothing clears the flag; `WindowInsetsHolder.resetState()` only runs when the
* holder's accessCount transitions 0 -> 1, which never happens in an app whose shell always
* reads insets.
*
* Filed upstream as b/552500419.
*
* [aCancelledImeAnimationMustNotWedgeTheAnimatedInset] FAILS on every version from 1.4.0 on, so it
* is [Ignore]d to keep CI green. It is not a test of Amethyst code — it is the upstream repro we
* attached to the bug. **Re-run it by hand after every Compose upgrade**: when it passes, the
* upstream fix has landed and [com.vitorpamplona.amethyst.ui.insets.SafeImeInsets] can be retired.
*
* [theAnimationTargetSurvivesTheWedge] documents the asymmetry that makes a workaround possible
* and is expected to PASS — `updateImeAnimationTarget` is called outside the guard. It stays
* enabled, because it guards the premise [com.vitorpamplona.amethyst.ui.insets.SafeImeInsets]
* depends on: if a future Compose release stopped keeping `imeAnimationTarget` current, our
* fallback would silently start reading a dead value too.
*/
class ComposeImeInsetWedgeTest {
@get:Rule val rule = createComposeRule()
private val keyboardHeight = 957
private fun imeInsets(bottom: Int): WindowInsetsCompat =
WindowInsetsCompat
.Builder()
.setInsets(WindowInsetsCompat.Type.ime(), Insets.of(0, 0, 0, bottom))
.setVisible(WindowInsetsCompat.Type.ime(), bottom > 0)
.build()
/** Compose's own listener for this view. Private class, but both interfaces it exposes are public. */
private fun listenerFor(view: View): Any {
val holderClass = Class.forName("androidx.compose.foundation.layout.WindowInsetsHolder")
val companion =
holderClass.getDeclaredField("Companion").run {
isAccessible = true
get(null)
}
val holder =
companion.javaClass
.getDeclaredMethod("getOrCreateFor", View::class.java)
.run {
isAccessible = true
invoke(companion, view)
}
return holderClass.getDeclaredField("insetsListener").run {
isAccessible = true
get(holder)!!
}
}
private fun anim() = WindowInsetsAnimationCompat(WindowInsetsCompat.Type.ime(), LinearInterpolator(), 250L)
private fun bounds() =
WindowInsetsAnimationCompat.BoundsCompat(
Insets.NONE,
Insets.of(0, 0, 0, keyboardHeight),
)
@OptIn(ExperimentalLayoutApi::class)
@Test
@Ignore("Fails by design until upstream fixes b/552500419 — re-run by hand on every Compose upgrade")
fun aCancelledImeAnimationMustNotWedgeTheAnimatedInset() {
var animated by mutableIntStateOf(-1)
lateinit var view: View
rule.setContent {
view = LocalView.current
val density = LocalDensity.current
animated = WindowInsets.ime.getBottom(density)
}
rule.waitForIdle()
val listener = listenerFor(view)
val onApply = listener as OnApplyWindowInsetsListener
val callback = listener as WindowInsetsAnimationCompat.Callback
// Baseline: with no animation in flight the inset tracks normally.
rule.runOnUiThread { onApply.onApplyWindowInsets(view, imeInsets(keyboardHeight)) }
rule.waitForIdle()
assertEquals("baseline: the inset must follow a plain dispatch", keyboardHeight, animated)
// A cancelled animation: prepared and started, but onEnd never arrives.
rule.runOnUiThread {
callback.onPrepare(anim())
callback.onStart(anim(), bounds())
}
rule.waitForIdle()
// The keyboard is gone and the window says so. The animated inset must follow.
rule.runOnUiThread { onApply.onApplyWindowInsets(view, imeInsets(0)) }
rule.waitForIdle()
assertEquals(
"WindowInsets.ime must still track the window after an animation was cancelled " +
"without onEnd; it is instead frozen at the keyboard height forever",
0,
animated,
)
}
@OptIn(ExperimentalLayoutApi::class)
@Test
fun theAnimationTargetSurvivesTheWedge() {
var target by mutableIntStateOf(-1)
lateinit var view: View
rule.setContent {
view = LocalView.current
val density = LocalDensity.current
target = WindowInsets.imeAnimationTarget.getBottom(density)
}
rule.waitForIdle()
val listener = listenerFor(view)
val onApply = listener as OnApplyWindowInsetsListener
val callback = listener as WindowInsetsAnimationCompat.Callback
rule.runOnUiThread { onApply.onApplyWindowInsets(view, imeInsets(keyboardHeight)) }
rule.waitForIdle()
assertEquals(keyboardHeight, target)
rule.runOnUiThread {
callback.onPrepare(anim())
callback.onStart(anim(), bounds())
}
rule.waitForIdle()
rule.runOnUiThread { onApply.onApplyWindowInsets(view, imeInsets(0)) }
rule.waitForIdle()
assertEquals(
"updateImeAnimationTarget is called outside the guard, so this reading stays truthful",
0,
target,
)
}
}
@@ -47,9 +47,22 @@ import kotlinx.coroutines.flow.collectLatest
*
* A real IME animation reports a new inset every frame, so any value comfortably above a few
* dropped frames never trips on one. In the stranded case nothing is emitted at all, so this is
* purely how long the stale padding stays on screen before it is corrected.
* purely how long the stale padding stays on screen before it is corrected — which is why this is
* kept as tight as the frame cadence safely allows rather than padded "to be sure".
*
* The gap that actually binds this is not the frame cadence but the dead time between the target
* flipping and the animation's first onProgress: until that first frame lands, "not started yet"
* and "wedged" look identical. Measured on a Pixel 8 (120Hz) over 12 real Gboard transitions:
* that dead time is 17-36ms (mean 24ms; closes 24-36ms, opens 17-23ms), after which every
* subsequent frame lands within 11ms across a ~264ms animation. 120ms therefore clears the worst
* observed start-up gap by ~3.3x while keeping the stale padding below the threshold where it
* reads as lag. The earlier 400ms was ~11x that gap, and once a window wedges its IME animation
* the full 400ms was paid on *every* dismissal.
*
* A value set too low degrades to a cosmetic snap — the padding jumps to its final position
* instead of animating — never to wrong padding, since the target is always the truthful reading.
*/
const val IME_STRAND_GRACE_MS = 400L
const val IME_STRAND_GRACE_MS = 120L
/**
* The animated IME inset ([WindowInsets.ime]), corrected for a Compose insets listener that has
@@ -151,6 +164,15 @@ internal suspend fun watchForStrandedIme(
* `imeAnimationTarget` is still marked experimental, but it is the whole point of this file: it is
* the only IME reading Compose keeps current while its listener is wedged, because
* `onApplyWindowInsets` publishes it before it consults the flag that gates everything else.
*
* The underlying defect is upstream, not ours: a cancelled IME animation never delivers `onEnd`,
* so `InsetsListener.runningAnimation` stays set and `composeInsets.update()` is never called
* again — `WindowInsets.ime` is dead for the life of the window. Filed as **b/552500419**;
* a regression in foundation-layout 1.4.0, still present in 1.12.0 and 1.13.0-alpha01. Compose's
* own self-heal is scoped to `SDK_INT == R`, and its other reset (`WindowInsetsHolder.resetState()`)
* only runs when the holder's accessCount goes 0 -> 1, which never happens in an app whose shell
* always reads insets. Retire this file once that bug is fixed — `ComposeImeInsetWedgeTest`
* (androidTest) is the canary that says when.
*/
@OptIn(ExperimentalLayoutApi::class)
@Composable