Skip to content

Prototype device verification status

Status when recorded

Host tests and JVM/Robolectric tests cover these features on fakes; the device checks below were still owed. The section headings of these features carried “not device-verified”.

  • Not device-verified features: flow and grid, carousel and pager indicator, pull-to-refresh, chat lists, swipe actions, loading and empty states, effects, motion v2, transition choreography, charts and lazy lists.
  • Frosted backdrop: the system layer blurring behind the window was not verified on a device.
  • Android Automotive: the floored top inset and whether the overlay stays visible under driving restrictions; checking them needs a userdebug or eng AAOS image, since user builds refuse driving-state emulation.
  • Custom shaders (needs-device-verify): real API 33+ compilation and rendering, procedural sources with an unused content input, and isolation of each node’s sampled content and uniforms. Also shader output clipped to node bounds, lifecycle/background and capture-hide sampling gaps, frame-budget fallback under sustained and spaced load, and status/inspect reporting after a real rendering failure. JVM tests use fake compilers and frame clocks and give no GPU/rendering proof.
  • Haptics and sound (needs-device-verify): an Android emulator tap appearing in feedbackLog and in dumpsys vibrator_manager, an iOS Simulator feedbackLog entry, and the feel on a physical device.
  • Lifecycle and safety: 3-button navigation with layer: "system", and a prototype shown with no application window in front (which Home does not hide). Device checks must cover keyguard timing, daemon death, pager page 3 across rotation, fold/display removal, and API 30/34/36 keyboard/cutout geometry.
  • Bottom sheet and the keyboard: the lift is unit-tested with fake bounds; the API 30/34/36 device check was open.
  • Chat lists: whether a fullscreen overlay receives IME insets (like the bottom-sheet lift, #10262); on iOS (#14671) the simulator acceptance of the pill, jump-to-end and keyboard behaviour.
  • Gesture actions: whether observe reports the TalkBack custom actions; a device capture was owed before the guide promises it.
  • Effects: goldens on API 30, 33 and 36 (effects_all_types, pending in PrototypeSpecContentScreenshotTest) and a real-GPU look at the AGSL effects. Robolectric covers decoding, API planning, the clock and budget on fakes, and bounds stability.
  • Rich text and linkify: advertise the capability only after renderer support. Verify mixed span styles, resolved concatenated observe text, per-span actions without URL navigation, Android semanticLinks and both selectors across windows, and iOS tapAt-only links plus platformNotes. Host tests do not verify device rendering.
  • Lazy lists: prepend, anchoring and per-stack-entry list scroll restore were not verified before the acceptance run; benchmark qualification was unverified. A 2000-row feed validated on the host in tens of milliseconds (benchmark/prototype-feed-limits.ts).
  • Strings and semantics: device acceptance had not been run.
  • Accessibility audit on iOS: the XCUITest accessibilityAudit could not target an injected prototype app, and the RTL change did not take effect on a simulator, so neither was verified.
  • Glass on iOS: verification pins accessibility settings and checks structure and bounds, not glass pixels. Group merging, morphs, unions, image extension and accessibility appearance still required iOS 26 simulator verification.
  • Saved prototype pinning: autoConfirm (#12266) was device-verification work because launcher dialogs differ.

Test budget

The repository’s 100 ms per-test unit budget is measured only from bun’s JUnit reporter (scripts/test-ts.sh, scripts/validate-bun-test-timings.sh). It does not cover Gradle or Robolectric tests, so the Compose and host tests for window and modal motion are not held to it.

Page process

The spec page was registered under How it Works in mkdocs.yml and linked from the design index.