iOS physical device: focused React Native overlay input can leave XCTest-backed commands timing out while screenshots/taps still work

[billsbooth]

Environment

• agent-device: 0.18.3 (npx -y agent-device --version)
• Host OS: macOS 26.5.2 (25F84)
• Xcode: 26.6 (17F113)
• Node: v26.4.0
• npm: 11.17.0
• Target: physical iPhone connected to the Mac
• App type: Expo React Native development-client app
• UI shape involved: focused text input inside a bottom-sheet / overlay style surface, plus a related custom numeric keypad surface with a decorative infinite Reanimated animation

Description

We are seeing intermittent but repeatable iOS physical-device automation stalls after a focused input inside a React Native bottom-sheet / overlay is active.

The app remains visibly responsive:

• agent-device screenshot still works.
• Raw coordinate presses still work.
• Pressing visible keyboard keys by coordinate can enter text.
• Pressing visible custom keypad keys by coordinate can update the amount display.

But XCTest-backed commands can start timing out in the same state:

• snapshot -i
• type
• fill
• keyboard dismiss
• drag/back-style recovery attempts

The observed error class is main thread execution timed out.

This feels like XCTest quiescence/idleness getting stuck after the overlay/input/keyboard state becomes active, rather than the app being frozen. One related surface also contains a decorative infinite Reanimated caret animation, which may make the problem easier to trigger, but the focused text-input overlay case can happen separately from that caret.

I saw #1105 and #1107, which sound adjacent: iOS capture can stall while the app is still healthy, and the merged fix bounds capture stalls / preserves sessions. This report is specifically from a physical iPhone workflow on the latest published npm release (0.18.3), where #1107 does not appear to be released yet. If #1107 is expected to cover this once released, confirmation would be helpful.

Reproducible Demo

I do not have a minimal public repo yet, but the minimal shape should be:

1. Create an Expo React Native dev-client app.
2. Add a screen with a bottom-sheet / overlay style component.
3. Put a text input in that overlay and focus it so the iOS keyboard is visible.
4. Optionally add a second overlay with a custom numeric keypad and a nonessential infinite Reanimated animation, such as a blinking caret.
5. Launch on a physical iPhone through agent-device.
6. Open the overlay and focus the input.
7. Run agent-device snapshot -i, type, fill, and keyboard dismiss against the focused state.

Expected:

• Snapshot/text-entry/keyboard commands either work, or fail quickly with actionable state while preserving the session.

Actual:

• Screenshots and coordinate taps still show the app is alive, but XCTest-backed commands can time out with main thread execution timed out.

Current workaround:

• Avoid repeated idle-dependent commands once the timeout appears.
• Capture screenshot evidence as visual truth.
• Press visible keyboard/keypad coordinates directly only to prove user-visible behavior.

Question:

• Is there a recommended agent-device pattern for focused React Native overlays on physical iOS devices, especially where XCTest waits for app idle but screenshots/taps still work?
• Should we expect fix: bound iOS capture stalls and make runner recovery session-preserving (#1105) #1107 to address this after the next npm release?
• Is there a way for agent-device to expose a faster "XCTest channel busy / app still alive" diagnostic or direct hint to use screenshots/coordinate presses in this state?

[thymikee]

Thanks for the detailed report. Triage: labeling this bug + ready-for-agent because it is specific enough to investigate and does not look like only a published-release gap.

Expected relation to #1107: that PR should improve snapshot stall containment and session preservation once released, but this report also includes type, fill, and keyboard dismiss timing out while screenshot and raw coordinate taps still work. That needs separate verification on current main, especially on a physical iPhone with a focused RN overlay input.

Suggested investigation path: compare 0.18.3 vs current main, reproduce with a focused bottom-sheet text input and optional animated custom keypad, capture daemon/runner logs around main thread execution timed out, and decide whether agent-device should surface a faster XCTest-channel-busy/app-still-alive hint that steers agents toward screenshot plus coordinate evidence when appropriate.

[thymikee]

fixed in 0.19.0