Our Android test kept retrying a tap on a “More” button that hadn’t moved. Only the animation behind it had changed. I help maintain SoloPi, and we changed what it checks just before tapping: let the background move, but stop if the button moves or an overlay covers it. On a Pixel 8, median runtime for this menu flow fell from 13.21s to 10.00s. All ten runs across five baseline–update pairs passed verification.
The runner targets the three dots in the upper right. The energy bubbles below animate independently. AI-localized illustration from a device capture.
This was in Ant Forest, a mini-app inside Alipay. We wanted to open its menu, visit reminder settings, and get back to the forest. There were two other ways this short trip could go wrong. The menu opened in a window only 508 pixels tall, small enough for our old window filter to discard it. And if a tap went through but the next screen loaded slowly, sending another tap could hit whatever had just appeared under the same spot.
The menu covers only part of the forest. SoloPi’s old size filter could discard this window even while it was open. English illustration adapted from the original device capture with AI.
We borrowed an idea from Jev’s browser automation work: once you have a candidate action, how much of the page do you actually need to check again? On Android, for an eligible native button, SoloPi now refreshes the target and checks its identity, bounds, window and nearby or overlapping nodes just before tapping. The bubbles can move. If the button moves or the accessibility tree shows an overlay covering it, the runner rejects the tap. Lists, scroll containers and WebViews still get the full-page check. We also use window ownership and focus to recognize small menus, and keep observing after a dispatched tap while a transition settles.
Did that save time on a real phone? On a Pixel 8 running Android 16, we alternated the baseline and the new runtime for five pairs per workflow. Both used the same deterministic action choices and the same independent checks. The timer started with the forest already open and included setup, verification, screenshots and cleanup.
All 20 runs passed verification. One result I found more interesting than the headline speedup: in a separate test where we changed only the waiting strategy, the settings interaction got 11.4% slower. These two workflows benefited from the combination. The phone runs used rules so that model variability stayed out of the timing comparison.
For those running Android UI tests: what change on screen makes you throw away a previously located target? I’m especially curious how you handle a small popup that appears between locating a button and tapping it.
Source code · SoloPi verification guide · Jev runtime that inspired the experiment