AppControlBench element-18 (gpt_high): reaching "Manage sessions" in Element's settings table cost 433s / 80 tool calls, ~40 of which were scroll attempts that all REPORTED SUCCESS while doing nothing:
scroll down 1 ×3, scroll up 1/3, scroll up --pixels 1200/2200/3000, gesture pan … ×2, scroll top, scroll down --pixels 1400/600, gesture fling down/up … — every one returned success ("Scrolled up by 1200px", "Flung down"), and the follow-up snapshot stayed byte-identical (Snapshot: 20 visible nodes (116 total) repeated ~15×). Only a raw swipe 200 250 200 700 finally moved the list.
Ask: no-op detection. Compare pre/post evidence (scroll offset, visible-node digest — the settle digest helper already exists) and either escalate to the drag-based path automatically or return an explicit "scroll had no effect" error with the working alternative in the hint. Silent false success is what turned one stuck list into 40 wasted calls.
Related context: iOS scroll oscillation vs gesture pan landing behavior was previously characterized in #1366's hint work; Android already exposes authoritative can-scroll-* attrs, iOS has no equivalent signal.
Transcript: AppControlBench gpt_high__agent-device/element-18/.
AppControlBench element-18 (gpt_high): reaching "Manage sessions" in Element's settings table cost 433s / 80 tool calls, ~40 of which were scroll attempts that all REPORTED SUCCESS while doing nothing:
scroll down 1×3,scroll up 1/3,scroll up --pixels 1200/2200/3000,gesture pan …×2,scroll top,scroll down --pixels 1400/600,gesture fling down/up …— every one returned success ("Scrolled up by 1200px", "Flung down"), and the follow-up snapshot stayed byte-identical (Snapshot: 20 visible nodes (116 total)repeated ~15×). Only a rawswipe 200 250 200 700finally moved the list.Ask: no-op detection. Compare pre/post evidence (scroll offset, visible-node digest — the settle digest helper already exists) and either escalate to the drag-based path automatically or return an explicit "scroll had no effect" error with the working alternative in the hint. Silent false success is what turned one stuck list into 40 wasted calls.
Related context: iOS
scrolloscillation vsgesture panlanding behavior was previously characterized in #1366's hint work; Android already exposes authoritative can-scroll-* attrs, iOS has no equivalent signal.Transcript: AppControlBench
gpt_high__agent-device/element-18/.