Skip to content

Commit f807005

Browse files
committed
docs(evidence): record rebased scroll and cancellation verification
1 parent dd287d8 commit f807005

1 file changed

Lines changed: 18 additions & 0 deletions

File tree

‎docs/evidence/android-long-scroll-2026-10-19.md‎

Lines changed: 18 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -85,3 +85,21 @@ Apple, web and Linux local runtimes use `bindLocalScrollInteractor`. WebDriver a
8585
The transferable optimization is to let that bounded loop inspect its captures instead of adding a public snapshot/stabilization request after every page. The external adapter can adopt the existing shared API for eligible local or provider targets. Android's 300 ms pacing, idle handling and input transport are not a cross-platform default to copy. The rejected 150 ms experiment also shows why whole-search progress matters more than individual gesture latency.
8686

8787
The first hosted Coverage job failed an existing import-budget assertion before coverage thresholds ran: flag metadata pulled `scroll-command.ts` and its execution helpers into the eager flag-registry closure, increasing it from 30 to 31 modules. A type-only policy link now checks the static schema bound against `SCROLL_UNTIL_MAX_PASSES` without evaluating that module. The existing eager-import test reproduces the failure before the fix and passes afterward. No tests were added for this correction.
88+
89+
## Review follow-up, 2026-10-20
90+
91+
Production head `dd287d874a7a4bae1216dfdaccf280a0a33f23ce` is rebased onto `620228b8ca67a69a108fa4fc5041fdd88880b792`. The range comparison preserves all earlier changes. Its only conflict moved main's physical-iOS gesture guidance into the extracted help topic.
92+
93+
Cancellation now belongs to `runScrollPasses`, shared by target searches and edge scrolling. It checks before each observation and gesture, and after awaited work. Edge discovery retries and settling captures also check cancellation. The settling sleep wakes on abort. This preserves the request's typed cancellation reason and prevents another pass from starting.
94+
95+
The signal-consumer audit distinguishes context types. Only the scroll runtime newly reads the signal populated by `contextFromFlags`. `runtimeExecutionFromContext` excludes that field. Existing Apple runner/native-helper contexts get their signal from platform request bindings. Backend capture contexts get theirs from the command surface. Limrun allocation gets its separate `LeaseLifecycleContext.signal` from the lease handler. Maestro and replay contexts also have their own construction paths. These consumers do not acquire new cancellation behavior from this change. Native macOS helper shutdown already sends SIGTERM with a one-second grace so its handler can release a held button.
96+
97+
Live verification used the same Pixel_7_CI and APK hashes above. The emulator restored an older development-build snapshot; the checksum-verified Release fixture and benchmark APK were restored before verification. The PR's built SDK and isolated state directory `/private/tmp/agent-device-scroll-review-20261020` exercised this source; the Device panel's installed CLI predates the flag.
98+
99+
- Exact benchmark search: 55 gestures, then a fresh target press and exact `Found Row 0512` assertion. Search, snapshot and press took 28.229 s in one run. This is a smoke measurement, not a new controlled comparison.
100+
- Our Release app: both routes reached and opened `Seasonal Footer Pick` in seven gestures. A one-pass search failed with `scroll_until_pass_limit` and `passes: 1`.
101+
- Canceling target search and `scroll bottom` after 1.6 seconds returned `request_canceled`. Request logs show three gestures for the target search and one for the edge search before cancellation. Both accepted a readable follow-up snapshot; all sessions closed.
102+
- Existing checks passed: 916 focused scroll/eager-import cases; 83 shared-loop, macOS helper and Limrun cases; eight Apple runner cancellation cases. No tests were added or changed in this follow-up.
103+
- Independent Opus review found no defects in the rebase, cancellation, signal-consumer audit, API-neutral validation text, or the 90-second timeout guidance.
104+
105+
Raw outputs and request logs remain in the task's temporary artifacts. The earlier controlled ABBA measurements remain attributed to their original heads above.

0 commit comments

Comments
 (0)