Five days of silence: a launchd job that only failed when a human ran it

Exit 127 at your own prompt does not read as a defect in the script. It reads as a typo. That is exactly the reflex the author describes: you assume you invoked it wrong, and you move on ^[5].

The scheduler's thin environment was already understood in that codebase, just not all of it. Phase 1 of the 528-line, 12-phase generator explicitly extends PATH with nvm and homebrew before doing any work ^[12][15], which is the code you write after launchd has failed to find a binary on you once. A missing LANG ^[1] is the same family of omission with none of the noise. Nothing is absent from disk. The scheduled process simply gets a different locale than the shell you tested in ^[2].

Then there is where the line sat. It was inside the failure notifier ^[6], and a notifier only executes after something else has already broken, so a run of green days tells you nothing at all about whether it works ^[13].

One seam in the writeup is worth naming. It reports the C-locale scheduled run passing and the interactive run dying ^[3][4], but the five-day outage requires the notification line to have died where launchd ran it, not at a terminal ^[6][7]. The post does not reconcile those two, and it does not show the failing line, so why a UTF-8 locale is the fatal one and C is not stays outside the evidence ^[19].

The monitoring conclusion survives either reading. The design's own definition of success is whether a stock file landed in content/article, deliberately cut loose from the Zenn deploy so a jammed publish cannot stop generation ^[10]. That part worked. The two daily slots kept firing across the window ^[9], roughly ten scheduled invocations ^[18], and every one of them was correct to say it had done its job ^[3]. The signal was accurate and had nothing to do with what had stopped.

The check that would have caught it is an alarm on absence rather than on error: nothing posted since yesterday, or a marker file older than some threshold. Against a stated target of 30 articles a month, one a day ^[11], five quiet days is about five posts not shipped ^[17], on the pipeline the author says sits under ¥1.2M a month of revenue ^[20]. Error alerts only cover failures that manage to reach the notifier intact. Everything upstream of it, including the notifier itself, needs the other kind.

The blast radius here is a property of writing style, not of the bug. Any script that logs or notifies in Japanese under launchd could stand on the same mine ^[14], which is how a single assumption about the environment ends up sitting in five repositories at once ^[8]. The author's own summary of the last six months is the useful part: a system begins rotting the moment it is built, and it does not tell you ^[16].