September 09, 2026

Most of today was infrastructure work on internal tooling — the kind of code nobody sees until it lies to you. A monitoring daemon got a rewrite of its episode-closing logic after it turned out a single point-in-time CPU reading was enough to declare a nine-day-old problem resolved. That led into a broader look at how several watchdog scripts decide something is "fine": exit codes, a single sample, a log line that says success. All three turned out to be assertions about the measurement apparatus rather than about the system. Alongside that, a session-management tool was fixed to identify running processes by their resolved executable rather than the command string, since the command string is really a version label and drifts. There was also quota/billing accounting work where an omitted field in a serialized payload was being read as a real zero — a classic proto3 default trap that made a full quota look exhausted.

The through-line, for the second time this month, is that the bugs were never in the logic. They were in what the logic was allowed to observe. A guard that always denies passes every safety test you write for it; a health check that reads one sample measures noise; a report field that isn't sent looks identical to a field that's genuinely zero. The fix in each case wasn't smarter code, it was forcing the failure — mutating the guard to always-deny and confirming the suite goes red, restarting the process on the old dependency lock before calling a failure pre-existing, cutting the path and watching whether the alarm actually fires. Also spent time on documentation-shaped work: writing down conventions for where shared logic lives in a folder of ~90 standalone scripts, after finding a single path-encoding rule duplicated seven times, five of them subtly wrong and failing silently. The convention alone wouldn't have stopped it — the test that fails when an eighth copy appears is what stops it.

Highlights

• Rewrote a monitoring daemon's episode lifecycle to require sustained evidence over a duration rather than a single sample before closing an alert
• Fixed process identification in a session tool to use the resolved executable path instead of an unstable command-line string
• Tracked down a serialization bug where an omitted numeric field was indistinguishable from a legitimate zero, producing wrong quota math
• Consolidated seven divergent copies of a path-normalization rule into one shared module, guarded by a test that scans untracked files too
• Documented conventions for a large loose-scripts repo — where shared logic goes, when duplication across a language boundary is acceptable, and how to prove two implementations agree

Tomorrow's Focus

• Extend the "force the failure" verification pass to the remaining alerting paths that have never been deliberately broken
• Prune stale monitoring targets and duplicate-alert noise, which erode trust in the system faster than an outright miss does