The Queue That Was Empty For Twenty-One Days
A triage queue reported nothing to do for three weeks. The generator was reading the wrong column after the table gained two new ones.
A triage script reported an empty queue for 21 days. Two columns had been added to the source table and the script still read by position, so it took the owner column and compared it against status values — never matching. Nothing errored, because a positional read of a wider table is still a valid read.
Symptom
A weekly triage script reported an empty queue. No items needed review. This continued for 21 days, which read as a quiet period rather than a fault — queues do empty sometimes.
They had not. When the table was finally opened by hand it held 20 open items against a tripwire threshold of 10, several of them more than two weeks old.
Root cause
The source is a markdown table. The script read fields by position.
Two weeks earlier the table gained two new columns in the middle. Every field after the insertion point shifted right by two. The script kept reading the same index, which now held the owner field, and compared its contents against a list of valid status values.
Nothing matched. An empty result set is a legal, unremarkable output, so the script returned it without complaint.
A second defect surfaced in the same pass: cells containing an escaped pipe character shifted their own row's columns independently of the header, which had been silently dropping one specific record from every run for longer than anyone could date.
Cost
Three weeks of a governance loop that everyone believed was running. The items did not disappear; they aged, and the ones that mattered had to be re-triaged from scratch against a context that had moved on.
The more expensive part is trust. Once a scheduled report is known to have lied, every previous quiet week becomes suspect, and there is no way to reconstruct which of them were genuine.
The failure class: a fix bound to the format of a file it does not own comes silently unbound when that file is refactored.
Fix
Two changes, both small.
The script now reads by header name, not by position. Adding, removing or reordering columns cannot move a field out from under it.
And it fails loudly. An empty queue is now distinguished from a queue it could not parse:
queue: 0 items (parsed 24 rows, 9 columns, header matched) — versus a hard error if the expected headers are absent.The escaped-pipe defect was fixed in the same pass by splitting on unescaped delimiters only.
Prevention
The generalisable rule is about empty results, not about markdown. An empty output is the most dangerous shape a report can take, because it is indistinguishable from success and it is what almost every parsing bug produces.
So: any scheduled check that can return zero must also report what it examined. Row count, column count, whether the schema it expected was found. A zero with provenance is information; a bare zero is an assumption wearing a number.
And where one program reads a file owned by another, the coupling is on names, never on positions — position is a promise the other file never made.