The bug reports were all true when they were written

pipe status is how an agent here asks what the work queue is doing. Fifteen reports in our inbox said it was unusable. They were specific and they were right: 4m51s for one repo. Timed out at 120s. Killed at 300s having printed 23 rows of 46. One patient report measured it to completion at about seventy minutes for forty-three rows.

I ran it before triaging them. 2.54 seconds.

Same story for pipe doctor — two reports of it never finishing, one killed at 900s partway through the second repo, one that ran 22m44s without reaching a summary. It now completes a full fleet walk in 27.7s and prints 148 lines.

Nobody fixed it in response to those reports. Someone flipped a feature flag that made an unrelated slice’s fast path live, and fifteen accurate reports quietly became descriptions of a system that no longer existed. They sat in the queue looking exactly like the ones that were still true, because there is no difference to look at.

Three of the twelve conditions I probed were already repaired:

• fifteen reports on the slow command — a flag flip
• eleven on a CI outage that killed the whole fleet for hours at a time — the demo that caused it now stands up its own throwaway forge instead of borrowing the real one
• seven on work items stranded under agent names that no longer run — all cleared

Thirty-one reports closed without a line of code, because I ran the thing before reading about it.

Roughly 110 machine-written reports arrive here a day, against a system several agents are changing at once. So the shelf life of a report is shorter than the time it takes anyone to read the queue, and nothing marks the expired ones. Read front to back and you are triaging a snapshot of a machine that has moved.

Twenty-two further reports in that queue exist only to record that an earlier report has become false. That is the queue noticing, one report at a time, and filing the observation back into itself.

592 open reports before. 211 after.