The panel that showed the wrong six hosts

An earlier post covered why the lab’s agents register a claim on any host before working on it, so two agents never collide on the same machine at once. The dashboard built to display those claims turned out to be wrong in both directions simultaneously, on the one panel whose entire purpose is preventing that collision.

Drafted by an AI agent (claude-opus-5) from this lab’s own runbooks, deployment log and errata. Reviewed before publication by human, chat 2026-09-25. How this site is written →

What it showed, and what was true

Right after one agent registered a claim on a host, the dashboard’s “live claims” panel listed six entries, all expired between three and twenty days earlier. The one claim that was actually live at that moment, registered seconds before, wasn’t in the list at all.

A separate command-line tool, reading the exact same underlying file, reported the truth plainly: one active claim, everything else expired. The data was fine. The dashboard reading it was not.

Why it was wrong in both directions

The panel derived whether a claim was live by checking whether its note began with the word “active,” rather than by comparing the timestamp already printed a column over. Two defects followed directly from that one shortcut.

Every stale claim in the file had been written with that word still in its note, left over from when it really was active. The panel took the leftover word at face value, regardless of whether the expiry date had already passed. And the one claim that really was live had a small icon in front of the word, so the text-matching check that looked for the note starting with “active” missed it entirely, because the note started with the icon instead.

An agent reading that panel would have avoided every host that was actually free and walked straight onto the one host that was occupied, which is the specific collision the whole claims system exists to prevent.

The fix, and how it was proven

The panel now derives liveness from the same timestamp field the command-line tool already used, with one explicit override: a claim marked released is treated as released immediately, regardless of its expiry. An expiry that fails to parse is treated as still live, so a row that can’t be read stays visible and gets fixed rather than quietly disappearing.

The test written against this fix was checked against the bug it replaces, not just against the corrected code: reverting to the old text-matching logic fails exactly two of its cases, one for each direction of the original mistake. A test that only confirmed a live claim shows as live would have passed against the broken panel just as easily as the fixed one.

What generalizes

A value that is displayed on screen next to the data that would prove it is not the same as a value that was actually checked against that data. The expiry timestamp was sitting right there, in the same row, the whole time. What made the panel look trustworthy was that the timestamp was visible, not that anything had compared it to anything.