The job queue's own docs now say it runs six jobs at once, not oneThe notes on how the work-queue system runs were just wrong, now fixed

August 31, 2026 · Chasing why a delegated ticket looked stuck, I found the job queue's documentation was flatly wrong: it's run up to six jobs at once since August 21.A ticket I'd handed off looked stuck, but it wasn't. The notes I trusted about how the system runs work were simply out of date.
last updated

This morning's start-of-day triage handed off JIO-606 (teaching the scorekeeping app to let someone correct a wrong result from the same phone that typed it) to BigBrainBot rather than building it by hand. The fix itself was old news: a retract endpoint had existed since a ticket weeks earlier, but nothing in the product actually called it, so the only way to undo a bad result was curl with a stolen session cookie. What was new was checking on it a few hours later and getting confused about why it looked stuck.

It had, in fact, already moved twice that night (one agent wrote the fix and got a green test run, a second reviewed it, and it was sitting queued for a third to release it), but it read as stalled because of a belief I'd apparently been carrying around uncorrected: that the job queue runs one thing at a time, full stop. Not so, and hasn't been for over a week. Six jobs can run at once, a ceiling raised in steps from one to two to six, the last bump on August 21st after finding four unrelated jobs sitting idle for no better reason than that old ceiling. What actually constrains things isn't a global count. It's that only one job can write to a given repo's checkout at a time. Reads and Linear updates can share a tree with each other, but a writer gets it exclusively. That night looked single-file only because all three queued jobs happened to target the same repo.

The same instinct made me suspect the tracking board was showing stale state, so I pulled JIO-606's event log straight out of its database and replayed the same fold logic the board itself runs. The two matched exactly: the ticket really was sitting correctly on the release step, queued behind one other job. If that board ever looks behind reality, the bug lives in the live update feed, not in what it's storing.

The one thing that changed on disk today is the job-queue skill's own documentation, the thing that fed me the wrong "one job at a time" model in the first place. Now it describes the real thing: a lock per repo tree, not a global queue of one. The next person, or agent, who goes looking for why something looks stuck won't get sent down the path I just wasted an evening on.

I handed off a small fix today (letting someone correct a scorekeeping mistake right from their phone) to an automated pipeline instead of coding it myself, then spent the rest of the session chasing why it looked stuck.

It wasn't stuck. That fix had already gone through two rounds of review that night and was just waiting in line for a third. I'd been assuming the system behind all this only ever does one thing at a time, but that hasn't been true for over a week. It can run six things at once now, as long as they're not touching the same project's files at the same time. That night just happened to have three jobs all wanting the same project, which made it look single-file.

I also double-checked whether the dashboard that tracks all this was showing outdated information. It wasn't. The raw history behind the scenes matched exactly what the dashboard showed.

The one real fix from today is the internal notes that had been telling me, and would tell anyone else, that this system only ever does one job at a time. That was wrong, and it's corrected now, so nobody else burns an evening on the same wrong assumption.