Work now passes between nine roles without coming back to meWork now passes between nine roles without coming back to me
August 17, 2026 · Work moves between nine roles now without coming back through me, and the one signal meant to vouch for a human was giving two obvious bots full marks.Work can now pass between nine kinds of helper without returning to me, and the check meant to spot real people was giving two obvious bots full marks.
last updated
Nine roles, and a day of watching them run
The fleet that picks up tickets was rebuilt around nine roles: triage, delivery, auditor, release, publisher, operations, architect, scout, and a registrar that is code rather than a session. Each has a Linear identity of its own, and a dispatched session now takes its charter from the identity the delegation arrived on rather than from a label on the ticket. Those were two separate facts before, and nothing compared them: an issue could carry one role's label while being delegated to a different actor, so the session read one thing and the board showed another.
On top of that sits the handoff. A session can end by naming the role the rest of the work belongs to, and the issue is re-dispatched under that role's charter without passing through me. Depth is capped at three hops and a cycle guard permits exactly one bounce. Delivery to auditor and back is a correction, a second lap is refused. Chained work is routed by triage now rather than by a static table, because by the time a blocker finishes, the evidence may have overtaken the plan that named its successor.
Seventy-three unattended sessions ran on the day: 26 triage, 13 delivery, 8 auditor, 2 release, 2 scout, 1 operations. The router ran twice as many sessions as the thing that writes code, which is the cost of deciding late and I would rather pay it than route on a guess. The best moment was in the vintage map: delivery parked a spend question on a finished fix, and the auditor refused it outright: gating reads on a suspect verdict would let one mis-compared row block every legitimate read for a month, an unbounded stall bought with a bounded overspend.
It broke where you would expect. Recording a finished session asserted that no other run was open, which had been true until handoffs started dispatching one; the fallback then wrote down the post-run assignee (me) seconds after custody had legitimately moved. Six issues came back to my name. The sweep meant to repair that made it worse: it stamped the correction from the issue's updated time, a few seconds behind the bad row, so the fix sorted underneath the thing it was fixing and every later attempt was swallowed as a duplicate. Zero corrections, a green health check, a board that never moved.
There is also now an operations sweep, which is the first source of work here that does not begin with a person. It found three scheduled jobs tracked and never loaded (one of them the only thing enforcing the retention window the mail tool promises its users), and nothing had ever filed a ticket about them, because nothing was looking.
Bot Detector was confident about things it could not know
One signal exists specifically as positive evidence of a person: does this account ask questions. It tested for a question mark in the raw comment body, so every ?context=3 in an account's own boilerplate read as curiosity. Live yesterday: RemindMeBot asked a question in 295 of its last 299 comments, RepostSleuthBot in 299 of 299. Links are stripped before the test now, and both score zero; link text is deliberately kept, because “does anyone know?” is a question its author wrote.
The bigger one was quieter. Roughly one author in five of a live thread got a grey “no data” badge blaming deletion, suspension or a typo, while the comment endpoint served their history perfectly well. The account index it consulted turns out to be a frozen March 2025 snapshot rather than a lagging one, so the blind spot is exactly the newest accounts, the ones most worth checking. A miss now builds a profile from the comment and post streams instead, with karma left unknown rather than zero.
Both were live under a fully green suite. They were found by pointing the tool at real accounts, which is also how TikTok and X were ruled out: X's one open route returns samples up to 554 days stale, and where signals fired they fired on artefacts: a dormancy flag for @POTUS off a pinned tweet.
Passing work along without me
The helpers that pick up my tickets were rebuilt around nine roles: one sorts, one builds, one checks the work, one publishes, one watches the machines. Each has its own identity now, so the board can show who is actually holding a job. A role used to be a sticker on a ticket and nothing more.
The new trick is that a role can finish by handing work to another role rather than handing it back to me. There is a limit of three hops, so two of them cannot play catch forever. Seventy-three of these sessions ran during the day, and the one whose whole job is deciding who does what ran twice as often as the one that writes code.
It broke the first time a ticket walked the whole chain. The step that records “this is done” assumed nothing else was running, which stopped being true the moment handing off existed, so six tickets landed back on my name seconds after they had properly moved on. The repair job was the funnier half: its correction was stamped slightly earlier than the mistake, so it filed underneath it and was ignored as a duplicate. Nothing moved, and everything reported itself perfectly healthy.
Sure about things it could not know
Bot Detector guesses how automated a Reddit account is. One check asks whether an account ever asks questions, on the theory that curiosity is a human thing. It was counting the question marks buried inside web links, so two well-known bots “asked a question” in 295 and 299 of their last 299 comments. The one check meant to vouch for humans was handing bots full marks.
The quieter problem: about one person in five in a real thread was reported as having no data at all, blamed on a deleted or misspelled account, when their comments were sitting right there. The list it checked against stopped being updated in March last year, so the people it could not see were precisely the newest ones. It works them out from their comments now.