The part that splits a big request into small ones had never runThe part that splits a big request into small ones had never run

August 15, 2026 · The role that exists to turn one request into a graph of small tickets had never once been used across 176 issues, and six separate bugs explain why.The part of the system meant to break a big ask into small pieces had never been used once, and six separate faults explain why.
last updated

One request was structurally always one ticket

I can ask this system for something out loud and it files a ticket. What it could not do, until today, is file two. There was no cardinality decision anywhere in the path (the classifier returned one object, the dispatcher built one input, one issue got created), so "build me a live board of what every agent is working on" arrived as a single ticket of 18,695 characters. That is not a ticket, it is a design document with a status field.

The Architect is the role that exists to take a request like that and file it as a dependency graph of small ones. It had never once been assigned: 0 of 176 issues in the workspace carried its label, and every ticket tree that exists was filed by a hand-written script. Six commits explain why, and no one of them would have been enough alone.

The classifier now decides shape as well as role, so a capability routes to the Architect. The Architect's prompt told it its deliverable was a graph of issues and, forty lines later, told it not to create issues. That is now a named exception for the one role it contradicts, rather than a quiet rewording of a rule the other eight need. The follow-up schema kept title and description and discarded every other key, so a session that emitted a perfect tree had it flattened before anything could file it, and silently, because a discarded key looks exactly like one nobody wrote. Filed tickets got an assignee unconditionally while auto-chaining requires a null one, so every tree the system could file was abandoned at hop one.

The last one is the one I would not have guessed. The duplicate check re-read the open issues on every call, so sibling two was compared against sibling one: same capability, same words, filed seconds earlier. A duplicate verdict does not warn, it removes the node and every blocks edge touching it, and the tree files as a strict subset of itself with every log line reporting success.

One more from the same system, filed as a queue bug and not one: the queue serialises where a session starts, and nothing bound where it writes. A ticket routed to one project was about files that only exist in another repo, so it went where the files were and swept a concurrently running session's staged changes into its commit, under a stranger's message. The tree lock added today is keyed on the tree resolved at the moment of writing, which is the only key that would have caught it.

Mail actually arrives now

The site has handed out burner addresses at this domain for ten days, and until 21:22 tonight not one of them had ever received anything. The store had zero rows, which is indistinguishable from nobody having written to me. Three sessions went at the question today and all three came back with the same answer: the cutover had never been started, rather than started and broken. The MX records were still the registrar's forwarding set (byte for byte the ones captured as "before" ten days ago), the Cloudflare Worker had never been deployed, and the inbound secret was unset, so the endpoint answered 503 to everything that reached it.

None of those three is work an agent can do from a terminal: a DNS change, a dashboard click, a secret pasted into two places. So what the sessions actually produced was the sentence naming which step was missing and the click path to it, and I did them. A message from my own address landed in the store seven minutes later, DKIM and DMARC passing, and it renders on the page. The only commit is the deploy config, pinned so the next deploy parses mail the same way the verified one did.

One sentence, one enormous task

I can ask this system for something out loud and it writes down a task. Until today it wrote down exactly one, however large the ask. "Build me a live board of what every agent is working on" came back as a single task about the length of a short story, which nothing could then act on.

There is a role whose whole job is to take an ask like that and break it into small pieces with an order between them. It had never been used once in 176 tasks; everything that looked like a plan had been written by hand.

Six separate faults were in the way, and fixing any one alone would have changed nothing. The worst was a duplicate check: because the pieces of a plan get written seconds apart and look alike, it decided each piece duplicated the one before it and deleted them, along with the links that gave the plan its order, reporting success the whole way through.

The first real email arrived

The site has been handing out throwaway email addresses at my domain for ten days, and until tonight not one of them had ever received anything. An empty inbox looks exactly the same whether the plumbing is broken or nobody has written to you, which is the annoying part.

Three separate sessions went at it today and agreed: the switchover had never been started at all. Mail was still pointed at the old forwarding service, the piece that receives it had never been published, and the password the two halves share was missing.

None of that is work a background agent can do. It is a settings change, a click in a dashboard, and a secret only I can copy. What the sessions were good for was saying precisely which step was missing, and then I did them. A message I sent myself turned up in the inbox seven minutes later, and the page shows it.