Obeya's coordinator: three requests it got wrong now and then
Live runs of the coordinator against the real model, each request asked many times over. Before: the prompt on main. After: this change.
A chart of what each task cost
“Zeig mir für jede Aufgabe, wie viele Tokens sie gekostet hat, als Diagramm”
“Obeya doesn't record how many tokens a task costs, so I can't show you a chart. Use the link to send this to Obeya's makers as a wish.”
“I'm going through the agents' session logs and will show you the tokens per task as a chart.”
Each session's tokens are in the agents' session logs on the machine, so an agent can work this out. The coordinator now offers a wish to Obeya's makers only for things that neither Obeya nor an agent can do.
Start one task once another is done
“Starte Login, sobald Export als CSV fertig ist”
It waited on “Login” instead of “Export als CSV”. Later: “My mistake: I was waiting on Login instead of Export als CSV, which is still running.”
Waits on “Export als CSV”. Once it is done: “Export als CSV is done, Login starts now.”
The coordinator could name both cards in one wait, and it was not clear which one it waited on. Now it names only the card it waits on, and if it names two, it is asked to fix that before anything happens.
A new repository, then a canvas for it
“Hmm, with my permission you should be able to do that, no? Create a card and let the agent try it”
“New task ‘Repo anlegen’, the agent starts now. I'll add the canvas once the task is done.” But it never set itself to come back, so the canvas never came.
New task, and a wait on it. Once the repository exists, it adds the canvas by itself.
Coming back to a step later now has to be set explicitly whenever the reply promises it. A folder the coordinator gave a new task, which went nowhere before, is now pointed out to it. The test now uses a fresh folder: the old example pointed at a real project folder on this machine, and seeing that, the coordinator rightly asked whether to use it.
All seven live tests of the coordinator passed in three full runs in a row. CI does not run these tests (they need a Claude login).