Skip to content

Something went wrong

The four things that go wrong most often, what each one actually is, and the page that fixes it — start here if you have a symptom rather than a question.

Start here if you have a symptom rather than a question. Each entry below names what the problem usually turns out to be and sends you to the page that covers it properly.

The most common report, and it has four different causes that look identical from the outside: a fixed per-turn cost ceiling, a five-minute clock, a step count, and stopping to ask you something. Only one of those is a fault, and the fix is different for each.

It stopped before it finished — how to tell which one you hit, and what to do about each.

Almost never one expensive thing. The usual causes are the tier the run was on, how many turns the conversation took, and a per-turn ceiling that is not the number you typed.

It cost more than I expected — a six-item self-check, in the order worth doing it.

The result is wrong, or it made something up

Section titled “The result is wrong, or it made something up”

A result that reads confidently is not the same as a result that holds up. There are specific signals worth checking, and they are quicker than re-reading the whole thing.

The result looks wrong — the signals that separate a checkable claim from an invented one.

I connected an app but it isn’t using it

Section titled “I connected an app but it isn’t using it”

This one is a real trap rather than a misunderstanding. Connecting an app does not put it into a run, and the home screen cannot add one — every app there reads Connect in Apps whether you have connected it or not.

A connected app isn’t being used — why, and the one place it does work.

  • A run failed and you were still chargedWhen a run fails.
  • A task is sitting there doing nothingWhen it takes too long covers stalled, paused and Needs you, which look alike and are not.
  • You want to stop something in flightStopping a run.
  • It refused to do somethingWhat it won’t do separates the refusals that are enforced from the ones it merely follows.
  • A top-up failedTopping up lists what each failure message means.
  • You are not sure what a receipt field meansReading a receipt.
  • You are not sure what a word in the interface means — the Glossary defines each of Else’s terms in one line, including the words it deliberately does not use.
  • It behaved differently the last time you used itUpdates is the weekly record of what changed, and a repair is named there as a repair.

If none of those match, the receipt and the run’s own step list are the record of what happened. They are accurate even when the summary at the end of a run is not.