Help centre · Troubleshooting

What should I do when a run ends in failure?

Read the error category, understand what it means, and know whether relaunching helps.

Open the run in Runs: the error category is shown at the top, and it is what tells you what to do. A failure is not necessarily a breakdown — most often the pipeline refused to deliver work it judged insufficient.

What do the error categories mean?

Check not passed — the produced code did not reach the minimum score, even after the correction attempts. The merge request exists, as a draft, with the objections as comments. Read them: often the issue was ambiguous, sometimes an internal convention is missing from your contexts.

Unanswered question — the reading agent needed a detail and nobody answered. See Diavi asked me a question.

Human intervention needed — the pipeline hit something it cannot decide alone. The merge request holds what could be done, and the comment says where it is stuck.

Response too long — the model hit its output ceiling mid-work. This happens on issues touching many files at once. Split the issue in two and launch again.

Infrastructure — an incident on our side. That run is not counted against your allowance. Launch again; if it recurs, write to us.

Should I relaunch?

Relaunching unchanged only helps if the failure was an infrastructure error. In every other case, change something first: sharpen the issue, split it, or add the missing convention to your contexts. Without a change, the second run meets the same wall as the first — and counts against your allowance.

Where do I see the detail?

The Logs tab shows the full trace, step by step. A glossary there explains the recurring terms.