Skip to content

Resume after an interruption

An agent can stop before the end. Nothing is restarted without you. This page says how to restart without redoing, or paying for, the same work twice.

What you see What happened What to do
“Échec : …” (Failure: …) An error interrupted the work Read the message, fix the cause, then Reprendre (Resume)
“Limite de travail atteinte.” (Work limit reached.) The agent used up its budget of steps Check the result, then Reprendre
“Interrompue au redémarrage. Vérifiez les actions dont le résultat est inconnu ; aucune action n’a été rejouée.” (Interrupted at restart. Check the actions whose result is unknown; no action was replayed.) The application closed during the work Inspect, then Reprendre
Réponse attendue (Reply expected) status The agent is asking you a question Reply in the conversation
Annulé (Cancelled) status You stopped or refused an action Reprendre is not offered: send a new message

Reprendre is above the message box of the conversation and in the Activité (Activity) panel.

  1. Read the failure message and fix what it points to: connection, quota, CLI version.
  2. Deal with the actions with an unknown result, if there are any. The agent can also do this itself when it restarts.
  3. Click Reprendre.

A new run restarts from the same request, with the same provider, the same model and the same mode. It receives the list of the recorded actions, and the agent is instructed not to redo them. The To-Do card is reused if it still exists. A Meshy or Tripo creation that has already been accepted is not submitted again: the agent retrieves its result.

Reprendre is no longer possible once another request has followed: “Une autre demande a déjà suivi cette exécution.” (Another request has already followed this run.) To abandon a paused run, use Clore la demande (Close the request) in Activité.

Every run is bounded. A pass counts at most the steps of the profile: 24 by default, adjustable from 1 to 30 under Étapes maximum (Maximum steps). The orchestrator gets 2 more steps per authorised delegation, a delegated specialist gets 8.

When a pass ends with work remaining, the next one starts by itself if the previous one made progress: “Poursuite automatique (passe 2/5).” (Automatic continuation (pass 2/5).) After 5 passes, or with no progress, the agent summarises and stops. A limit being reached is not a success: the task remains to be examined, with the notification “… : limite de travail atteinte. Vérifiez le résultat, puis poursuivez.” (…: work limit reached. Check the result, then continue.)

At the next launch:

  • a run in progress becomes a failure;
  • a change in progress becomes Résultat inconnu (Unknown result): it may have taken place;
  • a simple read is cancelled: it blocks nothing;
  • nothing is replayed automatically.

After a crash, Island Pilot also offers to send a report. See Logs and diagnostics.

During the work, a disconnection can produce these messages:

Message Advice
“Le CLI a dépassé dix minutes. Aucun appel n’est relancé automatiquement.” (The CLI went over ten minutes. No call is restarted automatically.) A step took too long. Check the CLI, then Reprendre
“Le CLI a interrompu la demande. Vérifiez sa connexion, le modèle et le quota de votre abonnement.” (The CLI interrupted the request. Check its connection, the model and your subscription’s quota.) Check these three points, then Reprendre
“La réponse du service a été interrompue. Une génération peut déjà avoir été acceptée ; vérifiez avant de relancer.” (The service’s reply was interrupted. A generation may already have been accepted; check before running it again.) Look in your account with the service before running it again

If the orchestrator is waiting for a specialist that was interrupted after your reply, Reprendre appears next to that specialist’s name. The orchestrator then restarts by itself.