Deep Research keeps going through server updates
A research run no longer starts over when we update the server. It hands over to the new worker and continues at the step where it stopped, without paying for finished steps again.
If you run Deep Research, a server update no longer throws your run back to the start or bills finished steps twice. No action needed.
Background jobs run in a separate worker. Every server update restarts it. Until today a Deep Research run that was busy at that moment began again from the plan, and on a day with many updates it never reached its report.
Text version
Old worker: stops at once, the finished steps are saved, the job goes back to the queue. Saved steps: plan, searches, pages, extractions and drafts, each billed once. New worker: replays the saved steps without new calls and continues with the step that was open.
What was wrong
Two faults added up. The worker was meant to let running jobs finish for up to five minutes, but a watchdog misread the quiet shutdown as a hang and ended the process after one minute. And the research job kept no record of its steps, so the next worker could only start over. Each restart also bought the finished model calls again.
What changed
Every finished step of a research run is now stored with its result. When an update begins, the run stops right away and returns to the queue. The next worker replays the stored steps without calling a model and continues live at the first open step. The cost of a step is recorded together with its result, so it is billed exactly once.
Runs that are interrupted again and again without progress now end with a clear message instead of looping. The watchdog no longer ends a healthy shutdown.
Other job types keep their existing checkpoints. We are moving them to the same hand-over next. Prices and plans did not change.