Persistence Beyond the Tab
Server ownership and stored progress let work outlive the browser that requested it.
Last verified against production: . See current limits and availability.
A build has an owner
Submitting a build records it in Postgres. An independent worker claims accepted jobs within the platform's two execution slots, with one build owning a project at a time. The web process reads the job and its progress, so closing the browser or restarting the web application does not stop the worker's execution.
Queued builds show position and allow Cancel or Edit before ownership is claimed. A running build shows saved steps and an activity indicator. Reopening a project reads its saved events; a dropped connection resumes from the last received event. The browser observes the existing run.
Stopping and recovering
Stop records a cancellation request, revokes access to the provider proxy and asks the execution owner to end the build's containers. The console remains Stopping while cleanup is being confirmed. Saved source and an earlier publication remain available.
The worker records source checkpoints and the preparation and outcome of deployment or publication. If it disappears with an uncertain external action, the run is held for reconciliation. An operator checks the recorded transaction or publication identity before resolving that state. A paused run is shown as paused; it does not keep claiming to work.
One agent turn currently has a 45-minute timeout. A build can include several stages and repair turns, but the release has no five-hour continuous-run acceptance. The limits and roadmap pages explain the remaining recovery and steering work.
Watch has a checkpoint and an outbox
The Watch daemon reads logs from its stored block checkpoint, matches enabled conditions and commits delivery records. Delivery is a separate loop with deduplication and bounded retries. A temporary Telegram or webhook error leaves a recorded status that can be inspected in Alerts.
Watch checks the checkpoint block hash. A replacement hash triggers a rewind of up to 64 blocks and retracts pending deliveries above that point before rescanning. Already delivered messages remain in the recipient's chat. These observations are sequencer-confirmed and can be replaced before finality.
Registry jobs have their own checkpoints and rollback behavior. The holder ledger uses a 300-block delay and logs deeper reorganizations for investigation; it does not share Watch's pending-delivery rollback. Check the job's coverage when using holder balances.