How jobs work

Everything Talos does is a queued, retryable job with live output.

There is no direct execution from an API endpoint. A request creates a job, the job is queued, and a worker picks it up and runs it against your infrastructure over SSH, Ansible, Terraform or a provider API. The output streams back while it runs and is kept afterwards.

This is why the audit trail is complete rather than best-effort: the record and the execution are the same object. What you watched is what is stored.

Retries

A failed job is retryable on its own. In a multi-stage pipeline this matters more than it sounds — a four-stage pipeline does not restart because stage three timed out, and a batch across forty benches re-runs the three that failed rather than all forty.

Gates

A job can be configured to pause for approval. The pause happens at the stage itself with the computed change already in hand, so the approver sees what will actually happen. The approver must be a different person and must independently hold the permission for the action.