VM and runner lifecycle

FARM watches Forgejo for waiting jobs and creates Incus VMs to provide capacity. Each VM registers one ephemeral runner. That runner accepts at most one job; FARM then removes the runner registration and VM.

flowchart LR
    A[Waiting job] --> B[VM startup]
    B --> C[Runner ready]
    C --> D[One job]
    D --> E[Runner and VM removed]

With runner_installation: controller, FARM configures the VM through cloud-init. With runner_installation: image, the image supplies the runner service and FARM starts it after writing the runtime configuration. Both modes need the Incus agent to deliver credentials to the VM.

FARM tracks lifecycle state in SQLite. Keep each controller ID and database assigned to one running FARM process. Reconciliation detects and cleans up stale FARM runners and VMs after interruptions.

See pool matching and scaling for how jobs trigger new VMs.