Every live instance across every user. Restart and delete go through the admin-vm-action edge function and are applied to the real owner’s container.

Every signed-up user from auth.users, via the admin-users edge function. ssh key = at least one row in ssh_keys; payment = active billing context has a Stripe subscription (solo user_billing or team team_billing).

Every pi-agent the routing layer can target. active hosts are eligible for new launches; toggling off keeps the host’s existing VMs alive but stops the picker from sending it new work.

Agent liveness + subsystem health for each Spark host (from host_status / host_health_history via admin_host_health). One lane per host and subsystem, one square per hour — the hour’s worst state. A silent agent shows as agent unreachable.

Open waitlist entries awaiting capacity — fifo per gpu type (same order as drain-waitlist).

pre-bookings

Open capacity reservations from /pre-book — a 1–5 day window per user per gpu type.

Daily aggregates computed by the admin_daily_stats RPC. Active users = distinct users with any vm overlapping that day.

Every prompt instance (via admin_feedback_responses RPC), with auth emails from admin_user_emails. shown = chip appeared, no answer yet; dismissed = closed without answering. Filter by outcome with the chips above the list; use view on a row to see the full question and possible answers (selected options highlighted).

Move a billing principal between the live and sandbox Stripe contexts (superadmin only). Switching flushes pending usage, then re-points billing_env. The principal sets up a card the first time they land in a new context; later flips reuse the retained binding. Stripe objects never move between accounts.