The cluster
The lobby is a fact: every Machine on this LAN. A cluster is a decision: Machines the owner has marked as belonging together. Each Machine is in exactly one cluster, or is a cluster of one. It is leaderless: every member holds the same list, the Cluster page is served by whichever box you opened, and nothing fails over. It is designed for four Machines on one switch.
Membership is cryptographic: each Machine's own certificate authority (behind its :443) is its cluster identity, and members accept only each other's certificates over mutual TLS. The same admin password is not membership.
Join and adopt
| Join this cluster? | the wizard's offer on a new box when the lobby shows a cluster. Type one member's admin password; the new box mints a one-time token and the member adds it. Neither password crosses between boxes |
| Adopt | on a member's Network view an unclaimed Machine has an Adopt button. Type this member's password once; the new box is claimed with it, added, named if you like, and its recovery phrase shown once |
| Add | a Machine already claimed, from the Cluster page. It needs that Machine's admin password: the list, not the password, makes a member |
A member that stops answering goes grey and stays listed until it returns, shows up in another cluster, or you remove it. Claiming the second box mints its own vault and recovery phrase.
$ pipeos cluster status $ pipeos cluster add miura $ pipeos cluster join corvette $ pipeos cluster adopt a4e0 giulia
The Cluster page
Admin only. One row per Machine, two lines: name, chassis id and role; verdict, activity (stream live, agent running, idle), disk use and release; one verdict pill for the cluster, live from each member's hourly check. From here you can see where every agent lives, tick members and flip a service on all of them, and reboot everything.
Reboot everything warns with the busy list first (a live stream, a running job, an open terminal, an image applied and not yet booted) and still lets you go ahead; it reboots the others first and the box you opened last.
Rolling updates
Updates are per Machine and hourly, but a member will not apply while another is mid-update or not green, so at least one is up at all times. Alerts stay one DM per box.
Agents on members
An agent is a scheduled job started on a member, where it lives for its whole life: its work on that member's /data, its secrets in that member's vault. No migration. Cluster → Start an agent on a Machine, or pipeos cluster start NAME --on ID|NAME|idlest; idlest is the awake member with nothing busy and the least load per CPU, then the fewest agents. Without a schedule it is a manual job: it runs now, then only when started again. A grey member's agents are shown as last known and never restarted elsewhere.
$ pipeos cluster agents $ pipeos cluster start indexer --on idlest --prompt "…" --cron "0 * * * *"
Secrets across members
Nothing propagates except a secret the owner copies on purpose: pre-shared with the shared with ticks on Secrets, or on request, approved once on a Machine that holds it. Caps can be set on the cluster as well as a box or an agent; the most restrictive wins, and every pause names its cap. See secrets and schedule and usage.
The LED and Wake
A healthy member that can still see another member blinks its place in the sorted member list on its cabled port's LED: one flash for the first member, two for the second. The Machine that is not blinking is the one to walk over to; pipeos blink status says why. The front power light has no software interface on these chassis, so the light is the port's, at the back.
A member that is off is remembered by its siblings, with Wake on their Network view and pipeos wake NAME; it needs the BIOS switch from hardware notes, and wakes through the on-board port, not the SFP+ card.
What stays per Machine
Users and passwords, services, sign-ins, storage, streaming, the support tunnel, the assistant: all per Machine; the Cluster page shows them side by side, but every setting is that box's own. Agents reaching another box's files use SMB, pipe or SSH. Agents on a cluster whose owner is on pipe are wired together as a cohort automatically.