Keeps Horde cluster membership scoped to the nodes that actually run Sagents.Supervisor, and keeps it current as nodes come and go.

This is the membership mechanism behind:

config :sagents, :distribution, :horde
config :sagents, :horde, members: :participation

Why this exists

Horde's built-in members: :auto derives membership from Node.list([:visible, :this]) — every connected BEAM node, regardless of whether it runs Horde. In a cluster that meshes multiple service roles into one Erlang cluster, that pulls unrelated nodes into the Sagents Horde cluster: it bloats the DeltaCrdt sync fan-out (a cause of registration timeouts) and, for membership added but never reporting :alive, leaves dead entries that are never pruned.

Membership here is instead derived from participation: every node that starts Sagents.Supervisor joins an OTP :pg group, and this process sets Horde's members to exactly the nodes in that group. Because a node runs Sagents.Supervisor only where the host application chose to (e.g. gated to a :web role), "nodes running Sagents" is "agent-hosting nodes" — no node-name predicate required. :pg removes a node's entry automatically on :nodedown, so dead nodes are pruned without any extra wiring.

What it manages

On startup and on every :pg join/leave it calls Horde.Cluster.set_members/2 on all three Sagents Horde instances so they stay consistent:

• Sagents.Registry
• Sagents.AgentsDynamicSupervisor
• Sagents.FileSystem.FileSystemSupervisor

It is started automatically by Sagents.Supervisor (together with its :pg scope) when members: :participation is configured; you do not start it yourself.

Partitioning

When config :sagents, :horde, partition: <value> is set, the :pg group is keyed by that partition ({:sagents_members, value}). A node only joins and monitors its own partition's group, so membership is isolated per partition — e.g. nodes in one Fly.io region never become members of another region's Horde cluster, even when all nodes share one connected BEAM cluster.