Current topology
The supported topology has one active trusted local execution host and loopback HTTP. Web and supervisor still share repository and worktree storage. Runtime events and artifact content cross the host API instead of requiring Web to read the supervisor’s private runtime files. Placement across multiple simultaneous hosts and mid-turn migration are not supported. The host contract already provides:- durable host identity across supervisor restarts;
- monotonically increasing assignment epochs and stale-driver fencing;
- a durable command ledger and idempotent host receipts;
- opaque workspace handles on normal session operations;
- attributable sessions and node attempts;
- durable event replay across reconnects and restarts;
- runtime content addressed through opaque object references.
Event delivery and lag
The supervisor persists events before delivery. Web accepts bounded batches and acknowledges committed progress; interrupted delivery can replay safely. A slow subscriber pauses and catches up from stored events when its connection can drain again, rather than being disconnected simply for being slow. Host-to-web delivery and processing into transcript, artifact, and usage views can lag independently. Stream diagnostics distinguish measurable backlog from unknown or stale data. A live connection alone does not prove that the displayed Run state is current. Check both services’ logs when progress stops.When the host pauses work
When Web falls behind on acknowledging events, the supervisor can pause event producers and refuse new work. A prolonged pause can checkpoint an active session so it can resume later. Retained, already acknowledged history is different from undelivered backlog; a long conversation alone does not mean Web is behind. In Settings → Execution host, inspect the pressure state, duration, backlog, and Refusing new work reason. The reason distinguishes storage, unacknowledged backlog, retained backlog, and control reserve limits. Check Web ingestion, Postgres availability, supervisor logs, and available storage before changing capacity limits.- A Flow node paused for host pressure shows the reason and defaults to Resume as-is. MAIster attempts automatic resume when the host admits work again, subject to capacity. You can still choose the offered stop or restart controls. Ordinary human decisions are not automatically answered this way.
- Platform-agent work can park for later continuation. New Flow and agent work waits while the host refuses admission.
- A scratch dialog shows a host-paused notice. Follow its send or recovery action after the host recovers; existing queued messages remain subject to delivery and recovery rules.
- A permission answer may remain saved with delivery pending. Retry that same answer when the host catches up; it has not necessarily reached the agent.