I spent part of today fixing a failure mode that matters more than flashy features in Open Relay / oly: cross-node notifications.
Open Relay is supposed to let me run long-lived CLI and agent sessions like managed services. I want to start a command once, detach, come back later, inspect logs, send input only when needed, and supervise the same workload from a browser or another machine. That gets much more interesting once sessions can run on connected secondary nodes instead of a single box.
But federation only works if the human checkpoints still arrive in the right place.
Before this fix, a session on a secondary node could hit an input-needed moment while the useful notification signal stayed stranded on the wrong side of the relay. Per-channel notification settings could drift too, which is exactly the sort of quiet inconsistency that makes a supervision layer feel flaky even when the core session is still alive.
So I tightened the notification path in a few practical ways:
- input-needed prompt events now relay from a secondary node back to the primary instead of dying locally
- notification channel enable/disable state stays in sync, so the control plane matches what the session will actually do
- the session detail page updates state more cleanly, so the web view has fewer stale edges when I jump back into a live session
That is the kind of work I care about in Open Relay. I am not trying to build another terminal toy. I am building a control plane for long-running interactive work: coding agents, approval-heavy CLIs, REPLs, installers, and anything else I do not want tied to one fragile terminal tab.
In that model, notifications are not decoration. They are the handoff contract between automation and a human operator. If a secondary worker needs attention but the primary never learns about it, the whole "run it like a managed service" story starts to crack.
That is why I keep spending time on the boring plumbing. Durable sessions need more than process persistence. They need trustworthy supervision signals, especially once the workload can move across machines.
If I ask people to trust Open Relay with real long-running agent work, I need the prompt and notification path to be as solid as the session path itself.