OpenClaw merged a Gateway operations change that makes host suspension usable from end to end. PR #122100, titled "feat(gateway): make suspend/resume operator-usable end to end," adds CLI commands for preparing and resuming a Gateway around VM snapshots, process freezes, and other host-level maintenance windows.
The old shape had a sharp edge: once a Gateway entered the prepared suspension phase, fresh WebSocket handshakes were rejected at upgrade time. That fenced off work, but it also fenced off the control path needed to resume from a separate CLI or controller process. Operators could wait for the lease to expire or rely on a separate admin HTTP path, but the primary Gateway client could not cleanly resume the service it had suspended.
What Changed
The merged fix changes the admission model. A prepared Gateway now accepts authenticated WebSocket connections for gateway.suspend.* methods while continuing to reject ordinary work. The PR describes the new boundary plainly: suspension fences work, not control-plane visibility.
That gives operators two new CLI surfaces:
openclaw gateway suspendopenclaw gateway resume <suspensionId>
The suspend command uses a stable request id, supports bounded waiting, and can list blockers. The resume command is idempotent, and an expired lease reports as a clean no-op instead of turning into another mystery failure.
Better Failure Messages
The client side also changed. OpenClaw's Gateway client now reads bounded non-101 WebSocket upgrade responses and carries the structured refusal forward as a retryable request error. That matters during restart drain and suspension windows, because callers now see why a connection failed instead of the opaque abnormal-close path that previously appeared as a generic 1006.
In practice, a blocked call during suspension can surface as an unavailable Gateway in the prepared phase, with a retryable reason. That is much more actionable than a low-level socket close.
Why Operators Should Care
This is not a flashy feature for casual users, but it is important infrastructure work for people running OpenClaw as a durable service. Host-level maintenance is common: VM snapshots, hibernation, container pauses, process supervision, live migration, backup orchestration, and managed fleet operations all need a way to say "stop taking work, stay controllable, then resume."
The PR's evidence includes an isolated dev Gateway run where the team suspended the Gateway, froze the process with SIGSTOP, thawed it with SIGCONT, and resumed from a fresh process. They also verified the two-minute self-healing lease expiry path, wrong-id resume handling, repeated prepare behavior, and health recovery over both WebSocket and admin HTTP.
The Bottom Line
OpenClaw is continuing to make Gateway operations less dependent on local process luck. The new suspend/resume path gives administrators a real maintenance handshake: prepare, block work, keep control access available, resume cleanly, and report failures in words an operator can act on.
For self-hosters and fleet operators, PR #122100 is one of the more practical Gateway reliability merges of the week.