Where jobs run
Your client submits and observes jobs. The runner starts the process, supervises it and retains the result. Closing the client leaves that supervision on the runner.
Client and runner responsibilities
| Component | Responsibility |
|---|---|
| Client CLI | Submit commands, list local history and retrieve remote status/output. |
| Client connection broker | Keep the client's authenticated endpoint and pooled connections alive between calls. |
| Runner service | Authorize requests and provide job status and retained output. |
| Job supervisor | Track the process and its descendants independently of the calling session. |
The CLI starts its local broker when needed. You do not need to run a foreground connection process or maintain a shell waiting loop for each runner.
ax exec nodes shows saved runners and snapshots of existing broker connections.
Use --refresh to probe their targets. ax exec info <name> actively queries
one runner and reports its selected direct or relay connection.
Disconnect and return
The runner must accept a request before it can execute it. Once accepted, a job can continue while the client is offline. Reconnect to retrieve its recorded state and retained output using the same job reference.
Linux/systemd and macOS/launchd use separate job services, so restarting the runner API does not itself stop those jobs. A host reboot can stop running work. Treat an unknown or lost outcome as something to reconcile before rerunning it.
The local job journal belongs to the submitting client identity. It is not an automatically synchronized history across all your devices. Remote output is read from the runner; an offline cached status is an earlier observation.
Execution hosts
Runners use Linux with systemd/cgroup v2 or macOS with launchd. A client can submit to either platform; SSH deployment uses a compatible local binary, a verified signed download, or a native bundle. Native Windows execution is not available; WSL uses the Linux backend. Mac jobs inherit their user's OS permissions and use process-group cancellation, without Linux's cgroup containment.
The runner must have the tools your command uses. Ordinary submission does not copy project files. ax can send a workspace explicitly before launching. ax does not translate paths or move a process to another machine when its runner goes offline.
Continue with service operation or jobs and results.