Runtime Lifecycle

RLMesh separates detaching a client from shutting down an endpoint. The process that created a server owns its lifetime; clients come and go without affecting it. Keeping the two operations distinct is what lets several clients share one environment endpoint, and what keeps a short-lived evaluation script from tearing down a long-lived server.

Clients detach

RemoteEnv.close() and RemoteVectorEnv.close() detach the client. They release the connection but do not normally stop the server — another client can connect afterwards, or the same script can reconnect later.

from rlmesh.numpy import RemoteEnv

with RemoteEnv("127.0.0.1:5555") as env:
    observation, info = env.reset()

The context manager form calls close() on exit, so a client never outlives the block that created it.

Servers shut down

Stopping an endpoint is a separate, deliberate action: call shutdown() on the EnvServer you created. The typical pattern starts the server, runs work against its address, and shuts down in a finally block so the endpoint goes away even when the work fails:

server = rlmesh.EnvServer(env, host="127.0.0.1", port=0)
server.start()

try:
    run_eval(server.address)
finally:
    server.shutdown()

See Serving environments for the full serving workflow, including the blocking serve() form.

Remote shutdown is opt-in

By default only the owning process can stop a server. Letting a remote client request shutdown is an owner-level decision made through server options:

options = rlmesh.ServeOptions(allow_remote_shutdown=True)
server = rlmesh.EnvServer(env, "127.0.0.1:5555", options=options)

Leave this off for shared endpoints. Enable it when a single evaluator owns the whole topology and should be able to wind it down from the client side.

Sandboxes own their session

Sandboxed environments invert the client rule. SandboxEnv.close() and SandboxVectorEnv.close() stop the sandbox session they own, because RLMesh started that session on your behalf. This is deliberately stronger than RemoteEnv.close(): when the handle goes away, so does the isolated process behind it.

See Sandboxed environments for when to reach for a sandbox instead of a plain remote connection.

Summary