Remote Control: host-targeted restart/update trigger for relay-only operators (re-file of the restart half of #53125)

Open 💬 2 comments Opened Jun 11, 2026 by reese-anschultz-ab

Problem

Operating a claude remote-control host purely over the relay (claude.ai/code or mobile), with no out-of-band shell or console, there is no in-band way to restart the server or apply an update:

  • The session you'd use is a worker spawned under the server, so killing the server to pick up a new binary also kills your only control channel, with nothing to bring it back. Recovery currently requires an external supervisor (e.g. a scheduled task respawning claude remote-control --name <name>), which stock installs don't have.
  • claude update updates the on-disk binary, but a long-lived server keeps running the version it was spawned with; there's no way to re-exec onto the new version without out-of-band access.

Key constraint: invocation must target the HOST

A plain typed /restart may not work here: it's unclear whether the web/mobile client interprets slash commands locally or forwards them to the host worker. If they're caught client-side, a typed command never reaches the box. So the affordance must be explicitly host-targeted:

  • a dedicated UI control in claude.ai/code and the mobile app ("Restart host" / "Update host"), and/or
  • a reserved command the client is required to forward to the host.

(Please confirm current slash-command routing for remote sessions - it determines the right surface.)

Requests

  1. Host restart - re-exec the remote-control server in place, preserving --name/flags, auto-reconnecting connected clients.
  2. Host update - detect a newer available version and, on explicit user confirmation, run the equivalent of claude update and re-exec onto it.

Design note

Both must be user-initiated (UI action / forwarded reserved command), never something a model turn can invoke on its own - a self-restart driven by model output is a trust/abuse surface.

Related / prior art

  • #53125 (closed, not_planned): asked for MCP reload and remote process restart. Only /mcp reconnect shipped; the process-restart half is unaddressed and the inactivity bot invited a re-file. This issue is that re-file, scoped to restart + update.
  • #36261: session persistence/reconnection across restart - the infrastructure a clean host restart would build on (complementary, not a trigger).
  • #63964: /remote-control silent failure on a stale build + surface version - related to update-awareness, but a bug report, not an update-and-re-exec action.
  • #30447: --headless daemonizable remote control - complementary recovery angle.

Environment

  • Claude Code v2.1.149 (host), native installer, Windows 11
  • claude remote-control --name <name>, default same-dir mode

View original on GitHub ↗

This issue has 2 comments on GitHub. Read the full discussion on GitHub ↗