Thanks, this matches what I traced on my side, and the lock/wake tip is a lighter recovery than the reboot I had been doing.
Same signature here with OpenClaw rather than Hermes: the ACP job starts, initialise and session/new both return, the session list comes back, and then r1 input produces no session/prompt at all. Nothing reaches the job’s stdin.
One deterministic trigger to add: restarting the OpenClaw gateway orphans the session the cloud is holding. After that, the r1 gets at most one reply, and then silence, and no amount of PTT brings it back until the device reconnects.
Two other causes produce the identical symptom but are fixable on the server, worth ruling out before blaming the cloud:
- rabbit-agent not running at all. The installer starts it with nohup and relies on an @reboot crontab line that is not always written. Mine had been dead for three weeks after a reboot. Check with
pgrep -f "rabbit-agent run". - OpenClaw 2026.8.x with more than one agent configured. The ACP bridge creates sessions under an unscoped key and the gateway refuses prompts on it, so the r1 sits on “listening” while
gateway healthsays OK.
Wrote both up with fixes here, and the triage script is at GitHub - GrewingM/r1-triage · GitHub if useful.