[NemoClaw][All Platform] sandbox disappears after host reboots

[eko-liu]

Description

[Description]
nemoclaw list
No sandboxes registered. Run nemoclaw onboard to get started.

[Environment]

Item	Version / detail
Device	macos and ALL platforms
Node.js	v22.x
OpenShell CLI	0.0.16
NemoClaw	v0.1.0
OpenClaw	2026.3.11 (29dc654)

[Steps to Reproduce]

1. nemoclaw onboard to finish the sandbox creation, nemoclaw list can see the sandbox
2. Restart the host or docker runtime (e.g. colima stop && colima start on mac os)
3. After host is ready, run nemoclaw list again

[Expected Result]
The sandobx onboarded before host reboot should remain.

[Actual Result]

nemoclaw list
No sandboxes registered. Run nemoclaw onboard to get started.

Steps to reproduce

No steps provided.

---

[NVB# 6035192]

[NVB#6035192]

[NVB#6039001]

[ericksoa]

Addressed by #1187 — nemoclaw list and nemoclaw <name> connect now recover sandbox inventory from the last onboard session and the live gateway after a reboot.

[tedy-y]

@ericksoa Hi, thanks for the quick effort.

In tags v0.0.2, we can see the existing sandbox via nemoclaw list after host rebooted.
However, the gateway (and the docker container) is gone and will be recreated if you run nemoclaw my-assistant status.
After gateway is ready, the nemoclaw my-assistant connect will still fail:

nemoclaw my-assistant connect
  Sandbox 'my-assistant' is not present in the live OpenShell gateway.
  Removed stale local registry entry.
  Run `nemoclaw list` to confirm the remaining sandboxes, or `nemoclaw onboard` to create a new one.
nemoclaw list

  Recovered sandbox inventory from the last onboard session.

  Sandboxes:
    my-assistant *
      model: nvidia/nemotron-3-super-120b-a12b  provider: nvidia-prod  CPU  policies: pypi, npm

  * = default sandbox

Can you double check?

[ericksoa]

Verified fixed in v0.0.6. PR #1466 (sandbox survival across gateway restarts) addresses both root causes: OpenShell v0.0.22 persists sandbox pods across gateway restarts via workspace PVC + deterministic node name, and NemoClaw auto-recovers the OpenClaw process inside the sandbox on connect/status. Full E2E validation: onboard → baseline inference → plant marker → gateway stop/start → sandbox survived → marker persisted → live inference works post-restart.

[PrachiShevate-nv]

I have verified this issue on Nemoclaw v0.0.14, the issue does not exist. Closing the nvbug.