[All Platforms][Sandbox] sandbox start fails to restore readiness and requires explicit recover

[mercl-lau]

Warning

Regression — the readiness-ordering defect was previously fixed in #7173 and reproduces again in v0.0.107.

Description

After a documented sandbox stop, nemoclaw {sandbox} start starts the container but fails to restore the sandbox to Ready or start the primary dashboard/API host forward. The operator must run an explicit nemoclaw {sandbox} recover; the first request after recovery may still use embedded fallback before normal routed requests resume.

Platform scope: User-confirmed as affecting all supported platforms; the detailed reproduction and logs below were captured on a DGX Station GB300.

Regression: Yes — the readiness-ordering defect was previously fixed in #7173 and reproduces again in v0.0.107.

OpenShell issue: No

Environment

Device:        DGX Station GB300
OS:            Ubuntu 24.04.4 LTS
Architecture:  aarch64
Node.js:       v22.23.2
npm:           10.9.8
Docker:        29.6.1
OpenShell CLI: 0.0.101
NemoClaw:      v0.0.107 (03ccdd7468b62ecfb7d0a72852dc1059cb11d90a)
OpenClaw:      2026.7.1

Steps to Reproduce

1. Install NemoClaw v0.0.107 with the official installer and onboard an OpenClaw sandbox until it reaches Ready.
2. Run nemoclaw {sandbox} stop.
3. Run nemoclaw {sandbox} start.
4. Observe that start reports startup recovery failure and does not start the primary dashboard/API host forward.
5. Run nemoclaw {sandbox} recover and check sandbox status.
6. Send three normal agent requests after recovery and observe the routing behavior.

Expected Result

nemoclaw {sandbox} start restores the stopped sandbox to Ready, restarts its gateway and primary dashboard/API host forward, and returns success without requiring a separate recovery command. The first request after start uses the configured inference route normally.

Actual Result

The container starts, but startup recovery fails because the recreated sandbox remains in Error rather than becoming Ready. The primary dashboard/API host forward is not started. Explicit nemoclaw {sandbox} recover is required; after recovery, the first request used embedded fallback before two normal routed requests succeeded.

Logs

$ nemoclaw {sandbox} stop
Sandbox '{sandbox}' stopped. Workspace state is preserved.

$ nemoclaw {sandbox} start
Container '{sandbox-container}' started.
Restoring sandbox startup state...
Error: Sandbox '{sandbox}' started, but startup recovery failed: the recreated sandbox did not become ready in OpenShell, so the primary dashboard/API host forward was not started.
Last OpenShell readiness error: Error: sandbox '{sandbox}' is not ready (phase: Error); wait for it to reach Ready state.
The existing sandbox was preserved. Run `nemoclaw {sandbox} recover`, then retry `nemoclaw {sandbox} start`.

$ nemoclaw {sandbox} recover
Probe complete: recovered OpenClaw gateway in '{sandbox}'.

[laitingsheng]

The fix in #8990 is merged as part of the v0.0.109 release.

This report is against v0.0.107 (03ccdd7), which predates the fix; v0.0.108 does not carry it either. Please retest on v0.0.109 or later, and reopen if the behaviour persists.