[N1x Linux][Sandbox] rebuild --yes destroys sandbox after registry-row recreate failure
- Dominant language
- TypeScript
- Stars
- 22.5k
- Forks
- 3.1k
- Avg merge
- 1d 43m
- Merged PRs (30d)
- 718
Description
## Description
On N1x Linux (RTX Spark N1x), running `nemoclaw my-assistant rebuild --yes` on a freshly installed and healthy sandbox fails with "the preserved source registry row changed" during the recreate phase. The rebuild deletes the existing sandbox but then cannot recreate it, leaving no working sandbox. The preserved backup is not restored automatically.
This is reproducible on both NemoClaw v0.0.124 and v0.0.125. It reproduces both when running rebuild directly and when adding a channel requires the sandbox to be rebuilt. The same rebuild operation succeeds on DGX Spark (GB10, FastOS 1.155.40) with v0.0.125.
Platform scope: Reproduced on N1x Linux only; DGX Spark rebuild succeeds with the same NemoClaw version. Other platforms not tested.
Regression: Unknown - earlier versions not tested on N1x Linux for rebuild.
OpenShell issue: No
Related: #10473 reported the same error text in an older channel stop/start sequence, but its failure left the original sandbox healthy. That earlier repro no longer occurs; this report fails after the original sandbox has been deleted.
## Environment
```text
Device: RTX Spark N1x Linux Laptop
OS: Ubuntu (N1x FASTOS 1.26.0, kernel 7.0.0-1015-nvidia, aarch64)
Architecture: aarch64
Node.js: bundled (not in system PATH)
npm: bundled (not in system PATH)
Docker: 29.2.1
OpenShell CLI: 0.0.116
NemoClaw: v0.0.125 (also reproduced on v0.0.124)
OpenClaw: v2026.7.1
```
## Steps to Reproduce
1. Fresh install NemoClaw v0.0.125 on N1x Linux via express install:
```bash
curl -fsSL https://www.nvidia.com/nemoclaw.sh | NEMOCLAW_INSTALL_TAG=v0.0.125 bash
```
2. Accept the license, choose managed vLLM option 1, and confirm express install.
3. Wait for the sandbox to become healthy (inference healthy, CUDA verified).
4. Run:
```bash
nemoclaw my-assistant rebuild --yes
```
5. The same failure also reproduces when adding a channel requires this rebuild.
## Expected Result
Rebuild should complete successfully, recreating the sandbox with the same configuration and restoring workspace state, similar to the behavior observed on DGX Spark.
## Actual Result
Rebuild deletes the old sandbox successfully but fails during recreate:
```text
[4/8] Setting up inference provider
[resume] Skipping inference (vllm-local / nvidia/Qwen3.6-35B-A3B-NVFP4)
Sandbox recreate error: Cannot resume sandbox 'my-assistant' replacement: the preserved source registry row changed.
Recreate failed after sandbox was destroyed.
Backup is preserved at: {home}/.nemoclaw/rebuild-backups/my-assistant/{timestamp}
```
The original sandbox is no longer available, and rebuild does not automatically recreate it or restore the preserved backup.
## Logs
```text
Full rebuild output:
Rebuild sandbox 'my-assistant'
Current: OpenClaw v2026.7.1
Target: OpenClaw v2026.7.1
Docker daemon: server 29.2.1
Docker CDI GPU support detected (/var/run/cdi/nvidia.yaml)
Docker can start bridge containers
Host DNS resolution works
Docker container DNS resolution works
Container runtime: docker
openshell CLI: openshell 0.0.116
Backing up sandbox state...
State backed up (12 directories, 1 files)
Backup: {home}/.nemoclaw/rebuild-backups/my-assistant/{timestamp}
Deleting old sandbox...
Old sandbox deleted
Creating new sandbox with current image...
[non-interactive] Agent: OpenClaw
NemoClaw Onboarding
(non-interactive mode)
(resume mode)
[1/8] Preflight checks
[resume] Skipping preflight (cached)
Docker daemon: server 29.2.1
Docker CDI GPU support detected (/var/run/cdi/nvidia.yaml)
Docker can start bridge containers
Host DNS resolution works
Docker container DNS resolution works
NVIDIA GPU detected; enabling OpenShell GPU passthrough.
[2/8] Starting OpenShell gateway
[resume] Skipping gateway (running)
[3/8] Configuring inference provider
[resume] Skipping provider_selection (vllm-local / nvidia/Qwen3.6-35B-A3B-NVFP4)
[resume] Reusing sandbox name: my-assistant.
[4/8] Setting up inference provider
[resume] Skipping inference (vllm-local / nvidia/Qwen3.6-35B-A3B-NVFP4)
Sandbox recreate error: Cannot resume sandbox 'my-assistant' replacement:
the preserved source registry row changed.
Recreate failed after sandbox was destroyed.
Backup is preserved at: {home}/.nemoclaw/rebuild-backups/my-assistant/{timestamp}
Cross-platform comparison:
- DGX Spark (GB10, FastOS 1.155.40, driver 580.173.02, NemoClaw v0.0.125): rebuild --yes PASSES
- N1x Linux (RTX Spark N1x, FastOS 1.26.0, driver 615.69, NemoClaw v0.0.125): rebuild --yes FAILS
- N1x Linux with NemoClaw v0.0.124: rebuild --yes also FAILS with same error
```
Contributor guide
Assessment
This issue has not been assessed yet.