Dear OpenClaw / NemoClaw Development Team,
I am submitting the attached OpenClaw / NemoClaw Engineering Bug Package Version 1.1 – Professional Edition, Developer Conference Edition (DCE) for technical review and developer investigation.
The package documents a reproducible onboarding problem encountered while installing OpenClaw/NemoClaw on an NVIDIA DGX Spark system.
Executive technical finding
The investigation systematically verified the principal infrastructure components involved in the onboarding process.
The following components were successfully validated:
- Node.js runtime
- NemoClaw CLI installation
- Docker Engine
- Docker permissions and user access
- OpenShell CLI
- OpenShell gateway
- NVIDIA GPU detection
- NVIDIA CDI GPU support
- Docker GPU passthrough
- available system memory
- available storage
- manual Docker image acquisition
- local Docker image inspection
The investigation nevertheless identified a reproducible failure during the automated NemoClaw onboarding workflow.
In particular, the onboarding process reaches the inference-provider stage and subsequently reports a Docker image pull operation that does not complete within the configured 900-second timeout.
An important comparative observation is that the relevant Docker image can be acquired successfully through a manual Docker pull on the same system.
This distinction is significant because it indicates that the basic Docker runtime and network/image acquisition capability are operational, while the failure appears to be associated with the way the onboarding workflow invokes, monitors, or handles the Docker pull operation.
NVIDIA Customer Care review
The investigation was also submitted to NVIDIA Customer Care.
NVIDIA’s response of 8 August 2026 recommended that the issue be brought to the OpenClaw/NemoClaw development team through the NVIDIA developer forum.
The NVIDIA response also summarized the investigation as a situation in which the principal subsystems pass validation while the automated onboarding workflow fails during the vLLM inference-provider image acquisition.
I have therefore prepared this package specifically to provide the development team with a structured engineering basis for further investigation.
Purpose of this submission
The objective is not to make an unsupported determination about the internal implementation of NemoClaw.
Instead, the purpose of the package is to provide sufficient evidence for the development team to investigate questions such as:
- Why does the onboarding workflow fail to complete the Docker image acquisition when a manual Docker pull succeeds?
- Does the onboarding workflow correctly detect an already available image?
- How is Docker pull progress monitored?
- Why is the operation allowed to continue until the 900-second timeout?
- Does the onboarding workflow handle large ARM64/NVIDIA images differently from the normal Docker CLI?
- Is the failure related to image detection, pull monitoring, timeout handling, or resume logic?
Submitted material
The accompanying Developer Conference Edition presents the investigation in a concise technical sequence:
Problem → System Architecture → Investigation → Evidence → Root-Cause Elimination → NVIDIA Review → Developer Recommendations → Final Assessment
The underlying Engineering Bug Package provides the supporting technical evidence and investigation history.
I would appreciate it if the OpenClaw/NemoClaw development team could review the evidence and determine whether the onboarding behaviour represents an implementation defect, an unsupported workflow condition, or another issue requiring correction or clarification.
I am willing to perform additional diagnostic tests on the DGX Spark system if the development team requires specific logs, commands, Docker diagnostics, or reproduction steps.
Thank you for reviewing this investigation.
Kind regards,
Adam Ammersken
DGX Spark user / Technical reporter
Bosnia and Herzegovina
Attachment:
How can I Upload this very important document
OpenClaw / NemoClaw Engineering Bug Package Version 1.1 – Professional Edition – Developer Conference Edition (DCE)