VDR 5 checklist: DGX Station & DGX Spark golden path

[sandl99]

VDR 5 Checklist: DGX Station and DGX Spark

This checklist reconciles the DGX Station and DGX Spark findings from VDR 5 with Will Curran's July 8, 2026 progress update. The tested evidence includes older releases, while the progress response uses .67 as the current validation target. A related pull request or closed issue is evidence of progress, not proof that the original platform-specific finding is fixed.

Status as of July 13, 2026:

• [x] The original VDR behavior is fixed, addressed, no longer applicable, or otherwise validated.
• [~] Relevant improvements landed, but the original Station or Spark behavior still needs a hands-on .67 retest.
• [ ] The original finding remains open, lacks an exact fix, or has not been proven fixed.

DGX Station

Remaining VDR findings

• Piped installer and sandbox-name flow do not complete cleanly out of box

  • The Station Playbook's curl | bash path did not reliably complete setup.
  • Re-running with --name <sandbox> was unclear or failed because arguments were being passed through a piped shell installer.
  • The working NEMOCLAW_SANDBOX_NAME=test-sandbox environment-variable path was not documented.
  • Acceptance: one supported copy-paste command completes setup or prints one unambiguous continuation command, including the sandbox name.

• Disk-space preflight before vLLM, image, and model downloads

  • #6757 — exact managed-vLLM storage-preflight follow-up and native sub-issue.
  • Status: Failed on DGX Station with Docker's containerd image store limited to 2 GiB.
  • The NGC image pull downloaded and verified multiple layers before failing with no space left on device; no capacity warning appeared before the pull.
  • Acceptance: preflight the actual Docker image store before an uncached pull and the model-cache filesystem before hf download; report available and required space plus remediation, and stop safely by default unless explicitly overridden.
  • Verify classic DockerRootDir and containerd image-store backends, plus the bounded-storage DGX Station reproduction.
  • Do not claim this fixed without backend-aware unit coverage and a clean bounded-storage Station retest.

• Host GPU detected, but sandbox reports GPU disabled

  • #5772 — related, closed.
  • #6172 — related, closed.
  • Acceptance: a clean current Station install detects and enables the expected sandbox GPU path.
  • Status: Passed, NemoClaw detected GPU.

• Managed local vLLM endpoint validation failed

  • #5744 — related, closed; generic Linux/A100 custom-local-vLLM proxy failure, not a Station-exact fix.
  • #6177 — related, closed.
  • #6314 — related, closed.
  • Acceptance: managed local vLLM configures, validates, and serves successfully out of box on a current DGX Station.
  • Status: Passed

• Piped installer : Onboard --resume --refresh contains bugs, bad UX

  • #6500 — a stopped registered sandbox blocks strict pre-upgrade backup during an installer rerun.
  • #6562 — canceling onboarding after sandbox-name capture can leave resume without its route reservation.
  • Status: Failed on the DGX Station manual retest at v0.0.81-4-gdeed1aaa3.
  • The stopped-sandbox path fails before onboarding after the current CLI is linked, with OpenShell installation still deferred.
  • The cancel/resume path replays uncheckpointed choices, then fails with sandbox route reservation 'tm' disappeared while onboarding was in progress and recommends the same nemoclaw onboard --resume recovery.
  • --refresh is not a supported flag.
  • Acceptance: installer reruns must provide one clear recovery path; resume must recover the original sandbox name and route reservation before later prompts, or fail fast with one actionable --fresh path.

• Failed preflight logs need stronger visual contrast

  • #6752 — exact VDR follow-up and native sub-issue.
  • Status: Failed on the DGX Station manual retest.
  • The port-8080 conflict heading and explanation render as plain white text instead of using warning emphasis.
  • With a silent listener holding port 8080, accepting Continue with onboarding? [y/N] can hang after ✓ openshell CLI: openshell 0.0.72 before the conflict remediation is printed.
  • Acceptance: use the shared warning/error presentation and fail promptly with actionable PID and alternate-port guidance; direct onboarding and the installer path must not hang.

Additional Station follow-up

• DGX Station Playbook still pins v0.0.55
  • #6755 — exact documentation follow-up and native sub-issue.
  • Status: Confirmed on July 13, 2026. The live official playbook calls v0.0.55 the currently recommended stable release and explicitly sets NEMOCLAW_INSTALL_TAG=v0.0.55.
  • The hosted installer defaults to the maintained lkg channel.
  • #6064 Closed corrected the DGX Spark playbook, not this Station page.
  • SWQA bug 6407748 tracks the stale-command concern.
  • Acceptance: remove the concrete version override and stale stable claim, use the supported hosted-installer path, and validate it on a clean Station.

Progress cited in Will's update

Related Station and managed-vLLM work landed through #4888, #4867, #4676, #5215, and #5612. The guided express-install prompt is also intended to steer Station users toward managed local vLLM and suggested policy defaults. These are progress signals; the unchecked and partially checked findings above remain until Station validation succeeds.

DGX Spark

Remaining VDR findings

• Optional Telegram integration can abort express onboarding

  • When Telegram was unreachable from the corporate network, setup aborted instead of skipping the optional integration.
  • Acceptance: skip the optional channel or clearly direct the user to continue interactively with nemoclaw onboard.
  • Retest on .67.

• Telegram configuration state and health diagnostics are misleading feat(messaging): telegram channels-status health probe #6887

  • Telegram could be shown as configured even when express setup did not finish.
  • nemoclaw <name> status did not show channel-specific health, while failed / fetch failed output lacked actionable token or network guidance.

• nemoclaw onboard --resume repeats completed prompts

  • #6932 — exact VDR follow-up and native sub-issue.
  • #6934 — implementation under review.
  • Status: reproduced after completing sandbox-name, Brave Search, and messaging setup, then interrupting at resource-profile selection; resume repeated those completed prompts and credentials.
  • #5961 — related, closed; it covers interrupted non-interactive onboarding, not the exact repeated-prompt behavior.
  • #6179 — related, closed.
  • Acceptance: resume preserves sandbox name, web-search selection, messaging selection and non-secret settings, resource profile, and all later completed choices. Validated web/messaging credentials are reused only through same-session exact OpenShell bindings; raw secrets are never persisted in session state.

• Empty resource-profile input does not select the displayed default

  • Pressing Enter at Choose [6]: should select profile 6 instead of failing validation.

• Brave Search prompt needs a usable skip/back path

  • A user without an API key could not leave the prompt with q or Escape and had to interrupt onboarding with Ctrl-C.
  • Coverage and related work exist, but the no-key path still needs a hands-on .67 check.

• Shields can block chat or re-enable unexpectedly

  • #5922 — exact match, closed.
  • #6182 — related, closed.
  • Retest that shields-up allows session/history operation without failing to create agents/main/sessions/, and that shields do not turn back on during a chat without user action.

• Working local Ollama endpoint can fail provider verification

  • #1060 — related, open.
  • Switching from an NVIDIA endpoint failed verification even though Ollama was reachable and inference worked with --no-verify.

• llama.cpp / custom OpenAI-compatible endpoint remains unproven

  • Onboarding appeared to succeed, but connecting to the sandbox failed.
  • This was outside the VDR 5 rating scope, but remains an unresolved Spark compatibility finding.

• Installed OpenClaw version is not proven current

  • VDR observed an older OpenClaw inside the sandbox than the then-current release. Validate the installed version and supported update path on .67.

• Gateway lifecycle and repeat-onboard cleanup remain unclear

  • There was no graceful gateway stop command; users had to kill the process.
  • A second onboard could fail because lingering OpenShell gateway or SSH-proxy processes still held ports 8080 or 18789.
  • Tunnel, dashboard, and gateway recovery improved, but this exact cleanup UX was not explicitly proven fixed.
  • Acceptance: provide a clear stop/cleanup command and make repeat onboarding recover automatically or explain exactly which prior process/state to clear.

• nemoclaw status unhealthy-probe wording is too opaque

  • Define which output is the "host-side delivery chain."
  • Explain that OpenShell-managed runtimes include NemoClaw sandboxes.
  • Replace bare #3975 with the repository and a link to NemoClaw Issue #3975.

• OpenClaw TUI displays noisy HEARTBEAT_OK messages

  • These consume space in the chat view and should be suppressed or reduced

• nemoclaw --help lacks default-sandbox guidance

  • Expose or document how users select the default sandbox

• Unknown-provider errors lack registration guidance

  • nemoclaw inference set is expected to reject unregistered providers, but the error should tell users to register them through nemoclaw onboard.

• Remote dashboard access lacks a copy-paste SSH forwarding command

  • Onboard output only showed the localhost URL.
  • Quickstart did not cross-reference a concrete SSH port-forwarding command with placeholders for the remote host and user.

Progress cited in Will's update

• Spark install/onboarding coverage: #5608.
• Messaging/channel coverage: #5535, #5576, and #5580.
• Brave Search coverage: #5552 and closed #5266.
• Tunnel, dashboard, and gateway recovery: #5587, #5379, #5389, #5609, #5181, and #5382.
• Shields coverage: #5482, #5337, #5195, and closed #5117.

Fixed, addressed, no longer applicable, or descoped

• Pip installation without a virtual environment has documented workarounds.
• Warm local latency is about 7–8 seconds and is comparable to the local Ollama baseline; the earlier 16-second concern improved.
• nemoclaw status now provides context for starting stopped cloudflared.
• Dead dashboard-forward recovery was added.
• Spark-specific setup guide/navigation findings are no longer applicable; the remaining onboarding step was added to the installation documentation.
• WhatsApp integration was descoped from VDR 5.
• Global npm installation no longer requires the reported sudo workaround.
• nemoclaw debug --quick no longer depends on failing ps and free commands inside the sandbox.
• Stale onboard-session output from nemoclaw debug was fixed.
• Onboard now selects another dashboard port when Remote SSH already holds 18789.
• Model/provider update guidance was added to inference documentation.
• nemoclaw connect installation/linking on Spark was fixed.
• Spark GPU detection was fixed.
• Long-output command-line guidance was added to the documentation.
• Sandbox-name validation now makes the no-spaces requirement clear.
• Documentation was corrected from openclaw nemoclaw status to the supported nemoclaw status path.

Validation rule

Do not close a [~] or [ ] item solely because a related issue or pull request closed. Close it only when the original behavior is either directly covered by code evidence or reproduced successfully on the target platform and release.

[hunglp6d]

[DGX Spark]
Brave Search prompt needs a usable skip/back path — resolved.

Verified hands-on: at the Brave Search API key: prompt, empty Enter now prints "Type back to choose a different option, or exit to quit" on every loop, and the prompt supports back (return to the web-search menu), exit (cleanly abort onboarding with a resume hint), and ? (show these options). The escape path is text-command based rather than q/Escape keys.

Fixed by #6032 (fixes #6025), shipped since v0.0.74. Verified on v0.0.81-dev (deed1aa)