Description

On DGX Spark with a pre-installed but outdated Ollama (v0.6.2), the express setup path detects the Spark platform correctly and triggers install-ollama provider, but because Ollama is already installed, the upgrade step (curl install.sh) is skipped. The old Ollama cannot load qwen3.6:35b (crashes with INVALIDARGUMENT in ggml.GraphSize), causing the model probe to time out repeatedly. The wizard loops back to model selection indefinitely with no escape — no version warning, no upgrade prompt, no actionable error message. The user is stuck and the sandbox is never created.

Two issues:

1. No Ollama version check — NemoClaw has no minimum Ollama version and does not detect that the installed version is too old to run the target model.
2. Model probe dead loop — when the probe fails, the wizard offers the same model list again instead of surfacing a diagnostic ("Ollama may need upgrading") or breaking out after N retries.

Environment

Device:        DGX Spark (NVIDIA GB10, unified memory 124609 MB)
OS:            Ubuntu 24.04.4 LTS (aarch64)
Architecture:  aarch64
Node.js:       v24.15.0
npm:           11.12.1
Docker:        Docker CE
OpenShell CLI: 0.0.44
NemoClaw:      v0.0.50
OpenClaw:      N/A (onboard not completed)
Ollama:        0.6.2 (deliberately downgraded from 0.24.0 to test upgrade path)

Steps to Reproduce

1. On DGX Spark, install an old Ollama: curl -fsSL https://ollama.com/install.sh | OLLAMA_VERSION=0.6.2 sh
2. Pull a model: ollama pull qwen3.6:35b (succeeds on 0.6.2 — pull works, inference does not)
3. Uninstall NemoClaw if present
4. Run: bash <(curl -fsSL https://www.nvidia.com/nemoclaw.sh)
5. Accept license, observe "Detected DGX Spark" and express setup prompt
6. Select Y for express setup
7. Observe onboard attempting to use Ollama with qwen3.6:35b

Expected Result

Either:

• Express setup detects Ollama 0.6.2 is too old and auto-upgrades via curl install.sh (the install-ollama code path already does this for fresh installs)
• Or at minimum: model probe failure produces an actionable message like "Ollama version 0.6.2 may be incompatible — upgrade with: curl -fsSL https://ollama.com/install.sh | sh" and exits cleanly

Actual Result

1. Express setup triggers correctly ("Detected DGX Spark")
2. Provider set to install-ollama, but because Ollama binary exists, the curl install.sh step is skipped
3. NemoClaw uses Ollama 0.6.2 directly — no version check
4. Model probe sequence:
   • "Could not verify 'tools' capability for 'qwen3.6:35b' — Ollama did not return capability metadata"
   • "Loading Ollama model: qwen3.6:35b"
   • "Selected Ollama model 'qwen3.6:35b' did not answer the local probe in time"
5. Wizard loops back to "Ollama starter models:" selection — infinite loop
6. Ollama service logs show crash:

ollama.service: Main process exited, code=exited, status=2/INVALIDARGUMENT

Crash in ggml.go:419 GraphSize → HeadCountKV — qwen3.6 GGUF format incompatible with 0.6.2

7. No sandbox created, no way to proceed without manually upgrading Ollama

Logs

NemoClaw onboard output:

[non-interactive] Provider: ollama
✓ Using Ollama on localhost:11434 (proxy on :11435)
Could not verify 'tools' capability for 'qwen3.6:35b' — Ollama did not return capability metadata; continuing.
Loading Ollama model: qwen3.6:35b
Selected Ollama model 'qwen3.6:35b' did not answer the local probe in time.
Choose a different Ollama model or select Other.

Ollama starter models:
  1) qwen2.5:7b
  2) nemotron-3-nano:30b
  3) qwen3.6:35b
  4) Other...
[loops back indefinitely]

Ollama service log (journalctl -u ollama):

ggml.go:419 GGML.GraphSize crash → HeadCountKV
ollama.service: Main process exited, code=exited, status=2/INVALIDARGUMENT
ollama.service: Failed with result 'exit-code'
ollama.service: Scheduled restart job, restart counter is at 5.
Listening on 127.0.0.1:11434 (version 0.6.2)
inference compute: library=cuda name="NVIDIA GB10" total="121.7 GiB" available="38.5 GiB"

NVB#6216882