Platform

WSL2 x86_64 (Windows 11 host) running Ubuntu 24.04.4 LTS in WSL2

Symptom

nemoclaw onboard completes steps [1/8]–[4/8] successfully with a running Windows-host Ollama, but silently exits after step [4/8] without running steps [5/8]–[8/8] and without creating any sandbox. A follow-up nemoclaw list shows no sandboxes, and <sandbox-name> status reports "does not exist".

Component area

CLI&UX / Onboarding / WSL2 + Windows-host Ollama

Steps to reproduce

Preconditions

• Windows 11 with WSL2 enabled.
• WSL2 Ubuntu 24.04.4 LTS.
• Docker Desktop installed and configured to use the WSL backend; Docker daemon running.
• NemoClaw installed in WSL via the official installer script (nemoclaw available on PATH).
• Ollama installed and running on the Windows host; reachable from WSL at http://host.docker.internal:11434.
• No existing NemoClaw sandboxes (verify with nemoclaw list).

Repro

# 1) Start onboarding fresh from WSL2
nemoclaw onboard --fresh --name ollama-wsl-test

# 2) Let preflight complete (step [1/8]) and gateway start (step [2/8])

# 3) At [3/8] "Configuring inference provider", choose the Windows-host Ollama:
#    When prompted with:
#      7) Local Ollama (WSL:11434) — running
#      8) Use Ollama on Windows host - running (suggested)
#    enter: 8

# 4) Confirm use of the Windows-host daemon:
#    "Use Ollama on the Windows host (already running)? [Y/n]:" Y

# 5) When asked to choose a model, select the starter model: 1   # qwen2.5:7b

# 6) At the review screen, keep web search disabled and set a sandbox name
#    (example: prachi-new), then accept:  Y
#    Apply this configuration? [Y/n]: Y

# 7) Observe step [4/8] output, then check sandboxes:
nemoclaw list
nemoclaw prachi-new status

Expected vs Actual result

Expected (from onboarding spec and behavior on non-WSL platforms)

For a WSL2 + Windows-host Ollama setup where Ollama is already running and reachable:

• At step [3/8], NemoClaw detects and uses the Windows-host Ollama daemon, printing a confirmation such as:

  ✓ Using Ollama on host.docker.internal:11434
  

• At step [4/8] "Setting up inference provider", NemoClaw:
  • Successfully sets the active gateway, e.g. ✓ Active gateway set to 'nemoclaw'.
  • Optionally logs non-fatal diagnostics (e.g., container reachability checks) but continues onboarding.
• Onboarding then proceeds to:
  • [5/8] Web search configuration.
  • [6/8] Messaging configuration.
  • [7/8] Sandbox build (image pulls, npm ci, etc.).
  • [8/8] Completion summary with clear "Onboarding complete" messaging and next steps.
• After onboarding completes:
  • nemoclaw list shows the newly created sandbox (e.g., ollama-wsl-test / prachi-new) with a Ready or Building status.
  • nemoclaw <sandbox-name> status returns details instead of "does not exist".

Actual

Real WSL2 + Windows-host Ollama case. Running:

nemoclaw onboard --fresh --name ollama-wsl-test

produces:

• [1/8] Preflight checks: all green (Docker running, WSL detected, GPU detected, memory OK, etc.).

• [2/8] Starting OpenShell gateway: gateway reported healthy.

• [3/8] Configuring inference provider:

  ⚠ Ollama is running on both WSL and the Windows host. Stop one to avoid
    duplicated GPU memory and model caches.
  
  Detected local inference option: Ollama
  Inference options:
  …
    7) Local Ollama (WSL:11434) — running
    8) Use Ollama on Windows host - running (suggested)
    9) Model Router (experimental)
  Choose : 8
  Use Ollama on the Windows host (already running)? [Y/n]: Y
  ✓ Using Ollama on host.docker.internal:11434
  

• Model selection and review:

  Ollama models:
    1) qwen2.5:7b
    2) Other...
  Choose model : 1
  Loading Ollama model: qwen2.5:7b
  Chat Completions API available — OpenClaw will use openai-completions.
  
  Sandbox name (1-63 characters, lowercase, starts with a letter,
    letters/numbers/internal hyphens only, ends with letter/number)
    [my-assistant]: prachi-new
  
  ──────────────────────────────────────────────────
  Review configuration
  ──────────────────────────────────────────────────
  Provider:        ollama-local
  Model:           qwen2.5:7b
  API key:         (not required for ollama-local)
  Web search:      disabled
  Managed tools:   none
  Messaging:       none
  Sandbox name:    prachi-new
  
  Note: Sandbox build typically takes 3–8 minutes on this host.
  ──────────────────────────────────────────────────
  Web search and messaging channels will be prompted next.
  Apply this configuration? [Y/n]: Y
  

• Step [4/8] "Setting up inference provider":

  [4/8] Setting up inference provider
  ──────────────────────────────────────────────────
  ✓ Active gateway set to 'nemoclaw'
  
  Local Ollama is responding on host.docker.internal, but the Docker container
  reachability check failed for http://host.openshell.internal:11435.
  This may be a Docker networking issue — the sandbox uses a different network
  path and may still work.
  Diagnostic: host-gateway resolved to: 172.17.0.1.
  Retried 3 times over ~4s.
  

Actual behavior immediately after this:

• nemoclaw onboard returns to the shell without showing steps [5/8]–[8/8].
• There is no additional error message explaining that sandbox creation failed.

Follow-up CLI checks:

nemoclaw list
# Output:
#   No sandboxes registered. Run `nemoclaw onboard` to get started.

nemoclaw prachi-new status
# Output:
#   Sandbox 'prachi-new' does not exist.
#   Run 'nemoclaw onboard' to create one.

So:

• Inference provider setup and gateway activation succeed.
• The "Windows-host Ollama" path is healthy (✓ Using Ollama on host.docker.internal:11434).
• But no sandbox resource is created, and onboarding silently stops after step [4/8].

Environment versions

• Host: local-qa@2u1g-b650-0151 (Windows 11 + WSL2 Ubuntu 24.04.4 LTS)
• OS (WSL): Ubuntu 24.04.4 LTS
• Kernel (WSL): 6.6.87.2-microsoft-standard-WSL2
• NemoClaw CLI: v0.0.37
• OpenShell CLI: 0.0.44
• Docker Desktop (WSL backend): 29.x (Docker daemon reachable from WSL)
• GPU: NVIDIA GeForce RTX 5090 (GPU passthrough detected in preflight)
• Inference: Ollama on Windows host, model qwen2.5:7b
• Web search: disabled in the configuration shown above

Logs / notes

Key log snippets:

• Confirmation that Windows-host Ollama is detected and used:

  Use Ollama on the Windows host (already running)? [Y/n]: Y
  ✓ Using Ollama on host.docker.internal:11434
  

• Step [4/8] finishing successfully but followed only by a container reachability warning, then silent exit:

  [4/8] Setting up inference provider
  ──────────────────────────────────────────────────
  ✓ Active gateway set to 'nemoclaw'
  
  Local Ollama is responding on host.docker.internal, but the Docker container
  reachability check failed for http://host.openshell.internal:11435.
  This may be a Docker networking issue — the sandbox uses a different network
  path and may still work.
  Diagnostic: host-gateway resolved to: 172.17.0.1.
  Retried 3 times over ~4s.
  

• No subsequent [5/8]–[8/8] output is printed, and no sandbox appears in nemoclaw list.

This behavior appears distinct from earlier WSL2 bugs where Ollama never starts and the wizard times out at step [3/8]; here, the Windows-host Ollama path is healthy and step [4/8] completes, but sandbox creation simply never happens and the user is left without any sandbox and without a clear error.