Description

When running nemoclaw onboard inside WSL2 with Ollama running on the Windows host at 127.0.0.1:11434 and unreachable from WSL, the inference menu never shows the expected "Restart Ollama on Windows host with 0.0.0.0 binding" variant. Instead it shows Local Ollama (WSL:11434) — running plus Install Ollama on Windows host (recommended), contradicting the WSL2 Ollama test's expectations.

Component area: Install / Onboard / Inference.

Environment

Host OS:             Windows 11
WSL2 distro:         Ubuntu 22.04 / 24.04
CPU arch:            x86_64
GPU:                 NVIDIA GeForce RTX 5070
Docker:              Docker CE; WSL Ubuntu uses Docker Desktop engine
nemoclaw (in WSL):   v0.0.46
openshell (in WSL):  0.0.39
Ollama:              installed on Windows host, running as Windows service,
                     bound to 127.0.0.1:11434 only (default Windows install)

Steps to Reproduce

Preconditions:

1. Confirm Ollama is running on Windows host, bound to loopback only. In Windows PowerShell:

   PS C:\Users\local-mercl> netstat -ano | findstr :11434
   TCP    127.0.0.1:11434     0.0.0.0:0    LISTENING    7652

2. NemoClaw CLI installed inside WSL Ubuntu, not yet onboarded:

   nemoclaw --version
   # nemoclaw v0.0.46

3. No prior NemoClaw onboarding in this WSL instance.
4. From WSL, host.docker.internal does not resolve by default:

   curl http://host.docker.internal:11434/api/tags
   # curl: (6) Could not resolve host: host.docker.internal

   This satisfies the spirit of "curl host.docker.internal must fail" — WSL cannot access Windows Ollama (failure is DNS resolution, not ECONNREFUSED).

Repro:

5. Inside WSL Ubuntu, run:

   nemoclaw onboard

6. Observe the preflight and inference menu output at step [3/8] Configuring inference (NIM).

Expected Result

Under documented WSL2 + Windows-Ollama-loopback conditions, the inference menu should:

• Include a rebinding variant, e.g.:

  "Restart Ollama on Windows host with 0.0.0.0 binding"

• Must NOT show:
  • "Use Ollama on Windows host – running" (implies reachability)
  • "Install Ollama on Windows host (recommended)" (Ollama is already installed)
  • "Start Ollama on Windows host (suggested)" (Ollama is already running)

Selecting "Restart Ollama on Windows host with 0.0.0.0 binding" should print:

• Concrete rebind instructions such as setting OLLAMA_HOST=0.0.0.0:11434 (or equivalent).
• A PowerShell command to restart the Windows Ollama service with that binding.

After the user reapplies the binding and restarts Ollama on Windows, a subsequent nemoclaw onboard in WSL should show something like:

"Use Ollama on Windows host – running (suggested)"

reflecting that WSL can now reach Ollama via the correct interface.

Actual Result

Inside WSL, the inference menu shows:

7) Local Ollama (WSL:11434) — running
8) Install Ollama on Windows host (recommended)

No option labelled "Restart Ollama on Windows host with 0.0.0.0 binding" appears, even though:

• Ollama is confirmed running on Windows 127.0.0.1:11434.
• WSL cannot reach host.docker.internal:11434.

Negative assertions in the test are not honored:

• "Install Ollama on Windows host (recommended)" is shown under these preconditions.

NemoClaw appears to treat Ollama as a "WSL-local" option (7) and offers Windows install (8), rather than exposing a dedicated rebinding helper.

Logs

WSL onboarding session:

local-mercl@2u1g-b650-0782:/mnt/c/Users/local-mercl$ nemoclaw onboard
NemoClaw Onboarding
===================
[1/8] Preflight checks
──────────────────────────────────────────────────
✓ Docker is running
⚠ Container DNS probe inconclusive (reason: no_output). docker run produced no output (timed out or failed to start)
  Proceeding. If the sandbox build later hangs at `npm ci`, see issue #2101.
✓ Container runtime: docker
ⓘ Running under WSL
✓ Container runtime resources: 16 vCPU / 62.4 GiB
✓ openshell CLI: openshell 0.0.39
✓ Port 8080 already owned by healthy NemoClaw runtime (OpenShell gateway)
✓ NVIDIA GPU detected (NVIDIA GeForce RTX 5070, 12227 MB)
✓ Docker CDI GPU support detected (/etc/cdi/nvidia.yaml)
✓ Sandbox GPU: enabled (auto)
✓ Memory OK: 63894 MB RAM + 16384 MB swap
NVIDIA GPU detected; enabling OpenShell GPU passthrough. Use --no-gpu to opt out.
[2/8] Starting OpenShell gateway
──────────────────────────────────────────────────
[reuse] Skipping gateway (running)
Reusing healthy NemoClaw gateway.
[3/8] Configuring inference (NIM)
──────────────────────────────────────────────────
Detected local inference option: Ollama
Inference options:
  1) NVIDIA Endpoints
  2) OpenAI
  3) Other OpenAI-compatible endpoint
  4) Anthropic
  5) Other OpenAI-compatible endpoint
  6) Google Gemini
  7) Local Ollama (WSL:11434) — running
  8) Install Ollama on Windows host (recommended)
  9) Model Router (experimental)
Choose [1]:

Under the documented WSL2 + Windows Ollama preconditions, NemoClaw's onboarding does not expose the expected "Restart Ollama on Windows host with 0.0.0.0 binding" variant, making the rebind-flow test impossible to execute as written.

NVB#6196866