[N1x][WSL2][Onboard] start-windows-ollama restarts Ollama but model listing fails

[JoyceChenNV]

Description

NemoClaw v0.0.115 on WSL2 ARM64 (N1X WoA). The fix from #10100 / PR #10120 partially works: the start-windows-ollama restart action correctly detects that Docker cannot reach Ollama, sets OLLAMA_HOST=0.0.0.0:11434, restarts Ollama, and confirms reachability on host.docker.internal:11434. However, the subsequent model listing fails with:

Error: Could not read Ollama models from host.docker.internal:11434 after 3 attempts.

This is a regression from the original #10100 behavior — the fix advances past the unreachable-daemon block but introduces a new failure at the model-listing step.

Platform scope: Reproduced on WSL2 ARM64 (N1X WoA) only; other platforms not tested

Environment

Device:        RTX Spark N1X (6144-core Blackwell RTX GPU)
CeleritOS:     developer-2.324.8 (NDUP)
OS:            Windows / WSL2 Ubuntu 24.04
Architecture:  aarch64
Driver:        616.62
GPU VRAM:      63936 MiB
Ollama:        0.33.0 (models: qwen3.5:9b, qwen3.6:35b)
Docker:        29.7.2 (Docker Desktop)
NemoClaw:      v0.0.115
OLLAMA_HOST:   not set (default 127.0.0.1 — intentionally unset to test #10120 fix)

Docker credential store note: credsStore: "desktop" returns "A specified logon session does not exist" for image pulls. curlimages/curl:8.10.1 was pre-pulled with DOCKER_CONFIG=$(mktemp -d) docker pull before onboarding.

Steps to Reproduce

1. Fresh N1X WoA machine with Ollama installed (default config, OLLAMA_HOST NOT set)
2. Pre-pull probe image: DOCKER_CONFIG=$(mktemp -d) docker pull curlimages/curl:8.10.1
3. In WSL:

   curl -fsSL https://www.nvidia.com/nemoclaw.sh | NEMOCLAW_INSTALL_TAG=v0.0.115 bash

4. Accept express install defaults

Expected Result

PR #10120 fix detects Docker unreachability → sets OLLAMA_HOST=0.0.0.0 → restarts Ollama → reads models → continues onboarding.

Actual Result

Steps 1–4 of the fix work correctly (detect, set, restart, verify reachability), but model listing fails:

[3/8] Configuring inference provider
[non-interactive] Provider: start-windows-ollama

⚠ Ollama is binding to 0.0.0.0 so the sandbox can reach it via Docker.
  This exposes the Ollama API to your local network (no auth required).
  On public WiFi, any device on the same network can send prompts to your GPU.
  See: CNVD-2025-04094, CVE-2024-37032

Starting Ollama on Windows host via WSL interop...
Waiting for Ollama to respond on host.docker.internal...
✓ Using Ollama on host.docker.internal:11434
Error: Could not read Ollama models from host.docker.internal:11434 after 3 attempts. Verify that Ollama is reachable at http://host.docker.internal:11434/, then retry onboarding.

Analysis

The start-windows-ollama action's awaitWindowsOllamaReady() confirms Ollama is reachable (✓), but the model-listing step that follows uses a different code path (possibly a Docker-based probe or a different HTTP client) that fails. Possible causes:

1. Timing: Ollama just restarted and models are not yet loaded when the listing runs
2. Docker credential store: the broken credsStore: "desktop" may block the Docker-based model-listing probe even though curlimages/curl:8.10.1 is cached
3. Probe path mismatch: awaitWindowsOllamaReady probes /api/tags via Docker and succeeds, but the model-listing step uses a different endpoint or transport

Related

• Fixes [N1x][WSL2][Onboard] Express install does not configure OLLAMA_HOST=0.0.0.0 for Windows-hosted Ollama #10100 (partial — restart action works, model listing does not)
• PR fix(onboard): verify Windows Ollama Docker reachability #10120
• NVBug 6657998