Summary

NemoClaw installs and runs successfully on Windows 11 + WSL2 Ubuntu 24.04 + Docker with cloud inference (Case 1 — working baseline). However, when attempting to use Windows-hosted Ollama as the local inference backend (Case 2), the NemoClaw sandbox cannot reach Ollama — even after proxy bypass — while the WSL host itself can reach Windows Ollama fine on both the WSL-facing host IP and the Windows LAN IP.

This relates to #305 (WSL2 support tracking), #315 (inference.local / local inference 403), and #246 (Ollama reasoning-model blank responses).

Environment

• OS: Windows 11 host
• WSL: WSL2 version 2.6.3.0, kernel 6.6.87.2-1
• Linux distro: Ubuntu 24.04.4 LTS (fresh install, set as WSL default)
• GPU: NVIDIA RTX 4060 Ti 8 GB (visible in WSL via /dev/dxg and nvidia-smi)
• Docker: CE 29.3.0, installed inside WSL using Docker's official Ubuntu repo
• Node.js: 22.x / npm 10.x (nvm-managed, Linux-native paths)
• NemoClaw: installed via curl -fsSL https://nvidia.com/nemoclaw.sh | bash
• Ollama: installed on Windows host only — NOT inside WSL or Docker
• Ollama bind: OLLAMA_HOST=0.0.0.0:11434 set at both User and Machine scope on Windows
• netstat confirmation: Ollama listening on 0.0.0.0:11434 and [::]:11434

What works (Case 1 — cloud inference)

• Sandbox reaches Ready phase successfully
• nemoclaw <sandbox> status returns correct provider/model info
• nemoclaw <sandbox> connect works
• openclaw tui inside the sandbox works and returns correct responses via NVIDIA Cloud API
• Validation: sent prompt Reply exactly: FORGE_PONG → received FORGE_PONG ✓

What fails (Case 2 — local Ollama)

From the WSL host shell, Ollama is reachable:

curl http://<WSL_HOST_IP>:11434/api/tags   → HTTP 200 ✓
curl http://<WINDOWS_LAN_IP>:11434/api/tags → HTTP 200 ✓

From inside the NemoClaw sandbox:

1. host.openshell.internal resolves correctly to the WSL/Docker bridge IP (172.17.0.1 placeholder)
2. Initial requests were intercepted by sandbox proxy variables and returned HTTP 403 Forbidden from the proxy:
   • HTTP_PROXY, HTTPS_PROXY, ALL_PROXY, GRPC_PROXY all set to an internal proxy endpoint inside the sandbox
3. After bypassing with curl --noproxy '*', direct calls to:
   • <WSL_HOST_IP>:11434 → timed out after 8s
   • <WINDOWS_LAN_IP>:11434 → timed out after 8s
4. host.openshell.internal:11434 → also did not return a working response

Additional signals

• Onboarding with NEMOCLAW_EXPERIMENTAL=1 did not surface an Ollama option — only "Local NIM container" (rejected due to 8 GB VRAM being insufficient) or "NVIDIA Cloud API"
• openshell inference set --no-verify --provider ollama-local --model <model> was not tested to completion due to the upstream network failure
• The model being tested (nemotron-3-nano:4b) is a reasoning-style model, which may also trigger the blank-response bug in [Bug] Ollama reasoning models return empty content — agent responses are blank #246 even if connectivity is fixed

What we already ruled out

• Ollama was not down
• Ollama was not bound only to localhost
• WSL host → Windows Ollama networking was not broken
• Docker installation was not broken
• NemoClaw installation was not broken (Case 1 cloud path works end-to-end)
• Sandbox creation was not broken

Questions for maintainers

1. Is Windows-hosted Ollama expected to be reachable from inside the NemoClaw sandbox on WSL2? If yes, what is the correct addressing pattern (host.openshell.internal, WSL host IP, Windows LAN IP, or something else)?
2. Is a host-side relay (e.g., socat forwarding on the WSL host to Windows Ollama) an expected or supported workaround for WSL2?
3. Is local Ollama on WSL2 currently considered supported, experimental-but-possible, or not yet ready in the current release?
4. What is the correct supported method to configure ollama-local today — through onboard, openshell inference set, or another route?
5. For first validation of local Ollama, is a plain chat model (e.g., llama3.2:3b) recommended over reasoning models given the open issue in [Bug] Ollama reasoning models return empty content — agent responses are blank #246?

One-line summary

NemoClaw works on WSL2 with cloud inference; Windows Ollama works and is reachable from the WSL host; but the NemoClaw sandbox cannot reach Windows-hosted Ollama directly after proxy bypass, and current project issues suggest the local-Ollama path on WSL2 is still experimental and possibly broken in the current release.