Description

Description

Template 6032119 validates that a GPU sandbox using NEMOCLAW_DOCKER_GPU_PATCH_NETWORK=host can reach the host's local Ollama provider. On vm2, onboarding completed and the sandbox preserved host.openshell.internal, but local Ollama inference from inside the sandbox failed with ECONNREFUSED to 127.0.0.1:11434.

Environment

Device: vm2, NVIDIA A100-SXM4-40GB
OS: Ubuntu 24.04.4 LTS
Architecture: x86_64
Docker: Docker version 29.5.2
OpenShell CLI: openshell 0.0.44
NemoClaw: nemoclaw v0.0.53
OpenClaw: OpenClaw v2026.5.22
Ollama: running on host 127.0.0.1:11434 with qwen2.5:0.5bSteps to Reproduce

1. Start Ollama on the host and pull qwen2.5:0.5b.
2. Run a GPU sandbox onboard with host-network patching enabled:
   NEMOCLAW_DOCKER_GPU_PATCH_NETWORK=host NEMOCLAW_PROVIDER=ollama NEMOCLAW_MODEL=qwen2.5:0.5b nemoclaw onboard --name gpu-host-net-test
3. Verify the sandbox reaches Ready and /etc/hosts contains host.openshell.internal.
4. From inside the sandbox, run an agent prompt that uses the local Ollama route.

Expected Result

The GPU host-network sandbox should reach the host Ollama service and complete an inference request successfully.Actual Result

The sandbox onboarded successfully and preserved host.openshell.internal, but local inference failed:

• ip route show failed with Operation not permitted
• agent prompt failed with ECONNREFUSED to 127.0.0.1:11434

Bug Details

[NVB#6240064]