Description

When the inference backend (Ollama) is stopped and a request is made to inference.local from inside the sandbox, the OpenShell gateway silently drops the connection with no HTTP response body. The user sees no error message — just a curl exit code 6 after ~5 seconds. The gateway should return a clear HTTP error response indicating the backend is unreachable.

Working State
With Ollama running, inference through the gateway works correctly:

# Inside sandbox — returns valid completion
curl -s https://inference.local/v1/chat/completions \
-H  Content-Type: application/json  \
-d '{ model : llama3.2:1b , messages :[{ role : user , content : say hello }]}'

Response:

{ id : chatcmpl-100 , object : chat.completion , created :1773620673, model : llama3.2:1b , choices :[{ index :0, message :{ role : assistant , content : Hello. How can I assist you today? }, finish_reason : stop }], usage :{ prompt_tokens :27, completion_tokens :10, total_tokens :37}}

Steps to reproduce

1. Complete nemoclaw setup successfully
2. Set inference route to ollama-local: openshell inference set --provider ollama-local --model llama3.2:1b
3. Verify inference works inside sandbox (pre-condition above)
4. Stop Ollama on the host: pkill ollama
5. Verify Ollama is down: curl -s http://localhost:11434/api/tags returns connection refused
6. Inside sandbox, re-run inference: curl -s https://inference.local/v1/chat/completions \ -H Content-Type: application/json \ -d '{ model : llama3.2:1b , messages :[{ role : user , content : say hello }]}'

Steps to reproduce

No steps provided.

[NVB# 5982629]

[NVB#5982629]