Description

[Description]
When running NemoClaw onboard in non‑interactive mode with NEMOCLAW_PROVIDER=install-vllm and NEMOCLAW_EXPERIMENTAL=1, the documented test case expects the same code path as the interactive vLLM installation: a “Local vLLM [experimental]” option should be available in the inference menu, and a vLLM container (e.g., nemoclaw-vllm) should be created. In practice, the interactive inference menu only shows options 1–8 (NVIDIA Endpoints, OpenAI, Anthropic, Gemini, Install Ollama, Model Router, etc.) and does not include a Local vLLM option at all, so vLLM cannot be selected or installed and no nemoclaw-vllm container is created. This means the non‑interactive install-vllm provider path does not actually expose or provision vLLM as expected, and the vLLM non‑interactive test case cannot be executed as written.

[Environment]
NemoClaw: v0.0.44
OpenShell CLI: 0.0.39
OpenClaw: 2026.4.24 (cbcfdf6)
Host OS: Linux (DGX Spark/Station class machine, x86_64)
Container runtime: Docker running and healthy
Network: outbound HTTPS allowed; no network errors during onboard

[Steps to Reproduce]

Pre‑conditions:

1. NemoClaw CLI installed and working.
2. Docker daemon running on the host.
3. Host has GPU resources appropriate for vLLM (same as used for other NemoClaw vLLM tests).

Steps (non‑interactive path):

1. On the host, run the non‑interactive vLLM onboard command:
   bash NEMOCLAW_EXPERIMENTAL=1 NEMOCLAW_PROVIDER=install-vllm \ NEMOCLAW_NON_INTERACTIVE=1 NEMOCLAW_ACCEPT_THIRD_PARTY_SOFTWARE=1 \ nemoclaw onboard --fresh --name vllm-ni-test
2. Capture the exit code:
   bash echo $?
3. Check whether any vLLM container was created:
   bash docker ps --filter name=nemoclaw-vllm --format '{{.Names}}'

Steps (interactive equivalence check):

[Expected]

• With NEMOCLAW_EXPERIMENTAL=1 and NEMOCLAW_PROVIDER=install-vllm set:
  • The non‑interactive onboard should follow the same code path as the interactive vLLM install option.
  • The interactive inference menu should include a “Local vLLM [experimental] (localhost:8000)” choice when experimental vLLM support is enabled.
  • Selecting or auto‑selecting that vLLM option should result in a vLLM container being created (e.g., visible via docker ps --filter name=nemoclaw-vllm).
  • The non‑interactive run should complete successfully (exit code 0) and leave a working vLLM-backed route.

[Actual]

• Running the non‑interactive command:
  bash NEMOCLAW_EXPERIMENTAL=1 NEMOCLAW_PROVIDER=install-vllm \ NEMOCLAW_NON_INTERACTIVE=1 NEMOCLAW_ACCEPT_THIRD_PARTY_SOFTWARE=1 \ nemoclaw onboard --fresh --name vllm-ni-test

  completes without exposing any way to confirm a vLLM install from the console output alone.

• 
  When checking for a vLLM container afterwards:
  bash

  docker ps --filter name=nemoclaw-vllm --format '{{.Names}}'

  no containers are listed; there is no nemoclaw-vllm (or similar) container running.

• 
  In the corresponding interactive onboard, the “Inference options” menu only shows:
  text

  1) NVIDIA Endpoints 2) OpenAI 3) Other OpenAI-compatible endpoint 4) Anthropic 5) Other Anthropic-compatible endpoint 6) Google Gemini 7) Install Ollama (Linux) 8) Model Router (experimental)

  There is no “Local vLLM [experimental] (localhost:8000)” entry at all, so vLLM cannot be selected from the menu.

• 
  As a result, the non‑interactive install-vllm provider setting does not correspond to any visible vLLM option in the interactive menu, and no vLLM container is created; the vLLM non‑interactive test case cannot be validated as specified.

[Impact / Notes]

• QA cannot verify the vLLM non‑interactive onboarding scenario (NEMOCLAW_PROVIDER=install-vllm) because there is no corresponding vLLM option in the interactive inference menu and no vLLM container is brought up.
• This suggests either:
  • The install-vllm provider flag is ignored or not wired to the current inference menu, or
  • The “Local vLLM [experimental]” option was removed/renamed without updating the test plan.
• Suggested fixes:
  • Ensure that when NEMOCLAW_EXPERIMENTAL=1 and NEMOCLAW_PROVIDER=install-vllm are set, the interactive menu exposes a “Local vLLM [experimental]” option and the non‑interactive path provisions a vLLM container, or
  • Update the test plan and documentation to reflect the current behavior (e.g., if vLLM experimental support has been disabled or replaced), and remove or rename the install-vllm provider flag accordingly.

Bug Details

[NVB#6188458]