Using openai-endpoint plugin without a runtime

[coryalder]

Feature

I've installed the openai-endpoint plugin (https://github.com/Mesh-LLM/openai-endpoint) via mesh-llm plugins install openai-endpoint. I've configured it in ~/.mesh-llm/config.toml:

[[plugin]]
    name = "openai-endpoint"
    url = "http://127.0.0.1:8080/v1"

but since I don't have a working runtime for my machine's GPU (#1203) it gives the following error when I run mesh-llm serve:

root@ollama:~# mesh-llm serve
no compatible native runtime found for Skippy ABI 0.1.35 on linux/x86_64: no compatible MeshLLM native runtime is installed or installable for MeshLLM 0.75.0 / Skippy ABI 0.1.35; run `mesh-llm runtime install` or inspect available runtimes with `mesh-llm runtime list --available`

It seems like there should be a way to tell mesh-llm to ignore the lack of a runtime and depend on the openai-endpoint to provide a model. Not sure if this is a documentation issue, or a missing feature.

[i386]

PR #1192 addresses the mesh-wide portion of this problem: plugin-only nodes can become discoverable and routable by peers. However, it does not address the startup failure described here: mesh-llm serve still initializes the native runtime before reading configured [[plugin]] entries. PR #1191 is unrelated. I recommend keeping this issue open until a follow-up change or regression test confirms that mesh-llm serve starts successfully with only openai-endpoint and no compatible native runtime. #1203 remains the underlying GPU-runtime issue.