Symptom

openclaw infer model run --model nvidia/nemotron-3-super-120b-a12b ...

fails with 404 page not found from NVIDIA NIM.

NIM requires the full nvidia/<model> form in the request body,
but parseStaticModelRef (line 28-52 of model-ref-shared-DcLkozLZ.js)
splits on the first / and yields:

• provider = "nvidia"
• modelId = "nemotron-3-super-120b-a12b" (bare — prefix stripped)

The OpenAI-completions client then posts
{"model": "nemotron-3-super-120b-a12b", ...} — wrong shape for NIM,
because nemotron-3-super-120b-a12b (without the nvidia/ prefix)
isn't a valid NIM model identifier.

Reproduction

1. Install plugin, register nvidia/<model> correctly in models.json
   (workaround for the related apiKey bug — see companion issue).
2. Invoke: openclaw infer model run --model nvidia/nemotron-3-super-120b-a12b --prompt hi
3. NIM responds with 404 page not found.

Direct curl against NIM with the full nvidia/<model> id in the body
returns 200 OK, confirming credentials and connectivity are fine — the
bug is purely in how the model ref gets split before reaching the wire.

Root cause

parseStaticModelRef assumes:

1. Provider id and model id are separable by /.
2. The wire body should contain only the model id (the suffix after the
   first /).

Both assumptions break for providers whose model ids embed the provider
name. NIM is one example; OpenRouter-style ids (e.g.
anthropic/claude-3-opus) have the same shape.

Suggested fix

A per-provider config option that tells the parser to send the full id
in the wire body when the parsed prefix matches a provider that needs
it. Something like:

"models": {
  "providers": {
    "nvidia": {
      "modelIdPrefix": "preserve"
    }
  }
}

Or a plugin-manifest flag: supportsBareModelRef: false.

Environment

• OpenClaw 2026.4.23 (commit a979721)
• Linux container
• Node 22.22.1

Workaround

Register the model under a different provider id (e.g.
nvidia-endpoints) so the parsed model id becomes the full
nvidia/<model> form (correct shape for NIM). Awkward but functional —
we ship this in production today. Worked example available in our
internal docs/operations/nvidia-routing.md writeup; happy to share
the relevant patches.

Reported by FortiumPartners after running into both bugs while wiring
NVIDIA NIM into a multi-agent NemoClaw deployment. Happy to test patches.