openclaw - 💡(How to fix) Fix [Bug]: Cross-provider fallback model ID gets provider prefix doubled (e.g. `nvidia/nvidia/nemotron-3-super-120b-a12b`) [2 comments, 2 participants]

Bug type

Regression (likely introduced by the fix for #62390)

Beta release blocker

No

Summary

buildQualifiedChatModelValue (or whatever function resolves fallback model IDs at agent runtime) now doubles the provider prefix when a fallback string already contains the qualifying nvidia/ prefix in the model ID. This appears to be a regression of the fix that landed for #62390 — the previous behavior dropped the prefix when one was already present, and the over-correction now unconditionally prepends.

Current behavior

With:

{
  "agents": {
    "defaults": {
      "model": {
        "primary": "openai-codex/gpt-5.5",
        "fallbacks": [
          "openai-codex/gpt-5.2",
          "nvidia/nemotron-3-super-120b-a12b"
        ]
      }
    }
  }
}

When the primary and gpt-5.2 fallback exhaust (e.g. ChatGPT plan rate-limit), the third fallback fires with this exact failure:

[model-fallback/decision] decision=candidate_failed
  candidate=nvidia/nvidia/nemotron-3-super-120b-a12b
  reason=model_not_found
  detail=Unknown model: nvidia/nvidia/nemotron-3-super-120b-a12b

The fallback string nvidia/nemotron-3-super-120b-a12b was already qualified — provider=nvidia, id=nemotron-3-super-120b-a12b. But the resolver appears to invoke buildQualifiedChatModelValue("nvidia/nemotron-3-super-120b-a12b", "nvidia") and now returns nvidia/nvidia/nemotron-3-super-120b-a12b instead of nvidia/nemotron-3-super-120b-a12b.

Workaround attempted (unsuccessful)

Removing the nvidia/ prefix in the fallback config:

"fallbacks": ["openai-codex/gpt-5.2", "nemotron-3-super-120b-a12b"]

makes the system inherit the primary's provider:

candidate=openai-codex/nemotron-3-super-120b-a12b
detail=Unknown model: openai-codex/nemotron-3-super-120b-a12b

So neither form succeeds. There appears to be no working syntax for cross-provider fallback in the current build.

Expected behavior

Either form should resolve to nvidia/nemotron-3-super-120b-a12b:

• nvidia/nemotron-3-super-120b-a12b → recognized as provider=nvidia, id=nemotron-3-super-120b-a12b, no prefix prepended
• nemotron-3-super-120b-a12b → looked up across all providers; nvidia plugin claims it (since models list shows it's owned by nvidia)

Reproduction

1. Configure primary + fallback chain crossing providers as above
2. Force primary + same-provider fallback to fail (rate-limit a ChatGPT plan or invalidate OAuth)
3. Observe the third fallback fail with Unknown model: <provider>/<provider>/<id>

Environment

• OpenClaw 2026.5.2 (8b2a6e5) — gateway running in agent-sandbox pod
• nvidia plugin enabled, NVIDIA_API_KEY set in env, Auth: yes shown in openclaw models list
• Relevant catalog entry: nvidia/nemotron-3-super-120b-a12b (text, 195k ctx)

Suspected location

ui/src/ui/chat-model-ref.ts buildQualifiedChatModelValue — same function modified for #62390. The fix may need to detect that the model ID already starts with ${provider}/ and skip prepending in that case, while still prepending when the slash separates non-provider segments (e.g. deepseek-ai/deepseek-v3.2 should still become nvidia/deepseek-ai/deepseek-v3.2).

Related issues

• #62390 (closed COMPLETED 2026-04-25) — fix likely introduced this regression
• #49369, #38256, #71552 — earlier prefix-handling issues in the same area