Bug type

Regression (worked before, now fails)

Beta release blocker

No

Summary

When agents.defaults.model.primary is set to a Nvidia NIM model id like nvidia/nemotron-3-super-120b-a12b, the gateway strips the nvidia/ prefix before the outbound request reaches integrate.api.nvidia.com/v1/chat/completions, causing HTTP 404 model_not_found. The Nvidia NIM API expects the full vendor-prefixed model id on the wire (same convention as OpenRouter, which is already handled correctly).

Root cause is in dist/model-ref-shared-DxJ4s7gE.js: parseStaticModelRef() splits the configured ref on the first / into provider + modelRaw, then calls normalizeStaticProviderModelId(provider, modelRaw). That function has a re-prefixing case for openrouter but no equivalent for nvidia, so the bare suffix becomes the runtime model.id and is what reaches pi-ai's openai-completions request body.

Steps to reproduce

1. npm i -g openclaw@2026.4.23 (Node 22)
2. Configure a provider entry under models.providers.nvidia:
   • baseUrl: https://integrate.api.nvidia.com/v1
   • api: openai-completions
   • models: [{ id: "nvidia/nemotron-3-super-120b-a12b", ... }]
   • valid nvapi-… key
3. Set agents.defaults.model.primary to nvidia/nemotron-3-super-120b-a12b
4. Trigger any agent run, e.g. openclaw agent --agent main -m "ping"

A direct curl https://integrate.api.nvidia.com/v1/chat/completions -H "Authorization: Bearer $NVAPI_KEY" -d '{"model":"nvidia/nemotron-3-super-120b-a12b","messages":[...]}' succeeds with the same key, confirming the API expects the full prefix.

Expected behavior

Agent completes successfully. Outbound request to integrate.api.nvidia.com carries "model": "nvidia/nemotron-3-super-120b-a12b" — matching the catalog entry's id and the documented Nvidia NIM model identifier.

Actual behavior

HTTP 404: 404 page not found, failoverReason: model_not_found. Gateway log shows the runtime/agent layer with model: "nemotron-3-super-120b-a12b" and sourceModel: "nemotron-3-super-120b-a12b" — both stripped of the nvidia/ prefix before pi-ai constructs the request body.

OpenClaw version

2026.4.23

Operating system

macOS 14

Install method

npm i -g openclaw (Node 22.22.2 via nvm)

Model

nvidia/nemotron-3-super-120b-a12b

Provider / routing chain

openclaw -> nvidia (integrate.api.nvidia.com/v1, openai-completions API)

Additional provider/model setup details

No response

Logs, screenshots, and evidence

{"event":"embedded_run_agent_end","isError":true,"error":"HTTP 404: 404 page not found","failoverReason":"model_not_found","model":"nemotron-3-super-120b-a12b","provider":"nvidia","sourceModel":"nemotron-3-super-120b-a12b","providerRuntimeFailureKind":"unknown"}


After applying the proposed fix locally, `openclaw agent --agent main -m "ping"` returns `status: ok` in ~17s with metadata showing `provider: "nvidia"`, `model: "nvidia/nemotron-3-super-120b-a12b"` — full prefix preserved end-to-end.

Impact and severity

Nvidia NIM provider is unusable out of the box for any model whose catalog id starts with nvidia/ (Nemotron variants in the bundled catalog). Catalog entries with non-nvidia/ vendor prefixes in the same provider (moonshotai/kimi-k2.5, minimaxai/minimax-m2.5, z-ai/glm5) happen to work because the inner slash survives the split. So the breakage is silent and partial — users may assume "only Nemotron is broken" rather than recognizing the underlying normalize-table gap.

Additional information

Suggested one-line fix (mirrors the existing OpenRouter case):

// dist/model-ref-shared-DxJ4s7gE.js, in normalizeStaticProviderModelId
if (provider === "openrouter" && !model.includes("/")) return `openrouter/${model}`;
if (provider === "nvidia" && !model.includes("/")) return `nvidia/${model}`;   // ← add
if (provider === "xai") return normalizeNativeXaiModelId(model);

The !model.includes("/") guard preserves correct behavior for already-prefixed catalog entries like moonshotai/kimi-k2.5 (which after the parser's split becomes moonshotai/kimi-k2.5 with the slash intact, so re-prefixing is skipped).

Cosmetic follow-up (separate, pre-existing): dist/server-startup-log-DMedWp7m.js builds the displayed model ref via naive ${provider}/${model} interpolation rather than the modelKey() helper that other call sites use. After applying the fix, the startup banner shows agent model: nvidia/nvidia/nemotron-3-super-120b-a12b — purely a display issue (runtime is correct), but worth swapping to modelKey() for consistency.