Description

nemoclaw <sandbox> inference set --provider nvidia-prod <model> (switching from a different provider) can configure the sandbox with api: "openai-responses" even though NVIDIA Build / NIM / Gemini do not expose /v1/responses. The agent then gets a hard 404 ("page not found") on every turn. The switch path runs no validation probe, so the misconfiguration is silent until first use.

Root cause

On a provider switch, resolveRuntimeInferenceApi (src/lib/actions/inference-route-api.ts) returns null because both its session and config branches are gated on currentProvider === provider. The caller runInferenceSet (src/lib/actions/inference-set.ts) then falls back to getPreferredInferenceApi(config), which reads models.providers.inference.api — the shared managed-provider block. If the previously-active provider was a compatible endpoint validated as openai-responses, that value is still there and is carried into the switch.

getSandboxInferenceConfig (src/lib/inference/config.ts) does not reset inferenceApi for nvidia-prod / nvidia-nim / gemini-api (unlike anthropic-prod, which forces anthropic-messages), so buildProviderConfig writes api: "openai-responses" for a provider that shouldSkipResponsesProbe (src/lib/validation.ts) already documents has no /v1/responses.

Steps to reproduce

1. Configure a sandbox with a compatible endpoint that validates as the Responses API (so models.providers.inference.api becomes openai-responses).
2. nemoclaw <sandbox> inference set --provider nvidia-prod <model>.
3. Inspect the generated openclaw.json: channels/models.providers.inference.api is openai-responses.
4. Send a turn: every request 404s against NVIDIA Build, which has no /v1/responses.

Expected

Switching to a provider without a /v1/responses endpoint always configures openai-completions, regardless of the prior provider's persisted api.

Fix

In getSandboxInferenceConfig, force inferenceApi = "openai-completions" when shouldSkipResponsesProbe(provider) is true (the canonical no-responses set: nvidia-prod, nvidia-nim, gemini-api), mirroring how anthropic-prod forces its api.