InferenceManage Inference

Use Shared Gateway Routes

View as Markdown

OpenShell exposes one live inference route per gateway. NemoClaw records each sandbox’s intended provider and model so compatible sandboxes can take turns using that route.

Change the Live Route During Onboarding

Onboarding a sandbox with another valid provider or model warns immediately before it re-points the shared route. The warning names every affected registered sandbox and its recorded provider and model. Those sandboxes use the new live route until another onboarding run or connection re-points it again.

When multiple sandboxes use the same provider name, their provider-global identity must match. For a custom provider, NemoClaw compares the normalized endpoint and API family. It also compares the credential environment-variable name for every provider. Onboarding exits before provider or route mutation when those fields differ.

Do not run openshell inference set directly on a shared NemoClaw gateway because that bypasses the registry compatibility checks.

Restore a Sandbox Route When You Connect

Connect to a sandbox when you want its recorded provider and model to become live again.

$nemo-deepagents <name> connect

Before changing the route, connect verifies the same provider-global identity used during onboarding. When the identity is compatible, connect warns and re-points the route to the sandbox’s recorded provider and model. When the identity differs or required metadata is incomplete, connect stops because a provider-and-model-only route change cannot safely reconstruct that configuration.

Inspect Recorded and Live Routes

Run sandbox status to compare the sandbox’s recorded route with the gateway’s live route.

$nemo-deepagents <name> status
$nemo-deepagents <name> status --json

Text output shows the provider and model recorded for the sandbox and adds a warning when the live route differs. JSON output exposes the two sources as recordedRoute and liveRoute and adds routeDrift when they differ. The schema-version 1 provider and model fields remain live-first when the gateway route is readable and fall back to the recorded route otherwise. When routeDrift.canConnect is false, connect cannot safely restore the recorded route because provider-global identity differs or required route or gateway metadata is incomplete.

Handle a Hard Route Conflict

NemoClaw fails closed when it cannot prove that a route change preserves the provider and sandbox contracts.

  • For a provider-global identity conflict, align the endpoint, API family, and credential environment-variable name, or remove a sandbox you no longer need.
  • For incomplete legacy custom-route metadata, back up and remove the affected sandbox, then re-onboard it with an explicit endpoint and API family.
  • For an invalid gateway binding, restore known-good gatewayName and gatewayPort metadata from a trusted backup, or back up and remove the sandbox before re-onboarding it.
  • For a runtime inference set conflict, align the registered sandbox routes or remove the conflicting sandbox before retrying.

Refer to Troubleshooting a Shared Inference Route Conflict for the recovery commands.