Health checks

When something feels off, check these surfaces in order. Each one tells you whether a layer of the system is doing its job.

Namespace learning status🔗

The namespace shows whether learning has completed and when the index was last updated. If learning never finishes or the counters sit at zero, agents have nothing to work from yet.

Endpoint sync status🔗

Applies if you brought your own hardware. Open Settings, then Your tailnets. Every endpoint shows a sync status:

• pending: the reconciler has not applied it yet, normally under a minute.
• applied: traffic can flow.
• failed: the endpoint shows an error snippet explaining why.

A failed endpoint means your models are unreachable and every task depending on them will stall.

Provider status🔗

Applies if you brought your own API key or hardware. Open Settings, then LLM providers. This is under the user menu in the top-right corner. It is organization-level, not workspace settings.

Check that your providers are enabled and assigned to at least one provider group, and that the groups cover the roles you expect them to serve. A provider that is missing a required capability is silently skipped for that role. It does not error; it is simply never chosen.

Token spend🔗

Open the namespace settings and choose LLM Token Spend. Do not be surprised to see platform models in your usage even though you brought your own model. Roles resolve through provider groups, and any role your providers do not cover falls back to the platform's group members. If you want a role on your model, check the group membership and the role's assignment under Role assignments.

Task activity🔗

Open a recent task and look at its progress.

• A task that starts and then fails repeatedly points at a model problem: a wrong model prefix, a capability the model does not really have, or an unreachable endpoint.
• A task that never starts at all points at a provisioning or pool problem.

If everything looks healthy🔗

If everything looks healthy and things still misbehave, collect what you see (page, time, what you clicked) and report it. That context is what makes issues fixable. See When you report a problem.