Health

:overview/:ov and :problems are projections over the same node, service, and etcd snapshots the rest of t9s already shows — they never make their own API calls, and evaluating them never changes what :nodes/:services/:etcd display.

Diagnoses are deterministic and cite evidence: each one has a severity, a short summary, the resource it’s about, and a stable rule identity, so “unhealthy” is always traceable back to why.

Active etcd alarms (NOSPACE, CORRUPT, and future alarm types) are treated as critical etcd-member-alarmed diagnoses — a NOSPACE alarm makes the cluster read-only until it is cleared — and are also shown in the :etcd view’s ALARMS column. From :etcd, A disarms the selected member’s alarms behind --enable-writes; disarming lifts the read-only restriction but does not reclaim disk, so free space first or the alarm returns. A CORRUPT alarm still requires restoring from a snapshot — disarming does not repair corruption. See Etcd operations.

:overview

A compact per-resource-kind breakdown — for example NODES: 2/3 healthy, 1 warning, 0 critical — plus the top critical diagnoses inline as a preview. Read-only, no per-row selection; use :problems to drill into individual issues.

:problems

A flat, filterable table of every current diagnosis: SEVERITY, KIND, RESOURCE, SUMMARY. Press / to filter like any other table.

Press Enter or d on a row to drill into the underlying resource:

• A node diagnosis opens that node’s detail view.
• An etcd member diagnosis opens the etcd list (etcd has no per-member detail yet, so this lands you on the list rather than failing to drill in at all).

r refreshes the underlying data; the health evaluation re-runs against the refreshed snapshot on the next render.

:healthcheck

:healthcheck/:hc streams the Talos server-side cluster health check. Unlike :overview/:problems, it is a live RPC: it opens a stream to the current Talos endpoint, waits for the cluster to converge, and prints the server’s progress lines (discovered nodes: …, waiting for …) as they arrive. A clean end of stream renders ✔ cluster is healthy; a failure renders ✘ cluster health check unavailable. Raw gRPC error text is never shown.

• The session endpoint must itself be a control-plane node — the server rejects a worker endpoint.
• The optional wait timeout defaults to 60s; override it with a Go duration, e.g. :healthcheck 5m. A longer timeout is appropriate for a cluster you expect to become ready. An invalid duration is treated as an unknown command.
• r reconnects, rebuilding the node lists from the current :nodes snapshot. A running check reflects the node set captured when it started; it does not follow live membership changes.
• Keys: / filter, C clear, r reconnect, q/Esc back.

Only IP addresses are sent to the server (the server rejects hostnames). If a list is empty the header shows ? for it; the server falls back to its own discovery only when both the control-plane and worker lists are empty. A worker-only or worker endpoint still fails server-side.

:healthcheck is an explicit, additional signal: it never changes :overview or :problems, and its verdict lives only in this view. Missing or unknown data never becomes “healthy”.

Missing or stale data

Missing, stale, or contradictory evidence produces unknown — it never silently reads as healthy. Objective resource exploration in :nodes/:services/:etcd stays fully useful even if health evaluation is incomplete.