Olares Dashboard

dashboard (overview + applications, AI-agent first)

Shared front door: load ../olares-shared/SKILL.md for suite routing, active-profile selection, platform entry points, and the auth proceed/stop gate. Load its auth reference only when login, profile switching, token storage, or auth recovery is actually needed.

Use olares-cli dashboard <verb> --help for syntax.

When to use

• Read CPU, memory, disk, network, pod, fan, or GPU metrics.
• Rank applications/workloads by resource use.
• Observe dashboard metrics over time or consume their stable JSON envelopes.

Mental model: dashboard answers "what's the resource usage and health". For inventory and lifecycle, route elsewhere. When the metrics reveal a problem (resource pressure, an app that's running but slow/unreachable), hand off to ../olares-doctor/SKILL.md for root-cause diagnosis.

Verb index

Envelope and capability semantics

• Pin automation to kind, raw, and meta.empty_reason, not table labels.
• No hardware, no integration, or a capability gate is usually an exit 0 empty envelope, not a command failure.
• Aggregate verbs (overview, overview disk|fan|gpu) exit non-zero only when every section carries meta.error; a partly degraded envelope is still exit 0 real data. The envelope reaches stdout either way, so read the per-section error rather than treating the exit code as the whole story.
• A whole-instance outage surfaces as HTTP 530 on every section — the Olares is unreachable, so confirm it is online before diagnosing dashboard itself.
• Fan is meaningful only on Olares One. Stop probing after the device gate says it is unavailable.
• A hidden GPU sidebar for a non-admin profile is advisory if data is still returned. vgpu_unavailable is transient evidence, not proof that the device has no GPU.

Watch and diagnosis

• Watch emits repeated snapshots; JSON mode is NDJSON. Ctrl-C ends observation, not any workload.
• A watch aborts after repeated iteration failures. Separate transient metric collection failure from actual resource pressure.
• Choose either a relative or absolute time window based on the user's question; do not combine them.
• Dashboard locates pressure; it does not explain pod scheduling, crashes, image pulls, or entrance failures. Route those to olares-doctor.

Safety and escalation

• Dashboard is read-only, but cross-user metrics require platform-admin authority. Do not switch profiles or broaden the target user without approval.
• Do not report an empty integration envelope as "healthy" or "zero usage"; preserve its empty reason.
• Stop and escalate persistent upstream metric errors after a bounded retry, including the affected section and time window.