olares-doctor

doctor (runtime diagnosis)

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.

This skill is a thin diagnostic router over Market, Cluster, and Dashboard. Load the shared application-state model when interpreting lifecycle states, TTLs, serialized downloads, or running. Use olares-cli doctor <verb> --help for syntax.

When to use

• An install/upgrade is stuck or never reaches running; an app won't start.
• An app crashes / restarts repeatedly (CrashLoopBackOff, exit codes, config errors).
• An image won't pull (ImagePullBackOff / ErrImagePull / wrong arch), or you want to find unused local images.
• An app is running but its entrance is unreachable / errors / times out.
• The system or an app is slow, or a GPU/resource binding is rejected (node-pressure).

Both catalog apps (installed via market) and your own dev apps (deployed via chart) route runtime failures here. Once the root cause is found, the fix for a dev app you authored is usually a chart edit — hand back to ../olares-chart/SKILL.md.

Mental model: doctor answers "why is this broken and what do I do next?" Diagnosis is read-only by default; the only mutation is the explicitly approved thirdleveldomain --force-dedupe repair. The four-skill develop->deploy->debug combo is chart + market + olares-shared + doctor.

Symptom routing

A model that is configured but does not answer is diagnosed one layer up first: olares-router separates the gateway, its access control and the model application's own download/engine state from the pod-level failures here, and routes back when the cause is below the application.

First, rule out the normal queue. Before declaring an install stuck, check whether another app is downloading — app-service runs one download at a time, so a pending row is often just queuing (see the appstate reference and the app-stuck reference).

Verb index

How doctor gathers evidence (orchestration, not ownership)

• Lifecycle state/source comes from olares-market.
• Pods, events, logs, and workloads come from olares-cluster.
• Pressure and utilization come from olares-dashboard.
• Namespace discovery follows the shared platform model.

Correlate evidence by time and object ownership. A Market timeout is not failure; running proves only entrance TCP reachability; a fresh install can settle at stopped after scheduling failure without a *Failed lifecycle state.

Safety and escalation

• Diagnosis is read-only by default. Do not restart, delete, scale, cancel, prune, or edit a chart as an automatic diagnostic step.
• thirdleveldomain --force-dedupe mutates Application resources. Show the proposed changes and obtain explicit approval first.
• An unused-image report is evidence, not authorization to remove images.
• System namespace evidence commonly requires admin visibility. On 403/404, use the app's own evidence and report the missing visibility; do not switch identities without approval.
• Stop when the app/user/namespace is ambiguous, required logs need a higher role, or the fix crosses into chart editing, lifecycle mutation, or host administration.*