使用方法
- 复制安装命令，在自己的智能体环境安装前先检查源码包。
- 从原始文档核对所需工具、凭据和支持的宿主。
- 说明输入材料与期望交付物，并对照来源检查输出。
npx skills add https://github.com/beclab/olares --skill olares-doctor开始前需了解
- 安装不包含模型订阅、API 额度或访问文件的权限。
- 仓库星数和安装量反映活跃情况，不等于独立质量或安全测试。
按源仓库内容呈现，保留标题、案例、代码、表格、链接以及原文引用的演示图片。
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
runningbut 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:doctoranswers "why is this broken and what do I do next?" Diagnosis is read-only by default; the only mutation is the explicitly approvedthirdleveldomain --force-deduperepair. The four-skill develop->deploy->debug combo ischart+market+olares-shared+doctor.
Symptom routing
| Symptom | Reference |
|---|---|
Install/upgrade stuck; never reaches running; sits in pending / downloading / installing / initializing; a fresh install ended in stopped | references/olares-doctor-app-stuck.md |
App crashes / restarts (CrashLoopBackOff, non-zero exit, CreateContainerConfigError, permission errors) | references/olares-doctor-app-crash.md |
Image won't pull (ImagePullBackOff / ErrImagePull / InvalidImageName / arch mismatch); finding unused local images | references/olares-doctor-image.md |
App is running but the entrance is unreachable / 5xx / times out / blank | references/olares-doctor-running-unhealthy.md |
System or app slow; resource pressure; GPU/compute binding rejected (node-pressure) | references/olares-doctor-resources.md |
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 isdownloading— app-service runs one download at a time, so apendingrow is often just queuing (see the appstate reference and the app-stuck reference).
Verb index
| Command | Purpose | Read when triggered |
|---|---|---|
images | Full local image inventory annotated with workload references; unused candidates | image diagnosis |
thirdleveldomain | Audit duplicate/reserved third-level domains; optional repair | domain audit and repair |
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-dedupemutates 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.
