AgentSkillsCN

cloudeka-helm

为 Cloudeka 环境量身定制 Kubernetes Helm 图表。当用户希望创建 Helm 图表、搭建新的 Kubernetes 部署，或在 Cloudeka 上部署应用时，可使用此技能。当用户提及“创建 Helm 图表”“搭建部署”“在 Cloudeka 上部署”“适用于 Cloudeka 的 Kubernetes 图表”“为我的应用创建 Helm 图表”等短语时，此技能将自动触发。无论是一般工作负载，还是 GPU 工作负载（ML/AI/LLM 推理），此技能均可轻松应对。

SKILL.md
--- frontmatter
name: cloudeka-helm
description: Generate Kubernetes Helm charts tailored for the Cloudeka environment. Use when the user wants to create a Helm chart, scaffold a new Kubernetes deployment, or deploy an application on Cloudeka. Triggers on phrases like "create a helm chart", "scaffold a deployment", "deploy on cloudeka", "kubernetes chart for cloudeka", "helm chart for my app". Handles both general workloads and GPU workloads (ML/AI/LLM inference).

Cloudeka Helm Chart Creator

Generate production-ready Helm charts for the Cloudeka environment with mandatory security constraints and optional GPU support.

Cloudeka Mandatory Constraints

Every chart MUST include:

yaml
securityContext:
  runAsUser: 1000
  runAsNonRoot: true
  allowPrivilegeEscalation: false

Resources MUST always be defined in values.yaml, never hardcoded in templates.

Default storage class: storage-nvme-c1.

Workflow

  1. •Ask user: app name, container image, ports, and what the app does
  2. •Determine if GPU is needed (ML/AI/LLM workload = yes)
  3. •Determine which K8s resources are needed (Deployment is always included)
  4. •Generate the full chart directory:
    • •Chart.yaml
    • •templates/_helpers.tpl
    • •templates/deployment.yaml (always)
    • •templates/service.yaml (if networking needed)
    • •templates/pvc.yaml (if persistent storage needed)
    • •templates/configmap.yaml (if config data needed)
    • •templates/hpa.yaml (if autoscaling needed)
    • •templates/ingress.yaml (if ingress needed)
    • •templates/serviceaccount.yaml (if service account needed)
    • •values.yaml
  5. •Populate values.yaml with sensible defaults

Template Reference

Read references/base-templates.md for all Helm template patterns (Deployment, Service, PVC, ConfigMap, HPA, Ingress, ServiceAccount, _helpers.tpl, Chart.yaml, values.yaml structure).

GPU Workloads

When the workload requires GPU (user mentions ML, AI, LLM, inference, training, CUDA, GPU):

  1. •Read references/gpu-config.md for GPU-specific configuration
  2. •Add to deployment: runtimeClassName: nvidia, GPU resources, shared memory volume (/dev/shm)
  3. •Add node affinity for GPU type selection (L40S or H100)
  4. •Add dekallm toleration
  5. •Use higher failureThreshold for probes (30-40) since GPU workloads take longer to initialize

When the workload does NOT require GPU, omit all GPU-specific config (no affinity, no tolerations, no runtimeClassName, no nvidia resources, no /dev/shm).

Image Registry

  • •Cloudeka registry: dekaregistry.cloudeka.id/<namespace>/<image>:<tag>
  • •Public registries (Docker Hub, ghcr.io, etc.) are also supported
  • •Default to Cloudeka registry when user doesn't specify

Output

Generate all files as actual files in the user's specified directory. After generation, list the created files and briefly explain what each values.yaml section controls.