name: resource-pressure-response
description: Respond to cluster resource pressure alerts
version: "1.0"

trigger:
  - alert:
      match:
        alertname: KubeMemoryOvercommit
  - alert:
      match:
        alertname: KubeCPUOvercommit
  - manual: true

defaults:
  model: sonnet

steps:
  - name: assess-pressure
    agent: prometheus-analyst
    model: sonnet
    task: |
      Assess current resource pressure:
      - Per-node CPU usage and requests vs limits
      - Per-node memory usage and requests vs limits
      - Identify nodes under most pressure
      - Check for OOM events in last hour

      Focus on Pi cluster constraints:
      - Pi 5 (8GB): Higher capacity
      - Pi 3 (1GB): Very limited, check if overloaded
    output: pressure_analysis

  - name: identify-hogs
    agent: k8s-diagnostician
    model: haiku
    task: |
      Identify resource-heavy workloads:
      - Top 5 pods by CPU usage
      - Top 5 pods by memory usage
      - Any pods exceeding their requests
      - Any pods with no limits set
    output: resource_hogs

  - name: check-scaling
    agent: argocd-operator
    model: haiku
    task: |
      Check if any deployments can be scaled:
      - List deployments with >1 replica
      - Check HPA configurations
      - Identify candidates for scale-down
    output: scaling_options

  - name: recommend-actions
    agent: k8s-orchestrator
    model: sonnet
    task: |
      Recommend resource optimization actions:

      Analysis:
      - Pressure: {{ steps.assess-pressure.output }}
      - Top consumers: {{ steps.identify-hogs.output }}
      - Scaling options: {{ steps.check-scaling.output }}

      Prioritize actions by impact and safety:

      [SAFE] - Can be auto-applied:
      - Clean up completed jobs/pods
      - Identify and report issues

      [CONFIRM] - Require approval:
      - Scale down non-critical deployments
      - Adjust resource limits
      - Evict low-priority pods

      [FORBIDDEN] - Never auto-apply:
      - Delete PVCs
      - Delete critical workloads
    output: recommendations

  - name: cleanup
    agent: k8s-diagnostician
    model: haiku
    task: |
      Perform safe cleanup actions:
      - Delete completed jobs older than 1 hour
      - Delete succeeded pods
      - Delete failed pods older than 24 hours

      Report what was cleaned up.
    output: cleanup_result
    confirm: false

outputs:
  - pressure_analysis
  - recommendations
  - cleanup_result