事件應變工作流程

# Incident Response Workflow
# AI triage with severity-based routing, auto-remediation, and post-mortem

osop_version: "2.0"
id: incident-response
name: "事件應變工作流程"

nodes:
  - id: alert_triggered
    type: api
    purpose: Receive alert from monitoring system (PagerDuty, Datadog, etc.)
    runtime:
      endpoint: webhook
      method: POST
      url: /api/incidents/ingest
    outputs: [alert_id, service, metric, threshold]

  - id: ai_triage
    type: agent
    purpose: AI classifies incident severity and identifies likely root cause
    runtime:
      provider: anthropic
      model: claude-sonnet-4-20250514
    inputs: [alert_id, service, metric, recent_deploys, runbook]
    outputs: [severity, root_cause_hypothesis, recommended_action]
    timeout_sec: 30
    explain:
      what: Analyzes alert context, recent deploys, and historical patterns
      why: Faster triage reduces MTTR — AI handles P3/P4 without waking humans

  - id: severity_router
    type: system
    subtype: router
    purpose: Route incident based on AI-assessed severity level
    inputs: [severity]
    outputs: [route]

  - id: auto_remediate
    type: cli
    purpose: Execute automated remediation for low-severity incidents
    runtime:
      os: linux
      command: |
        ./runbooks/${service}/remediate.sh \
          --action ${recommended_action} \
          --alert-id ${alert_id}
    inputs: [service, recommended_action, alert_id]
    outputs: [remediation_result]
    timeout_sec: 300
    retry_policy:
      max_retries: 2
      backoff_sec: 15
    explain:
      what: Runs pre-approved remediation scripts (restart, scale, rollback)
      why: P3/P4 incidents with known fixes should resolve without human intervention

  - id: escalate
    type: api
    purpose: Page on-call engineer for high-severity incidents
    runtime:
      endpoint: pagerduty
      method: POST
      url: https://api.pagerduty.com/incidents
    inputs: [alert_id, severity, root_cause_hypothesis]
    outputs: [incident_id, responder]
    security:
      credentials: [PAGERDUTY_API_KEY]

  - id: human_investigate
    type: human
    purpose: On-call engineer investigates and applies manual fix
    role: sre_oncall
    inputs: [incident_id, root_cause_hypothesis, service]
    outputs: [actual_root_cause, fix_description]
    timeout_sec: 7200

  - id: apply_fix
    type: cli
    purpose: Deploy hotfix or configuration change to resolve the incident
    runtime:
      os: linux
      command: |
        ./deploy.sh --hotfix --service ${service} \
          --change "${fix_description}"
    inputs: [service, fix_description]
    outputs: [fix_deployed]

  - id: post_mortem
    type: human
    purpose: Conduct blameless post-mortem and document learnings
    role: engineering_manager
    inputs: [incident_id, actual_root_cause, fix_description]
    outputs: [post_mortem_doc, action_items]
    explain:
      what: Team reviews timeline, root cause, and improvement actions
      why: Post-mortems drive systemic reliability improvements

  - id: update_runbook
    type: agent
    purpose: AI updates runbook with new remediation steps from post-mortem
    runtime:
      provider: anthropic
      model: claude-sonnet-4-20250514
    inputs: [post_mortem_doc, action_items, service]
    outputs: [updated_runbook]

edges:
  - from: alert_triggered
    to: ai_triage
    mode: sequential

  - from: ai_triage
    to: severity_router
    mode: sequential

  # Low severity — auto-remediate
  - from: severity_router
    to: auto_remediate
    mode: conditional
    condition: severity in ["P3", "P4"]

  # High severity — escalate to human
  - from: severity_router
    to: escalate
    mode: conditional
    condition: severity in ["P1", "P2"]

  - from: escalate
    to: human_investigate
    mode: sequential

  - from: human_investigate
    to: apply_fix
    mode: sequential

  # Auto-remediation failure falls back to escalation
  - from: auto_remediate
    to: escalate
    mode: fallback
    condition: remediation_result == "failed"

  # Triage failure falls back to direct escalation
  - from: ai_triage
    to: escalate
    mode: error

  - from: apply_fix
    to: post_mortem
    mode: sequential

  - from: post_mortem
    to: update_runbook
    mode: sequential