A plan is not permission.

Plan contents

Every plan should be understandable before it reaches an executor.

• Proposed typed action and target resource
• Supporting evidence and observed failure
• Expected impact and risk
• Estimated blast radius and affected services
• Verification criteria and recovery behavior
• Idempotency key and audit context

Policy evaluation

The policy engine evaluates environment, resource scope, action type, risk, confidence, blast radius, cooldowns, concurrency, and maintenance windows. Generic apply, patch, and delete operations require explicit administrator confirmation.

Investigation
  → typed remediation plan
  → policy evaluation
  → denied | approval required | permitted
  → restricted executor

Approval and execution

The operator reviews what will change, why it was recommended, which resources are affected, which policy allowed or denied it, and how success will be measured. Approved actions are sent to a separately authenticated restricted executor.

Verification and recovery

Tracey checks workload readiness, replica recovery, rollout state, error-rate changes, latency changes, and minimum evidence requirements. If verification detects a regression, the permitted recovery path runs and is verified again.

A successful Kubernetes API response is not proof that production recovered.

Operating modes

Start with Observe or Recommend. Production teams should begin with Approval, then move to Guarded autopilot only for reversible, allowlisted actions with proven verification and rollback criteria.