AIAnalysis Approval Policy

The AIAnalysis controller uses a Rego policy to determine whether a proposed remediation requires human approval. This page documents the policy schema, input contract, and customization.

Overview

Property	Value
ConfigMap name	aianalysis-policies
Key	approval.rego
Mount path	/etc/aianalysis/policies/
Required	Yes -- chart fails at install if neither policies.content nor policies.existingConfigMap is set

Provisioning

Option A: Inline content via --set-file (recommended)

helm install kubernaut charts/kubernaut/ \
  --set-file aianalysis.policies.content=my-approval.rego \
  ...

Option B: Pre-existing ConfigMap

kubectl create configmap aianalysis-policies \
  --from-file=approval.rego=my-approval.rego \
  -n kubernaut-system

helm install kubernaut charts/kubernaut/ \
  --set aianalysis.policies.existingConfigMap=aianalysis-policies \
  ...

Input Contract

The approval policy receives this input from the AIAnalysis controller:

Field	Type	Description
environment	string	Namespace environment (production, staging, development, etc.)
confidence	float	LLM investigation confidence score (0.0--1.0)
confidence_threshold	float	Configurable threshold (default 0.8, via aianalysis.rego.confidenceThreshold)
remediation_target	object	LLM-identified remediation target (kind, name, namespace)
target_resource	object	Original alert target resource
detected_labels	map	Detected workload labels (snake_case keys: stateful, git_ops_managed, pdb_protected, hpa_enabled, helm_managed, network_isolated, service_mesh)
failed_detections	array	Detection fields that failed (e.g., ["gitOpsManaged"])
warnings	array	Investigation warnings

Output Contract

The policy must produce these outputs:

Output	Type	Description
require_approval	boolean	true to require human approval, false to auto-approve
reason	string	Human-readable explanation for the decision

Default Behavior

The reference policy (charts/kubernaut/examples/approval.rego) implements:

• Production environments: Always require approval (controlled via kubernaut.ai/environment=production namespace label)
• Sensitive resources (Node, StatefulSet): Always require approval regardless of environment
• Missing remediation target: Always require approval (safety default)
• Non-production: Auto-approved unless critical safety conditions are met

Risk Factors

The reference policy uses scored risk factors for reason generation:

Score	Condition
90	Missing remediation target
85	Sensitive resource kind (Node/StatefulSet)
80	Production + sensitive resource
70	Production environment

The highest-scoring factor determines the approval reason.

Customization

Always Require Approval

package aianalysis.approval
import rego.v1
default require_approval := true
default reason := "All remediations require manual approval"

Auto-Approve Everything (testing only)

package aianalysis.approval
import rego.v1
default require_approval := false
default reason := "Auto-approved (testing mode)"

Environment-Specific Thresholds

require_approval if {
  input.environment == "staging"
  input.confidence < 0.9
}

CRD Safety Gate

Block automated CRD modifications and require human approval. CRD changes cascade to all CRs of that type, making them high-risk for automated remediation:

require_approval if {
    input.remediation_target.kind == "CustomResourceDefinition"
}

risk_factors contains {"score": 95, "reason": "CRD modification — cascades to all CRs of this type"} if {
    input.remediation_target.kind == "CustomResourceDefinition"
}

For GitOps-managed clusters, combine with detected_labels — see CRD Safety Policy for the full pattern.

Hot-Reload

The approval policy supports hot-reload via fsnotify (~60s kubelet sync delay). If the new policy has a syntax error, the previous policy is kept and an error is logged.

Reference File

A complete reference policy is available in the chart: charts/kubernaut/examples/approval.rego