• Read-only Kubernetes API access via your kubeconfig: pod/node/event listing, NodeClaim/NodePool/EC2NodeClass objects, and Karpenter controller logs. It never mutates cluster state.
• Read-only AWS API calls (optional): ec2:DescribeInstanceTypes, ec2:DescribeSpotPriceHistory. No writes, no IAM, no instance access.

~/.kexplain/<cluster>/ contains harvested controller logs, Karpenter-related events, and object snapshots. These include node names, private IPs, pod names, and EC2 instance IDs. That is infrastructure metadata, not workload data or secrets. Treat the store directory with the same sensitivity as your kubeconfig. It is intentionally kept outside the repo; .gitignore also blocks it and all *.jsonl harvests as a second layer.

• Harvest stores, logs from provisioning runs (infra/*.log), since they contain account IDs and instance IDs.
• Kubeconfigs, AWS credentials, .env files, private keys.

This is a diagnostic tool; if you find an issue (e.g. the store capturing more than described above), please open an issue.