Input and output contain message text in plaintext. The tool does not contact a model or network service. Use approved local paths, and review generated bundles before sharing because every level may preserve sensitive content from the canonical message. Output is no-overwrite by default; use --force only after checking the exact target.