What problem does it solve? AI agents and developers can accidentally run destructive shell commands like rm -rf, DROP TABLE, or git push --force, causing irreversible data loss or production impact. This Skill adds a PreToolUse safety hook that inspects every Bash command before it runs. ## Core Features & Use Cases - Two-tier protection: MEDIUM-tier patterns (recursive deletes, SQL DROP/TRUNCATE, force-push, git reset --hard, kubectl delete, docker prune) trigger an overridable warning, while HIGH-tier catastrophes (recursive delete of / or $HOME, force-push to the default branch) are hard-denied. - Evasion resistance: Detects shell obfuscation via IFS word-splitting and base64-to-shell decoding, parses hook payloads with a real JSON parser, and fails closed on unparseable input. - Configurable rules: Add custom warn patterns via global or per-project careful-patterns.txt files, which can only add rules, never suppress baseline warnings. - Use Case: Activate safety mode before debugging a live production system so any accidental kubectl delete or DROP TABLE prompts for confirmation first. ## Quick Start Ask the AI to enable careful mode so every destructive Bash command is checked and warned about before running.
