The Agent Skills Directory

[SAFE]: The skill implements robust data protection policies and a strict blacklist of sensitive files.
Evidence: Found in references/safe-to-delete.md (Blacklist section) and SKILL.md (Highest Discipline section).
Details: The skill explicitly forbids the deletion of source code (src/), credentials (.env, .ssh, .pem, .key), databases (*.db, *.sqlite), and version history (.git). It also classifies infrastructure-related repositories as protected assets even if they have low activity.
[COMMAND_EXECUTION]: The skill uses standard system commands for filesystem analysis and management.
Evidence: git log, du -sh, mv, and rm are referenced in SKILL.md and references/archive-protocol.md.
Details: These commands are used to identify large directories and move or delete them. The risk of accidental or malicious data loss is mitigated by the instruction that the agent must list every path and action for user confirmation before executing any command. The skill uses quoted variables (e.g., "<项目>") to prevent simple path injection vulnerabilities.
[SAFE]: The skill includes extensive self-correction and boundary testing in its evaluation suite.
Evidence: evals/evals.json contains test cases specifically designed to verify the agent refuses to delete files without confirmation, even when told 'I trust you' or 'don't ask.'

ray-cleanup