The Agent Skills Directory

[COMMAND_EXECUTION]: The skill executes diagnostic scripts using the standard platform executor pattern.
Evidence: Operations in SKILL.md are defined using python3 .plan/execute-script.py for all audit functions.
[REMOTE_CODE_EXECUTION]: The scripts use dynamic path resolution to load internal runtime libraries within the skill bundle.
Evidence: scripts/permission_doctor.py and scripts/permission_common.py modify sys.path to include relative directories containing the claude_runtime module.
[PROMPT_INJECTION]: The skill processes untrusted data from user-defined settings files, which serves as an indirect prompt injection surface.
Evidence: The tool reads and analyzes content from .claude/settings.json, ~/.claude/settings.json, and marshal.json during the audit process.
[SAFE]: The skill implements proactive security audits by detecting dangerous permission patterns in configuration files.
Evidence: The SUSPICIOUS_PATTERNS list in scripts/permission_doctor.py flags risky configurations such as privilege escalation, root directory writes, and dangerous shell commands.

tools-permission-doctor