The Agent Skills Directory

[COMMAND_EXECUTION]: The skill invokes local CLI tools (git and claude) using subprocess.run. These calls are protected by resolving absolute paths, using allowlisted environments, and employing file-descriptor-based directory changes to mitigate path-based vulnerabilities.
[DATA_EXFILTRATION]: While the skill shares repository data with an external model, it implements extensive safeguards. It automatically filters out sensitive files and directories (e.g., .ssh, .aws) and performs a deep regex-based content scan to identify and block credentials before they are sent.
[PROMPT_INJECTION]: The skill addresses potential indirect prompt injection by using a system prompt that explicitly instructs the sub-agent to treat repository content as evidence only and to ignore any instructions found within the files.
[SAFE]: The skill exhibits professional-grade security engineering, including protections against Time-of-Check to Time-of-Use (TOCTOU) attacks and automated contract verification that monitors the sub-agent's behavior for compliance with safety rules.

consult