The Agent Skills Directory

[SAFE]: The skill performs a legitimate audit of the agent's memory files. It includes sophisticated safety mechanisms such as atomicity in snapshots using os.O_EXCL and restrictive file permissions (0o600).
[SAFE]: Data access is limited to the local filesystem for the purpose of reading memories and verifying them against project repositories. There is no evidence of network exfiltration or unauthorized data exposure.
[SAFE]: Command execution is limited to standard development tools (git, gh) and internal Python scripts for file management. The instructions include rigorous escaping for shell variables and glob patterns to prevent injection vulnerabilities.
[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted content from memory files and session transcripts. While this presents an attack surface where a malicious memory could attempt to influence the subagents, the risk is mitigated by the structured JSON output required from subagents and the human-in-the-loop approval process before any edits are applied.

audit-memory