The Agent Skills Directory

[COMMAND_EXECUTION]: The skill employs standard command-line utilities including git, grep, awk, wc, and stat to perform quantitative analysis on project metadata, file sizes, and modification timestamps. This behavior is restricted to local auditing of the agent's own state.
[DATA_EXPOSURE]: The skill is designed to read platform-specific configuration files such as ~/.claude/scheduled_tasks.json and ~/.codex/config.toml to identify existing automation tasks. While these files are sensitive, the skill does not include any network operations or data exfiltration logic.
[REMOTE_CODE_EXECUTION]: Several external GitHub repositories and blogs are cited as reference material for 'memory hygiene'. The skill explicitly instructs the agent and user not to execute code from these sources verbatim, framing them strictly as educational resources.
[COMMAND_EXECUTION]: A specialized Python script is included for forensic analysis of session logs (.jsonl files). The script uses the json module to safely parse historical data and extract reports without using eval(), exec(), or other dynamic execution sinks on the ingested data.
[SAFE]: The skill utilizes a user-in-the-loop pattern, placing all destructive actions and prompt modifications behind an AskUserQuestion requirement, ensuring that the user maintains control over behavioral changes to their autonomous loops.

clean-my-loop