The Agent Skills Directory

[DYNAMIC_EXECUTION]: The learn_capture.py script contains a plugin architecture that allows loading local Python modules via the --sink-module argument. This facilitates custom data sinks for learning outcomes but requires the agent to be explicitly directed to a local file path.
[COMMAND_EXECUTION]: The skill interacts with Git through a whitelisted wrapper (gitio.py). This wrapper restricts usage to a specific set of read-only subcommands and explicitly prohibits flags or shell redirection characters that could modify the repository.
[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted documentation as part of its audit function. The potential risk of the agent following instructions embedded in these documents is mitigated by system instructions that enforce a read-only, observation-focused mode for its inspectors.
[SAFE]: The skill exhibits high security standards, including the use of os.path.realpath and os.path.commonpath to prevent path traversal in enforce_apply.py, and the use of secure low-level primitives like O_NOFOLLOW in history.py to prevent symlink attacks.

doc-steward