The Agent Skills Directory

Command Execution Safeguards: The skill implements robust measures to prevent command injection when interacting with the Git CLI. It specifies the use of --literal-pathspecs and the -- delimiter to distinguish options from path arguments. Additionally, it requires that all paths be safely quoted for the active shell and mandates stopping the process if a path cannot be represented safely.
Data Exposure Prevention: To protect against accidental leakage of sensitive information, the instructions state that path lists presented to the user should not include file contents or suspected sensitive values. This reduces the risk of exposing credentials or secrets that might be present in file names or metadata.
Restricted Execution Environment: The skill explicitly restricts its operations to local repositories, specifically stating that no 'push' or other remote operations should occur. This effectively mitigates risks associated with unintended data exfiltration to external servers.
Human-in-the-loop Protocol: A mandatory two-step confirmation process is enforced, requiring the user to verify both the selection of paths for staging and the final set of staged changes before a commit is created. This prevents the agent from acting autonomously in a way that could lead to unintended repository modifications.

git-commit