The Agent Skills Directory

[COMMAND_EXECUTION]: The skill executes various git commands using a user-provided range. It includes defensive instructions to validate the range string, specifically checking for shell metacharacters and strings starting with dashes (e.g., '--output=...') to prevent command injection vulnerabilities.
[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted data from git repositories, such as diff contents, filenames, and commit messages. While it includes robust boundaries, this remains an attack surface. 1. Ingestion points: Untrusted data enters the agent context through 'git diff --name-only', 'git diff --numstat', and direct file reading during the binding pass (SKILL.md). 2. Boundary markers: The skill contains an explicit 'Passive data' section that warns the agent to treat diff text, commit subjects, and file contents as non-instructional data that should not change the execution rules. 3. Capability inventory: The skill performs local file system analysis via git and can suggest commands (refuters) for the user to run manually, but it does not execute them automatically. 4. Sanitization: Input validation is performed on the range parameter to ensure it is a valid git revision or range shape before being passed to subshells.

select-sample