The Agent Skills Directory

[COMMAND_EXECUTION]: The skill installs a bash script (block-dangerous-git.sh) as a PreToolUse hook. This script is designed to intercept and validate shell commands before they are executed by the AI agent. This is a security-enhancing feature intended to prevent destructive Git operations like git push --force or git reset --hard.
[DATA_EXPOSURE]: The bundled sidecar.sh utility manages project metadata and 'review stamps' within the ~/.claude/sidecars/ directory. It performs local file system operations and Git commands to maintain a private record of project states, which is consistent with the skill's purpose of managing infrastructure safety.
[EXTERNAL_DOWNLOADS]: To verify if a pull request review is still current, the hook uses the GitHub CLI (gh) to fetch pull request metadata from GitHub. These network operations are limited to authenticated GitHub API calls to retrieve specific OIDs for verification.
[INDIRECT_PROMPT_INJECTION]: The hook processes data from GitHub pull request bodies to extract reviewed-at markers. The analysis indicates that this data is parsed strictly using sed to find a specific 40-character hexadecimal pattern (a Git SHA). This strict validation prevents instructions embedded in a PR body from influencing the hook's logic or bypassing the security checks.

git-guardrails-claude-code