The Agent Skills Directory

[COMMAND_EXECUTION]: The skill uses Bash to execute local Git commands and repository-specific validation scripts (e.g., Makefiles, CI configurations). These operations are used to stage changes, create commits, and verify the integrity of the build process within the local environment.
[DATA_EXPOSURE]: Reads the local repository state, including diffs, history, and configuration files. It may also use the GitHub CLI (gh) to retrieve issue or pull request context to improve the quality of generated commit messages.
[SAFE]: Implements a sophisticated transactional backup mechanism. Before making any changes, it creates a snapshot of the working tree and index, storing the data in the system's temporary directory (/tmp) and as persistent Git references (refs/prepare-review-commits/). This ensures that the original state can be restored even if the agent's process is interrupted.
[SAFE]: Utilizes the disable-model-invocation: true flag, a platform-level restriction that prevents the AI agent from autonomously activating the skill. This ensures the skill only runs when explicitly requested by a human user.
[SAFE]: Explicitly prohibits remote-mutating operations such as git push, git fetch, or git rebase against remote branches, maintaining a local-only scope for all its operations.
[SAFE]: Employs indirect prompt injection defenses by requiring specific manual triggers and performing validation checks (such as git apply --check) before applying generated patches to the repository index.

prepare-review-commits