The Agent Skills Directory

[COMMAND_EXECUTION]: The skill performs its primary functions by executing shell commands, specifically git and gh. It constructs these commands using data from the conversation, user arguments, and the local repository state. While the instructions mandate shell-quoting for literal paths (git --literal-pathspecs add -- <paths>), the reliance on subprocess execution for state-changing operations is an inherent capability that requires monitoring.
[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted data from the local file system and git state, creating a surface for potential indirect prompt injection.
Ingestion points: Untrusted data enters the context through SKILL.md in the Commit section (reading current branch, status, staged/unstaged diffs, and untracked contents) and the Pull request section (inspecting committed branch diffs).
Boundary markers: The instructions do not define explicit delimiters or warnings to the agent to ignore instructions embedded within the file contents or diffs being processed.
Capability inventory: The skill possesses capabilities to write to the repository (git add, git commit), modify remote state (git push), and interact with external APIs (gh pr create).
Sanitization: The skill implements shell-quoting for pathspec handling to prevent basic command injection via filenames, but it lacks sanitization or filtering for the content of the diffs or files that could influence the LLM's generation of commit messages or PR descriptions.
[EXTERNAL_DOWNLOADS]: The installation instructions in README.md direct users to install the skill via npx skills add shihyuho/skills, which involves fetching content from the vendor's repository. Additionally, the SKILL.md file mentions the invocation of an external skill, git-commit-co-author, as part of the commit workflow.

commit-push-pr