The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted data from the user's repository which could contain malicious instructions.
Ingestion points: Reads output of git diff, git status, and package.json in SKILL.md to analyze changes.
Boundary markers: Instructions explicitly require the agent to review changes against a checklist (references/diff-review-checklist.md) and scan for secrets before committing, providing a manual verification step.
Capability inventory: The skill can stage files, create commits, push to remote repositories, and publish packages to the npm registry.
Sanitization: Relies on the agent's ability to identify and manually exclude 'suspected secrets' or 'stray files' as instructed in the 'Analyze Changes' section.
[COMMAND_EXECUTION]: Executes shell commands to manage the Git workflow and run package scripts.
Evidence: Uses git commit, git push, and npm publish to manage the repository state.
Evidence: Executes project-specific scripts such as <pm> test, <pm> build, and <pm> typecheck as part of the pre-publish verification process.
[DATA_EXFILTRATION]: Performs network operations to push code and publish packages to external registries.
Evidence: The skill executes git push and npm publish to move local data to remote registries (e.g., GitHub, npm). This behavior is the stated primary purpose of the skill and is gated by user arguments ('push' or 'publish'). The skill includes explicit instructions to scan for and exclude secrets or credentials from being committed and pushed.

commit