The Agent Skills Directory

[COMMAND_EXECUTION]: The skill employs dynamic context markers in its main instruction file to gather local environment information, such as git status and branch details. These commands are benign, run locally, and do not incorporate external inputs.
[DATA_EXFILTRATION]: Interaction with the GitHub CLI (gh) is used for legitimate pull request management tasks. This activity is governed by a 'backlog throttle' that halts execution if too many PRs are already open, preventing automated spam.
[SAFE]: A comprehensive exclusion policy in reference/exclusions.md protects critical security and configuration files (e.g., GitHub workflows, agent settings, and git hooks) from being modified by the skill.
[SAFE]: The skill relies on locally installed verification tools (like shellcheck, markdownlint, and pwsh) and project-specific test suites to ensure all structural changes are behavior-preserving.
[PROMPT_INJECTION]: The skill processes repository files (Ingestion points: files in glob-scoped lanes) which could serve as an indirect prompt injection surface. Boundary markers and sanitization are absent, but the skill's risks are minimized by its focus on constrained structural 'watch-for' patterns and the requirement for successful build/test verification (Capability inventory: file editing, PR creation).

tidy