The Agent Skills Directory

[COMMAND_EXECUTION]: The skill utilizes a Python script (scripts/review.py) that executes multiple shell commands using subprocess.run, including git, grok, and claude. These commands are used to inspect the repository state and invoke external LLM interfaces. The command arguments are dynamically constructed from the repository paths and git ranges provided as input to the skill.- [INDIRECT_PROMPT_INJECTION]: The skill ingests untrusted code and history from the target repository to create a review subject. This content is then passed to external LLM sessions (Grok or Claude Code). Malicious instructions embedded in the repository's files or commit messages could attempt to deceive the reviewer or influence the audit results.
Ingestion points: Local source files and git history from the repository being reviewed.
Boundary markers: The prompt generator uses markdown headers and instruction blocks to separate user intent from untrusted repository content.
Capability inventory: The skill can execute local commands (git), read any file in the repository, and communicate with external LLM CLIs.
Sanitization: Uses list-based arguments for subprocess.run and sanitizes paths using Path.resolve().- [DYNAMIC_EXECUTION]: The skill dynamically assembles prompt files and subject diff files at runtime. These generated artifacts serve as the primary input for the independent review processes executed via external CLI tools.

ship-check