The Agent Skills Directory

[COMMAND_EXECUTION]: The skill orchestrates local command-line tools including git, claude, and codex. This execution is central to the skill's purpose of capturing repository state and delegating review tasks. The provided runner script (scripts/run_claude_review.py) uses subprocess.Popen with argument lists (avoiding shell interpolation) to manage these processes safely.
[INDIRECT_PROMPT_INJECTION]: The skill is designed to process external, potentially untrusted code and documents, which creates an indirect injection surface. It addresses this through several defensive layers:
Instructions explicitly direct the agent to treat all artifact content as data, not instructions.
The reviewer model is launched with --safe-mode, disabling customizations and automatic instruction discovery.
Tool access for the reviewer is strictly limited to Read, Glob, and Grep, preventing the delegated model from modifying files or executing shell commands.
[CREDENTIALS_UNSAFE]: The skill includes documentation and logic for managing model authentication via environment variables like ANTHROPIC_API_KEY. The runner script provides an --auth-mode subscription feature that explicitly strips these sensitive variables from the reviewer's environment to prevent unintended API billing or credential exposure during the review process.
[EXTERNAL_DOWNLOADS]: The skill references official documentation and CLI tools from Anthropic and OpenAI. These are well-known services relevant to the skill's core functionality and are documented for user reference.

cross-model-review