The Agent Skills Directory

[COMMAND_EXECUTION]: The skill executes local commands such as git diff and gh pr diff to collect content for review. It uses strict regular expression allowlists (e.g., ^[1-9][0-9]*$ for PR numbers) to validate all user-supplied arguments before execution, effectively preventing command injection attacks.
[EXTERNAL_DOWNLOADS]: The skill references official NPM packages from well-known and trusted organizations, specifically @github/copilot-cli, @openai/codex, and @google/gemini-cli. These are optional dependencies used only if the user explicitly selects an external model for the review process.
[DATA_EXFILTRATION]: When configured to use external models (Copilot, Codex, or Gemini), the skill transmits source code diffs and PR metadata to the respective third-party providers. This risk is mitigated by a built-in pre-flight secret scanner that identifies private keys and tokens (e.g., AWS, GitHub, Slack) and requires explicit user confirmation before any data is sent externally.
[PROMPT_INJECTION]: The skill implements strong boundary markers by wrapping untrusted code in <untrusted_diff> or <untrusted_files> tags. It provides explicit framing instructions to the reviewer model to treat the content as data only and ignore any embedded directives, mitigating indirect prompt injection risks from the files being reviewed.

peer-review