The Agent Skills Directory

[PROMPT_INJECTION]: The skill performs indirect prompt injection analysis on itself. It processes untrusted code changes via git diff and file content without implementing boundary markers or instructions to ignore embedded commands. This creates a surface where code being reviewed could contain instructions (e.g., in comments) designed to subvert the agent's findings or recommendations.
Ingestion points: Untrusted source code is ingested via git diff and direct file access in SKILL.md.
Boundary markers: None. The delegated prompts for the code-reviewer and architect roles interpolate the code scope directly into instructions without delimiters.
Capability inventory: The skill uses git diff and has the ability to call external validation tools like mcp__x__ask_codex.
Sanitization: No sanitization or escaping is performed on the code content before it is processed by the LLM.
[EXTERNAL_DOWNLOADS]: The skill uses the mcp__x__ask_codex tool to consult an external model (Codex) for code validation. This involves sending the code under review to an external service provider. While this is a documented feature for high-stakes reviews, it represents an external data flow that users should be aware of.
[COMMAND_EXECUTION]: The skill executes git diff to identify changes in the local repository. This is a standard operation for a code review tool but involves spawning a subprocess to interact with the version control system.

code-review