The Agent Skills Directory

[COMMAND_EXECUTION]: The skill executes a non-standard binary named codex using the review command. This tool is granted access to the entire codebase through stdin and various command flags (e.g., --uncommitted, --base). Because codex is not a standard system utility and is not a known official CLI tool for this purpose, its execution presents a significant risk of data exfiltration or unauthorized code processing if an untrusted or malicious version of the binary is present on the user's system.
[EXTERNAL_DOWNLOADS]: The instructions direct the user to install the required tool via brew install codex or by visiting the OpenAI GitHub repository. While referencing a well-known organization, the specific repository does not currently host the described CLI tool, and the codex package available in standard package managers like Homebrew refers to an unrelated indexing utility. This discrepancy creates a risk where a user might be misled into installing and granting codebase access to an incorrect or potentially malicious third-party binary.
[INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest the entire application codebase and pass it to an external model for review, which establishes a surface for indirect prompt injection. If the code being reviewed contains malicious instructions or adversarial patterns designed to influence the reviewer, it could lead to biased or harmful analysis results.
Ingestion points: The full codebase is passed to the tool via stdin or diff flags in SKILL.md.
Boundary markers: No explicit delimiters or instructions to ignore embedded commands are included in the shell pipeline.
Capability inventory: The skill utilizes shell execution for the codex binary and directory creation via mkdir.
Sanitization: No validation or sanitization of the codebase content is performed before it is processed by the external tool.

codex-review