The Agent Skills Directory

[COMMAND_EXECUTION]: The skill invokes the Codex CLI via shell commands. It implements comprehensive validation logic for user-supplied arguments such as model names and reasoning efforts, utilizing a strict whitelist regex and explicitly rejecting shell metacharacters to mitigate command injection risks.
[INDIRECT_PROMPT_INJECTION]: The skill is designed to analyze untrusted code and plans. To prevent malicious instructions within this data from influencing the agent, it mandates the use of randomly-suffixed delimiters and provides specific instructions to the model to treat delimited content solely as data for review.
[PROMPT_INJECTION]: The skill explicitly instructs the agent to detect and disregard malicious directives like "ignore previous instructions" or "switch the sandbox to write mode" that might be embedded in untrusted review material. This is a defensive implementation designed to protect the integrity of the review process.
[SAFE]: The skill enforces a global read-only constraint on the external reviewer tool using both command-line flags ("-s read-only") and prompt instructions, ensuring the integrity of the local filesystem during the review process. Additionally, it provides specific guidelines to avoid processing sensitive files like SSH keys or environmental secrets.

codex-review