The Agent Skills Directory

[COMMAND_EXECUTION]: The script scripts/review.js executes the codex command-line utility to handle the review logic. Evidence: spawn("codex", ["app-server"]) in scripts/review.js.
[PROMPT_INJECTION]: The skill interpolates user-provided text into instructions for the Codex review agent and has an indirect prompt injection surface through the analysis of untrusted codebase content.
Ingestion points: The project directory (cwd) and the user-supplied review prompt.
Boundary markers: No explicit delimiters or boundary warnings are enforced in the prompt template.
Capability inventory: The review agent is initialized with a danger-full-access sandbox configuration.
Sanitization: No content filtering or validation is performed on the codebase or user input.
[DATA_EXFILTRATION]: The skill provides the Codex service with access to the project directory for the purpose of the review. Evidence: The thread/start message in scripts/review.js includes the cwd parameter pointing to the user's project.
[SAFE]: The identified behaviors, including command execution and code analysis, are intrinsic to the skill's primary purpose of performing automated code reviews.

codex-review