The Agent Skills Directory

[COMMAND_EXECUTION]: Uses the codex exec command to perform an AI-driven review of skill proposals. The command is executed with a --sandbox read-only flag to limit its access to the environment.
[COMMAND_EXECUTION]: Executes make link-skills in the workspace to manage symbolic links for the skills after they have been edited or renamed.
[COMMAND_EXECUTION]: Uses mktemp to create temporary files for storing analysis data and audit results during the review process.
[DATA_EXFILTRATION]: Reads local SKILL.md files and transmits metadata (names, descriptions, triggers) to the external Codex service for review. This behavior is the primary intended function of the skill.
[PROMPT_INJECTION]: Processes the contents of other SKILL.md files, which presents a surface for indirect prompt injection. This is mitigated by the AskUserQuestion gate, which ensures that no changes are applied to the filesystem or agent configuration without explicit manual approval from the user.

cmd-skills-review