The Agent Skills Directory

[COMMAND_EXECUTION]: The skill is designed to invoke the codex CLI tool, which autonomously generates and executes shell commands based on LLM output.
[PRIVILEGE_ESCALATION]: The skill documents the danger-full-access mode and the --dangerously-bypass-approvals-and-sandbox flag. These features allow the underlying tool to perform network operations, access system-level files outside the workspace (such as /etc/), and execute commands without standard user confirmation or sandboxing protections.
[INDIRECT_PROMPT_INJECTION]: The skill's primary function is to ingest and process untrusted repository data including code, comments, and diffs.
Ingestion points: Workspace files and git diffs are read and passed to the codex exec command.
Boundary markers: The skill includes 'Security & Trust Boundaries' instructions and specific review handling protocols to separate data from instructions.
Capability inventory: The skill can invoke Bash(codex:*), Write, and Read tools, enabling arbitrary code execution and file modifications.
Sanitization: The instructions emphasize that repository content should be treated as data and that the agent must stop for confirmation before applying review findings.
[EXTERNAL_DOWNLOADS]: The skill provides instructions to install the Codex CLI using npm i -g @openai/codex or brew install codex. These installation sources belong to a well-known and trusted organization.

codex-skill