The Agent Skills Directory

[COMMAND_EXECUTION]: The skill constructs and executes shell commands to interact with the codex and git command-line interfaces. Scripts such as scripts/codex-ask.sh and scripts/codex-review-changes.sh use safe array expansion patterns to prevent command injection when passing user-defined tasks or file paths to the underlying tools.- [EXTERNAL_DOWNLOADS]: References the external codex CLI tool, provided by OpenAI, as a mandatory dependency. Documentation directs users to established sources for installation, such as Homebrew or the official GitHub repository, which are recognized as safe sources for development tools.- [DATA_EXFILTRATION]: The skill is designed to transmit repository data, including file contents and git diffs, to the external OpenAI Codex service for analysis. This data flow is documented, requires the presence of an authenticated codex installation, and is subject to user-configurable sandbox and approval policies designed to mitigate unauthorized access to local resources.- [PROMPT_INJECTION]: As an analysis tool for repository content, the skill represents an indirect prompt injection surface when processing untrusted code.
Ingestion points: Untrusted data enters the agent context via git diff outputs and file reads performed in scripts/codex-ask.sh and scripts/codex-review-changes.sh.
Boundary markers: The skill uses structured templates (e.g., templates/delegation-prompt.md) with explicit headers like 'Task:' and 'Context:' to separate instructions from data.
Capability inventory: The skill has the ability to execute CLI commands (codex, git) and read local files.
Sanitization: The tool relies on the codex CLI's native read-only sandbox mode and user confirmation gates rather than performing complex input sanitization on the reviewed code.

codex-cli