The Agent Skills Directory

[COMMAND_EXECUTION]: The script scripts/codex-review.sh invokes codex exec with the --yolo flag enabled via the CODEX_YOLO environment variable. This configuration is designed to execute model-generated actions or commands without interactive user confirmation.\n- [PROMPT_INJECTION]: The skill is susceptible to indirect prompt injection because untrusted task descriptions are interpolated directly into prompts sent to the external Codex service in scripts/codex-review.sh.\n
Ingestion points: Command-line arguments passed as the DESCRIPTION variable to scripts/codex-review.sh.\n
Boundary markers: None. User-controlled content is embedded directly in the prompt text without delimiters.\n
Capability inventory: Execution of shell commands via codex exec --yolo and access to the local filesystem.\n
Sanitization: No escaping, filtering, or validation is performed on the input content before interpolation.\n- [EXTERNAL_DOWNLOADS]: The skill requires the global installation of the @openai/codex CLI package via npm, which serves as the interface for executing model-generated tasks.\n- [COMMAND_EXECUTION]: The load_config function in scripts/common.sh uses the source command on the project's .codex-review/config.env file. This pattern allows for arbitrary shell command execution if the configuration file is modified by a malicious process or actor.

codex-review