The Agent Skills Directory

[REMOTE_CODE_EXECUTION]: The skill provides a command to fetch and execute an installation script from OpenAI's official domain (chatgpt.com) using a piped shell execution pattern.
[DYNAMIC_EXECUTION]: The skill generates and executes content at runtime through several mechanisms:
It creates a local shell wrapper script in the user's home directory (~/.local/bin/codex) to enable the CLI if it is bundled with the ChatGPT application.
It utilizes temporary files (mktemp) to assemble and pass multi-line shell prompts and repository work orders to the CLI.
[INDIRECT_PROMPT_INJECTION]: The skill processes potentially untrusted external data (base repository code, CI output, and git history) while maintaining high-privilege tool capabilities.
Ingestion points: Accesses files and configuration (e.g., AGENTS.md) within the active project directory (SKILL.md).
Boundary markers: Uses a delegated sub-agent architecture for task execution, although it notes that saturated sessions may conflate data with instructions (SKILL.md).
Capability inventory: Involves extensive shell command execution, repository modification, and unrestricted execution permissions via the --yolo flag (SKILL.md).
Sanitization: Strictly mandates that the parent agent (Claude) must review all diffs, verify test results, and perform final verification before code is merged or landed (SKILL.md).
[COMMAND_EXECUTION]: The skill instructs the agent to use the --yolo flag with the Codex CLI, which bypasses typical approval prompts for command execution within the delegated environment.
[OBFUSCATION]: The skill includes logic to decode hex-encoded model identifiers from the system environment (e.g., echo <hex> | xxd -r -p) to determine model provenance and determine whether to delegate tasks. This is a functional requirement and not a deceptive technique.

codex-first