The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill facilitates the download of configuration and installation scripts from OpenAI's official chatgpt.com domain. These downloads include install.sh for Unix-like systems and install.ps1 for Windows, which are standard for the Codex CLI setup.
[REMOTE_CODE_EXECUTION]: The installation process involves executing scripts fetched from the chatgpt.com service. The skill directs the agent to obtain host-policy authorization before execution and prefers installing to user-local directories like ~/.local/bin.
[COMMAND_EXECUTION]: The core logic involves running the codex binary with various subcommands and flags. This includes executing model-generated prompts via codex exec, performing code reviews via codex review, and managing environments through worktree and session subcommands.
[INDIRECT_PROMPT_INJECTION]: The skill ingests user-defined prompt strings and passes them to the codex exec command for processing. While this creates a vulnerability surface for indirect injection attacks, the skill incorporates several defensive measures.
Ingestion points: User-supplied prompts used in SKILL.md and referenced in references/streaming.md.
Boundary markers: The skill defines specific sandbox modes (read-only, workspace-write) to limit the impact of model-generated shell commands. It also mandates 'waivers' for bypassing standard approval steps.
Capability inventory: The underlying codex tool has significant capabilities, including shell execution, file system modification, and network access (implied by MCP and plugin management).
Sanitization: Prompt content is literal-quoted when building the command-line arguments to prevent shell meta-character injection during the initial process launch.

codex