The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The documentation recommends installing the Codex CLI tool from chatgpt.com via shell scripts or package managers. These references target official domains associated with OpenAI.
[REMOTE_CODE_EXECUTION]: The README provides installation commands that pipe remote scripts directly into the shell (sh and PowerShell) from chatgpt.com. These patterns are used for the initial setup of the required CLI tool from a trusted source.
[COMMAND_EXECUTION]: The scripts/run.ts script uses spawnSync to execute the codex binary found on the system's PATH or specified via the CODEX_BIN environment variable. This is used to launch the codex exec command which performs the GUI automation.
[PROMPT_INJECTION]: The skill facilitates indirect prompt injection by design. The run.ts script interpolates user-provided text from the --task and --extra arguments directly into the instructions sent to the Codex sub-agent.
Ingestion points: The task and extraInstruction parameters in scripts/run.ts are sourced from command-line arguments.
Boundary markers: No specific delimiters are used to wrap the user-supplied task within the sub-agent prompt.
Capability inventory: The sub-agent (codex exec) has capabilities for GUI interaction (clicking, typing), file system access (configurable sandbox), and process inspection.
Sanitization: The script performs shell quoting for display purposes in dry-runs, but the content of the instructions is passed to the sub-agent without filtering.

codex-computer-use