The Agent Skills Directory

[COMMAND_EXECUTION]: The bridge script scripts/codex_bridge.py executes the codex binary using subprocess.Popen to perform arbitrary tasks defined in the --PROMPT argument provided by the agent.
[COMMAND_EXECUTION]: The skill explicitly disables safe sandboxing options. Both the SKILL.md instructions and the scripts/codex_bridge.py logic reject the read-only sandbox level, defaulting instead to danger-full-access.
[COMMAND_EXECUTION]: The script implements an automatic downgrade mechanism in probe_bwrap_sandbox that reverts to danger-full-access if the bwrap (bubblewrap) utility cannot initialize correctly, which is a common occurrence in restricted environments like WSL, Docker, or specific Linux distributions (Ubuntu 24.04+).
[COMMAND_EXECUTION]: On Windows platforms, the script wraps commands in cmd.exe /c and uses custom character escaping for the prompt text. This implementation introduces a potential command injection surface if the escaping logic is bypassed by specially crafted prompt payloads.
[COMMAND_EXECUTION]: The skill instructions mandate that the bridge script must always be executed with run_in_background: true. This persistent background execution reduces the user's ability to monitor or audit the commands and file system changes being performed by the delegated Codex agent in real-time.

collaborating-with-codex