The Agent Skills Directory

[COMMAND_EXECUTION]: The skill executes the codex CLI tool to perform tasks. The wrapper script scripts/codex-session.sh uses codex exec and codex exec resume to run arbitrary instructions provided as strings. While these are documented CLI features, they allow for substantial automated system interaction.
[PRIVILEGE_ESCALATION]: The skill supports the danger-full-access sandbox mode for the Codex CLI. As documented in the skill's own README.md and SKILL.md, this mode removes sandboxing and approvals, granting the external model full system access. Although it is labeled as a 'last resort' and requires user confirmation, the skill provides the functional path to bypass standard security boundaries.
[INDIRECT_PROMPT_INJECTION]: The skill acts as a bridge between Claude and Codex. It is designed to ingest codebase data (files, project structure) and pass it to Codex. This creates an attack surface where malicious instructions embedded in the codebase being analyzed could potentially influence the behavior of the Codex subagent, leading to unintended file edits or command execution.
[COMMAND_EXECUTION]: The skill instructions in SKILL.md advise the agent to run git status and git diff to verify changes made by Codex. While this is a safety best practice for auditing, it confirms that the skill is intended to perform persistent, automated modifications to the local environment.

codex-delegate