The Agent Skills Directory

[COMMAND_EXECUTION]: The skill is designed to construct and execute complex shell commands using a 'codex' CLI tool. It includes patterns for autonomous execution via the --full-auto flag and pipes user-provided prompts directly into shell commands (echo "prompt" | codex exec ...).\n- [PRIVILEGE_ESCALATION]: The skill explicitly promotes the use of the --sandbox danger-full-access flag to permit 'network or broad access'. Although it includes a directive to ask the user for permission, this capability grants the tool significant system privileges. Furthermore, the skill instructs the agent to append 2>/dev/null to all commands, which suppresses the stderr stream, effectively concealing warnings, errors, or logs from the user's view.\n- [INDIRECT_PROMPT_INJECTION]: The skill possesses a significant attack surface for indirect prompt injection.\n
Ingestion points: The skill ingests user prompts via stdin and processes code in the workspace for analysis and refactoring tasks (SKILL.md).\n
Boundary markers: There are no mentioned boundary markers or 'ignore' instructions to separate processed code/data from the agent's control logic.\n
Capability inventory: The skill has the capability to execute subprocesses (codex exec), write to the workspace (--sandbox workspace-write), and access the network (--sandbox danger-full-access).\n
Sanitization: There is no mention of sanitizing or escaping the data/code processed by the tool before it influences the agent's actions.

codex