The Agent Skills Directory

[COMMAND_EXECUTION]: The skill utilizes the codex CLI to perform various tasks, including local edits and network operations. It instructs the agent to use profiles such as --profile edit (workspace-write) and --profile full (danger-full-access), which grants broad filesystem and network permissions to the sub-process. The skill consistently mandates the use of 2>/dev/null in all commands, which suppresses standard error output and may conceal operational warnings or security-related failures from the user.
[INDIRECT_PROMPT_INJECTION]: The skill is designed to delegate complex reasoning and code generation to external models, then ingest the resulting output, creating a surface for indirect instructions.
Ingestion points: Results from the codex workers are written to a temporary file created via mktemp and subsequently read into the agent's context.
Boundary markers: The instructions do not define specific delimiters or instructions for the agent to ignore potential malicious prompts embedded in the external model's output.
Capability inventory: The agent is authorized to execute go test, go build, and perform workspace writes based on the task delegation, providing a path for instructions in the output to be executed.
Sanitization: While the skill provides a best practice of verifying output via git diff --stat to catch unexpected file modifications, it lacks explicit sanitization or filtering of the text content returned by the external model before the agent processes it.

codex