The Agent Skills Directory

[COMMAND_EXECUTION]: The skill facilitates the execution of the local codex CLI binary using subcommands like exec, resume, review, and apply. The instructions emphasize using non-interactive mode (exec) to ensure compatibility with the agent's environment.
[DATA_EXFILTRATION]: The skill documents the use of sandbox policies (read-only, workspace-write) to control filesystem access. It correctly defaults to read-only for analysis tasks and only recommends workspace-write when the user explicitly requests file modifications.
[PROMPT_INJECTION]: As an integration for a reasoning model, the skill is inherently a surface for indirect prompt injection if malicious instructions are present in the workspace files it analyzes. However, it incorporates human-in-the-loop safety via the approval_policy=on-request configuration.
Ingestion points: User prompts and workspace files (accessed via @path syntax or the -C working directory flag).
Boundary markers: Encourages the use of approval_policy=on-request to gate tool actions, though it does not specify explicit content delimiters in the instructions.
Capability inventory: The CLI tool can modify the workspace filesystem (workspace-write) and execute shell commands (shell_tool feature).
Sanitization: The skill acts as a direct pass-through for instructions to the CLI tool without performing local sanitization, relying on the underlying model's guardrails and CLI sandboxing.

codex