The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: Recommends the installation of the @openai/codex CLI tool, which is a package maintained by a trusted organization.
[COMMAND_EXECUTION]: Utilizes the codex binary and standard utilities like git and mktemp to orchestrate coding subtasks. These actions are aligned with the skill's purpose of managing a subagent workflow.
[DATA_EXFILTRATION]: Includes explicit instructions to prevent the agent from reading or copying sensitive authentication files, such as ~/.codex/auth.json, which mitigates the risk of credential theft.
[PROMPT_INJECTION]: The skill implements a delegation pattern where output from a subagent is ingested into the parent context.
Ingestion points: Task results are written to a temporary file and read using the cat command.
Boundary markers: The skill relies on natural language instructions for success criteria rather than technical delimiters.
Capability inventory: Delegated tasks include file system modifications within the workspace and optional network access via command-line flags.
Sanitization: The instructions mandate that a human user reviews all changes made by the subagent before the task is finalized.

codex-subagent