The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill possesses a potential attack surface by ingesting and acting upon untrusted data from the repository environment.
Ingestion points: The agent reads CODEX_AGENTS.yml for configuration, INDEX.md files for progress tracking, and general repository instructions.
Boundary markers: There are no explicit instructions for the agent to treat external file content with lower trust or to use delimiters to prevent command hijacking.
Capability inventory: The skill can write to the filesystem, execute arbitrary shell commands for testing, and delegate tasks to sub-agents.
Sanitization: The skill does not describe processes for validating or sanitizing configuration schema values before application.
[DYNAMIC_EXECUTION]: The skill is designed to generate and modify executable code and tests as its primary function.
It performs runtime verification by executing the code it produces or modifies to ensure it meets requirements.
It uses "experiments" and "discriminating checks" which involve executing dynamically generated logic.
[COMMAND_EXECUTION]: The skill frequently interacts with the system shell to perform its tasks.
It executes repository-required checks, regression suites, and environment inspections.
It manages delegated tasks and integration with the orchestrator, involving process coordination.

implement