The Agent Skills Directory

Command Execution: The skill is built around executing local CLI commands using the plugin-eval utility. It uses various subcommands like analyze, start, and benchmark to process local skill directories. This is a standard pattern for developer-oriented tools integrated into an agent environment.
Local File System Access: The workflow specifies resolving paths in the user's home directory (~/.codex/skills/) and local repository paths to locate target skills. This access is limited to identifying and analyzing skill configuration files.
Indirect Prompt Injection Surface: The skill evaluates external SKILL.md files which are ingested into the analysis workflow.
Ingestion points: Files are read at <skill-path> during the plugin-eval analyze step in SKILL.md.
Boundary markers: There are no explicit instructions to the agent to treat the content of the analyzed files as untrusted data.
Capability inventory: The skill can execute shell commands via the plugin-eval tool.
Sanitization: No explicit sanitization of the analyzed file content is mentioned in the instructions, though the use of a dedicated analysis tool typically handles data parsing.

evaluate-skill