The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill fetches plugins and assets from the vendor's GitHub repository (github.com/romeodiaz/guava-stack) and utilizes the npx utility to execute the skills package from the npm registry for management tasks.
[COMMAND_EXECUTION]: It performs several shell operations to manage tool configurations, including claude plugin and codex plugin commands for marketplace updates, and executes environment-specific CLI tools like agy, claude, and codex to verify model accessibility.
[INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest data from external user-controlled 'hubs' which serve as the primary source of truth for the agent's instructions and context.
Ingestion points: Untrusted data enters the agent context from files like context.md, rules.md, and the skills/ directory within the configured hub (Notion, Google Drive, OneDrive, or local storage).
Boundary markers: The skill contains explicit instructions to ignore any content within imported data that attempts to override safety rules, bypass approvals, or change system access.
Capability inventory: The skill possesses the capability to modify local agent configuration files (~/.claude/settings.json, ~/.codex/config.toml), update standing instruction files (CLAUDE.md, AGENTS.md), and execute shell commands.
Sanitization: Imported skills are subject to an integrity check via the skills.sh audit database, or a mandatory review by a 'head-chef' model to explain the skill's behavior to the user before it is enabled.
[DYNAMIC_EXECUTION]: The skill dynamically loads and executes instruction sets (skills) from the hub folder. It also performs runtime installation and updates of agent plugins using marketplace CLI commands.

guava-hub