The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted external data which could contain malicious instructions designed to influence the agent's behavior.
Ingestion points: Codebase areas, PR diffs, build outputs, and live web content via browser navigation tools (SKILL.md, resources/execution-protocol.md).
Boundary markers: No specific delimiters or "ignore" instructions are defined for processing external code or web data to prevent instructions within that data from being followed.
Capability inventory: The agent can execute shell commands (audit tools, curl) and perform browser interactions including script evaluation (resources/execution-protocol.md).
Sanitization: No explicit sanitization or filtering of the ingested content is mentioned before it is processed by the agent.
[COMMAND_EXECUTION]: The skill is designed to execute various shell commands for diagnostic, auditing, and testing purposes.
Evidence: Instructions include running npm audit, bandit -r ., lighthouse <url>, and starting local servers via bun run dev or uv run manage.py runserver (SKILL.md, resources/execution-protocol.md).
[DYNAMIC_EXECUTION]: The skill utilizes runtime script execution within a browser environment to verify application state.
Evidence: The runtime verification protocol employs evaluate_script(function) to inspect DOM state, global variables, and computed styles (resources/execution-protocol.md).

oma-qa