The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted text data captured from browser environments (accessibility tree or visible text) and provides this content to an LLM-based model for evaluation. Ingestion points: scripts/jev-call.mjs ingests untrusted page state via standard input or from a file path specified by the --state-file command-line argument. Boundary markers: The untrusted state is included directly as a JSON property state.page_text in the request body to the Jev API, lacking specific delimiters or instructions for the receiving model to ignore embedded markers. Capability inventory: Within scripts/jev-call.mjs, the skill uses the global fetch API to transmit data to the Jev API at api.typesafe.ai. Sanitization: There is no sanitization or filtering of the ingested page content before it is forwarded to the AI model, creating a surface where malicious UI content could attempt to manipulate the verification judgment.
[COMMAND_EXECUTION]: The skill defines a Node.js utility script (scripts/jev-call.mjs) which is executed by the agent to perform network-based assertions and process semantic verdicts.

jev-assert