Jev Kit
A set of Jev-powered tools for Claude Code, including guarded tool calls, document classification, search, browser control, and review.
A set of Jev-powered tools for Claude Code, including guarded tool calls, document classification, search, browser control, and review.
Jev Kit includes several Jev-backed agent tools. In its document classifier, it sends up to 6,000 characters of page text to a first Choice over document families, then a second Choice over members of the chosen family with that earlier decision in state. The Python classifier reads the returned choice and confidence, applies its gates, and can request human review.
Choose the kind of document this page is, or the escape option when none fits.
The current document page text, truncated to 6,000 characters; taxonomy entries supply the candidate families.
Choose a specific document type within the decided family, or the escape option.
The same page text plus already_decided, which states the selected family and its definition; the member choices come from that family.
Take text from the document page and load the caller-supplied taxonomy.
Ask Jev for a document family and check the choice and confidence against the configured gate.
For a supported family, ask a second Jev Choice with the first decision in state, then check the narrower answer and its confidence.
Return the taxonomy label when supported; otherwise report an unmatched or low-confidence result for review.
Its two-stage document taxonomy keeps choices focused. Both stages include a not-in-this-list option, so a page outside the taxonomy need not be forced into a wrong class. The code treats low confidence as a review signal rather than a final label.