Jev Unpack hook

A Codex and Claude Code Stop hook that detects multi-topic, multi-ask agent messages and forces the agent to present issues one at a time, adding relevant context and missing information. It targets the out-of-the-blue handovers agents send when they need clarification, feedback or approval.

The README argues a Skill instructing the agent to spot such handovers is quickly forgotten after a few turns, while running a hook on every message with a general model is expensive and slow. Jev is used because it identifies which messages need pacing, fast and cheaply.

Two structured Choice questions: how many Asks the message contains (Zero, One, Two_or_more) and how many Points (Single_point, Two_or_three_points, More_than_three_points), run on the most recent stable official release, jev-latest.

Following the confidence-gated routing pattern: Two_or_more Asks plus More_than_three_points at 0.80 confidence or higher blocks the stop and supplies one-issue-at-a-time continuation instructions; the same classification at 0.50–0.80 blocks with a cautious confirmation prompt; otherwise the hook returns continue true.

No. The hook allows events with stop_hook_active to continue immediately, preventing a loop when the agent invokes the Stop hook again for its continuation.

Codex or Claude Code, a TypeSafe API key, and python3 version 3.9 or newer. The README's try-it section walks through copying the hook command, and short messages are ignored before any Jev call is made.

lucaminudel on GitHub, MIT licensed, with badges marking Codex and Claude Code support and Jev v1.13+. The README links a smharter.ai blog post on reducing cognitive fatigue in AI-to-human handover as the motivation.