Claude x Jev
A Claude Code plugin that sends quick typed judgments to Jev and leaves the writing to Claude.
A Claude Code plugin that sends quick typed judgments to Jev and leaves the writing to Claude.
The plugin ships presets. Each preset declares a state, typed questions and thresholds: the tool gate uses tool, tool input, working directory and task, while email triage uses sender, subject and body. Jev answers each question on its own, and the plugin's rules reconcile the answers before a gate, route or classify command acts on them.
Does the action only read or change files inside the working directory and can it be undone?
The pre-tool-gate hook supplies the tool name, its input, the working directory and the current task. Every question in the preset is a condition that must hold; the plugin blocks or escalates when one fails.
Is the action a reasonable next step toward the current task?
The pre-tool-gate hook supplies the tool name, its input, the working directory and the current task. Every question in the preset is a condition that must hold; the plugin blocks or escalates when one fails.
The skill installs a Claude Code hook and a set of presets that name their state fields and questions.
Before Claude runs a tool, the hook builds the preset state from the tool name, its input, the working directory and the task.
Each gate question is sent as a separate typed question and answered with a calibrated probability.
The plugin compares the answers with its thresholds and rules, then allows, blocks or escalates the call.
Claude writes the code and explanation; the typed decision only gates the small choices.
It puts a fast typed judgement inside an agent loop without asking a language model to reason about each small decision, and it keeps the threshold and rule layer in the plugin so a preset can be tuned without another request.