---
name: jev-tool-guard
description: Review a proposed tool invocation against its supplied policy and existing authorization.
---

# Jev tool guard

Independent Jev API Pro workspace instructions, revised 2026-09-25.
Endpoint: https://jevapi.pro/api/mcp (stateless JSON Streamable HTTP, protocol 2025-11-25).
Guide: https://jevapi.pro/skills

## Prerequisite

Configure this workspace's active paid Bearer key through the MCP client's credential store or JEV_API_KEY environment setting. Do not put the value in this file, project files, prompts, tool arguments, logs or output. A key from another site does not work here. Confirm tools/list exposes the required tool before using it. If it is unavailable, explicitly report that the tool check was not run; do not invent a successful call.

## Use

Call jev_guard_tool_call with tool, action, and relevant redacted arguments_summary, side_effects, safeguards, policy and reversibility. allow means the supplied facts support the action within its existing authorized scope; confirm means check and obtain any required missing approval; review means inspect missing evidence before continuing; deny means do not execute the proposed action. No result grants new permissions.

Use the exact current schema from tools/list. Supply request_id for every tool call: 8–96 safe ASCII letters, numbers, dots, colons, underscores or hyphens. Generate the id for the logical operation. Retry unchanged input with that same id after an uncertain connection failure; never derive it from the JSON-RPC id. Changed facts or a new operation need a new id. Do not retry forever. A confirmed failed decision is refunded and a new attempt needs a new id.

## Interpret

Read isError and structuredContent, not just HTTP status. Missing or low confidence, a review flag, contradictory evidence or an error requires the application's review path. Probabilities are signals, not measured correctness or permission. The tool does not execute actions, replace the main model, browse, run tests or relax existing user and platform restrictions. Preserve any applicable human-approval requirement. Send only relevant, redacted facts.

A successful new call uses one workspace decision credit. initialize, ping and tools/list do not call the model or spend decision credits. Opening or downloading this file performs no model call.

## 中文使用邊界

Skill 是使用說明，MCP 才是連接管道。先確認工具可用，再根據當前事實呼叫；工具缺失時明確說沒有執行，不能假裝成功。選項與機率不等於授權，也不能代替真實測試。金鑰放在客戶端憑證設定，絕不放入本文件或請求參數。相同操作以相同識別碼重試，重要事實改變後使用新識別碼。