shipwithjev

Catalog / Tools & apps

0344GitHub

jev-agent-skill

Claude Code/ZCode skill that offloads classify, screen, score, and compliance-check judgments to Jev via OpenCode Zen's free tier; ships a retry-hardened zero-dependency caller…

yuyang2230/jev-agent-skillREADME ↗
# jev-agent-skill — 给 AI 编码代理装上「免费决策前置反射弧」





把**高频小判断**(分类 / 初筛 / 打分 / 核查)从主模型卸载给 [Jev](https://docs.typesafe.ai/concepts/system-one)(TypeSafe System One 决策模型,OpenCode Zen 免费档),主模型专心生成,判断走免费通道。适配 Claude Code / ZCode 及任何带 skills 目录的 agent。

| | 主模型做判断 | Jev 做判断 |
|---|---|---|
| 单次成本 | 数百~数千 token 推理 + 进上下文被反复计费 | ~400 输入 token,**$0**,不进上下文 |
| 输出 | 自由文本,需解析 | 类型化 schema(choice/noul/score),零格式错误 |
| 并发 | 串行 | 一次请求多问并行 |

实测(2026-09-20,10+ 次真实调用):单次输入 283–694 token、`cost: 0`、端到端 ~2s;一次请求并行 4 问正常;中文 state/criteria 正常。**详细对比(tokens 消耗 & 内容质量,含 3 连发一致性实测)见 [benchmark 报告](references/benchmark-jev-vs-mainmodel.md)。**

## 为什么省 token

批量场景是收益大头:50 条评论初筛,主模型直读 ≈ 几万字进上下文(且此后**每一轮对话都在为它付费**);走本 skill,原始条目留在脚本里,回上下文的只有 50 行结论。

## 快速开始

```bash
# 1. 拿一个 OpenCode Zen key(免费档即可),放进 ~/.jev/zen.key 或 export ZEN_API_KEY=...
# 2. 把本仓库 clone 到 agent 的技能目录
git clone https://github.com/yuyang2230/jev-agent-skill.git ~/.claude/skills/jev   # Claude Code
# ZCode: ~/.zcode/skills/jev
# 3. 冒烟
echo '{"state":"Payment system down 3h, users complaining","questions":{"is_urgent":{"type":"noul","instructions":"Urgent?"}}}' | python ~/.claude/skills/jev/jev.py
# -> is_urgent: 0.95
```

装好后,agent 读到 SKILL.md 的触发规则,遇到分类/初筛/打分/核查类任务会**自动**调 `jev.py`,不需要你点名。

## 触发场景

| 任务形态 | 原语 |
|---|---|
| 分类/路由:工单、评论、留言、日志归哪类、给谁 | `choice` |
| 批量初筛:哪些值得细看、哪些丢弃 | `noul` ×N |
| 打分/排序:按明确标准,权重在代码里组合 | `score` |
| 核查:陈述是否有证据、输出是否达标、有无违禁意 | `noul` |

不适用:生成文字/代码、需跨多轮上下文综合的决策、一次性琐碎判断。

## API 坑位(全部实战踩过)

| 症状 | 原因/修法 |
|---|---|
| 之前能用的 payload 突然 `HTTP 500` | Zen 网关瞬时故障(<1s 快速失败),重试 1–2 次(`jev.py` 已内置) |
| Python 直连 `HTTP 403` | WAF 拦裸 urllib UA,带 `User-Agent: curl/8.9.1` |
| Windows 下 `surrogates not allowed` | GBK 管道污染,读 `sys.stdin.buffer` 自己按 UTF-8 解 |
| key 正常却 `HTTP 401` | 模型名没权限/拼错,用 `jev-1.13-free` 不是 `jev-lates

Also filed under Tools & apps

  1. 0573

    Support chat: agent offered a human escalation path

    A frustrated customer hits a bot loop twice; the judge checks whether a human handoff was offered.

    everyai-com · Tools & apps

  2. 0572

    Subscription cancel: effective date is explicit

    An agent cancels a subscription; the judge checks the exact date the cancellation takes effect.

    everyai-com · Tools & apps

  3. 0571

    Sales quote: numbers match the official price list

    An agent quotes $29/seat against a price list; the judge checks the quote matches exactly.

    everyai-com · Tools & apps

  4. 0570

    Refund chat: agent stated the amount before issuing it

    A support agent issues a $42.50 refund; the judge checks the exact amount was stated before the refund was confirmed.

    everyai-com · Tools & apps