jev-skill：给 Agent 装上 Jev 决策感知，9 个可安装 Skill 覆盖分诊/路由/代码审查

jev-skill 是一个 SKILL.md 格式的 Agent Skill 集合，今天刚开源（2026-09-20），把 TypeSafe 的 Jev 决策 API 打包成 9 个可安装的场景专用 Skill，覆盖 90 个标注决策场景，14 条真实 API I/O 示例。目标场景：Claude Code、Codex、OpenCode 等 Agent 系统。

仓库：github.com/wuyoscar/jev-skill | License：MIT | Stars：34

---

为什么要给 Agent 接 Jev

普通 LLM 做分类/打分/路由时，结果是生成出来的——答案随 prompt 微小变化而漂移，置信度来自自评而非校准概率。Jev 是 TypeSafe 专门训练的决策模型，输出是校准过的概率分布，不生成自由文本。

jev-skill 的作用：把这个能力封装成 Agent 随时可调用的 Skill——Agent 不需要自己判断”这条消息紧不紧急”、“现在该调哪个工具”，转而调 Jev 拿一个有置信度数字的结论。

---

9 个 Skill 对应 90 个场景

Skill	典型用途
jev	通用自定义决策检查点
jev-triage	消息/反馈分类、紧急程度评级
jev-documents	证据片段选取、论点支持/反驳判定
jev-ui	浏览器/桌面下一步操作决策
jev-route	工具/模型/专家路由
jev-context	上下文相关性评分、何时压缩上下文
jev-code-review	代码变更优先级排序
jev-find-code	仓库导航、定位目标文件
jev-simulation	模拟世界中的法律动作选择

---

三种决策类型

Jev 的输入是结构化 JSON，输出三种格式：

Choice：从候选列表中选一个

{
  "state": "上下文描述",
  "questions": {
    "next_step": {
      "type": "choice",
      "instructions": "Agent 下一步应该做什么？",
      "criteria": {
        "inspect_input": "查看输入数据",
        "retry_call": "重试 API 调用",
        "escalate": "转人工"
      }
    }
  }
}

Noul：独立是/否，返回校准概率

// 输出示例（来自真实 API 调用）：
// stuck = true, P = 0.88
// → "Agent 有 88% 概率陷入循环，建议检查输入"

Score：按评分量表返回档位 + 期望值

// urgency = 1.29/2（0~2 三档：低/中/高）

---

安装方式

把这条 prompt 贴进 Claude Code：

Install Jev Skills for my current agent, including the general skill and all 
scenario skills. Read and follow this installation guide, then verify the 
installation: https://raw.githubusercontent.com/wuyoscar/jev-skill/main/docs/install.md

Agent 会自动读取安装指引、写入 Skill 文件、离线验证。不需要 npm / Node.js / Vercel。

需要设置环境变量：

export OPENROUTER_API_KEY="sk-or-..."  # TypeSafe Jev 通过 OpenRouter 调用

命令行直接调用：

jev-decide --state "用户报告支付失败" \
           --question "urgency" \
           --type score

---

没有 API Key 时的处理规则

这一点值得专门说：jev-skill 明确规定，若 OPENROUTER_API_KEY 缺失，Agent 禁止静默降级。只有两个合法路径：

• A：提示用户获取 OpenRouter key，等拿到再调真实 Jev
• B：用宿主 Agent 模拟决策，但必须在输出中标注 mode: agent_simulation 和 jev_called: false

“悄悄用 GPT 代替 Jev 然后假装是 Jev 的结果”——这条被明确禁止。校准概率是 Jev 的核心价值，用模拟结果冒充会破坏 Agent 对置信度数字的信任。

---

14 条真实 API 记录

仓库包含 14 条真实 Jev API 调用的 I/O 记录，来自合成示例（非作者自己的 Agent 轨迹）。几个有代表性的：

场景	问题	Jev 输出
Agent 循环检测	是否陷入循环？	stuck=true, P=0.88
客服工单路由	归类为哪个队列？	queue=bug, urgency=1.29/2
文档核查	哪条来源支持这个论点？	source=s2, P=0.97, claim_support=contradicted

---

不足之处

1. 需要付费 API：Jev 通过 OpenRouter 调用，没有本地权重可下载。与 Kev（jaredpalmer/kev）是两个不同项目——Kev 是本地开源 LoRA，jev-skill 是 TypeSafe 商业 Jev 的 Agent 接入层。

2. “Jevify” 适配器校准待验证：仓库提到 Jevify（用其他模型适配 Jev 接口）明确标注为”API 兼容性≠等价校准”，社区对此的基准测试被标记为未核实。

3. 仓库极新：今天刚开源，34 stars，社区验证极少。

4. Jev 不生成文字：它只做分类/打分/路由。开放式规划、文本生成、工具执行仍然由宿主 Agent 完成；Jev 只负责”要不要做”和”做哪个”这类判断。

5. 无会话记忆：每次 Jev 调用都是独立的，不跟踪调用间的状态。

---

怎么看这件事

jev-skill 做的事情很具体：把一个需要自己写 prompt 工程才能接进去的决策 API，包装成 Agent 可以直接 import 的 Skill。9 个场景覆盖了 Agent 系统里最常见的几类判断需求——什么时候升级、调哪个工具、这段代码改动有多紧急。

价值在于标准化，而不是新功能：这些决策 Agent 本来也会做，只是现在有了一套带置信度的一致接口，而且明确区分了”真 Jev 的概率”和”Agent 自己猜的概率”。

今天刚开源，观望一段时间再决定是否接入。

开源代码与 Skill 仅供学习研究。调用真实 Jev API 会产生 OpenRouter 费用，请查看 TypeSafe 定价。

---

jev-skill: Plug Jev Decision Intelligence into Your Agent

jev-skill is a SKILL.md-format agent skill collection, open-sourced today (2026-09-20). It wraps TypeSafe’s Jev decision API into 9 domain-specific installable skills covering 90 labeled decision scenarios and 14 real API I/O examples. Works with Claude Code, Codex, and OpenCode.

Repo: github.com/wuyoscar/jev-skill | License: MIT | Stars: 34

---

Why Jev for Agent Decisions

Standard LLMs doing classification/scoring/routing produce outputs that drift with prompt phrasing, and their “confidence” is self-reported, not calibrated. Jev is TypeSafe’s purpose-trained decision model that outputs calibrated probability distributions without free-text generation. jev-skill packages this into skills agents can call directly — instead of asking itself “how urgent is this ticket?”, the agent calls Jev and gets a number it can act on.

---

9 Skills, 90 Scenarios

Skill	Purpose
jev	Custom decision checkpoints
jev-triage	Message classification, urgency scoring
jev-documents	Evidence span selection, claim support/contradiction
jev-ui	Browser/desktop next-action decisions
jev-route	Tool/model/specialist routing
jev-context	Context relevance, when to compact
jev-code-review	Code change prioritization
jev-find-code	Repository navigation
jev-simulation	Legal action choices in simulated worlds

---

Three Output Types

• Choice: pick from a defined candidate list
• Noul: calibrated yes/no probability (stuck=true, P=0.88)
• Score: graded level with expected value (urgency=1.29/2)

---

Installation

Paste this prompt into Claude Code:

Install Jev Skills for my current agent, including the general skill and all 
scenario skills. Read and follow this installation guide, then verify the 
installation: https://raw.githubusercontent.com/wuyoscar/jev-skill/main/docs/install.md

No npm, Node.js, or Vercel required. Set OPENROUTER_API_KEY to call real Jev.

---

The No-Key Rule

When OPENROUTER_API_KEY is absent, the agent is forbidden from silent fallback. Only two legal paths: (A) prompt the user to get a key, or (B) simulate with the host agent but explicitly label the output mode: agent_simulation, jev_called: false. Pretending a regular LLM result is a Jev calibrated probability is explicitly prohibited — it breaks the trust premise of the whole setup.

---

Limitations

1. Paid API required — Jev via OpenRouter, no open weights available locally. This is unrelated to jaredpalmer/kev, which is a separate open-source local LoRA model.
2. “Jevify” adapters: API compatibility doesn’t equal calibration equivalence. Community benchmarks for this are flagged as unverified.
3. Repo is brand new — 34 stars, minimal community validation.
4. Jev doesn’t generate text — it only classifies, scores, and routes. Planning and execution stay with the host agent.
5. No session memory — each Jev call is stateless.

---

Bottom Line

jev-skill standardizes the interface between agent systems and calibrated decision outputs. The value is in providing a consistent, labeled boundary between “what Jev says with a probability” and “what the agent guessed” — including the explicit rule that you can’t blur that line when the API key is absent. Very new repo; worth watching before integrating.

Skills for research and learning only. Real Jev API calls via OpenRouter incur charges — check TypeSafe pricing before deploying.