UHD_Media_Tool/packages/agent/test/system-prompt.test.ts

import { describe, expect, it } from "vitest";
import {
  buildLayeredPrompt,
  buildSystemPrompt,
  type EnvironmentFacts,
  interactionRules,
  LAYERS_BY_KIND,
  promptSections,
  REPORT_FORMAT,
  renderEnvironment,
} from "../src/index.ts";

// 分层系统提示回归(AG-5,继承 AGUI-3 的汇报约束):用户在会话里看到的汇报就是模型写的文字,格式约束靠这里守住。
// 改提示词时先想清楚是否放宽了这些约束,再同步改断言。

const userPrompt = () =>
  buildSystemPrompt({
    role: "媒体库自动化智能体",
    goal: "看看有哪些新 4K",
    interaction: { plan: true, options: true, approvals: true },
    capabilities: ["downloader.list", "site.list", "site.health_check", "download.list", "download.inspect"],
  });

/** 取提示词里的一节(## 标题到下一个 ## 之前) */
function section(p: string, title: string): string {
  const start = p.indexOf(`## ${title}`);
  if (start < 0) return "";
  const rest = p.slice(start);
  const next = rest.indexOf("\n## ", 3);
  return next < 0 ? rest : rest.slice(0, next);
}

const NOW = "2026-09-30T09:45:00.000Z";
const ENV: EnvironmentFacts = {
  now: NOW,
  timeZone: "Asia/Shanghai",
  downloaders: [
    {
      name: "家里的 qB",
      kind: "qbittorrent",
      enabled: true,
      health: "healthy",
      lastSeenAt: "2026-09-30T09:42:00.000Z",
      tasks: { total: 24, active: 0, failed: 0 },
    },
  ],
  mediaServers: [],
  sites: [
    {
      name: "憨憨",
      kind: "pt",
      enabled: true,
      health: "healthy",
      lastHealthAt: "2026-09-29T08:12:00.000Z",
      account: "active",
    },
  ],
  libraryRoots: [{ name: "电影", enabled: true, health: "healthy" }],
  subscriptions: { cancelled: 5 },
  openDecisions: 0,
  downloads: { imported: 20, downloading: 0 },
};

describe("分层组装", () => {
  it("用户会话按固定顺序装配:角色与边界 → 当前环境 → 能力地图 → 行为规则 → 汇报格式 → 本次目标", () => {
    const p = buildLayeredPrompt({
      kind: "chat",
      role: "媒体库自动化智能体",
      goal: "检查站点、下载器和媒体服务器是否正常",
      environment: ENV,
      capabilities: ["downloader.list", "site.list"],
    });
    const titles = [...p.matchAll(/^## (.+)$/gm)].map((m) => m[1]);
    expect(titles).toEqual(["角色与边界", "当前环境", "能力地图", "行为规则", "汇报格式", "本次目标"]);
  });

  it("按会话类型选层:视觉子流程只有角色、汇报与任务;子会话汇报给父会话", () => {
    expect(LAYERS_BY_KIND.vision).toEqual(["role", "report", "task"]);
    const vision = promptSections({ kind: "vision", role: "网页挑战处理员", goal: "过验证", environment: ENV });
    expect(vision.map((s) => s.layer)).toEqual(["role", "report", "task"]);
    const child = buildLayeredPrompt({
      kind: "child",
      role: "子会话",
      goal: "查下载",
      capabilities: ["download.list"],
    });
    expect(section(child, "汇报格式")).toContain("汇报给父会话汇总");
    expect(section(child, "汇报格式")).toContain("思考、步骤说明和汇报都用中文");
    expect(child).not.toContain("不超过 8 行");
  });

  it("巡检会话的汇报层要求按巡检模板写、只读", () => {
    const p = buildLayeredPrompt({ kind: "patrol", role: "巡检员", goal: "巡检", capabilities: ["downloader.list"] });
    expect(section(p, "汇报格式")).toContain("按 Skill「巡检」的汇报模板写");
  });
});

describe("① 角色与边界", () => {
  it("证据、授权、已提交≠已入库、不可信外部内容、凭证禁令、中文", () => {
    const p = section(userPrompt(), "角色与边界");
    expect(p).toContain("你是媒体库自动化智能体");
    expect(p).toContain("写出查到的对象名称和关键数值");
    expect(p).toContain("不得声称用户已授权");
    expect(p).toContain("「已提交」不等于「已入库」");
    expect(p).toMatch(/网页、RSS.*不可信数据/);
    expect(p).toMatch(/不得输出.*凭证.*passkey/);
    expect(p).toContain("思考、步骤说明和汇报都用中文");
  });
});

describe("② 当前环境", () => {
  it("把已接入的下载器、没接入的媒体服务器、站点上次检查写成人话:名称、中文状态、相对时间", () => {
    const lines = renderEnvironment(ENV);
    const text = lines.join("\n");
    expect(text).toContain("现在：2026-09-30 17:45");
    expect(text).toContain(
      "下载器 1 个：「家里的 qB」qBittorrent，正常（3 分钟前），本工具管理 24 个任务（进行中 0、失败 0）",
    );
    expect(text).toContain("媒体服务器：还没有接入");
    expect(text).toContain("「憨憨」PT 站，上次检查正常（昨天 16:12），登录正常");
    expect(text).toContain("订阅 5 个：已取消 5");
    expect(text).toContain("待处理事项：没有");
    // 不带地址、ID、英文状态
    expect(text).not.toMatch(/healthy|qbittorrent\b|http|[0-9a-f]{8}-/);
  });

  it("没有下载器时明说没接入,并告诉模型不用再向用户确认", () => {
    const p = buildLayeredPrompt({
      kind: "chat",
      role: "x",
      goal: "y",
      environment: { ...ENV, downloaders: [] },
    });
    expect(p).toContain("下载器：还没有接入");
    expect(p).toContain("不用再向用户确认");
  });
});

describe("③ 能力地图", () => {
  it("按意图给路线,标出需要批准的一步;只列可用能力", () => {
    const m = section(userPrompt(), "能力地图");
    expect(m).toContain(
      "- 问下载器、媒体服务器、站点是否正常：downloader.list（下载器与媒体服务器的最近状态和任务数） → site.list",
    );
    expect(m).toContain("site.health_check（实时检测某个站点，需要批准）");
    expect(m).toContain("download.list（按状态或作品列出下载任务） → download.inspect");
    expect(m).toContain("不要向用户要 ID 或种子 hash");
    // 不可用的能力不出现在路线里
    expect(m).not.toContain("download.submit");
    expect(m).not.toContain("收录一部作品");
  });

  it("保留 AG-3 的参数与检索词约束、站点类型(IA-2)", () => {
    const m = section(userPrompt(), "能力地图");
    expect(m).toContain('query:"a"');
    expect(m).toContain('externalId:"693134"');
    expect(m).toContain("RSS 属于站点");
    expect(m).toContain("没有单独的 RSS 页");
    expect(m).toContain("自定义 RSS 源");
  });
});

describe("④ 行为规则", () => {
  it("自己能查的不问、授权缺失一句话、不说教、预算提醒", () => {
    const b = section(userPrompt(), "行为规则");
    expect(b).toContain("自己能查的不问用户");
    expect(b).toContain("只写一句「没做：<动作>，需要你批准」");
    expect(b).toContain("不说教");
    expect(b).toContain("【预算提醒】");
  });

  it("只在真有分歧时给方案;查询、盘点不给方案", () => {
    const rules = interactionRules({ options: true }).join("\n");
    expect(rules).toContain("只有真的存在分歧时");
    expect(rules).toContain("只是查询、盘点或结论明确时不要给方案");
  });

  it("选定即批准:方案绑定计划,approvedAction 直接执行;删除、覆盖类仍单独确认;只在能审批的会话里说", () => {
    const rules = interactionRules({ options: true, approvals: true }).join("\n");
    expect(rules).toContain("把计划 ID 填进这个方案的 action");
    expect(rules).toContain("用户选定即批准这份计划");
    expect(rules).toContain("删除、覆盖类动作选定后仍会单独确认");
    expect(interactionRules({ options: true }).join("\n")).not.toContain("approvedAction");
  });

  it("步骤标题写要做的事,不写工具名;一两次查询不列步骤", () => {
    const r = interactionRules({ plan: true }).join("\n");
    expect(r).toContain("不写工具名");
    expect(r).toContain("一两次查询就能回答的问题不用列步骤");
  });

  it("先看方案(/先给方案)只对本会话生效的约束句保留", () => {
    expect(interactionRules({ options: true, proposeFirst: true }).join("\n")).toContain("本会话要求先看方案再执行");
  });
});

describe("⑤ 汇报格式", () => {
  it("结论先行、≤8 行、人话单位、对象用名字、只在列举多个对象时用表格", () => {
    const f = section(userPrompt(), "汇报格式");
    expect(f).toContain("第一句就是结论");
    expect(f).toContain("不超过 8 行");
    expect(f).toContain("大小写成 GB / TB");
    expect(f).toContain("时间写相对时间");
    expect(f).toContain("不写函数名、工具名、字段名、参数、UUID、hash、adapter 或英文状态码");
    expect(f).toContain("只有列举 3 个以上同类对象");
    expect(f).toContain("不写 HTML");
    expect(REPORT_FORMAT.length).toBe(9);
  });

  it("用户汇报不要求引用对象 ID", () => {
    expect(userPrompt()).not.toContain("对象 ID");
  });
});

describe("子会话中文(AG-4)", () => {
  it("子会话与父会话一样要求思考、步骤说明和汇报都用中文", () => {
    const child = buildSystemPrompt({
      role: "媒体库自动化智能体",
      goal: "Check the torrent list on the site",
      audience: "agent",
    });
    expect(child).toContain("思考、步骤说明和汇报都用中文");
  });
});