mirror of
https://github.com/ThearchyHelios/AgentLab.git
synced 2026-11-15 21:12:01 +00:00
ThearchyHelios
3631 lines
241 KiB
JavaScript
3631 lines
241 KiB
JavaScript
// 问数据页的检查：会话、轮次、运行全部在浏览器这一侧伪造，页面走的是真的
|
||
// chat store / ChatPage / AssistantStream，只有后端换成了脚本。
|
||
//
|
||
// 一轮问数据在服务端跑（docs/design/orchestration-round/00-总方案.md §1）：页面只发起（POST /api/chat/turns）、只播放
|
||
// （GET /api/chat/turns/{id}/stream）。这里伪造的服务端（startTurn / driveTurn，照 backend/app/chat/job.py 的做法）用原来那几份
|
||
// 剧本把一轮跑完：生成（copilotScript）→ 发起运行（launchScript）→ 等运行的剧本放完（fakeStream）→ 复核（reviewScript）→
|
||
// 照服务端的写法拼那一行（end）。log.gens / runStarts / reviews 照旧记下服务端这一侧做了什么，log.patches 用来守「页面不再写
|
||
// 这一轮的结局」。
|
||
//
|
||
// 为什么不真跑：沙箱没有模型密钥，而且要撞上的恰恰是真跑很难稳定复现的那几种
|
||
// 时刻——两个会话同时在跑、答案超过 2000 字被事件截断、跑到一半服务重启、
|
||
// 历史取得慢或者取失败。这里每一种都由脚本精确地摆出来。
|
||
//
|
||
// 守的事：
|
||
// 1. 忙不忙按会话算：A 在跑时 B 的按钮是「发送」，停止只停 A 自己，删除保护
|
||
// 对准正在跑的那个；
|
||
// 2. run.finished 的 output 被截断时，显示用服务端按库里完整成果写的那一份；页面不 PATCH 这一轮；
|
||
// llm.token 不进 events；
|
||
// 3. 服务重启：stream.end status=interrupted 且没有待审批，这一轮收成
|
||
// 「服务重启，本轮已中断」，给「继续运行」，不再转圈；
|
||
// 4. 加载中 / 加载失败 / 真的是空的 三种状态分开，加载失败时不能在空态上提问；
|
||
// 5. 已取消的、人驳回这类原样续上还会失败的轮次没有「继续运行」，续跑被拒时说人话；
|
||
// 6. 1440×900 下长会话的输入框和发送键完整可见，整页不多滚；
|
||
// 7. 首屏例句不拿系统表造句；可信度信息刷新后还在，且写进单独的 meta，不夹在 review 里；
|
||
// 8. 从库里恢复的半路轮次（老轮次、服务重启后停着的）：交给服务端接着收尾（mode=resume），核对撞上 500 有出路、
|
||
// 补交付不走实时计时，结局照服务端写的那一行摆；
|
||
// 9. 老轮次的长答案没取全：刷新后补全；2000 字按码点数，不按 UTF-16；新轮次按库里的完整成果交付；
|
||
// 10. 发起那一下没到服务端：报错不糊成一整行粗体，不许诺不存在的自动重试；重跑挂起的轮次
|
||
// 顺手取消旧运行；断在建图阶段的老轮次页面不替它写库，重试本轮交给服务端；
|
||
// 11. 左栏：没打开过的会话也标得出状态；筛选框在有筛选词时不消失；删掉的会话进回收站，
|
||
// 能预览、恢复、彻底删除（要确认）；回收站里只能看，不发起任何运行，正在跑的不给
|
||
// 彻底删除（清空也跳过它）；
|
||
// 12. 限定数据源：点一下只查这个库，请求里带上 datasource_ids，库不在了说人话、补救是放开范围；
|
||
// 库列表没取回来时，从上一问接过来的范围照样看得见；
|
||
// 13. 画布助手的「在现有工作流上改 | 从头生成」分段控件；助手出错时同一句报错只说一遍；
|
||
// 14. 助手反问（normal-user-apps N11）：只问不改的一轮记成「助手反问」，问题作为这一轮的内容、落库进历史，不再一直停在
|
||
// 「正在理解需求」；输入框提示回答助手的问题；刷新后照样标「助手反问」，不当成没查库的回答；
|
||
// 15. 回答上的「做成应用」带着这次运行进蓝图（建运行起点的蓝图），不再提取工作流、跳画布的发布弹窗；
|
||
// 16. 服务端这一轮已经自查修正过（收到 check 的结论）还没过的：不再整轮重新生成一遍，直接说没过、给补救；助手只写了思考、
|
||
// 一个操作都没给（error.next=retry）的照服务端的决定自动重来一次，还不行再照实报。
|
||
// 17. 口径账本（next-round 02）：交付了的轮次底部写「正在比对口径」，按退避再取到比对出结果、比完不再取；一行摘要
|
||
// （几个数来自哪份口径）、按口径正式出具和对账按钮照服务端给的动作摆；出具、对账的写请求在浏览器层拦下并核对请求体，
|
||
// 出具过的正式运行挂在轮次下面（档位、打开运行）。
|
||
// 18. 过程里被驳回的那一步（node.failed 带 reason=rejected）：停止的剪影、中性色、行尾写「被驳回」，不画成完成的绿勾；
|
||
// 19. 答案身上的召回标记（账本接口的 recall）写在轮次底部：「已按口径「X」vN 核对，不影响」、依据有变（目录 / 表结构，几项、
|
||
// 按新目录新出现几条错误级问题），看的人手上有说的是这个答案的召回待办时链到待办；没比对的答案照样写依据有变；做完的
|
||
// 对账卡上写「差在哪」、问说明有没有帮助（说明有用 / 看不懂）。
|
||
// 20. 一轮在服务端跑：实时流断了按收到的最后一个 seq 重连，补齐、不重复；刷新、关掉再打开，服务端那一轮照跑，页面从
|
||
// 这一次尝试的第一条接着播；停止发给服务端的这一轮（建图中也停得下）；发起、重试、继续运行的请求体照实；
|
||
// 使用者（user 角色）不给「继续运行」（服务端照 POST /api/runs/{id}/continue 的老规矩不让它续跑）；两个标签页同时打开
|
||
// 停在半路的一轮，服务端只接着收尾一次，后到的那个接着看同一轮。
|
||
//
|
||
// 所有写请求都被拦在浏览器里，不碰任何库。
|
||
// 跑之前前端得起着（./scripts/dev.sh），默认连 5273。对别的实例（比如一份沙箱拷贝）跑时
|
||
// 带上地址：AGENTLAB_WEB=http://localhost:<前端端口> node scripts/check-chat.mjs
|
||
import { chromium } from '../frontend/node_modules/playwright-core/index.mjs'
|
||
import { mkdirSync, readFileSync } from 'node:fs'
|
||
import { ISSUE, RECALL, RECONCILE, WHOLE, WHOLE_LABELS, ledgerOf, metricDetail, reconciledOf } from './ledger-fixtures.mjs'
|
||
|
||
const WEB = process.env.AGENTLAB_WEB ?? 'http://localhost:5273'
|
||
const CHROME = process.env.CHROME_PATH
|
||
?? '/Applications/Google Chrome.app/Contents/MacOS/Google Chrome'
|
||
const SHOTS = process.env.SHOTS ?? '/tmp/agentlab-check-chat'
|
||
mkdirSync(SHOTS, { recursive: true })
|
||
|
||
// 只跑其中几段：ONLY=busy,restart THEMES=dark node scripts/check-chat.mjs
|
||
const ONLY = (process.env.ONLY ?? '').split(',').filter(Boolean)
|
||
const THEMES = (process.env.THEMES ?? 'light,dark').split(',').filter(Boolean)
|
||
/**
|
||
* 一节一节地跑：某一节里等待超时、元素找不到，只记成这一节失败，收掉它开的页面，
|
||
* 接着跑下一节。以前一处超时就让整个脚本崩掉，后面几百项一项都不跑
|
||
*/
|
||
const opened = new Set()
|
||
async function section(name, title, fn) {
|
||
if (ONLY.length && !ONLY.includes(name)) return
|
||
console.log(`\n=== ${title} ===`)
|
||
// 伪造的服务端：上一段里还在跑的几轮就此放下，不再往下写（各段互不牵连）
|
||
srv.gen += 1
|
||
try {
|
||
await fn()
|
||
} catch (e) {
|
||
check(`${title} 中途出错`, false, String(e?.message ?? e).split('\n')[0])
|
||
} finally {
|
||
for (const c of opened) await c.close().catch(() => {})
|
||
opened.clear()
|
||
}
|
||
}
|
||
|
||
let failed = 0
|
||
const check = (name, cond, detail = '') => {
|
||
console.log(` ${cond ? '✓' : '✗'} ${name}${detail ? ` — ${detail}` : ''}`)
|
||
if (!cond) failed++
|
||
}
|
||
|
||
// ---------------------------------------------------------------- 伪造的数据
|
||
|
||
const iso = (minsAgo) => new Date(Date.now() - minsAgo * 60_000).toISOString()
|
||
const YESTERDAY_NOON = () => Math.round((Date.now() - new Date(new Date().setHours(12, 0, 0, 0) - 86_400_000).getTime()) / 60_000)
|
||
// 「今天」同理：零点刚过的那一分钟里，「1 分钟前」已经是昨天
|
||
const TODAY_RECENT = () => Math.min(1, Math.floor((Date.now() - new Date().setHours(0, 0, 0, 0)) / 60_000))
|
||
/**
|
||
* 会话的时间在每次被读到（列表接口序列化、展开拷贝）时现算。以前模块加载时就算好了：
|
||
* 整套跑十几分钟，23:4x 开跑、跑到会话列表那段已经过了零点，「今天」的全成了「昨天」，
|
||
* 分组检查误报。minsAgo 可以是函数：「昨天中午」「今天刚才」要按读的那一刻算
|
||
*/
|
||
const conv = (id, title, turns, minsAgo, extra = {}) => {
|
||
const ago = typeof minsAgo === 'function' ? minsAgo : () => minsAgo
|
||
return {
|
||
id, title, kind: 'chat', workflow_id: null, archived: false,
|
||
get created_at() { return iso(ago() + 5) },
|
||
get last_active_at() { return iso(ago()) },
|
||
turn_count: turns, last_question: title,
|
||
last_status: turns ? 'done' : null, last_run_id: null, ...extra,
|
||
}
|
||
}
|
||
const LONG = (n) => Array.from({ length: n }, (_, i) =>
|
||
`${i + 1}. 这一段是很长的结论正文，用来把会话撑到超过一屏，检查输入框会不会被挤出视口。`).join('\n')
|
||
const FULL = '这是完整答案的开头。' + '正文'.repeat(1920) + '【完整结尾】' // 3,850 字左右
|
||
// 后端按码点截在 2000：带 BMP 之外的字符时，JS 的 length 会比 2000 多
|
||
const ASTRAL_2000 = '旧'.repeat(1995) + '📊'.repeat(5)
|
||
const turn = (id, question, extra = {}) => ({
|
||
id, seq: 0, question, answer: '', explanation: '', graph: null, run_id: null,
|
||
status: 'done', error: '', review: null, created_at: iso(30), ...extra,
|
||
})
|
||
// 坏参数定义的后端原文（tools/custom.py broken_tool_message）。BROKEN_TOOL 是老说法：库里老运行的 error 存的是它，
|
||
// 前端要继续认得；BROKEN_TOOL_NOW 是后端现在的说法。「先去改参数定义」那段两份各走一遍
|
||
const BROKEN_TOOL = '自定义工具「lookup_order」的参数定义格式不对：参数 store 要写成 {"type": "string"} 这样的对象，'
|
||
+ '不能直接写 "string"。到「工具」页把它的参数定义改好再运行'
|
||
const BROKEN_TOOL_NOW = '自定义工具「lookup_order」的参数定义格式有误：参数 store 应写成 {"type": "string"} 这样的对象，'
|
||
+ '不能直接写 "string"。到「工具」页修改参数定义后再运行'
|
||
const GRAPH = {
|
||
nodes: [
|
||
{ id: 'in', type: 'input', position: { x: 0, y: 0 }, data: { label: '问题', config: { fields: [{ name: 'question' }] } } },
|
||
{ id: 'ag', type: 'agent', position: { x: 0, y: 100 }, data: { label: '查数', config: { max_steps: 12 } } },
|
||
{ id: 'out', type: 'output', position: { x: 0, y: 200 }, data: { label: '成果', config: {} } },
|
||
],
|
||
edges: [{ source: 'in', target: 'ag' }, { source: 'ag', target: 'out' }],
|
||
}
|
||
|
||
/**
|
||
* 问数据的答案（可点击证据第二期）：input → agent → report → output，报告直接引用 agent 查过的
|
||
* 单元格、整表。文档是 compose_doc 真跑出来的，片段接口的答复带查询步骤（窗口、高亮、遮罩）
|
||
*/
|
||
const EVQ = JSON.parse(readFileSync(new URL('../frontend/src/run/__tests__/evidence-query.json', import.meta.url), 'utf8'))
|
||
const EVQ_DOC = { ...EVQ.doc, run_id: 'run-evq' }
|
||
/** 整形节点按 JSON 解析失败的两种原话（backend/app/engine/nodes/io.py 的 _json_error） */
|
||
const JSON_TEMPLATE_ERROR = '模板渲染出来的不是合法 JSON（第 1 行第 934 列附近）。检查模板里的引号、逗号，字符串值要用 | json 过滤器输出'
|
||
const JSON_UPSTREAM_ERROR = '上游「查数」输出的不是合法 JSON（第 934 列附近），常见原因是字符串里有没转义的英文引号；'
|
||
+ '让模型交结构化数据请用 output_schema + cite_fields，别用整形节点解析它写的文字'
|
||
/**
|
||
* 上面两句是老说法（库里老运行的 error 存的是它们），前端要继续认得；下面两句是后端现在的原文
|
||
*（io.py 的 TEMPLATE_HINT、_json_error 里上游那一句），两组都走一遍
|
||
*/
|
||
const JSON_TEMPLATE_ERROR_NOW = '模板渲染出来的不是合法 JSON（第 1 行第 934 列附近）。请检查模板中的引号和逗号，字符串值需使用 | json 过滤器输出'
|
||
const JSON_UPSTREAM_ERROR_NOW = '上游「查数」输出的不是合法 JSON（第 934 列附近），常见原因是字符串中有未转义的英文引号；'
|
||
+ '如需模型输出结构化数据，请配置「结构化输出 Schema」并开启「按出处核对字段」，不要用「数据整形」节点解析模型写的文字'
|
||
/** 同一句上游说法，节点名里带着别的规则的关键词（超时）：不能被当成超时、又把「继续运行」给回来 */
|
||
const JSON_SLOW_ERROR = JSON_UPSTREAM_ERROR.replace('上游「查数」', '上游「查询超时订单」')
|
||
const EVQ_SEG = (text, nth = 0) => Object.values(EVQ.segments).filter((d) => d.segment.text === text)[nth]?.segment.id
|
||
|
||
const CONVS = {
|
||
busyA: conv('c0busya', '会话 A：跑得很久的那个', 0, TODAY_RECENT),
|
||
idleB: conv('c0idleb', '会话 B：已经答完的那个', 1, 3),
|
||
longC: conv('c0longc', '会话 C：长会话，检查布局', 6, 60),
|
||
slowD: conv('c0slowd', '会话 D：历史取得慢', 2, 90),
|
||
failE: conv('c0faile', '会话 E：历史取失败', 3, 120),
|
||
// 「昨天」按日历算：写死 30 小时前的话，凌晨 6 点前跑会落进前天，分组检查就挂了
|
||
histF: conv('c0histf', '会话 F：取消过、失败过、不可用的历史', 4, YESTERDAY_NOON),
|
||
emptyG: conv('c0emptg', '新对话', 0, 60 * 24 * 3),
|
||
truncH: conv('c0trunc', '会话 H：长答案', 0, 60 * 24 * 10),
|
||
// 下面四个在「列表」那段里没打开过：状态只能从列表接口的 last_status 来
|
||
restI: conv('c0resti', '会话 I：跑到一半服务重启', 0, 60 * 24 * 12, { last_status: 'suspended' }),
|
||
buildJ: conv('c0buildj', '会话 J：建流程就失败', 0, 60 * 24 * 13, { last_status: 'error' }),
|
||
waitK: conv('c0waitk', '会话 K：停在审批上', 0, 60 * 24 * 14, { last_status: 'waiting', last_run_id: 'run-wait' }),
|
||
writeL: conv('c0writel', '会话 L：模型正在出字', 0, 60 * 24 * 15, { last_status: 'running' }),
|
||
metaM: conv('c0metam', '会话 M：刷新之后的可信度', 1, 60 * 24 * 16),
|
||
killN: conv('c0killn', '会话 N：服务被强杀', 0, 60 * 24 * 17),
|
||
suspO: conv('c0suspo', '会话 O：三天前被重启打断的', 1, 60 * 24 * 3, { last_status: 'suspended' }),
|
||
unkP: conv('c0unkp', '会话 P：核对时后端出错', 1, 60 * 24 * 18),
|
||
lateQ: conv('c0lateq', '会话 Q：跑完了、没来得及交付', 1, 60 * 24 * 3),
|
||
termR: conv('c0termr', '会话 R：停在半路的几轮', 6, 60 * 24 * 19),
|
||
clipS: conv('c0clips', '会话 S：长答案没取全', 0, 60 * 24 * 20),
|
||
clipT: conv('c0clipt', '会话 T：刷新之后补全长答案', 3, 60 * 24 * 21),
|
||
launchU: conv('c0launu', '会话 U：发起运行失败', 0, 60 * 24 * 22),
|
||
scopeV: conv('c0scopv', '会话 V：限定数据源', 1, 60 * 24 * 23),
|
||
scopeW: conv('c0scopw', '会话 W：上一问限定过数据源', 1, 60 * 24 * 24),
|
||
buildX: conv('c0buildx', '会话 X：两天前的问题，别处刚又跑了一遍', 2, 60 * 24 * 25),
|
||
fixY: conv('c0fixy', '会话 Y：绑的自定义工具参数定义写坏了', 1, 60 * 24 * 26),
|
||
}
|
||
/** 回收站里的：列表接口只有带 include_archived 才给 */
|
||
const TRASHED = {
|
||
trash1: conv('c0trash1', '删掉的对话甲', 3, 60 * 24 * 2, { archived: true }),
|
||
trash2: conv('c0trash2', '删掉的对话乙', 1, 60 * 24 * 40, { archived: true }),
|
||
}
|
||
/** 只在「回收站里只能看」那段出现，免得别的段数回收站里有几个时数错 */
|
||
const TRASH_MORE = {
|
||
// 各种能补救的轮次都有一个：没跑的、停在审批上的、跑挂了的、结论不可用的
|
||
trash3: conv('c0trash3', '删掉的对话丙：各种没跑完的', 4, 60 * 24 * 4, { archived: true, last_status: 'done' }),
|
||
// 删掉之后又跑起来了（比如在「记录」里通过了审批）：它底下有一个真在跑的运行
|
||
trash4: conv('c0trash4', '删掉的对话丁：还在跑', 1, 60 * 24 * 5,
|
||
{ archived: true, last_status: 'running', last_run_id: 'run-trash-live' }),
|
||
}
|
||
/** 只在「护栏」那段出现：放宽步数只在步数用满、而且还有得放宽时给（后端 engine/guards.py） */
|
||
const GRAPH8 = { ...GRAPH, nodes: GRAPH.nodes.map((n) => (n.type === 'agent'
|
||
? { ...n, data: { ...n.data, config: { max_steps: 8 } } } : n)) }
|
||
const guardTurn = (id, graph, reason, detail) => turn(id, '查一下各门店的销量', {
|
||
answer: '只查了一部分的结论', graph, run_id: `run-${id}`,
|
||
review: {
|
||
verdict: 'annotated', note: detail, answer: null, retry: true, severity: 'broken',
|
||
signals: [{ kind: 'step_limit', detail, severity: 'broken', ...(reason ? { reason } : {}) }],
|
||
},
|
||
meta: { v: 1, runId: `run-${id}`, runClass: 'exploratory', runStatus: 'succeeded' },
|
||
})
|
||
const GUARD_CONVS = {
|
||
stall: conv('c0gstall', '护栏：连续几步没进展收的尾', 1, 60 * 24 * 30),
|
||
steps: conv('c0gsteps', '护栏：写死 8 步、用满了', 1, 60 * 24 * 31),
|
||
dflt: conv('c0gdflt', '护栏：跟随默认步数', 1, 60 * 24 * 32),
|
||
}
|
||
/** 只在「先去改参数定义」那段出现：同一种失败，后端现在的说法（BROKEN_TOOL_NOW） */
|
||
const FIX_CONVS = {
|
||
fixZ: conv('c0fixz', '会话 Z：参数定义有误（后端现在的说法）', 1, 60 * 24 * 27),
|
||
}
|
||
/**
|
||
* 只在「被驳回」那段出现（05-已知遗留 (a)）：一轮的运行被审批人驳回结束（服务端 cancelled + end_reason=rejected，
|
||
* 落库时 meta.outcome 是 cancelled、meta.rejected 记着那一句）。rejR 会打开、重新运行一次；rejS 不打开，状态只能从
|
||
* 列表接口的 last_status 来
|
||
*/
|
||
const REJECT_CONVS = {
|
||
rejR: conv('c0rejr', '会话：退款那一问', 1, 60 * 24 * 35, { last_status: 'rejected', last_run_id: 'run-rej2' }),
|
||
rejS: conv('c0rejs', '会话：没打开过的那一问', 1, 60 * 24 * 36, { last_status: 'rejected' }),
|
||
}
|
||
/** 只在「助手反问」那段出现：库里存着一轮以反问收尾的（answer 列里是助手问的那句话，meta.asked 标着） */
|
||
/** 只在「刷新后接着看」那段出现：一开始是空的，问的那一轮在伪造的服务端里跑，刷新、关掉再打开时从那里取回来 */
|
||
const REPLAY_CONVS = {
|
||
replay: conv('c0replay', '会话：刷新后接着看', 0, 60 * 24 * 41),
|
||
}
|
||
const ASKBACK_CONVS = {
|
||
askBack: conv('c0askbk', '助手反问过的会话', 1, 60 * 24 * 34),
|
||
}
|
||
/** 只在「口径账本」那段出现：库里存着一轮带证据的答案（evidence-doc 的周报），交付了，等着比对口径 */
|
||
const LEDGER_CONVS = {
|
||
led: conv('c0ledgr', '问数据：上周的周报', 1, 60 * 24 * 32),
|
||
}
|
||
/** 只在「问数据的证据」那段出现：库里存着一轮带证据的答案，刷新后回运行取 _evidence */
|
||
const ASK_CONVS = {
|
||
askZ: conv('c0askz', '问数据：各区域销售额和最大的一单', 1, 60 * 24 * 33),
|
||
}
|
||
const DETAIL = {
|
||
c0replay: [],
|
||
c0askbk: [turn('tab1', '上周销售额是多少', { answer: '销售额按下单日还是支付日算？', meta: { v: 1, asked: true } })],
|
||
c0ledgr: [turn('tl1', '上周的周报', {
|
||
answer: '（周报原文）', graph: GRAPH, run_id: 'run-led',
|
||
meta: { v: 1, runId: 'run-led', runClass: 'exploratory', runStatus: 'succeeded', evidence: true },
|
||
})],
|
||
c0askz: [turn('tz9', '上周各区域的销售额，最大的一单是多少', {
|
||
answer: EVQ.output.answer, graph: GRAPH, run_id: 'run-evq',
|
||
meta: { v: 1, runId: 'run-evq', runClass: 'exploratory', runStatus: 'succeeded', evidence: true },
|
||
})],
|
||
c0gstall: [guardTurn('tg1', GRAPH8, 'stall', 'Agent 连续 3 步没有获得新信息，已根据已查到的内容收尾。可能是提示词中的目标无法查到，或工具参数持续有误，请检查 Agent 最后几次的工具调用。收尾时仍未给出结论。')],
|
||
c0gsteps: [guardTurn('tg2', GRAPH8, 'steps', 'Agent 已用完 8 步上限。如需完整结果，请调大节点的「最大步数」或设置中的默认步数；如果大部分步数用于逐张表查询结构，可在提示词中指明要查的表。收尾时仍未给出结论。')],
|
||
c0gdflt: [guardTurn('tg3', GRAPH, null, 'Agent 已用完 12 步上限，仍未给出结论。请调大节点的「最大步数」；如果大部分步数用于逐张表查询结构，可在提示词中指明要查的表。')],
|
||
c0busya: [],
|
||
c0idleb: [turn('tb1', 'B 的问题', { answer: 'B 的答案', graph: GRAPH, run_id: 'run-b1' })],
|
||
c0longc: Array.from({ length: 6 }, (_, i) =>
|
||
turn(`tc${i}`, `第 ${i + 1} 个问题`, { answer: LONG(14), graph: GRAPH, run_id: `run-c${i}` })),
|
||
c0slowd: [turn('td1', 'D 的第一问', { answer: 'D 的答案一' }), turn('td2', 'D 的第二问', { answer: 'D 的答案二' })],
|
||
c0faile: [turn('te1', 'E 的第一问', { answer: 'E 的答案' })],
|
||
c0histf: [
|
||
// 用户点过停止的老数据：status=error + 「已取消」，运行在后端是 cancelled
|
||
turn('tf1', '被我停下的那一轮', { status: 'error', error: '已取消', graph: GRAPH, run_id: 'run-cancelled' }),
|
||
// 运行失败了：可以接着跑，但这次后端会拒绝（比如别处已经把它续跑完了）
|
||
turn('tf2', '跑挂了的那一轮', { status: 'error', error: 'KeyError: 查询超时，数据库没有在 30 秒内返回', graph: GRAPH, run_id: 'run-failed' }),
|
||
// 人工驳回：原样接着跑只会再被驳回一次（lib/explain 的 continuable=false），不给「接着跑」
|
||
turn('tf5', '被人驳回的那一轮', { status: 'error', error: '人工驳回：数字对不上', graph: GRAPH, run_id: 'run-rejected' }),
|
||
// 当时就拆好、落进 meta 的运行失败（带着原话 source，没记 fix 的老写法）：流里照这份画，
|
||
// 不拿原话再讲一遍；直达入口按原话认出来（401 → 去模型接入）
|
||
turn('tf6', '密钥失效的那一轮', {
|
||
status: 'error', error: '当时记下的标题：这把密钥被拒了', graph: GRAPH, run_id: 'run-auth',
|
||
meta: { v: 1, runId: 'run-auth', runStatus: 'failed', failure: {
|
||
title: '当时记下的标题：这把密钥被拒了', reason: '当时记下的原因：对方回了 401。',
|
||
hint: '当时记下的怎么办：换一把密钥再来。', continuable: false,
|
||
source: 'AuthenticationError: Error code: 401 - invalid x-api-key' } },
|
||
}),
|
||
// 没有运行、没有图，却有答案：reply，没查库。老数据没有 meta，靠推断
|
||
turn('tf3', '上一轮的数字是多少', { answer: '上一轮查到的是 42。' }),
|
||
// 复核判为不可用，原因是步数用满
|
||
turn('tf4', '步数用满的那一轮', {
|
||
answer: '只查了一半的结论', graph: GRAPH, run_id: 'run-broken',
|
||
review: {
|
||
verdict: 'annotated', note: 'Agent 已用完 12 步上限，仍未给出结论。', answer: null, retry: true, severity: 'broken',
|
||
signals: [{ kind: 'step_limit', detail: 'Agent 已用完 12 步上限，仍未给出结论。请调大节点的「最大步数」；如果大部分步数用于逐张表查询结构，可在提示词中指明要查的表。', severity: 'broken' }],
|
||
},
|
||
meta: { v: 1, runId: 'run-broken', runClass: 'exploratory', runStatus: 'succeeded', queries: 3, ms: 41000 },
|
||
}),
|
||
],
|
||
c0emptg: [],
|
||
c0trunc: [],
|
||
c0resti: [],
|
||
c0buildj: [],
|
||
c0waitk: [],
|
||
c0writel: [],
|
||
c0killn: [],
|
||
// 三天前服务重启打断、跑了 42 秒的那一轮。error 是老前端落库时的原话（现在落的是「服务重启，本轮已中断」），
|
||
// 留着旧说法：认中断靠 meta.outcome，界面上照样按现在的说法显示
|
||
c0suspo: [turn('to1', '三天前没跑完的问题', {
|
||
status: 'error', error: '服务重启，这一轮中断了', graph: GRAPH, run_id: 'run-susp', created_at: iso(60 * 24 * 3),
|
||
meta: { v: 1, runId: 'run-susp', runStatus: 'interrupted', outcome: 'suspended', ms: 42000 },
|
||
})],
|
||
// 出具档位、运行类别、耗时、查库次数只存在 meta 里：刷新后得从这里读回来，
|
||
// 而不是再去问一遍运行
|
||
c0metam: [turn('tm1', '上个月各产线的出勤率', {
|
||
answer: '一线 96.2%，二线 91.0%。', graph: GRAPH, run_id: 'run-meta',
|
||
meta: {
|
||
v: 1, runId: 'run-meta', runClass: 'exploratory', runStatus: 'succeeded', ms: 12300, queries: 2,
|
||
issuance: { tier: 'degraded', gaps: ['叙述模板渲染为空'], matched_numbers: 2, metrics_checked: 3, unmatched_numbers: [] },
|
||
},
|
||
})],
|
||
// 上次停在半路（轮次还是 running），核对时后端回 500
|
||
c0unkp: [turn('tp1', '核对时撞上后端出错的那一轮', {
|
||
status: 'running', graph: GRAPH, run_id: 'run-unk', created_at: iso(60 * 24 * 2),
|
||
})],
|
||
// 运行三天前就跑完了，交付没赶上：补交付期间不能按实时运行从三天前起算计时
|
||
c0lateq: [turn('tq1', '三天前跑完、没交付的问题', {
|
||
status: 'running', graph: GRAPH, run_id: 'run-late', created_at: iso(60 * 24 * 3),
|
||
})],
|
||
// 轮次都还是 running，运行早已各有结局：查到之后要落库，不能每次进来都重新核对
|
||
c0termr: [
|
||
turn('tr1', '在别处被取消的那一轮', { status: 'running', graph: GRAPH, run_id: 'run-t-cancel' }),
|
||
turn('tr2', '在别处失败的那一轮', { status: 'running', graph: GRAPH, run_id: 'run-t-failed' }),
|
||
turn('tr3', '服务重启挂起的那一轮', { status: 'running', graph: GRAPH, run_id: 'run-t-susp' }),
|
||
turn('tr4', '运行记录被删掉的那一轮', { status: 'running', graph: GRAPH, run_id: 'run-404-t' }),
|
||
turn('tr5', '还停在审批上的那一轮', { status: 'running', graph: GRAPH, run_id: 'run-t-wait' }),
|
||
turn('tr6', '在别处失败的那一轮（后端现在的说法）', { status: 'running', graph: GRAPH, run_id: 'run-t-failed-now' }),
|
||
],
|
||
c0clips: [],
|
||
c0clipt: [
|
||
// 上次交付时 GET 运行失败，只拿到事件里截断的那份：meta 记着 partial，刷新后要补全。
|
||
// 多个键拼起来比 2000 长，光看长度是认不出来的
|
||
turn('tt1', '多个成果键的长答案', {
|
||
answer: `${FULL.slice(0, 2000)}\n附注：另一个成果键`, graph: GRAPH, run_id: 'run-multi',
|
||
meta: { v: 1, runId: 'run-multi', runStatus: 'succeeded', clipped: 'partial' },
|
||
}),
|
||
// 恰好 2000 个码点（UTF-16 是 2005）、运行也没了：补不回来，但要标出来
|
||
turn('tt2', '旧版本被截断、运行没了的长答案', { answer: ASTRAL_2000, graph: GRAPH }),
|
||
turn('tt3', '旧版本被截断、运行记录删掉了的长答案', { answer: ASTRAL_2000, graph: GRAPH, run_id: 'run-404-clip' }),
|
||
],
|
||
c0launu: [],
|
||
c0scopv: [turn('tv1', '上一问', { answer: '上一问的答案', graph: GRAPH, run_id: 'run-b1' })],
|
||
c0scopw: [turn('tw0', '上一问只查了一个库', {
|
||
answer: '上一问的答案 W', graph: GRAPH, run_id: 'run-b1',
|
||
meta: { v: 1, runId: 'run-b1', runStatus: 'succeeded', scope: [{ id: 'ds-plant', name: 'factory' }] },
|
||
})],
|
||
// 两天前问的、刚被重试过：run_id 列还指着上一次的运行（清不掉），meta.runId 是 null。
|
||
// 有没有动静按 meta.at 算——取的那一刻才定，脚本跑得再久也是「2 分钟前」「20 分钟前」
|
||
c0buildx: () => [
|
||
turn('tx7', '两分钟前在别的页面里重试的问题', {
|
||
status: 'running', graph: GRAPH, run_id: 'run-failed', created_at: iso(60 * 24 * 2),
|
||
meta: { v: 1, runId: null, at: Date.now() - 2 * 60_000 },
|
||
}),
|
||
turn('tx8', '重试之后二十分钟没动静的问题', {
|
||
status: 'running', graph: GRAPH, run_id: 'run-failed', created_at: iso(60 * 24 * 2),
|
||
meta: { v: 1, runId: null, at: Date.now() - 20 * 60_000 },
|
||
}),
|
||
],
|
||
c0trash1: [turn('tx1', '回收站里的问题', { answer: '回收站里的答案', graph: GRAPH, run_id: 'run-b1' })],
|
||
c0trash2: [],
|
||
c0trash3: [
|
||
turn('ty0', '搭好了没跑的问题', { graph: GRAPH }),
|
||
turn('ty1', '停在审批上的问题', { status: 'running', graph: GRAPH, run_id: 'run-trash-wait' }),
|
||
turn('ty2', '跑挂了的问题', {
|
||
status: 'error', error: '查询超时', graph: GRAPH, run_id: 'run-failed',
|
||
meta: { v: 1, runId: 'run-failed', runStatus: 'failed' },
|
||
}),
|
||
turn('ty3', '步数用满、结论不可用的问题', {
|
||
answer: '只查了一半的结论', graph: GRAPH, run_id: 'run-broken',
|
||
review: {
|
||
verdict: 'annotated', note: 'Agent 已用完 12 步上限，仍未给出结论。', answer: null, retry: true, severity: 'broken',
|
||
signals: [{ kind: 'step_limit', detail: 'Agent 已用完 12 步上限，仍未给出结论。请调大节点的「最大步数」；如果大部分步数用于逐张表查询结构，可在提示词中指明要查的表。', severity: 'broken' }],
|
||
},
|
||
meta: { v: 1, runId: 'run-broken', runStatus: 'succeeded' },
|
||
}),
|
||
],
|
||
c0trash4: [turn('tz1', '删掉之后还在跑的问题', { status: 'running', graph: GRAPH, run_id: 'run-trash-live', created_at: iso(10) })],
|
||
// 要改的是工具库里的参数定义，不在这一轮的流程里：先去改，再接着跑（没记 fix 的老写法，按原话认）
|
||
c0fixy: [turn('ty1', '查一下订单状态', { status: 'error', error: BROKEN_TOOL, graph: GRAPH, run_id: 'run-badtool' })],
|
||
c0fixz: [turn('tfz1', '查一下订单状态', { status: 'error', error: BROKEN_TOOL_NOW, graph: GRAPH, run_id: 'run-badtool-now' })],
|
||
c0rejr: [turn('trj1', '这个月的退款有多少', { status: 'error', error: '已取消', graph: GRAPH, run_id: 'run-rej2',
|
||
meta: { v: 1, runId: 'run-rej2', runStatus: 'cancelled', outcome: 'cancelled', rejected: { sub: '张工 驳回了「查数」：数字对不上' } } })],
|
||
c0rejs: [],
|
||
}
|
||
/**
|
||
* 成果带逐段证据的一轮（报告撰写节点 + _evidence）。夹具是后端 compose_doc 真跑出来的，只用
|
||
* 通用名。复核回的是一次改写：有证据时复核只能加说明，答案不能被换掉
|
||
*/
|
||
const EVIDENCE = JSON.parse(readFileSync(new URL('../frontend/src/run/__tests__/evidence-doc.json', import.meta.url), 'utf8'))
|
||
const RUNS = {
|
||
'run-evid': { status: 'succeeded', output: EVIDENCE.output, error: null },
|
||
'run-evq': { status: 'succeeded', output: EVQ.output, error: null },
|
||
'run-led': { status: 'succeeded', output: EVIDENCE.output, error: null },
|
||
'run-cancelled': { status: 'cancelled', output: {}, error: null },
|
||
'run-failed': { status: 'failed', output: {}, error: '查询超时' },
|
||
'run-rejected': { status: 'failed', output: {}, error: '人工驳回：数字对不上' },
|
||
'run-auth': { status: 'failed', output: {}, error: 'AuthenticationError: Error code: 401 - invalid x-api-key' },
|
||
'run-broken': { status: 'succeeded', output: { answer: '只查了一半的结论' }, error: null },
|
||
'run-late': { status: 'succeeded', output: { answer: '补交付的答案：三条记录。' }, error: null,
|
||
started_at: iso(60 * 24 * 3 - 1), finished_at: iso(60 * 24 * 3 - 1.5) },
|
||
'run-t-cancel': { status: 'cancelled', output: {}, error: null },
|
||
// 鉴权失败的后端原文（core/errors.py）：run-t-failed 是老说法（库里老运行的 error），run-t-failed-now 是现在的说法
|
||
'run-t-failed': { status: 'failed', output: {}, error: '鉴权没通过（401）：对方拒绝了这把密钥' },
|
||
'run-t-failed-now': { status: 'failed', output: {}, error: '鉴权失败（401）：服务方拒绝了当前 API Key' },
|
||
'run-multi': { status: 'succeeded', output: { answer: FULL, note: '附注：另一个成果键' }, error: null },
|
||
'run-trash-live': { status: 'running', output: {}, error: null },
|
||
'run-badtool': { status: 'failed', output: {}, error: BROKEN_TOOL },
|
||
'run-badtool-now': { status: 'failed', output: {}, error: BROKEN_TOOL_NOW },
|
||
'run-rej2': { status: 'cancelled', end_reason: 'rejected', output: {}, error: '人工驳回：数字对不上' },
|
||
}
|
||
const SOURCES = [
|
||
{ id: 'ds-mig', name: 'shop', kind: 'mysql', description: '' },
|
||
{ id: 'ds-plant', name: 'factory', kind: 'oracle', description: '示例工厂库：采购、物料、生产、设备' },
|
||
]
|
||
/** 别处（另一个标签页、别人）刚加上的库：只有重新取列表才看得到 */
|
||
const EXTRA_SOURCE = { id: 'ds-new', name: 'warehouse', kind: 'postgres', description: '' }
|
||
const TABLES = {
|
||
'ds-mig': ['__prisma_migrations', 'schema_migrations', 'admin_audit_log', 'app_log', 'test_sales', 'user', 'orders'],
|
||
'ds-plant': ['ANALYTICS.test_sales', 'ANALYTICS.job_etl_audit', 'ANALYTICS.v_demo_table'],
|
||
}
|
||
|
||
// ---------------------------------------------------------------- 伪造的后端
|
||
|
||
const log = {
|
||
patches: [], cancels: [], continues: [], runStarts: [], archived: [], posts: [], approvalGets: [], runGets: [],
|
||
deleted: [], gens: [], sourceGets: 0, catalogPatches: [], reviews: [],
|
||
/** 问数据的一轮：发起（POST /api/chat/turns 的请求体）、停止、实时流的每一次连接（从哪个 seq 之后） */
|
||
turnStarts: [], turnCancels: [], streams: [],
|
||
}
|
||
/** 新服务端复核给的重来那句话：前端得原样交回去，不自己拼 */
|
||
const DECIDE_RETRY = '（服务端拼好的重来说法）照服务端的决定重新回答：有几家门店'
|
||
const ctl = {
|
||
slowMs: 0, failE: true, wsClosed: {}, approvals: [], thinkGo: false, writingGo: false, writingAt: 0,
|
||
/** 列表只给前 8 个：筛选框刚好出现，删一个就不够 8 个了 */
|
||
few: false, run500: true, reviewDelay: {}, launchFail: true, scope400: false,
|
||
/** 目录修改提案第一次保存回 409（别人刚改过这张表） */
|
||
patchConflict: false,
|
||
/** 回收站里多摆两个（TRASH_MORE） */
|
||
trashMore: false,
|
||
/** 列表里多摆「护栏」那段的三个会话（GUARD_CONVS） */
|
||
guardConvs: false,
|
||
/** 列表里多摆「问数据的证据」那段的会话（ASK_CONVS） */
|
||
askConvs: false,
|
||
/** 列表里多摆「助手反问」那段的会话（ASKBACK_CONVS） */
|
||
askBack: false,
|
||
/** 列表里多摆「先去改参数定义」那段的会话（FIX_CONVS） */
|
||
fixConvs: false,
|
||
/** 数据源列表取不回来 */
|
||
sourcesFail: false,
|
||
/** 数据源列表里多一个别处刚加的库 */
|
||
extraSource: false,
|
||
/** 列表里多摆「被驳回」那段的两个会话（REJECT_CONVS） */
|
||
rejectConvs: false,
|
||
/** 列表里多摆「口径账本」那段的会话（LEDGER_CONVS） */
|
||
ledgerConvs: false,
|
||
/** 伪造的服务端：运行的剧本放完了（`${runId}@${after}`）、继续运行过的运行、等复核最多多久（毫秒） */
|
||
wsDone: {}, continued: {}, reviewWaitMs: 25_000,
|
||
/** POST /api/chat/turns 根本没到后端（网断了） */
|
||
startFail: false,
|
||
/** 接下来几次实时流连接直接断掉（测断线重连补齐） */
|
||
dropStreams: 0,
|
||
/** 对话详情里带上服务端写的那几行（测刷新后接着看） */
|
||
liveDetail: false,
|
||
/** 助手每条操作之间隔多久（毫秒） */
|
||
opGapMs: 10,
|
||
/** 列表里多摆「刷新后接着看」那段的会话（REPLAY_CONVS）；那段里建图、运行各停在一道闸上 */
|
||
replayConvs: false, replayBuildGo: false, replayRunGo: false,
|
||
/** 「停止」那段：建图停在一道闸上 */
|
||
stopBuildGo: false,
|
||
}
|
||
/** 会话的归档 / 删除状态。各段各自 resetDb，免得上一段删掉的会话影响下一段 */
|
||
const db = { archived: new Set(), purged: new Set() }
|
||
const resetDb = () => {
|
||
db.archived = new Set([...Object.values(TRASHED), ...Object.values(TRASH_MORE)].map((c) => c.id))
|
||
db.purged = new Set()
|
||
}
|
||
resetDb()
|
||
const allConvs = () => [...Object.values(CONVS), ...Object.values(TRASHED), ...(ctl.trashMore ? Object.values(TRASH_MORE) : []),
|
||
...(ctl.guardConvs ? Object.values(GUARD_CONVS) : []), ...(ctl.askConvs ? Object.values(ASK_CONVS) : []),
|
||
...(ctl.askBack ? Object.values(ASKBACK_CONVS) : []), ...(ctl.replayConvs ? Object.values(REPLAY_CONVS) : []),
|
||
...(ctl.fixConvs ? Object.values(FIX_CONVS) : []), ...(ctl.rejectConvs ? Object.values(REJECT_CONVS) : []),
|
||
...(ctl.ledgerConvs ? Object.values(LEDGER_CONVS) : [])]
|
||
.filter((c) => !db.purged.has(c.id))
|
||
.map((c) => ({ ...c, archived: db.archived.has(c.id) }))
|
||
let runSeq = 0
|
||
|
||
async function fakeApi(route) {
|
||
const req = route.request()
|
||
const url = new URL(req.url())
|
||
const path = url.pathname.replace(/^\/api/, '')
|
||
const method = req.method()
|
||
const json = (body, status = 200) => route.fulfill({ status, json: body })
|
||
const body = () => { try { return req.postDataJSON() } catch { return {} } }
|
||
|
||
if (path === '/conversations' && method === 'GET') {
|
||
if (url.searchParams.get('kind') === 'canvas') return json([])
|
||
const rows = allConvs()
|
||
const live = rows.filter((c) => !c.archived)
|
||
const shown = ctl.few ? live.slice(0, 8) : live
|
||
return json(url.searchParams.get('include_archived') === 'true' ? [...shown, ...rows.filter((c) => c.archived)] : shown)
|
||
}
|
||
if (path === '/conversations' && method === 'POST') return json({ ...conv('c0new', '新对话', 0, 0), turns: [] }, 201)
|
||
let m = path.match(/^\/conversations\/([^/]+)$/)
|
||
if (m && method === 'GET') {
|
||
const id = m[1]
|
||
if (id === 'c0slowd') await new Promise((r) => setTimeout(r, ctl.slowMs))
|
||
if (id === 'c0faile' && ctl.failE) return json({ detail: '数据库暂时不可用' }, 500)
|
||
const known = allConvs().find((c) => c.id === id)
|
||
if (!known && id !== 'c0new') return json({ detail: '这个对话不存在，可能已经被删了' }, 404)
|
||
let turns = (typeof DETAIL[id] === 'function' ? DETAIL[id]() : DETAIL[id]) ?? []
|
||
if (ctl.liveDetail) {
|
||
// 服务端写的那几行（刷新、关掉再打开时取回来的就是它们）：夹具里有的换成服务端现在的，新问的接在后面
|
||
const live = [...srv.turns.values()].filter((T) => T.conversation_id === id && T.gen === srv.gen).map(rowOut)
|
||
turns = [...turns.map((t) => live.find((x) => x.id === t.id) ?? t), ...live.filter((x) => !turns.some((t) => t.id === x.id))]
|
||
}
|
||
return json({ ...(known ?? conv('c0new', '新对话', 0, 0)), turns })
|
||
}
|
||
if (m && method === 'PATCH') {
|
||
const b = body()
|
||
log.archived.push([m[1], b])
|
||
if (b.archived === true) db.archived.add(m[1])
|
||
if (b.archived === false) db.archived.delete(m[1])
|
||
return json({ ...allConvs().find((c) => c.id === m[1]), ...b })
|
||
}
|
||
if (m && method === 'DELETE') {
|
||
log.deleted.push(m[1])
|
||
db.purged.add(m[1])
|
||
return route.fulfill({ status: 204, body: '' })
|
||
}
|
||
m = path.match(/^\/conversations\/([^/]+)\/turns$/)
|
||
if (m && method === 'POST') {
|
||
log.posts.push(m[1])
|
||
return json(turn(`srv-${m[1]}-${log.posts.length}`, body().question, { status: 'running' }), 201)
|
||
}
|
||
m = path.match(/^\/conversations\/([^/]+)\/turns\/([^/]+)$/)
|
||
if (m && method === 'PATCH') { log.patches.push({ conv: m[1], turn: m[2], body: body() }); return json(turn(m[2], '', body())) }
|
||
|
||
// 问数据的一轮（伪造的服务端，见 startTurn / driveTurn）
|
||
if (path === '/chat/turns' && method === 'POST') {
|
||
const b = body()
|
||
if (ctl.startFail) {
|
||
log.turnStarts.push({ ...b, failed: true })
|
||
return route.abort('connectionrefused')
|
||
}
|
||
const got = await startTurn(b)
|
||
return json(got.json, got.status)
|
||
}
|
||
m = path.match(/^\/chat\/turns\/([^/]+)(\/stream|\/cancel)?$/)
|
||
if (m) {
|
||
const T = srv.turns.get(decodeURIComponent(m[1]))
|
||
if (!T) return json({ detail: '该轮对话不存在，可能已被删除' }, 404)
|
||
if (m[2] === '/stream' && method === 'GET') return streamTurn(route, T, Number(url.searchParams.get('after') ?? 0))
|
||
if (m[2] === '/cancel' && method === 'POST') {
|
||
log.turnCancels.push(T.id)
|
||
return cancelTurn(T) ? json({ ok: true }) : json({ detail: '这一轮已经结束，无需停止', code: 'chat_turn_finished' }, 409)
|
||
}
|
||
if (!m[2] && method === 'GET') return json({ id: T.id, conversation_id: T.conversation_id, job: { ...T.job }, turn: rowOut(T) })
|
||
}
|
||
if (path === '/copilot/generate-stream' && method === 'POST') {
|
||
// 画布助手照旧走这条流（问数据的一轮在伪造的服务端里走同一份剧本，见 copilotScript）
|
||
const b = body()
|
||
log.gens.push(b)
|
||
const got = copilotScript(b)
|
||
if (got.status) return json(got.json, got.status)
|
||
return route.fulfill({ status: 200, contentType: 'text/event-stream',
|
||
body: got.ops.map((o) => `data: ${JSON.stringify(o)}\n\n`).join('') })
|
||
}
|
||
if (path === '/datasources/src-shop/catalog/orders/patch' && method === 'POST') {
|
||
log.catalogPatches.push(body())
|
||
if (ctl.patchConflict) {
|
||
ctl.patchConflict = false
|
||
return json({ detail: '表「orders」的数据目录刚被修改过，请重新载入后再提交' }, 409)
|
||
}
|
||
return json({ table_name: 'orders', in_schema: true, notes: {}, version: 4, updated_at: null, updated_by: null,
|
||
structure: null, system_notes: false, usage: 0 })
|
||
}
|
||
if (path === '/datasources/src-shop/catalog/orders/patch/preview' && method === 'POST') {
|
||
log.catalogPatches.push({ preview: body() })
|
||
return json({ table_name: 'orders', version: 3, problems: [], changes: [
|
||
{ path: 'columns.status.codes', before: { 1: '已支付', 2: '已退款' }, before_status: 'confirmed',
|
||
after: { 1: '已支付', 2: '已退款', 9: '作废' }, value: { 9: '作废' }, reason: '用户说明 status=9 表示作废', state: 'change' }] })
|
||
}
|
||
if (path === '/datasources/src-shop/catalog/orders/impact' && method === 'GET') {
|
||
return json({ table: 'orders', templates: [] })
|
||
}
|
||
if (path === '/copilot/review' && method === 'POST') return json(await reviewScript(body()))
|
||
if (path === '/runs' && method === 'POST') {
|
||
const got = launchScript(body())
|
||
if (got.abort) return route.abort('connectionrefused')
|
||
if (got.status) return json(got.json, got.status)
|
||
return json({ id: got.id, status: 'queued', run_class: 'exploratory', output: {}, usage: {} })
|
||
}
|
||
m = path.match(/^\/runs\/([^/]+)\/(cancel|continue)$/)
|
||
if (m && method === 'POST') {
|
||
if (m[2] === 'cancel') { log.cancels.push(m[1]); return json({ ok: true }) }
|
||
log.continues.push(m[1])
|
||
if (m[1] === 'run-restart' || m[1] === 'run-susp') return json({ id: m[1], status: 'running' })
|
||
return json({ detail: '这次运行已完成，无需继续运行：只有失败的运行或被服务重启打断的运行可以继续运行。如需再次运行，请重新发起。' }, 409)
|
||
}
|
||
m = path.match(/^\/runs\/([^/]+)$/)
|
||
if (m && method === 'GET') {
|
||
log.runGets.push(m[1])
|
||
const got = runRecord(m[1])
|
||
return json(got.json, got.status)
|
||
}
|
||
if (path.startsWith('/approvals')) {
|
||
log.approvalGets.push(Date.now())
|
||
const runId = url.searchParams.get('run_id')
|
||
return json(ctl.approvals.filter((a) => !runId || a.run_id === runId))
|
||
}
|
||
m = path.match(/^\/runs\/([^/]+)\/events$/)
|
||
if (m && method === 'GET' && m[1] === 'run-rej2') {
|
||
// 被驳回的那一轮：「查数」里的动作确认被驳回，那一步的 node.failed 带 reason=rejected（服务端 compiler 的写法）
|
||
return json([ev(1, 'run.started', null, { nodes: 3 }), ev(2, 'node.started', 'ag', { node_type: 'agent', label: '查数' }),
|
||
ev(3, 'node.failed', 'ag', { duration_ms: 900, error: '人工驳回：数字对不上', reason: 'rejected', kind: 'action',
|
||
decision: 'reject', actor: '张工', note: '数字对不上' }),
|
||
ev(4, 'run.cancelled', null, { reason: 'rejected', node_id: 'ag', label: '查数', kind: 'action', decision: 'reject',
|
||
actor: '张工', note: '数字对不上' })])
|
||
}
|
||
if (m && method === 'GET') {
|
||
return json([ev(1, 'run.started', null, { nodes: 3 }), ev(2, 'node.started', 'ag', {}),
|
||
ev(3, 'node.finished', 'ag', { duration_ms: 700 }), ev(4, 'run.finished', null, { output: {} })])
|
||
}
|
||
if (path === '/datasources' && method === 'GET') {
|
||
log.sourceGets++
|
||
return ctl.sourcesFail ? json({ detail: '数据源列表暂时取不到' }, 500)
|
||
: json(ctl.extraSource ? [...SOURCES, EXTRA_SOURCE] : SOURCES)
|
||
}
|
||
m = path.match(/^\/datasources\/([^/]+)\/schema$/)
|
||
if (m) return json({ tables: TABLES[m[1]] ?? [] })
|
||
|
||
if (method === 'GET') return route.continue()
|
||
return route.abort() // 其余写操作一律不放
|
||
}
|
||
|
||
/**
|
||
* 复核的剧本（服务端 api/copilot.review_of_run 的结果），按运行 id 挑；有的故意慢（reviewDelay、lateReview）。问数据的一轮由伪造的
|
||
* 服务端（fakeTurns）调它，/copilot/review 照旧答它
|
||
*/
|
||
async function reviewScript(b) {
|
||
const wait = ctl.reviewDelay[b.run_id]
|
||
if (wait) await new Promise((r) => setTimeout(r, wait))
|
||
if (ctl.lateReview && String(b.run_id).startsWith('run-decide')) {
|
||
// 复核慢过问数据页等的上限（检查里调短成几百毫秒）：答案先交出去，这一次复核的用量晚到了照样要记进这一轮
|
||
await new Promise((r) => setTimeout(r, ctl.lateReview))
|
||
log.reviews.push(b)
|
||
return { verdict: 'ok', note: '', answer: null, retry: false, severity: '', signals: [], retry_instruction: null,
|
||
usage: { cost_usd: 0.004, unpriced_calls: 0, calls: 1 } }
|
||
}
|
||
if (String(b.run_id).startsWith('run-decide')) {
|
||
// 新服务端：第一次判不可用、给出重来的那句话；重来之后的那次没问题、不再重来。各带复核的用量
|
||
log.reviews.push(b)
|
||
const first = b.attempt === 0
|
||
return { verdict: first ? 'annotated' : 'ok', note: first ? '查到一半就停了' : '', answer: null, retry: first,
|
||
severity: first ? 'broken' : '', signals: first ? [{ kind: 'step_limit', detail: '用满了步数', severity: 'broken' }] : [],
|
||
retry_instruction: first ? DECIDE_RETRY : null,
|
||
usage: { cost_usd: first ? 0.003 : 0, unpriced_calls: 0, calls: first ? 1 : 0 } }
|
||
}
|
||
// 老后端的复核不认 _evidence，照样回一次改写
|
||
if (b.run_id === 'run-evid' || b.run_id === 'run-evq') {
|
||
return { verdict: 'rewritten', note: '检索降级过，结论请对照原始数据', answer: '改写后的答案：销售额大约四万多',
|
||
original: b.run_id === 'run-evq' ? EVQ.output.answer : EVIDENCE.output['周报'], retry: false, severity: 'degraded',
|
||
signals: [{ kind: 'retrieval_degraded', detail: '检索退回关键词', severity: 'degraded' }] }
|
||
}
|
||
return { verdict: 'ok', note: '', answer: null, retry: false, severity: '', signals: [] }
|
||
}
|
||
|
||
/** 发起一次运行（服务端 engine/launch.launch_run）：按问题挑这次运行的 id，记下发起时的图和来源。被拒的给 {status, json} */
|
||
function launchScript(b) {
|
||
const conversation = String(b.input?.question ?? '')
|
||
// 发起请求根本没到后端（网断了、后端正在重启）
|
||
if (conversation.includes('启动失败') && ctl.launchFail) return { abort: true }
|
||
// 发起就被拒：绑定的工具不存在（后端 runs.TOOL_MISSING，{detail, code} 的形状）
|
||
if (conversation.includes('工具不在')) {
|
||
return { status: 422, json: { detail: '绑定的工具不存在：「查数」（调用工具）绑定的 db_query__nope。请到「数据」页接入，或在节点中重新选择',
|
||
code: 'run_tool_missing' } }
|
||
}
|
||
const id = conversation.includes('整形解析失败') ? (conversation.includes('现在的说法')
|
||
? (conversation.includes('上游') ? 'run-jsonup-now' : 'run-jsonerr-now')
|
||
: conversation.includes('节点名带超时') ? 'run-jsonslow'
|
||
: conversation.includes('上游') ? 'run-jsonup'
|
||
: conversation.includes('改写') ? 'run-jsonauth' : 'run-jsonerr')
|
||
: conversation.includes('问数据带证据') ? 'run-evq'
|
||
: conversation.includes('带证据') ? 'run-evid'
|
||
: conversation.includes('长答案') ? (conversation.includes('取不到运行') ? 'run-noget' : 'run-long')
|
||
: conversation.includes('刷新后接着看') ? 'run-replay'
|
||
: conversation.includes('审批') ? 'run-wait'
|
||
: conversation.includes('重启') ? 'run-restart'
|
||
: conversation.includes('正在写') ? 'run-writing'
|
||
: conversation.includes('强杀') ? 'run-killed'
|
||
: conversation.includes('步数用满') ? `run-steps-${++runSeq}`
|
||
: conversation.includes('照服务端的决定') ? `run-decide-${++runSeq}`
|
||
: `run-busy-${++runSeq}`
|
||
log.runStarts.push({ id, graph: b.graph, origin_ref: b.origin_ref })
|
||
return { id }
|
||
}
|
||
|
||
/** 库里的运行（GET /api/runs/{id}）：{status, json} */
|
||
function runRecord(id) {
|
||
if (id === 'run-long') return { status: 200, json: { id, status: 'succeeded', run_class: 'exploratory', output: { answer: FULL }, usage: {} } }
|
||
if (id === 'run-noget' || (id === 'run-unk' && ctl.run500)) return { status: 500, json: { detail: '服务端错误（500），详情请查看服务日志' } }
|
||
if (id === 'run-unk') return { status: 200, json: { id, status: 'cancelled', run_class: 'exploratory', output: {}, usage: {} } }
|
||
if (id.startsWith('run-404')) return { status: 404, json: { detail: '运行不存在' } }
|
||
if (RUNS[id]) return { status: 200, json: { id, run_class: 'exploratory', usage: {}, ...RUNS[id] } }
|
||
if (id.startsWith('run-c') || id === 'run-b1') return { status: 200, json: { id, status: 'succeeded', run_class: 'exploratory', output: {}, usage: {} } }
|
||
if (id.startsWith('run-decide')) {
|
||
return { status: 200, json: { id, status: 'succeeded', run_class: 'exploratory', output: { answer: '三家门店。' },
|
||
usage: { cost_usd: 0.02, calls: 2, input_tokens: 6000, output_tokens: 400 } } }
|
||
}
|
||
return { status: 200, json: { id, status: 'interrupted', run_class: 'exploratory', output: {}, usage: {} } }
|
||
}
|
||
|
||
/**
|
||
* 助手生成过程的剧本（服务端 api/copilot.answer_ops 交出的那一串操作），按问法挑。开流之前就被拒的给 {status, json}。
|
||
* 画布助手经 /copilot/generate-stream 收它；问数据的一轮由伪造的服务端（fakeTurns）照它跑
|
||
*/
|
||
function copilotScript(b) {
|
||
if (ctl.scope400 && b.datasource_ids?.length) {
|
||
// 和后端 copilot._sources 同一个形状：detail 是一句话，code 是稳定的机读码
|
||
// reworded：后端换了说法，只剩机读码认得出来
|
||
return { status: 400, json: { detail: ctl.scope400 === 'reworded' ? '这一轮圈定的库一个也找不到了：可能已经被删掉或停用'
|
||
: '限定的数据源都已不可用，可能已被删除或停用。请取消限定后重试，或到「数据」页确认数据源存在且已启用', code: 'datasource_scope_empty' } }
|
||
}
|
||
if (String(b.instruction ?? '').includes('建流程就失败')) {
|
||
// 后端的 error 操作是三段：人话、怎么办、原始异常
|
||
const ops = [
|
||
{ op: 'heartbeat', phase: 'planning', elapsed_ms: 400 },
|
||
{ op: 'error', message: '模型返回的工作流结构有误：第 2 个节点的类型「sql」不存在',
|
||
hint: '请重试，或把需求描述得更具体',
|
||
detail: "2 validation errors for GraphSpec\nnodes.1.type\n Input should be 'input', 'output', 'llm' [type=literal_error]" },
|
||
]
|
||
return { ops }
|
||
}
|
||
if (String(b.instruction ?? '').includes('自查没修好')) {
|
||
// 服务端自查交回模型改过、还是没过（check failed），final 带着同一批问题：前端不该再整轮重来
|
||
const ops = [
|
||
{ op: 'heartbeat', phase: 'planning', elapsed_ms: 300 },
|
||
...GRAPH.nodes.map((n) => ({ op: 'add_node', node: n })),
|
||
{ op: 'check', status: 'failed', issues: ['「查数」节点还没有选择工具'] },
|
||
{ op: 'final', graph: GRAPH, explanation: '', issues: [{ level: 'error', node_id: 'ag', message: '「查数」节点还没有选择工具' }] },
|
||
]
|
||
return { ops }
|
||
}
|
||
if (String(b.instruction ?? '').includes('只写了思考')) {
|
||
// 生成过程一个操作都没给（deepseek 偶发只出思考）：服务端的 error 带 next=retry（backend/app/chat/turn.py 的 next_step）。
|
||
// 重来那一次（指令里缀了「上一次没有成功」）照常直接回答；「两次都」那一问每次都只写思考
|
||
const again = String(b.instruction).includes('上一次没有成功') && !String(b.instruction).includes('两次都')
|
||
const spent = { by_model: { m: { input_tokens: 500, output_tokens: 40, calls: 1, cost_usd: 0.001, unpriced_calls: 0 } },
|
||
input_tokens: 500, output_tokens: 40, calls: 1, cost_usd: 0.001, unpriced_calls: 0 }
|
||
const ops = again ? [
|
||
{ op: 'heartbeat', phase: 'planning', elapsed_ms: 200 },
|
||
{ op: 'reply', text: '已经查过了：一共 3 家门店。' },
|
||
{ op: 'usage', ...spent },
|
||
] : [
|
||
{ op: 'heartbeat', phase: 'planning', elapsed_ms: 200 },
|
||
{ op: 'thinking', delta: '先想一想要查哪张表……' },
|
||
{ op: 'error', message: '助手本轮未完成：模型没有给出任何修改，也没有说明原因', hint: '画布未修改。可以重试，或把需求描述得更具体',
|
||
detail: '', reason: '模型没有给出任何修改，也没有说明原因', code: 'copilot_empty', next: 'retry' },
|
||
{ op: 'usage', ...spent },
|
||
]
|
||
return { ops }
|
||
}
|
||
if (String(b.instruction ?? '').includes('照服务端的决定') || String(b.instruction ?? '').includes(DECIDE_RETRY)) {
|
||
// 新服务端：final 带 next（backend/app/chat/turn.py），最后一条是这一轮建图的用量（answer_ops 的 usage）。
|
||
// 「只建不跑」那一问：autorun 没写 false，next 却是 ready——前端得照 next 走
|
||
const ready = String(b.instruction ?? '').includes('只建不跑')
|
||
const ops = [
|
||
{ op: 'heartbeat', phase: 'planning', elapsed_ms: 300 },
|
||
...GRAPH.nodes.map((n) => ({ op: 'add_node', node: n })),
|
||
{ op: 'done', explanation: '一张三步的图' },
|
||
{ op: 'check', status: 'passed', repaired: 0, warnings: [] },
|
||
{ op: 'final', graph: GRAPH, issues: [], explanation: '一张三步的图', autorun: true, next: ready ? 'ready' : 'run' },
|
||
{ op: 'usage', by_model: { m: { input_tokens: 4000, output_tokens: 300, calls: 1, cost_usd: 0.012, unpriced_calls: 0 } },
|
||
input_tokens: 4000, output_tokens: 300, calls: 1, cost_usd: 0.012, unpriced_calls: 0 },
|
||
]
|
||
return { ops }
|
||
}
|
||
if (/刷新后接着看|停止测试/.test(String(b.instruction ?? ''))) {
|
||
// 建到一半停在闸上：刷新、关掉再打开、停止都要在建图中途发生
|
||
const gate = String(b.instruction).includes('停止测试') ? 'stopBuildGo' : 'replayBuildGo'
|
||
const ops = [
|
||
{ op: 'heartbeat', phase: 'planning', elapsed_ms: 300 },
|
||
{ op: 'plan', summary: '查一下再回答' },
|
||
...GRAPH.nodes.slice(0, 2).map((n) => ({ op: 'add_node', node: n })),
|
||
{ gate },
|
||
{ op: 'add_node', node: GRAPH.nodes[2] },
|
||
{ op: 'done', explanation: '一张三步的图' },
|
||
{ op: 'final', graph: GRAPH, issues: [], explanation: '一张三步的图', autorun: true, next: 'run' },
|
||
]
|
||
return { ops }
|
||
}
|
||
if (String(b.instruction ?? '').includes('先问清楚')) {
|
||
// 助手要先问清楚才能动手（服务端 copilot_gen 的 ask）：问题转出去，这一轮到此为止——没有 final，也没有 reply
|
||
const ops = [
|
||
{ op: 'heartbeat', phase: 'planning', elapsed_ms: 300 },
|
||
{ op: 'ask', question: '销售额按下单日还是支付日算？' },
|
||
]
|
||
return { ops }
|
||
}
|
||
if (String(b.instruction ?? '').includes('表示作废')) {
|
||
// 用户说了一条数据事实：服务端核对后转出目录修改提案，只回答、不建图
|
||
const ops = [
|
||
{ op: 'heartbeat', phase: 'planning', elapsed_ms: 300 },
|
||
{ op: 'catalog_patch', source: 'shop', source_id: 'src-shop', table: 'orders', table_label: '订单', version: 2, changes: [
|
||
{ path: 'columns.status.codes', before: { 1: '已支付' }, before_status: 'confirmed', after: { 1: '已支付', 9: '作废' },
|
||
value: { 9: '作废' }, reason: '用户说明 status=9 表示作废', state: 'change' }] },
|
||
{ op: 'reply', text: '已整理成数据目录的修改建议，确认后保存即可。' },
|
||
]
|
||
return { ops }
|
||
}
|
||
const ops = [
|
||
{ op: 'heartbeat', phase: 'planning', elapsed_ms: 800 },
|
||
{ op: 'heartbeat', phase: 'looking', elapsed_ms: 1000 },
|
||
// 助手按需查看了表（look 事件）：过程里多一行「查看了 N 张表」，查看记录存进这一轮的 meta.looks。
|
||
// 这个库还没有数据目录（catalog_tables: 0）：那一行下面提示先去起草目录
|
||
{ op: 'look', id: 'L1', by: 'model', need: '查订单和明细', status: 'ok', elapsed_ms: 1200, objects: [
|
||
{ kind: 'table', source: 'shop', source_id: 'src-shop', named: [], picked: ['orders', 'order_items'], joined: [],
|
||
missing: [], total: 48, catalog_tables: 0 }] },
|
||
{ op: 'plan', summary: '查一下再回答' },
|
||
...GRAPH.nodes.map((n) => ({ op: 'add_node', node: n })),
|
||
{ op: 'done', explanation: '一张三步的图' },
|
||
{ op: 'final', graph: GRAPH, issues: [], explanation: '一张三步的图' },
|
||
]
|
||
return { ops }
|
||
}
|
||
|
||
/** 每条运行的事件剧本。done=true 时剧本放完就发 stream.end */
|
||
const ev = (seq, type, node_id, data = {}) => ({ seq, type, node_id, data, ts: Date.now() / 1000 })
|
||
function script(runId, after = 0) {
|
||
if (runId === 'run-noget') {
|
||
// 事件里的成果被切过，交付时 GET 运行又失败了：只能先交这一份，并且要记下来
|
||
return {
|
||
events: [
|
||
ev(1, 'run.started', null, { nodes: 3 }),
|
||
ev(2, 'node.finished', 'ag', { duration_ms: 900 }),
|
||
ev(3, 'run.finished', null, { output: { answer: FULL.slice(0, 2000), note: '附注' }, output_truncated: true }),
|
||
],
|
||
end: 'succeeded',
|
||
}
|
||
}
|
||
if (runId === 'run-jsonerr' || runId === 'run-jsonup' || runId === 'run-jsonauth' || runId === 'run-jsonslow'
|
||
|| runId === 'run-jsonerr-now' || runId === 'run-jsonup-now') {
|
||
// 用户真实踩到的：整形节点按 JSON 解析上游的文字失败。旧说法（让人查模板）和新说法（点名上游）各一种；
|
||
// 对照：说法被改写过的（鉴权失败）原话还得留着
|
||
const error = runId === 'run-jsonerr' ? JSON_TEMPLATE_ERROR : runId === 'run-jsonup' ? JSON_UPSTREAM_ERROR
|
||
: runId === 'run-jsonerr-now' ? JSON_TEMPLATE_ERROR_NOW : runId === 'run-jsonup-now' ? JSON_UPSTREAM_ERROR_NOW
|
||
: runId === 'run-jsonslow' ? JSON_SLOW_ERROR
|
||
: 'AuthenticationError: Error code: 401 - invalid x-api-key'
|
||
const detail = "JSONDecodeError: Expecting ',' delimiter: line 1 column 934 (char 933)\n出错位置前后的原文：…\"gmv\": 45678.5, \"note\": \"含\"⟨此处⟩促销\"…"
|
||
return {
|
||
events: [
|
||
ev(1, 'run.started', null, { nodes: 3 }),
|
||
ev(2, 'node.started', 'ag', { node_type: 'agent', label: runId === 'run-jsonslow' ? '查询超时订单' : '查数' }),
|
||
ev(3, 'node.finished', 'ag', { duration_ms: 900 }),
|
||
ev(4, 'node.started', 'tf', { node_type: 'transform', label: '解析取数结果为变量' }),
|
||
ev(5, 'node.failed', 'tf', { error, detail, duration_ms: 3 }),
|
||
ev(6, 'run.failed', null, { error, detail }),
|
||
],
|
||
end: 'failed',
|
||
}
|
||
}
|
||
if (runId === 'run-evq') {
|
||
// 问数据：agent 查库（新运行的 tool.end 多一个 query_artifact）→ 按出处抽取字段 → 报告核对 → 成果
|
||
return {
|
||
events: [
|
||
ev(1, 'run.started', null, { nodes: 4 }),
|
||
ev(2, 'node.started', 'ag', { node_type: 'agent', label: '查数' }),
|
||
ev(3, 'llm.start', 'ag', { model: 'm' }),
|
||
ev(4, 'tool.start', 'ag', { tool: 'db_query__shop', call_id: 'q1', args: { sql: 'SELECT region, SUM(amount) AS amount FROM orders GROUP BY region' } }),
|
||
ev(5, 'tool.end', 'ag', { tool: 'db_query__shop', call_id: 'q1', duration_ms: 30, artifact: 'tool-snap', query_artifact: 'query-snap', rows: 4 }),
|
||
ev(6, 'llm.end', 'ag', { model: 'm', duration_ms: 800 }),
|
||
ev(7, 'llm.start', 'ag', { model: 'm', structured: true, purpose: 'cite_fields', message_count: 6 }),
|
||
ev(8, 'llm.end', 'ag', { agent: '查数', model: 'm', purpose: 'cite_fields', duration_ms: 900 }),
|
||
ev(9, 'log', 'ag', { level: 'warn', code: 'agent_field_mismatch', fields: ['order_cnt'],
|
||
message: '有 1 个字段与查询快照不一致，已按快照取值：order_cnt 模型给出 1240，快照为 1234' }),
|
||
ev(10, 'node.finished', 'ag', { duration_ms: 2100 }),
|
||
ev(11, 'node.started', 'write', { node_type: 'report', label: '写答案' }),
|
||
ev(12, 'report.checked', 'write', EVQ.report_checked),
|
||
ev(13, 'node.finished', 'write', { duration_ms: 900 }),
|
||
ev(14, 'run.finished', null, { output: EVQ.output }),
|
||
],
|
||
end: 'succeeded',
|
||
}
|
||
}
|
||
if (runId === 'run-evid') {
|
||
return {
|
||
events: [
|
||
ev(1, 'run.started', null, { nodes: 3 }),
|
||
ev(2, 'node.started', 'write', { node_type: 'report', label: '写周报' }),
|
||
ev(3, 'report.checked', 'write', EVIDENCE.report_checked),
|
||
ev(4, 'node.finished', 'write', { duration_ms: 900 }),
|
||
ev(5, 'run.finished', null, { output: EVIDENCE.output }),
|
||
],
|
||
end: 'succeeded',
|
||
}
|
||
}
|
||
if (runId === 'run-long') {
|
||
return {
|
||
events: [
|
||
ev(1, 'run.started', null, { nodes: 3 }),
|
||
ev(2, 'node.started', 'ag', {}),
|
||
...Array.from({ length: 60 }, (_, i) => ev(3 + i, 'llm.token', 'ag', { delta: '字' })),
|
||
ev(70, 'node.finished', 'ag', { duration_ms: 1200 }),
|
||
ev(71, 'run.finished', null, { output: { answer: FULL.slice(0, 2000) }, output_truncated: true }),
|
||
],
|
||
end: 'succeeded',
|
||
}
|
||
}
|
||
if (runId === 'run-replay') {
|
||
// 跑到一半停在闸上：关掉再打开时运行还在跑
|
||
return {
|
||
events: [
|
||
ev(1, 'run.started', null, { nodes: 3 }),
|
||
ev(2, 'node.started', 'ag', { node_type: 'agent', label: '查数' }),
|
||
{ gate: 'replayRunGo' },
|
||
ev(3, 'node.finished', 'ag', { duration_ms: 900 }),
|
||
ev(4, 'run.finished', null, { output: { answer: '刷新之后接着跑完了：两家门店。' } }),
|
||
],
|
||
end: 'succeeded',
|
||
}
|
||
}
|
||
if (runId === 'run-restart' && after >= 4) {
|
||
// 接着跑：从断点往下，这次跑完
|
||
return {
|
||
events: [
|
||
ev(5, 'run.resumed', null, { from: 'ag', message: '从「查数」继续运行' }),
|
||
ev(6, 'node.started', 'ag', { resumed: true }),
|
||
ev(7, 'tool.start', 'ag', { tool: 'db_query', call_id: 'q2', args: { sql: 'select 1' } }),
|
||
ev(8, 'tool.end', 'ag', { tool: 'db_query', call_id: 'q2', rows: 3 }),
|
||
ev(9, 'node.finished', 'ag', { duration_ms: 800 }),
|
||
ev(10, 'run.finished', null, { output: { answer: '接着跑完了：三条记录。' } }),
|
||
],
|
||
end: 'succeeded',
|
||
}
|
||
}
|
||
if (runId === 'run-killed') {
|
||
// 进程被强杀：连结束标记都来不及发，连接直接断了。前端自己重连，
|
||
// 重启后的后端回放完补一条 stream.end status=interrupted
|
||
if (after >= 2) return { events: [], end: 'interrupted' }
|
||
return { events: [ev(1, 'run.started', null, { nodes: 3 }), ev(2, 'node.started', 'ag', {})], end: 'close' }
|
||
}
|
||
if (runId === 'run-restart') {
|
||
return {
|
||
events: [
|
||
ev(1, 'run.started', null, { nodes: 3 }),
|
||
ev(2, 'node.started', 'ag', {}),
|
||
ev(3, 'tool.start', 'ag', { tool: 'db_query', call_id: 'q1', args: { sql: 'select 1' } }),
|
||
ev(4, 'log', null, { level: 'warn', code: 'server_shutdown', message: '服务正在关停' }),
|
||
],
|
||
end: 'interrupted',
|
||
}
|
||
}
|
||
if (runId === 'run-wait') {
|
||
// 停在审批上：事件流不关，后端也不发 stream.end
|
||
return {
|
||
events: [
|
||
ev(1, 'run.started', null, { nodes: 3 }),
|
||
ev(2, 'node.started', 'ag', {}),
|
||
ev(3, 'human.requested', 'ag', { interrupt_id: 'ap-1', payload: { title: '确认一下再查' } }),
|
||
ev(4, 'run.interrupted', null, { node_id: 'ag' }),
|
||
],
|
||
end: null,
|
||
}
|
||
}
|
||
if (runId === 'run-writing') {
|
||
// 先想、再写，写完不收尾：一次还在出字的模型调用。gate 处等脚本放行
|
||
// 第一段停在半句上：半句不该出现在头部，要等它写完
|
||
const tokens = [...Array.from({ length: 12 }, () => '字'.repeat(100)), '字'.repeat(34)] // 1,234 字
|
||
return {
|
||
events: [
|
||
ev(1, 'run.started', null, { nodes: 3 }),
|
||
ev(2, 'node.started', 'ag', {}),
|
||
ev(3, 'llm.start', 'ag', { model: 'm' }),
|
||
ev(4, 'llm.thinking.delta', 'ag', { delta: '先看看 factory 里有哪些表。' }),
|
||
ev(5, 'llm.thinking.delta', 'ag', { delta: '然后按月汇总出勤' }),
|
||
{ gate: 'thinkGo' },
|
||
ev(6, 'llm.thinking.delta', 'ag', { delta: '率，再和上个月比。' }),
|
||
{ gate: 'writingGo' },
|
||
...tokens.map((delta, i) => ev(10 + i, 'llm.token', 'ag', { delta })),
|
||
],
|
||
end: null,
|
||
}
|
||
}
|
||
if (runId.startsWith('run-decide')) {
|
||
return {
|
||
events: [
|
||
ev(1, 'run.started', null, { nodes: 3 }),
|
||
ev(2, 'node.started', 'ag', {}),
|
||
ev(3, 'node.finished', 'ag', { duration_ms: 900 }),
|
||
ev(4, 'run.finished', null, { output: { answer: '三家门店。' } }),
|
||
],
|
||
end: 'succeeded',
|
||
}
|
||
}
|
||
if (runId.startsWith('run-steps')) {
|
||
return {
|
||
events: [
|
||
ev(1, 'run.started', null, { nodes: 3 }),
|
||
ev(2, 'node.started', 'ag', {}),
|
||
ev(3, 'node.finished', 'ag', { duration_ms: 900 }),
|
||
ev(4, 'run.finished', null, { output: { answer: '放宽步数之后查全了' } }),
|
||
],
|
||
end: 'succeeded',
|
||
}
|
||
}
|
||
// 忙的那条：开了头就停住，流不关，像一次还在跑的运行
|
||
return { events: [ev(1, 'run.started', null, { nodes: 3 }), ev(2, 'node.started', 'ag', {})], end: null }
|
||
}
|
||
|
||
async function fakeStream(ws) {
|
||
const runId = new URL(ws.url()).pathname.split('/')[3]
|
||
ctl.wsClosed[runId] = false
|
||
let closed = false
|
||
ws.onClose(() => { closed = true; ctl.wsClosed[runId] = true })
|
||
const after = Number(new URL(ws.url()).searchParams.get('after') ?? 0)
|
||
const { events, end } = script(runId, after)
|
||
for (const e of events) {
|
||
if (e.gate) {
|
||
await until(() => ctl[e.gate], 15000)
|
||
continue
|
||
}
|
||
if (e.seq <= after) continue
|
||
if (closed) return
|
||
if (runId === 'run-writing' && e.type === 'llm.token' && !ctl.writingAt) ctl.writingAt = Date.now()
|
||
// 审批和中断同时产生：后端在停下的那一刻才建出这条待审批
|
||
if (runId === 'run-wait' && e.type === 'run.interrupted') {
|
||
ctl.approvals = [ctl.pendingApproval]
|
||
ctl.interruptedAt = Date.now()
|
||
}
|
||
ws.send(JSON.stringify(e))
|
||
await new Promise((r) => setTimeout(r, 15))
|
||
}
|
||
if (closed) return
|
||
if (end === 'close') ws.close()
|
||
else if (end) ws.send(JSON.stringify({ type: 'stream.end', status: end, data: { status: end } }))
|
||
// 剧本放完了：伪造的服务端据此认这次运行到了终态（真服务端按库里的状态，事件先落库）
|
||
if (end) ctl.wsDone[`${runId}@${after}`] = true
|
||
}
|
||
|
||
// ---------------------------------------------------------------- 伪造的服务端一轮
|
||
|
||
/**
|
||
* 问数据的一轮在服务端跑（backend/app/chat/job.py；接口 backend/app/api/chat_turns.py）。页面只发起（POST /api/chat/turns）、
|
||
* 只看（GET /api/chat/turns/{id}/stream 先回放、再推实时，断了按 seq 重连），这里照服务端的做法把一轮跑完：
|
||
* 生成（copilotScript 的剧本，原样记进 log.gens）→ final.next → 发起运行（launchScript，记进 log.runStarts）→ 等运行的剧本放完
|
||
* （fakeStream）→ 复核（reviewScript，最多等 ctl.reviewWaitMs，晚到的照记用量）→ 按复核重来一次 → 照服务端的写法拼那一行（end）。
|
||
* 实时流是长轮询：没有新事件就等一会儿再答，答完连接就关；页面按收到的最后一个 seq 马上接着连（和服务端断线时一样）。
|
||
*/
|
||
const srv = { turns: new Map(), seq: 0, launched: new Set(), gen: 0 }
|
||
const ACTIVE_JOB = new Set(['queued', 'running'])
|
||
const turnActive = (T) => ACTIVE_JOB.has(T.job.status) || !!T.late
|
||
const sleep = (ms) => new Promise((r) => setTimeout(r, ms))
|
||
const BUSY_TEXT = '这段对话正有一轮在进行：等它结束或先停止，再提问'
|
||
|
||
function emit(T, type, data = {}) {
|
||
const e = { seq: T.events.length + 1, type, data, ts: Date.now() / 1000 }
|
||
T.events.push(e)
|
||
T.job.last_seq = e.seq
|
||
return e.seq
|
||
}
|
||
|
||
/** 那一行（GET /api/conversations/{id} 里的一轮、end 里带的 turn） */
|
||
const rowOut = (T) => ({ ...T.row, meta: T.row.meta ? { ...T.row.meta } : T.row.meta, job: { ...T.job } })
|
||
|
||
const addTotal = (a, b) => {
|
||
const out = { ...(a ?? {}) }
|
||
for (const k of ['cost_usd', 'unpriced_calls', 'calls', 'unfinished_calls', 'input_tokens', 'output_tokens']) {
|
||
const v = Number(b?.[k] ?? 0)
|
||
if (Number.isFinite(v) && v) out[k] = Math.round((Number(out[k] ?? 0) + v) * 1e6) / 1e6
|
||
}
|
||
return out
|
||
}
|
||
const retireRun = (usage) => {
|
||
const { run, ...rest } = usage ?? {}
|
||
return run && Object.keys(run).length ? { ...rest, earlier: addTotal(rest.earlier, run) } : rest
|
||
}
|
||
const answerText = (output) => Object.entries(output ?? {})
|
||
.filter(([k, v]) => !k.startsWith('_') && v != null && v !== '').map(([, v]) => (typeof v === 'string' ? v : JSON.stringify(v))).join('\n')
|
||
|
||
/** 用户点重试：上一次的结局归档（同服务端 chat/record.attempt_record） */
|
||
function attemptRecord(row, n) {
|
||
const meta = row.meta ?? {}
|
||
const review = row.review ?? {}
|
||
const base = { n, at: Date.now(), runId: meta.runId !== undefined ? meta.runId : row.run_id ?? null,
|
||
...(row.answer ? { answer: row.answer.length > 4000 ? `${row.answer.slice(0, 4000)}…` : row.answer } : {}) }
|
||
const signals = (review.signals ?? []).map((x) => x.detail).join('\n') || undefined
|
||
if (meta.outcome === 'cancelled' && meta.rejected) {
|
||
const sub = meta.rejected.sub ?? ''
|
||
return { ...base, outcome: 'rejected', summary: `被驳回${sub ? `：${sub}` : ''}` }
|
||
}
|
||
if (meta.outcome === 'cancelled' || (row.status === 'error' && row.error === '已取消')) return { ...base, outcome: 'cancelled', summary: '已手动停止' }
|
||
if (meta.outcome === 'suspended') return { ...base, outcome: 'suspended', summary: '服务重启，运行中断' }
|
||
if (row.status === 'error') {
|
||
const f = meta.failure
|
||
return { ...base, outcome: 'failed', summary: f ? (f.reason ? `${f.title}：${f.reason}` : f.title) : row.error || '发生错误',
|
||
...(meta.cause ? { cause: meta.cause } : {}) }
|
||
}
|
||
if (review.severity === 'broken') return { ...base, outcome: 'unusable', summary: `结论不可用：${review.note ?? ''}`, detail: signals }
|
||
if (review.severity === 'degraded') return { ...base, outcome: 'partial', summary: `存在缺口：${review.note ?? ''}`, detail: signals }
|
||
return { ...base, outcome: 'done', summary: '已给出答案' }
|
||
}
|
||
const attemptReason = (a) => (a.detail ? `${a.summary}（${String(a.detail).slice(0, 300)}）` : a.summary)
|
||
|
||
/** 种一轮：库里已经有的（夹具里的那一行），或者新问的 */
|
||
function seedTurn(conversationId, id, row) {
|
||
const T = {
|
||
id, conversation_id: conversationId, row: { ...row, meta: row.meta ? { ...row.meta } : null },
|
||
job: { status: 'succeeded', stage: null, mode: 'ask', run_id: null, attempt: 0, last_seq: 0, attempt_seq: 0, error: null },
|
||
events: [], cancelled: false, late: null, runId: null, usage: {}, looks: [],
|
||
}
|
||
srv.turns.set(id, T)
|
||
return T
|
||
}
|
||
function knownRow(conversationId, turnId) {
|
||
const rows = typeof DETAIL[conversationId] === 'function' ? DETAIL[conversationId]() : DETAIL[conversationId] ?? []
|
||
return rows.find((t) => t.id === turnId) ?? null
|
||
}
|
||
const scopeOf = (ids) => (ids ?? []).map((id) => {
|
||
const hit = [...SOURCES, EXTRA_SOURCE].find((x) => x.id === id || x.name === id)
|
||
return hit ? { id: hit.id, name: hit.name } : { id, name: id }
|
||
})
|
||
|
||
/** 这一轮还在跑时每一步之间看一眼：停了、这一段检查结束了，就不再往下写 */
|
||
const gone = (T, epoch) => T.cancelled || T.epoch !== epoch || T.gen !== srv.gen
|
||
|
||
async function startTurn(b) {
|
||
const mode = b.mode ?? (b.retry_of ? 'retry' : 'ask')
|
||
let T
|
||
if (mode === 'ask') {
|
||
log.posts.push(b.conversation_id)
|
||
const id = `srv-${b.conversation_id}-${++srv.seq}`
|
||
const scope = scopeOf(b.datasource_ids)
|
||
T = seedTurn(b.conversation_id, id, turn(id, b.question, { status: 'running', meta: { v: 1, ...(scope.length ? { scope } : {}) } }))
|
||
} else {
|
||
T = srv.turns.get(b.retry_of)
|
||
// 夹具里的那一行：每一段检查从库里原样的那一行起（上一段、上一套主题对它做过什么都不算）
|
||
const row = !T || T.gen !== srv.gen ? knownRow(b.conversation_id, b.retry_of) : null
|
||
if (row) T = seedTurn(b.conversation_id, b.retry_of, row)
|
||
if (!T) return { status: 404, json: { detail: '该轮对话不存在，可能已被删除' } }
|
||
}
|
||
const others = [...srv.turns.values()].filter((x) => x !== T && x.conversation_id === b.conversation_id
|
||
&& ACTIVE_JOB.has(x.job.status) && x.job.stage !== 'waiting' && x.gen === srv.gen)
|
||
if (mode !== 'resume' && others.length) return { status: 409, json: { detail: BUSY_TEXT, code: 'chat_turn_busy' } }
|
||
if (mode !== 'ask' && ACTIVE_JOB.has(T.job.status)) return { status: 409, json: { detail: BUSY_TEXT, code: 'chat_turn_busy' } }
|
||
const meta = { ...(T.row.meta ?? {}) }
|
||
const runId = meta.runId !== undefined ? meta.runId : T.row.run_id ?? null
|
||
const now = Date.now()
|
||
const request = { mode, started_ms: now }
|
||
let attempts = [...(meta.attempts ?? [])]
|
||
const scope = b.datasource_ids ? scopeOf(b.datasource_ids) : meta.scope ?? []
|
||
if (mode === 'continue') {
|
||
log.continues.push(runId)
|
||
if (runId !== 'run-restart' && runId !== 'run-susp') {
|
||
return { status: 409, json: { detail: '这次运行已完成，无需继续运行：只有失败的运行或被服务重启打断的运行可以继续运行。如需再次运行，请重新发起。' } }
|
||
}
|
||
ctl.continued[runId] = true
|
||
delete ctl.wsDone[`${runId}@4`]
|
||
}
|
||
if (mode === 'resume' && ctl.run500 && runId === 'run-unk') {
|
||
// 服务端核对那一下出错（查库超时之类）：页面说没核对上，给「重新核对」
|
||
return { status: 500, json: { detail: '服务端错误（500），详情请查看服务日志' } }
|
||
}
|
||
if (mode === 'continue' || mode === 'resume') {
|
||
request.follow = runId
|
||
if (T.row.status !== 'running' && typeof meta.ms === 'number') request.started_ms = now - meta.ms
|
||
const rec = runRecord(String(runId ?? ''))
|
||
if (T.row.status === 'running' && rec.json?.started_at) request.started_ms = Date.parse(rec.json.started_at)
|
||
if (mode === 'resume' && ['succeeded', 'failed', 'cancelled'].includes(rec.json?.status) && rec.json?.finished_at) {
|
||
request.ended_ms = Date.parse(rec.json.finished_at)
|
||
}
|
||
T.row = { ...T.row, status: 'running', error: '', meta: { ...meta, scope: scope.length ? scope : undefined } }
|
||
} else {
|
||
if (['rerun', 'max_steps', 'run'].includes(mode) && !(T.row.graph?.nodes?.length)) {
|
||
return { status: 409, json: { detail: '这一轮没有可运行的工作流，请重试本轮', code: 'chat_turn_invalid' } }
|
||
}
|
||
if (runId && mode !== 'run') {
|
||
// 服务重启挂起的旧运行没人会再去续它：顺手取消（停在审批上的不算）
|
||
const rec = runRecord(runId)
|
||
const pending = ctl.approvals.some((a) => a.run_id === runId && a.status === 'pending')
|
||
if (rec.json?.status === 'interrupted' && !pending) log.cancels.push(runId)
|
||
}
|
||
if (mode !== 'run') {
|
||
const archived = attemptRecord(T.row, attempts.length + 1)
|
||
attempts.push(archived)
|
||
if (mode === 'retry') {
|
||
const reason = attemptReason(archived)
|
||
request.base = !!T.row.graph?.nodes?.length
|
||
request.instruction = request.base
|
||
? `这个工作流上一次没有给出可用的结果：${reason}。\n\n请针对这个原因改它（比如调大最大步数、改正工具参数、换个查询方式），改完继续回答原来的问题：${T.row.question}`
|
||
: `${T.row.question}\n\n（上一次没有成功：${reason}。请针对这个原因调整，重新给出完整的工作流。）`
|
||
}
|
||
if (mode === 'max_steps') request.max_steps = b.max_steps
|
||
}
|
||
const usage = mode === 'run' ? meta.usage : retireRun(meta.usage)
|
||
T.row = { ...T.row, status: 'running', error: '', answer: mode === 'run' ? T.row.answer : '', review: {}, run_id: null,
|
||
meta: mode === 'run' ? { ...meta, runId: null } : { v: 1, ...(attempts.length ? { attempts } : {}),
|
||
...(scope.length ? { scope } : {}), ...(usage && Object.keys(usage).length ? { usage } : {}), runId: null } }
|
||
}
|
||
T.epoch = (T.epoch ?? 0) + 1
|
||
T.gen = srv.gen
|
||
T.cancelled = false
|
||
T.late = null
|
||
T.request = request
|
||
T.scope = scope
|
||
T.job = { status: 'running', stage: null, mode, run_id: request.follow ?? null, attempt: 0, last_seq: T.events.length,
|
||
attempt_seq: 0, error: null }
|
||
T.job.attempt_seq = emit(T, 'turn.attempt', {
|
||
mode, attempts, scope, usage: T.row.meta?.usage ?? {}, started_ms: request.started_ms,
|
||
...(request.ended_ms ? { ended_ms: request.ended_ms } : {}), ...(request.max_steps ? { max_steps: request.max_steps } : {}),
|
||
})
|
||
log.turnStarts.push({ ...b, _turn: T.id })
|
||
void driveTurn(T, T.epoch).catch((e) => console.log(` · 伪造的服务端出错：${e?.stack ?? e}`))
|
||
return { status: 202, json: { turn_id: T.id, turn: rowOut(T) } }
|
||
}
|
||
|
||
function stage(T, name) {
|
||
T.job.stage = name
|
||
emit(T, 'stage', { stage: name })
|
||
}
|
||
|
||
function saveRow(T, fields, meta) {
|
||
const merged = { ...(T.row.meta ?? {}), v: 1 }
|
||
for (const [k, v] of Object.entries(meta ?? {})) {
|
||
if (v === undefined || v === null) delete merged[k]
|
||
else merged[k] = v
|
||
}
|
||
merged.at = Date.now()
|
||
T.row = { ...T.row, ...fields, meta: merged }
|
||
}
|
||
|
||
/** 这一轮到此为止：写那一行、发 end（同服务端 ChatTurnExecution._conclude） */
|
||
function settle(T, outcome, fields, meta, output = null, run = null) {
|
||
const ms = Math.max(0, (T.request.ended_ms ?? Date.now()) - T.request.started_ms)
|
||
saveRow(T, fields, { looks: T.looks.length ? T.looks : undefined, ms, ...meta })
|
||
emit(T, 'end', { outcome, turn: rowOut(T), output, run })
|
||
T.job = { ...T.job, status: 'succeeded', stage: null }
|
||
}
|
||
|
||
async function driveTurn(T, epoch) {
|
||
const req = T.request
|
||
const mode = req.mode
|
||
const autoRetry = mode === 'ask' || mode === 'retry'
|
||
let attempts = 0
|
||
let instruction = req.instruction ?? T.row.question
|
||
let base = req.base ? T.row.graph : null
|
||
let direct = ['rerun', 'max_steps', 'run'].includes(mode) ? T.row.graph : null
|
||
if (mode === 'max_steps') {
|
||
direct = { ...direct, nodes: direct.nodes.map((n) => (n.type === 'agent'
|
||
? { ...n, data: { ...n.data, config: { ...(n.data?.config ?? {}), max_steps: req.max_steps } } } : n)) }
|
||
T.row = { ...T.row, graph: direct }
|
||
}
|
||
let follow = req.follow ?? null
|
||
T.usage = { ...(T.row.meta?.usage ?? {}) }
|
||
T.looks = []
|
||
while (true) {
|
||
if (!direct && !follow) {
|
||
stage(T, 'build')
|
||
const gen = { instruction, base_graph: base, intent: 'answer', conversation_id: T.conversation_id,
|
||
...(T.scope?.length ? { datasource_ids: T.scope.map((x) => x.id) } : {}) }
|
||
log.gens.push(gen)
|
||
const got = copilotScript(gen)
|
||
if (got.status) {
|
||
const message = got.json.detail
|
||
return settle(T, 'error', { status: 'error', error: message, review: {} },
|
||
{ cause: { kind: 'refused', message, code: got.json.code ?? null }, runId: null, usage: T.usage })
|
||
}
|
||
const ops = []
|
||
T.beatPhase = undefined // 每一次生成从头算：第一拍心跳照落
|
||
for (const op of got.ops) {
|
||
if (op.gate) { await until(() => ctl[op.gate] || gone(T, epoch), 15000); continue }
|
||
if (gone(T, epoch)) return
|
||
// 心跳只落换了阶段的那一拍（服务端 chat/job.ChatDriver.on_op）；同一阶段接着的不落库、不带 seq（chat/log.beat）：只记
|
||
// 最近的一拍，流上发一条不带 seq 的；之后落了新事件就作废
|
||
if (op.op === 'heartbeat' && op.phase === T.beatPhase) T.beat = { op, at: T.events.length }
|
||
else {
|
||
if (op.op === 'heartbeat') T.beatPhase = op.phase
|
||
emit(T, 'op', op)
|
||
}
|
||
if (op.op !== 'thinking' && op.op !== 'heartbeat') ops.push(op)
|
||
if (op.op === 'look') T.looks.push(op)
|
||
if (op.op === 'usage') T.usage.copilot = addTotal(T.usage.copilot, op)
|
||
if (op.op === 'final' && !op.asked) T.row = { ...T.row, graph: op.graph, explanation: op.explanation ?? '' }
|
||
await sleep(ctl.opGapMs ?? 10)
|
||
}
|
||
if (gone(T, epoch)) return
|
||
const terminal = ops.filter((o) => ['final', 'reply', 'error'].includes(o.op)).at(-1)
|
||
if (!terminal) {
|
||
const ask = ops.filter((o) => o.op === 'ask').at(-1)
|
||
const q = String(ask?.question ?? '')
|
||
return settle(T, 'asked', { status: 'done', answer: q, review: {} },
|
||
{ asked: { question: q, options: ask?.options ?? [] }, runId: null, usage: T.usage }, q ? { answer: q } : null)
|
||
}
|
||
if (terminal.op === 'error') {
|
||
if (terminal.next === 'retry' && attempts + 1 < 2) {
|
||
attempts += 1
|
||
const reason = terminal.reason ?? String(terminal.message ?? '').replace(/^助手本轮未完成：/, '')
|
||
instruction = base ? `这个工作流上一次运行没有成功：${reason}。\n\n请针对这个原因改它（比如调大最大步数、改正工具参数、换个查询方式），改完继续回答原来的问题：${T.row.question}`
|
||
: `${T.row.question}\n\n（上一次没有成功：${reason}。请针对这个原因调整，重新给出完整的工作流。）`
|
||
emit(T, 'turn.retry', { kind: 'empty', reason, base: !!base, attempt: attempts, usage: T.usage })
|
||
continue
|
||
}
|
||
const message = String(terminal.message ?? '助手本轮未完成')
|
||
return settle(T, 'error', { status: 'error', error: message, review: {} },
|
||
{ cause: { kind: 'copilot', message, hint: terminal.hint, detail: terminal.detail }, runId: null, usage: T.usage })
|
||
}
|
||
if (terminal.op === 'reply') {
|
||
const text = String(terminal.text ?? '')
|
||
return settle(T, 'reply', { status: 'done', answer: text, review: {} }, { noQuery: true, runId: null, usage: T.usage },
|
||
{ answer: text })
|
||
}
|
||
const blockers = (terminal.issues ?? []).filter((i) => i.level === 'error').map((i) => i.message)
|
||
const selfChecked = ops.some((o) => o.op === 'check')
|
||
const asked = ops.filter((o) => o.op === 'ask').at(-1)
|
||
const next = terminal.next ?? (blockers.length ? (selfChecked ? 'fail' : 'retry') : terminal.autorun === false ? 'ready' : 'run')
|
||
if (next === 'fail' || next === 'retry') {
|
||
return settle(T, 'error', { status: 'error', error: `生成的工作流没有通过校验：${blockers.join('；')}`, graph: terminal.graph,
|
||
explanation: terminal.explanation ?? '', review: {} },
|
||
{ cause: { kind: 'blocked', issues: blockers, self_checked: selfChecked }, runId: null, usage: T.usage })
|
||
}
|
||
if (next === 'ready') {
|
||
const q = String(asked?.question ?? '').trim()
|
||
return settle(T, 'ready', { status: 'done', graph: terminal.graph, explanation: terminal.explanation ?? '', review: {},
|
||
...(q ? { answer: q } : {}) },
|
||
{ runId: null, usage: T.usage, ...(q ? { asked: { question: q, options: [] } } : {}) }, q ? { answer: q } : null)
|
||
}
|
||
direct = terminal.graph
|
||
}
|
||
|
||
stage(T, 'run')
|
||
let runId = follow
|
||
if (!runId) {
|
||
const launched = launchScript({ graph: direct, input: { question: T.row.question, history: '' }, origin: 'chat',
|
||
origin_ref: { conversation_id: T.conversation_id, turn_id: T.id } })
|
||
if (launched.abort || launched.status) {
|
||
const message = launched.json?.detail ?? '运行没有发起'
|
||
return settle(T, 'error', { status: 'error', error: message, review: {} },
|
||
{ cause: { kind: 'start', message, code: launched.json?.code ?? null, status: launched.status ?? 0 }, runId: null, usage: T.usage })
|
||
}
|
||
runId = launched.id
|
||
srv.launched.add(runId)
|
||
// 同一个运行 id 各段、各套主题会再用：这一次从头算（剧本还没放、没有接着跑过）
|
||
delete ctl.wsDone[`${runId}@0`]
|
||
delete ctl.continued[runId]
|
||
delete T.usage.run
|
||
T.runId = runId
|
||
T.job.run_id = runId
|
||
saveRow(T, { run_id: runId, status: 'running' }, { runId, runClass: 'exploratory', runStatus: 'queued' })
|
||
emit(T, 'run', { run_id: runId, status: 'queued', run_class: 'exploratory' })
|
||
} else {
|
||
T.runId = runId
|
||
}
|
||
const graphRun = direct
|
||
direct = follow = null
|
||
const result = await followRun(T, epoch, runId)
|
||
if (gone(T, epoch)) return
|
||
const rec = runRecord(runId)
|
||
const run = { id: runId, status: result.status, run_class: 'exploratory', usage: rec.json?.usage ?? {},
|
||
error: result.error ?? null, end_reason: result.endReason ?? null }
|
||
T.usage.run = addTotal({}, rec.json?.usage)
|
||
const meta = { runId, runClass: 'exploratory', runStatus: result.status, usage: T.usage }
|
||
if (result.status === 'gone') {
|
||
return settle(T, 'error', { status: 'error', error: '运行记录不见了', review: {} }, { ...meta, cause: { kind: 'gone' }, runStatus: 'gone' })
|
||
}
|
||
if (result.status === 'rejected') {
|
||
return settle(T, 'cancelled', { status: 'error', error: '已取消', review: {} },
|
||
{ ...meta, runStatus: 'cancelled', outcome: 'cancelled', rejected: result.rejected }, null, { ...run, status: 'cancelled' })
|
||
}
|
||
if (result.status === 'cancelled') {
|
||
return settle(T, 'cancelled', { status: 'error', error: '已取消', review: {} }, { ...meta, outcome: 'cancelled' }, null, run)
|
||
}
|
||
if (result.status === 'suspended') {
|
||
return settle(T, 'suspended', { status: 'error', error: '服务重启，本轮已中断', review: {} },
|
||
{ ...meta, outcome: 'suspended', runStatus: 'interrupted' }, null, { ...run, status: 'interrupted' })
|
||
}
|
||
if (result.status === 'failed') {
|
||
return settle(T, 'error', { status: 'error', error: result.error ?? '运行失败', review: {} },
|
||
{ ...meta, cause: { kind: 'run', error: result.error ?? '', ...(result.detail ? { detail: result.detail } : {}) } }, null, run)
|
||
}
|
||
|
||
stage(T, 'review')
|
||
// 成果带逐段证据时复核只能加说明（服务端 review_of_run 按 locked 不改写）：剧本里老后端那种改写在这里收成「加说明」
|
||
const locked = !!(result.output ?? {})._evidence
|
||
const reviewing = reviewScript({ run_id: runId, question: T.row.question, attempt: attempts }).then((r) => (locked && r?.answer
|
||
? { ...r, answer: null, original: null, verdict: r.verdict === 'rewritten' ? 'annotated' : r.verdict } : r))
|
||
const review = await Promise.race([reviewing, sleep(ctl.reviewWaitMs ?? 25_000).then(() => null)])
|
||
if (gone(T, epoch)) return
|
||
if (review) {
|
||
emit(T, 'review', { result: review })
|
||
T.usage.review = addTotal(T.usage.review, review.usage)
|
||
if (autoRetry && review.retry_instruction) {
|
||
attempts += 1
|
||
T.usage = retireRun(T.usage)
|
||
T.runId = null
|
||
saveRow(T, { run_id: null }, { runId: null, runStatus: null, usage: T.usage })
|
||
const reason = (review.signals ?? []).map((x) => x.detail).join('；') || review.note
|
||
emit(T, 'turn.retry', { kind: 'review', reason, base: !!graphRun?.nodes?.length, attempt: attempts, usage: T.usage })
|
||
instruction = review.retry_instruction
|
||
base = graphRun?.nodes?.length ? graphRun : null
|
||
continue
|
||
}
|
||
} else {
|
||
// 等不及：先交原答案；那一次复核晚到了照样把用量记进这一轮
|
||
T.late = reviewing.then((r) => {
|
||
if (T.epoch !== epoch) return
|
||
if (r?.usage) {
|
||
const usage = { ...(T.row.meta?.usage ?? {}), review: addTotal(T.row.meta?.usage?.review, r.usage) }
|
||
saveRow(T, {}, { usage })
|
||
emit(T, 'review.late', { usage: r.usage })
|
||
}
|
||
}).finally(() => { if (T.epoch === epoch) T.late = null })
|
||
}
|
||
const output = result.output ?? {}
|
||
const rewritten = review && review.verdict !== 'ok' && !output._evidence ? review.answer : null
|
||
const issuance = output._issuance
|
||
return settle(T, 'done', { status: 'done', answer: rewritten || answerText(output), review: review ?? {} },
|
||
{ ...meta, runStatus: 'succeeded', queries: result.queries,
|
||
...(issuance ? { issuance } : {}), ...(output._evidence ? { evidence: true } : {}) }, output, { ...run, status: 'succeeded' })
|
||
}
|
||
}
|
||
|
||
/**
|
||
* 等这次运行到终态（同服务端 _follow：停在审批上照旧等人，没有待审批的断点收成中断）。这一段检查里发起的运行按它的剧本：
|
||
* 剧本有结局的等实时流放完再收尾，没结局的（在跑、等审批、出字中）一直等到被停止；库里早就有的运行（老轮次接着收尾）按
|
||
* GET 运行的状态
|
||
*/
|
||
async function followRun(T, epoch, runId) {
|
||
const after = ctl.continued[runId] ? 4 : 0
|
||
const sc = script(runId, after)
|
||
const rec = runRecord(runId)
|
||
if (rec.status === 404) return { status: 'gone' }
|
||
const pending = () => ctl.approvals.some((a) => a.run_id === runId && a.status === 'pending')
|
||
const terminal = (status) => {
|
||
const fin = sc.events.filter((e) => e.type === 'run.finished').at(-1)?.data?.output
|
||
const failed = sc.events.filter((e) => e.type === 'run.failed').at(-1)?.data
|
||
const queries = sc.events.filter((e) => e.type === 'tool.start' && String(e.data?.tool ?? '').startsWith('db_query')).length
|
||
const db = rec.status === 200 && rec.json.status === 'succeeded' && Object.keys(rec.json.output ?? {}).length ? rec.json.output : null
|
||
if (status === 'succeeded') {
|
||
// 成果以库里那份为准（事件里那份为了推送体积被切过）
|
||
return { status, queries, output: db ?? (runId === 'run-noget' ? { answer: FULL, note: '附注' } : fin ?? {}) }
|
||
}
|
||
if (status === 'failed') return { status, error: failed?.error ?? rec.json?.error ?? '', detail: failed?.detail }
|
||
if (status === 'cancelled' && rec.json?.end_reason === 'rejected') {
|
||
return { status: 'rejected', endReason: 'rejected', rejected: { sub: rec.json?.error || '被驳回' } }
|
||
}
|
||
return { status }
|
||
}
|
||
/** 停在审批上：等人（stage waiting），审批过了接着等 */
|
||
const waitApproval = async () => {
|
||
if (T.job.stage !== 'waiting') stage(T, 'waiting')
|
||
while (!gone(T, epoch) && pending()) await sleep(80)
|
||
if (!gone(T, epoch)) stage(T, 'run')
|
||
}
|
||
const deadline = Date.now() + 120_000
|
||
if (srv.launched.has(runId) || ctl.continued[runId]) {
|
||
const ended = sc.end === 'close' ? 'interrupted' : sc.end
|
||
if (ended) {
|
||
await until(() => ctl.wsDone[`${runId}@${after}`] || gone(T, epoch), 3000)
|
||
if (ended !== 'interrupted') return terminal(ended)
|
||
if (!pending()) return { status: 'suspended' }
|
||
}
|
||
// 剧本里不结束的：一直在跑，停在审批上的等人
|
||
while (!gone(T, epoch) && Date.now() < deadline) {
|
||
if (pending()) await waitApproval()
|
||
await sleep(80)
|
||
}
|
||
return { status: 'running' }
|
||
}
|
||
// 库里早就有的运行
|
||
while (!gone(T, epoch) && Date.now() < deadline) {
|
||
const status = runRecord(runId).json?.status
|
||
if (['succeeded', 'failed', 'cancelled'].includes(status)) return terminal(status)
|
||
if (status === 'interrupted') {
|
||
if (!pending()) return { status: 'suspended' }
|
||
await waitApproval()
|
||
}
|
||
await sleep(80)
|
||
}
|
||
return { status: 'running' }
|
||
}
|
||
|
||
/** 停止（服务端 chat/job.cancel）：取消生成、取消运行，那一行收成已取消 */
|
||
function cancelTurn(T) {
|
||
if (!ACTIVE_JOB.has(T.job.status)) return false
|
||
T.cancelled = true
|
||
if (T.runId) log.cancels.push(T.runId)
|
||
saveRow(T, { status: 'error', error: '已取消', review: {} }, { outcome: 'cancelled', runStatus: T.runId ? 'cancelled' : undefined })
|
||
emit(T, 'end', { outcome: 'cancelled', turn: rowOut(T), output: null,
|
||
run: T.runId ? { id: T.runId, status: 'cancelled', run_class: 'exploratory', usage: {} } : null })
|
||
T.job = { ...T.job, status: 'cancelled', stage: null }
|
||
return true
|
||
}
|
||
|
||
/** 实时流（长轮询）：没有新事件就等一会儿；这一轮不在进行中、也都给完了，补一条 stream.end */
|
||
async function streamTurn(route, T, after) {
|
||
log.streams.push({ turn: T.id, after })
|
||
if (ctl.dropStreams > 0) {
|
||
// 断线：这一次连接什么都没给就断了（网络抖一下）。页面按收到的最后一个 seq 退避重连
|
||
ctl.dropStreams -= 1
|
||
return route.abort('connectionreset')
|
||
}
|
||
await until(() => T.events.length > after || !turnActive(T) || T.gen !== srv.gen, 6000)
|
||
const fresh = T.events.filter((e) => e.seq > after)
|
||
let body = fresh.map((e) => `data: ${JSON.stringify(e)}\n\n`).join('')
|
||
if (turnActive(T) && T.beat && T.beat.at === T.events.length) {
|
||
body += `data: ${JSON.stringify({ type: 'op', data: T.beat.op, ts: Date.now() / 1000 })}\n\n`
|
||
}
|
||
if (!turnActive(T) && after + fresh.length >= T.events.length) {
|
||
body += `data: ${JSON.stringify({ type: 'stream.end', status: T.job.status, data: { status: T.job.status } })}\n\n`
|
||
}
|
||
return route.fulfill({ status: 200, contentType: 'text/event-stream', body: body || ': ping\n\n' })
|
||
}
|
||
|
||
// ---------------------------------------------------------------- 浏览器
|
||
|
||
const browser = await chromium.launch({ executablePath: CHROME })
|
||
|
||
async function open(theme = 'light', { width = 1440, height = 900, reducedMotion = 'no-preference', settings = null } = {}) {
|
||
const ctx = await browser.newContext({ viewport: { width, height }, colorScheme: theme, reducedMotion })
|
||
opened.add(ctx)
|
||
await ctx.addInitScript((t) => localStorage.setItem('agentlab.theme', t), theme)
|
||
const page = await ctx.newPage()
|
||
const errors = []
|
||
page.on('pageerror', (e) => errors.push(e.message))
|
||
watchLoad(page)
|
||
await page.route('**/api/settings', async (route) => {
|
||
if (route.request().method() !== 'GET') return route.abort()
|
||
// 主题以设置为准（沙箱里存的是 light），这里按要看的那套改掉
|
||
return route.fulfill({ json: { ui: { theme }, run: {}, limits: { max_agent_steps: 25 }, copilot: {}, embedding: {},
|
||
...(settings ?? {}) } })
|
||
})
|
||
await page.route(/\/api\/(conversations|chat\/turns|copilot\/(generate-stream|review)|runs|approvals|datasources)(\/|\?|$)/, fakeApi)
|
||
await page.routeWebSocket(/\/api\/runs\/[^/]+\/stream/, fakeStream)
|
||
return { page, ctx, errors }
|
||
}
|
||
|
||
/** 记下页面加载时出的岔子（页面报错、模块没取到），等不到 store 时拿来说清卡在哪 */
|
||
const loadTrouble = new WeakMap()
|
||
function watchLoad(page) {
|
||
const list = []
|
||
loadTrouble.set(page, list)
|
||
const isModule = (url) => /^\/(src|node_modules|@vite|@id|@fs)\//.test(new URL(url).pathname)
|
||
page.on('pageerror', (e) => list.push(`页面报错「${e.message.split('\n')[0]}」`))
|
||
page.on('requestfailed', (r) => { if (isModule(r.url())) list.push(`模块没取到 ${new URL(r.url()).pathname}`) })
|
||
page.on('response', (r) => { if (r.status() >= 400 && isModule(r.url())) list.push(`模块 ${r.status()} ${new URL(r.url()).pathname}`) })
|
||
}
|
||
/**
|
||
* 等页面把 store 挂到 window 上。main → App → store 全是静态 import，load 之前就求值完了
|
||
* （实测单跑、四个检查并行跑，load 那一刻都已挂出），所以等不到不是「慢」，是模块图断了
|
||
* 或页面卡死。只有「模块没取到」（开发服务器那一下没答上）重载一次，并打一行出来；
|
||
* 页面报错、vite 报错层、主线程卡死是真问题，说清卡在哪，不只留一句 Timeout
|
||
*/
|
||
async function ready(page, fn, what, { arg = null, timeout = 15000, retry = true } = {}) {
|
||
if (await page.waitForFunction(fn, arg, { timeout }).then(() => true, () => false)) return
|
||
const trouble = loadTrouble.get(page) ?? []
|
||
if (retry && trouble.length && trouble.every((t) => t.startsWith('模块'))) {
|
||
console.log(` · ${what} 没挂出来（${trouble.slice(0, 2).join('；')}），重载一次`)
|
||
trouble.length = 0
|
||
await page.reload()
|
||
return ready(page, fn, what, { arg, timeout, retry: false })
|
||
}
|
||
const probe = page.evaluate(() => {
|
||
const overlay = document.querySelector('vite-error-overlay')
|
||
if (overlay) return `vite 报错层「${(overlay.shadowRoot?.querySelector('.message')?.textContent ?? '').trim().slice(0, 120)}」`
|
||
return `页面停在 ${location.pathname}，#root 下 ${document.getElementById('root')?.childElementCount ?? 0} 个元素`
|
||
}).catch((e) => `读不到页面：${String(e.message).split('\n')[0]}`)
|
||
const hung = new Promise((r) => setTimeout(() => r('页面主线程 3 秒没有响应'), 3000))
|
||
const where = await Promise.race([probe, hung])
|
||
throw new Error(`${what} ${timeout / 1000} 秒没挂出来：${[where, ...trouble.slice(0, 3)].join('；')}`)
|
||
}
|
||
const goto = async (page, id) => {
|
||
await page.goto(`${WEB}/chat/${id}`)
|
||
await ready(page, () => !!window.__chat, '问数据页的 store（window.__chat）')
|
||
}
|
||
const shows = (page, text, timeout = 6000) =>
|
||
page.getByText(text).first().waitFor({ timeout }).then(() => true, () => false)
|
||
const chatState = (page) => page.evaluate(() => window.__chat.getState())
|
||
/** 等脚本这一侧的某个条件成立（伪造后端记下的请求、事件流的开关） */
|
||
const until = async (cond, ms = 5000) => {
|
||
const end = Date.now() + ms
|
||
while (!cond() && Date.now() < end) await new Promise((r) => setTimeout(r, 50))
|
||
return cond()
|
||
}
|
||
const send = async (page, q) => {
|
||
const box = page.getByRole('textbox', { name: '向数据提问' })
|
||
await box.fill(q)
|
||
await box.press('Enter')
|
||
}
|
||
|
||
for (const theme of THEMES) {
|
||
console.log(`\n######## ${theme} ########`)
|
||
|
||
await section('hero', '首屏：输入框是主角，例句不拿系统表造句', async () => {
|
||
const { page, ctx, errors } = await open(theme)
|
||
await goto(page, 'c0emptg')
|
||
check('空会话显示首屏', await shows(page, '问你的数据'))
|
||
const hints = await page.locator('button[title="填入输入框，修改后发送"]').allInnerTexts()
|
||
check('有例句', hints.length > 0, hints.join(' | '))
|
||
check('例句里没有系统表',
|
||
hints.every((h) => !/__prisma|_migrations|test_|etl_audit|_log\b/.test(h)), hints.join(' | '))
|
||
check('描述里的业务词造了句', hints.some((h) => /factory：/.test(h)), hints.join(' | '))
|
||
const ph = await page.getByRole('textbox', { name: '向数据提问' }).getAttribute('placeholder')
|
||
check('占位符简短', !!ph && ph.length <= 12, ph ?? '')
|
||
const box = await page.getByRole('textbox', { name: '向数据提问' }).boundingBox()
|
||
check('输入框在首屏中部，不贴底', !!box && box.y > 200 && box.y + box.height < 700, JSON.stringify(box))
|
||
check('数据源入口指向 /data', await page.locator('a[href="/data"]').count() > 0)
|
||
const postsBefore = log.posts.length
|
||
await page.locator('button[title="填入输入框，修改后发送"]').first().click()
|
||
await page.waitForTimeout(200)
|
||
check('点例句只是填进输入框', (await page.getByRole('textbox', { name: '向数据提问' }).inputValue()).length > 0
|
||
&& log.posts.length === postsBefore)
|
||
await page.screenshot({ path: `${SHOTS}/hero-${theme}.png` })
|
||
check('没有运行时报错', errors.length === 0, errors[0] ?? '')
|
||
await ctx.close()
|
||
|
||
// 知识库、工具没取回来（system-2 终验）：能力条不能把「没取回来」说成「0 个」
|
||
const off = await open(theme)
|
||
await off.page.route(/\/api\/(tools|kb\/collections)(\?|$)/, (route) => route.fulfill({ status: 502, body: 'bad gateway' }))
|
||
await goto(off.page, 'c0emptg')
|
||
await shows(off.page, '问你的数据')
|
||
// 导航里也有这两个链接（「知识」「工具」），只看首屏能力条上的
|
||
const strip = off.page.locator('a[href="/knowledge"], a[href="/tools"]').filter({ hasText: /知识库|工具\s*(\d|—|加载失败)/ })
|
||
// 请求还在路上时写「—」，等它落定
|
||
await off.page.waitForFunction(() => [...document.querySelectorAll('a[href="/knowledge"], a[href="/tools"]')]
|
||
.every((a) => !a.textContent.includes('—')), null, { timeout: 8000 }).catch(() => {})
|
||
const said = (await strip.allInnerTexts()).join(' | ')
|
||
check('知识库、工具没取回来：不写「0 个」', !/知识库 0 个|工具 0 个/.test(said), said)
|
||
check('知识库、工具没取回来：说加载失败', /知识库加载失败/.test(said) && /工具加载失败/.test(said), said)
|
||
await off.page.screenshot({ path: `${SHOTS}/hero-catalog-failed-${theme}.png` })
|
||
await off.ctx.close()
|
||
})
|
||
|
||
await section('states', '三种状态：加载中 / 加载失败 / 空', async () => {
|
||
const { page, ctx, errors } = await open(theme)
|
||
ctl.slowMs = 2500
|
||
await page.goto(`${WEB}/chat/c0slowd`)
|
||
await ready(page, () => !!window.__chat, '问数据页的 store（window.__chat）')
|
||
await page.waitForTimeout(600)
|
||
check('取历史时画骨架', await page.locator('[aria-busy="true"]').count() > 0)
|
||
check('取历史时不显示首屏', !(await page.getByText('问你的数据').count()))
|
||
const blocked = await page.getByRole('button', { name: '发送' }).isDisabled()
|
||
check('取历史时发不出去', blocked)
|
||
await page.screenshot({ path: `${SHOTS}/loading-${theme}.png` })
|
||
check('历史回来了', await shows(page, 'D 的答案二', 5000))
|
||
ctl.slowMs = 0
|
||
|
||
ctl.failE = true
|
||
const before = log.posts.length
|
||
await goto(page, 'c0faile')
|
||
check('取失败时给出错误和重试', await shows(page, '数据库暂时不可用'))
|
||
check('取失败时不显示首屏', !(await page.getByText('问你的数据').count()))
|
||
await send(page, '失败时问一句')
|
||
await page.waitForTimeout(400)
|
||
check('取失败时不能在空态上提问', log.posts.length === before, `${log.posts.length - before} 次开轮`)
|
||
await page.screenshot({ path: `${SHOTS}/load-failed-${theme}.png` })
|
||
ctl.failE = false
|
||
await page.getByRole('button', { name: '重试' }).first().click()
|
||
check('重试后历史回来', await shows(page, 'E 的答案'))
|
||
check('草稿还在', (await page.getByRole('textbox', { name: '向数据提问' }).inputValue()) === '失败时问一句')
|
||
ctl.failE = true
|
||
|
||
// 取失败（500、断网）留在原地给重试；只有「这个对话不存在」（404）才把人带走
|
||
await page.goto(`${WEB}/chat/c0gone`)
|
||
check('指向不存在的对话：说清楚为什么换了地方', await shows(page, '该对话不存在'))
|
||
check('落到一个真实的对话', await page.waitForURL(/\/chat\/c0[a-z]+$/, { timeout: 5000 }).then(() => !page.url().endsWith('c0gone'), () => false), page.url())
|
||
check('没有运行时报错', errors.length === 0, errors[0] ?? '')
|
||
await ctx.close()
|
||
})
|
||
|
||
await section('layout', '布局：长会话里输入框和发送键完整可见（1440×900，另看 1024×700）', async () => {
|
||
for (const [width, height] of [[1440, 900], [1024, 700]]) {
|
||
const { page, ctx } = await open(theme, { width, height })
|
||
await goto(page, 'c0longc')
|
||
await shows(page, '第 6 个问题')
|
||
await page.waitForTimeout(500)
|
||
const m = await page.evaluate(() => {
|
||
const ta = document.querySelector('textarea')?.getBoundingClientRect()
|
||
const btn = document.querySelector('button[aria-label="发送"]')?.getBoundingClientRect()
|
||
return { ta: ta && [ta.top, ta.bottom], btn: btn && [btn.top, btn.bottom], h: innerHeight,
|
||
scroll: document.documentElement.scrollHeight, wide: document.documentElement.scrollWidth, w: innerWidth }
|
||
})
|
||
const at = `${width}×${height}`
|
||
check(`${at} 输入框底边在视口内`, !!m.ta && m.ta[1] <= m.h, JSON.stringify(m))
|
||
check(`${at} 发送键完整可见`, !!m.btn && m.btn[0] >= 0 && m.btn[1] <= m.h, JSON.stringify(m.btn))
|
||
check(`${at} 整页不多滚`, m.scroll <= m.h && m.wide <= m.w, `${m.wide}×${m.scroll} / ${m.w}×${m.h}`)
|
||
await page.screenshot({ path: `${SHOTS}/long-${width}-${theme}.png` })
|
||
await ctx.close()
|
||
}
|
||
|
||
// 窄屏（390×844）：240px 的会话列表以前照常占宽，正文一行只剩一两个字、整页横向溢出
|
||
const { page, ctx } = await open(theme, { width: 390, height: 844 })
|
||
await goto(page, 'c0longc')
|
||
await shows(page, '第 6 个问题')
|
||
await page.waitForTimeout(400)
|
||
const narrow = await page.evaluate(() => {
|
||
const root = document.querySelector('button[aria-label="展开对话列表"]')?.closest('.relative')
|
||
const body = document.querySelector('[data-stream-scroll]')?.getBoundingClientRect()
|
||
const ta = document.querySelector('textarea')?.getBoundingClientRect()
|
||
return { list: !!document.querySelector('aside[aria-label="对话列表"]'), toggle: !!root,
|
||
over: root ? root.scrollWidth - root.clientWidth : null, body: body && Math.round(body.width),
|
||
ta: ta && Math.round(ta.right), w: innerWidth }
|
||
})
|
||
check('390 宽：会话列表默认收起', !narrow.list && narrow.toggle, JSON.stringify(narrow))
|
||
check('390 宽：问数据这一栏不横向溢出', narrow.over === 0, JSON.stringify(narrow))
|
||
check('390 宽：正文宽度够读（≥ 260px）', (narrow.body ?? 0) >= 260, JSON.stringify(narrow))
|
||
check('390 宽：输入框右边在视口内', !!narrow.ta && narrow.ta <= narrow.w, JSON.stringify(narrow))
|
||
await page.screenshot({ path: `${SHOTS}/narrow-390-${theme}.png` })
|
||
await page.getByRole('button', { name: '展开对话列表' }).click()
|
||
const drawer = page.locator('[data-conversation-drawer] aside[aria-label="对话列表"]')
|
||
check('390 宽：展开是浮在正文上面的抽屉', await drawer.waitFor({ timeout: 3000 }).then(() => true, () => false))
|
||
const bodyAfter = await page.evaluate(() => Math.round(document.querySelector('[data-stream-scroll]')?.getBoundingClientRect().width ?? 0))
|
||
check('390 宽：抽屉打开时正文不被挤窄', bodyAfter === narrow.body, `${narrow.body} → ${bodyAfter}`)
|
||
await page.screenshot({ path: `${SHOTS}/narrow-390-drawer-${theme}.png` })
|
||
await drawer.locator('div.group', { hasText: '会话 B' }).locator('button').first().click()
|
||
await page.waitForURL(/\/chat\/c0idleb$/, { timeout: 5000 }).catch(() => {})
|
||
await page.waitForTimeout(300)
|
||
check('390 宽：挑了一个对话，抽屉收回去', page.url().endsWith('/chat/c0idleb')
|
||
&& await page.locator('aside[aria-label="对话列表"]').count() === 0, page.url())
|
||
await ctx.close()
|
||
})
|
||
|
||
await section('busy', '忙不忙按会话算', async () => {
|
||
const { page, ctx, errors } = await open(theme)
|
||
const cancelsBefore = log.cancels.length
|
||
await goto(page, 'c0busya')
|
||
await send(page, '一个会跑很久的问题')
|
||
check('A 在跑：按钮是停止', await page.getByRole('button', { name: '停止这一轮' })
|
||
.waitFor({ timeout: 8000 }).then(() => true, () => false))
|
||
// 停止键在建图时就出现了；要等运行真的起来、事件流接上，才谈得上"别的会话别碰它"
|
||
await page.waitForFunction(() => !!window.__chat.getState().byConversation.c0busya?.at(-1)?.run?.id,
|
||
null, { timeout: 8000 }).catch(() => {})
|
||
const runA = (await chatState(page)).byConversation.c0busya.at(-1).run?.id
|
||
await until(() => ctl.wsClosed[runA] === false)
|
||
await page.getByRole('button', { name: /^会话 B/ }).click()
|
||
await shows(page, 'B 的答案')
|
||
check('切到 B：按钮是发送，不是停止',
|
||
await page.getByRole('button', { name: '发送' }).count() === 1
|
||
&& await page.getByRole('button', { name: '停止这一轮' }).count() === 0)
|
||
check('B 里说清 A 还在跑', await shows(page, '正在运行，这里可以照常提问'))
|
||
const rowA = page.locator('aside[aria-label="对话列表"] div.group', { hasText: '会话 A' })
|
||
check('列表里 A 标着运行中', (await rowA.innerText()).includes('运行中'))
|
||
await rowA.hover()
|
||
check('A 正在跑，不能删', await rowA.getByRole('button', { name: /删除/ }).isDisabled())
|
||
const rowB = page.locator('aside[aria-label="对话列表"] div.group', { hasText: '会话 B' })
|
||
await rowB.hover()
|
||
check('B 空闲，可以删', !(await rowB.getByRole('button', { name: /删除/ }).isDisabled()))
|
||
await page.screenshot({ path: `${SHOTS}/busy-elsewhere-${theme}.png` })
|
||
check('在 B 里什么都没停：A 的事件流还开着', ctl.wsClosed[runA] === false, runA)
|
||
check('A 的运行没被取消', log.cancels.length === cancelsBefore)
|
||
|
||
await page.locator('[role=status]', { hasText: '正在运行，这里可以照常提问' }).getByRole('link', { name: '查看', exact: true }).click()
|
||
await page.getByRole('button', { name: '停止这一轮' }).click()
|
||
await until(() => log.cancels.length > cancelsBefore && ctl.wsClosed[runA] === true)
|
||
check('停止只取消 A 自己的运行', log.cancels.slice(cancelsBefore).join() === runA, log.cancels.slice(cancelsBefore).join())
|
||
check('A 的事件流关了', ctl.wsClosed[runA] === true)
|
||
const a = (await chatState(page)).byConversation.c0busya.at(-1)
|
||
check('A 这一轮收成已取消', a.phase === 'cancelled', a.phase)
|
||
check('已取消的轮次没有「继续运行」', !(await page.getByRole('button', { name: '继续运行' }).count()))
|
||
check('已取消的轮次给「重新运行本轮」', await page.getByRole('button', { name: '重新运行本轮' }).count() === 1)
|
||
const b = (await chatState(page)).byConversation.c0idleb
|
||
check('B 的轮次没被动过', b.length === 1 && b[0].phase === 'done')
|
||
await page.waitForTimeout(300) // 「查看」刚切过来，卡片的入场动效还没播完
|
||
await page.screenshot({ path: `${SHOTS}/cancelled-${theme}.png` })
|
||
check('没有运行时报错', errors.length === 0, errors[0] ?? '')
|
||
await ctx.close()
|
||
})
|
||
|
||
await section('truncated', '长答案：事件里截断了，用 GET 运行的完整版', async () => {
|
||
const { page, ctx, errors } = await open(theme)
|
||
const patchesBefore = log.patches.length
|
||
await goto(page, 'c0trunc')
|
||
await send(page, '给我一份长答案')
|
||
check('答完了', await page.waitForFunction(
|
||
() => window.__chat.getState().byConversation.c0trunc?.at(-1)?.phase === 'done', null, { timeout: 10000 })
|
||
.then(() => true, () => false))
|
||
const t = (await chatState(page)).byConversation.c0trunc.at(-1)
|
||
check('显示的是完整答案', String(t.output?.answer ?? '').length === FULL.length, `${String(t.output?.answer ?? '').length} / ${FULL.length}`)
|
||
check('llm.token 没进 events', !t.events.some((e) => e.type === 'llm.token'), `${t.events.length} 条`)
|
||
// 一轮在服务端跑（docs/design/orchestration-round §1）：那一行由服务端写，页面不再 PATCH 这一轮的结局
|
||
check('页面不写这一轮的结局（不 PATCH 轮次）', log.patches.slice(patchesBefore).length === 0,
|
||
JSON.stringify(log.patches.slice(patchesBefore).map((p) => Object.keys(p.body))))
|
||
const meta = t.meta
|
||
check('服务端写的那一行带着可信度信息（运行、类别），页面照它摆', meta?.runId === 'run-long' && meta?.runClass === 'exploratory',
|
||
JSON.stringify(meta))
|
||
// 后端有了单独的 meta 列：review 只放复核结论，不再夹带
|
||
check('meta 单独存，不夹在 review 里', !t.reviewRaw || !('meta' in t.reviewRaw), JSON.stringify(Object.keys(t.reviewRaw ?? {})))
|
||
check('完整答案取到了，meta 里没有「没取全」', !meta?.clipped, JSON.stringify(meta?.clipped))
|
||
await page.getByText('展开全部').first().click().catch(() => {})
|
||
check('完整结尾在页面上', await shows(page, '【完整结尾】'))
|
||
// 建图时助手查看了哪些表：收在「规划」里，点开是一行，再点开看表名
|
||
check('建图的 look 事件原样记进这一轮', t.ops.some((o) => o.op === 'look'))
|
||
check('查看记录随那一行存进 meta.looks（look §2.5.8）', Array.isArray(meta?.looks) && meta.looks.length === 1
|
||
&& meta.looks[0].objects?.[0]?.source === 'shop', JSON.stringify(meta?.looks))
|
||
const turnCard = page.locator('[data-turn]').last()
|
||
await turnCard.locator('button', { hasText: '规划' }).first().click()
|
||
const ctxRow = turnCard.locator('[data-step-code="copilot_look"]')
|
||
check('点开「规划」看到查看了几张表', await ctxRow.waitFor({ timeout: 3000 }).then(() => true, () => false)
|
||
&& (await ctxRow.innerText()).includes('查看了 2 张表'))
|
||
await ctxRow.locator('button').first().click()
|
||
check('再点开看到每个库一行的表名', await ctxRow.getByText('「shop」：按需求挑出 orders、order_items')
|
||
.waitFor({ timeout: 3000 }).then(() => true, () => false))
|
||
// 真实运行里数据源一张目录都没有，挑表和 SQL 检查都用不上目录：这一行下面提示起草，并给去目录页的入口
|
||
const noCatalog = ctxRow.locator('[data-step-next]')
|
||
check('没有数据目录的大库：提示先起草目录', await noCatalog.count() === 1
|
||
&& (await noCatalog.innerText()).includes('这个数据源还没有数据目录，起草后挑表和 SQL 检查会更准'),
|
||
await noCatalog.count() ? await noCatalog.innerText() : '没有这一句')
|
||
check('……入口直达这个数据源的目录页', await ctxRow.locator('a[data-fix="catalog"]').getAttribute('href').catch(() => null)
|
||
=== '/data/catalog/src-shop')
|
||
check('没有运行时报错', errors.length === 0, errors[0] ?? '')
|
||
await ctx.close()
|
||
})
|
||
|
||
await section('evidence', '成果带逐段证据：复核只加说明，不改写答案（可点击证据第一期）', async () => {
|
||
const { page, ctx, errors } = await open(theme)
|
||
// 文档记着的运行要和成果的运行对得上（run-evid），不然前端按「别的运行写的」退回普通文本
|
||
await page.route(/\/api\/artifacts\//, (route) => route.fulfill({ json: { id: EVIDENCE.doc_artifact,
|
||
content: { ...EVIDENCE.doc, run_id: 'run-evid' } } }))
|
||
await page.route(/\/api\/runs\/run-evid\/evidence(\/.*)?(\?.*)?$/, (route) => route.fulfill({ json: EVIDENCE.graph }))
|
||
const patchesBefore = log.patches.length
|
||
await goto(page, 'c0trunc')
|
||
await send(page, '写一份带证据的周报')
|
||
check('答完了', await page.waitForFunction(
|
||
() => window.__chat.getState().byConversation.c0trunc?.at(-1)?.phase === 'done', null, { timeout: 10000 })
|
||
.then(() => true, () => false))
|
||
const t = (await chatState(page)).byConversation.c0trunc.at(-1)
|
||
check('复核的改写没有落到答案上：成果还是报告原文和它的证据标注',
|
||
t.output?.['周报'] === EVIDENCE.output['周报'] && !!t.output?._evidence && !t.output?.answer, JSON.stringify(Object.keys(t.output ?? {})))
|
||
check('复核的说明照样摆出来，改写的文字丢掉', t.review?.note?.includes('检索降级过') && t.review.answer === null
|
||
&& t.review.verdict === 'annotated' && !t.rawOutput, JSON.stringify(t.review))
|
||
await page.waitForSelector('[data-evidence-doc]', { timeout: 5000 }).catch(() => {})
|
||
check('页面上是逐段可点的报告，不是改写后的文字', await page.locator('[data-evidence-doc] [data-seg]').count() > 0
|
||
&& !(await shows(page, '改写后的答案')))
|
||
check('没有「复核改写过此答案」的对照', await page.getByText('复核改写过此答案').count() === 0)
|
||
check('页面不写这一轮的结局（不 PATCH 轮次）', log.patches.slice(patchesBefore).length === 0)
|
||
check('meta 记下这一轮有证据（刷新后回运行那里取标注）', t.meta?.evidence === true, JSON.stringify(t.meta))
|
||
check('没有运行时报错', errors.length === 0, errors[0] ?? '')
|
||
await ctx.close()
|
||
})
|
||
|
||
await section('askdata', '问数据的答案带逐段证据：实时、刷新后补回、复核守卫、窄屏抽屉（可点击证据第二期）', async () => {
|
||
const routeEvidence = async (page) => {
|
||
// 文档记着的运行要和成果的运行对得上（run-evq），片段按运行认
|
||
await page.route(/\/api\/artifacts\//, (route) => route.fulfill({ json: { id: EVQ.doc_artifact, content: EVQ_DOC } }))
|
||
await page.route(/\/api\/runs\/run-evq\/evidence(\/.*)?(\?.*)?$/, (route) => {
|
||
const m = new URL(route.request().url()).pathname.match(/\/segments\/([^/]+)$/)
|
||
if (!m) return route.fulfill({ json: EVQ.graph })
|
||
const body = EVQ.segments[decodeURIComponent(m[1])]
|
||
return body ? route.fulfill({ json: body }) : route.fulfill({ status: 404, json: { detail: '没有这个片段', code: 'evidence_segment_not_found' } })
|
||
})
|
||
}
|
||
{
|
||
const { page, ctx, errors } = await open(theme)
|
||
await routeEvidence(page)
|
||
const patchesBefore = log.patches.length
|
||
await goto(page, 'c0trunc')
|
||
await send(page, '问数据带证据：各区域销售额，最大的一单是多少')
|
||
check('答完了', await page.waitForFunction(
|
||
() => window.__chat.getState().byConversation.c0trunc?.at(-1)?.phase === 'done', null, { timeout: 10000 })
|
||
.then(() => true, () => false))
|
||
const t = (await chatState(page)).byConversation.c0trunc.at(-1)
|
||
check('复核的改写没落到答案上：成果还是报告原文和 _evidence', t.output?.answer === EVQ.output.answer && !!t.output?._evidence
|
||
&& t.review?.answer === null, JSON.stringify({ keys: Object.keys(t.output ?? {}), review: t.review?.verdict }))
|
||
await page.waitForSelector('[data-evidence-doc] [data-seg]', { timeout: 6000 }).catch(() => {})
|
||
const turn = page.locator('[data-turn]').last()
|
||
check('答案是逐段可点的报告', await turn.locator('[data-evidence-doc] [data-seg]').count() > 10)
|
||
check('没有出具契约：文档自己给计数条', (await turn.locator('[data-evidence-tally]').innerText().catch(() => '')).includes('数字有出处'))
|
||
const text = await turn.innerText()
|
||
check('执行过程里有「按出处抽取字段」和字段对不上的提醒', text.includes('按出处抽取字段') && text.includes('有 1 个字段与查询快照不一致'),
|
||
text.replace(/\s+/g, ' ').slice(0, 200))
|
||
check('整表的每格都是可点的片段', await turn.locator('[data-evidence-doc] table td [data-seg]').count() === 9)
|
||
await turn.locator(`[data-seg="${EVQ_SEG('1,288')}"]`).click()
|
||
await page.waitForSelector('[data-evidence-panel] [data-ev-query]', { timeout: 5000 }).catch(() => {})
|
||
const panel = page.locator('[data-evidence-panel]')
|
||
check('宽屏：点开数字从侧边弹出，看得到查询步骤', await panel.getAttribute('data-evidence-panel') === 'side'
|
||
&& await panel.locator('[data-ev-query]').count() === 1)
|
||
check('……被引用的格高亮，SQL 在，遮罩说明在', (await panel.locator('td[data-highlight="cell"]').innerText().catch(() => '')) === '1288'
|
||
&& (await panel.locator('[data-ev-sql]').innerText().catch(() => '')).includes('FROM orders')
|
||
&& (await panel.locator('[data-ev-mask-note]').innerText().catch(() => '')).includes('不是安全边界'))
|
||
await page.waitForTimeout(300)
|
||
await page.screenshot({ path: `${SHOTS}/askdata-evidence-${theme}.png` })
|
||
await page.keyboard.press('Escape')
|
||
check('页面不写这一轮的结局；服务端写的 meta 记下这一轮有证据', log.patches.slice(patchesBefore).length === 0
|
||
&& t.meta?.evidence === true, JSON.stringify(t.meta))
|
||
check('没有运行时报错（实时）', errors.length === 0, errors[0] ?? '')
|
||
await ctx.close()
|
||
}
|
||
resetDb()
|
||
ctl.askConvs = true
|
||
try {
|
||
const { page, ctx, errors } = await open(theme)
|
||
await routeEvidence(page)
|
||
const gets = log.runGets.length
|
||
await goto(page, 'c0askz')
|
||
await page.waitForSelector('[data-evidence-doc] [data-seg]', { timeout: 8000 }).catch(() => {})
|
||
check('刷新后补回：库里只有文字，回运行取回 _evidence，答案照样逐段可点',
|
||
log.runGets.slice(gets).includes('run-evq') && await page.locator('[data-evidence-doc] [data-seg]').count() > 10)
|
||
await page.locator(`[data-evidence-doc] [data-seg="${EVQ_SEG('8.7%')}"]`).click()
|
||
await page.waitForSelector('[data-evidence-panel] [data-ev-sources]', { timeout: 5000 }).catch(() => {})
|
||
const panel = page.locator('[data-evidence-panel]')
|
||
check('……点开指标：口径卡来源、输入来源、两次查询都在', (await panel.locator('[data-ev-caliber-from]').innerText().catch(() => '')).includes('销售周报')
|
||
&& await panel.locator('[data-ev-source]').count() === 2 && await panel.locator('[data-ev-query]').count() === 2)
|
||
check('没有运行时报错（刷新后）', errors.length === 0, errors[0] ?? '')
|
||
await ctx.close()
|
||
|
||
const narrow = await open(theme, { width: 800, height: 900 })
|
||
await routeEvidence(narrow.page)
|
||
await goto(narrow.page, 'c0askz')
|
||
await narrow.page.waitForSelector('[data-evidence-doc] [data-seg]', { timeout: 8000 }).catch(() => {})
|
||
await narrow.page.locator(`[data-evidence-doc] [data-seg="${EVQ_SEG('1,288')}"]`).click()
|
||
await narrow.page.waitForSelector('[data-evidence-panel] [data-ev-query]', { timeout: 5000 }).catch(() => {})
|
||
const drawer = narrow.page.locator('[data-evidence-panel]')
|
||
const sc = await narrow.page.evaluate(() => ({ sw: document.documentElement.scrollWidth, vw: innerWidth }))
|
||
check('窄屏：面板从底部抽出，里面是查询步骤', await drawer.getAttribute('data-evidence-panel') === 'drawer'
|
||
&& await drawer.locator('[data-ev-query]').count() === 1)
|
||
check('……整页没有横向滚动', sc.sw <= sc.vw, JSON.stringify(sc))
|
||
await narrow.page.waitForTimeout(300)
|
||
await narrow.page.screenshot({ path: `${SHOTS}/askdata-drawer-${theme}.png` })
|
||
await narrow.page.keyboard.press('Escape')
|
||
await narrow.page.waitForTimeout(200)
|
||
check('……Esc 收起抽屉', await narrow.page.locator('[data-evidence-panel]').count() === 0)
|
||
check('没有运行时报错（窄屏）', narrow.errors.length === 0, narrow.errors[0] ?? '')
|
||
await narrow.ctx.close()
|
||
} finally {
|
||
ctl.askConvs = false
|
||
}
|
||
})
|
||
|
||
await section('nodeerror', '节点报错说一遍：原因、怎么办不在展开区里再贴一遍原话（整形节点解析 JSON 失败）', async () => {
|
||
for (const [ask, said] of [['整形解析失败：旧说法', JSON_TEMPLATE_ERROR], ['整形解析失败：上游说法', JSON_UPSTREAM_ERROR],
|
||
['整形解析失败：节点名带超时', JSON_SLOW_ERROR], ['整形解析失败：现在的说法', JSON_TEMPLATE_ERROR_NOW],
|
||
['整形解析失败：上游、现在的说法', JSON_UPSTREAM_ERROR_NOW]]) {
|
||
const { page, ctx, errors } = await open(theme)
|
||
await goto(page, 'c0trunc')
|
||
await send(page, ask)
|
||
await page.waitForFunction(() => window.__chat.getState().byConversation.c0trunc?.at(-1)?.phase === 'error', null, { timeout: 10000 })
|
||
.catch(() => {})
|
||
const row = page.locator('[data-turn]').last().locator('[data-node-id="tf"][data-step-status="failed"]').first()
|
||
await row.locator('button[aria-expanded]').first().click()
|
||
await page.waitForTimeout(250)
|
||
const shown = await row.innerText()
|
||
const head = said.split(/[。；，]/)[0]
|
||
const times = shown.split(head).length - 1
|
||
check(`「${ask}」：节点那一行把报错说一遍，展开后不再整段贴一遍原话`, times === 1, `出现 ${times} 次：${shown.replace(/\s+/g, ' ').slice(0, 240)}`)
|
||
const tail = said.split(/[；]|。/).at(-1).slice(0, 12)
|
||
check(`「${ask}」：怎么办那半句照样在`, shown.includes(tail), tail)
|
||
const resume = await page.getByRole('button', { name: /^继续运行/ }).count()
|
||
const canvas = await page.getByRole('button', { name: '在画布里打开' }).count() + await page.locator('[data-fix="canvas"]').count()
|
||
check(`「${ask}」：原样继续运行只会再失败一次，不给「继续运行」，给去画布改的入口`, resume === 0 && canvas > 0,
|
||
`继续运行 ${resume} 个 · 画布入口 ${canvas} 个`)
|
||
await row.locator('summary', { hasText: '技术细节' }).click().catch(() => {})
|
||
check(`「${ask}」：原始异常和出错位置前后的原文收在技术细节里`, (await row.innerText()).includes('JSONDecodeError')
|
||
&& (await row.innerText()).includes('出错位置前后的原文'))
|
||
if (said === JSON_UPSTREAM_ERROR) await page.screenshot({ path: `${SHOTS}/nodeerror-json-${theme}.png` })
|
||
check('没有运行时报错', errors.length === 0, errors[0] ?? '')
|
||
await ctx.close()
|
||
}
|
||
const { page, ctx } = await open(theme)
|
||
await goto(page, 'c0trunc')
|
||
// 节点名是用户起的，里面带着超时、额度、连不上这些别的规则的关键词：照样认作 JSON 解析失败
|
||
// 老说法、后端现在的说法各喂一遍
|
||
const named = await page.evaluate(async (saids) => {
|
||
const { explainRunError } = await import('/src/lib/explain.ts')
|
||
return saids.flatMap((said) => ['查询超时订单', '额度查询', '连不上的库', '服务器 500 统计'].map((t) => {
|
||
const x = explainRunError(said.replace('上游「查数」', `上游「${t}」`))
|
||
return { t, title: x.title, continuable: x.continuable, fix: x.fix }
|
||
}))
|
||
}, [JSON_UPSTREAM_ERROR, JSON_UPSTREAM_ERROR_NOW])
|
||
check('节点名里带超时 / 额度 / 连不上 / 500：照样认作「上游输出的不是合法 JSON」，不给继续运行',
|
||
named.length === 8 && named.every((x) => x.title.startsWith(`上游「${x.t}」输出的不是合法 JSON`) && x.continuable === false && x.fix === 'canvas'),
|
||
JSON.stringify(named.filter((x) => x.continuable !== false)))
|
||
await send(page, '整形解析失败：改写过的说法')
|
||
await page.waitForFunction(() => window.__chat.getState().byConversation.c0trunc?.at(-1)?.phase === 'error', null, { timeout: 10000 })
|
||
.catch(() => {})
|
||
const row = page.locator('[data-turn]').last().locator('[data-node-id="tf"][data-step-status="failed"]').first()
|
||
await row.locator('button[aria-expanded]').first().click()
|
||
await page.waitForTimeout(250)
|
||
const t = await row.innerText()
|
||
check('对照：说法改写过（鉴权失败）时，原话还在展开区里', t.includes('模型鉴权失败') && t.includes('invalid x-api-key'),
|
||
t.replace(/\s+/g, ' ').slice(0, 200))
|
||
await ctx.close()
|
||
})
|
||
|
||
await section('restart', '服务重启：interrupted 且没有待审批，收尾而不是转圈', async () => {
|
||
const { page, ctx, errors } = await open(theme)
|
||
await goto(page, 'c0resti')
|
||
await send(page, '跑到一半服务重启')
|
||
check('说清是服务重启打断的', await shows(page, '服务重启，本轮已中断', 8000))
|
||
await page.waitForTimeout(300)
|
||
const t = (await chatState(page)).byConversation.c0resti.at(-1)
|
||
check('这一轮收成挂起', t.phase === 'suspended', t.phase)
|
||
// 只数这一轮自己和它在左栏的那一行：左栏里夹具的会话 L 本来就标着运行中，它转是对的
|
||
const spinning = await page.locator('[data-turn] .animate-spin').count()
|
||
+ await page.locator('aside[aria-label="对话列表"] div.group', { hasText: '会话 I' }).locator('.animate-spin').count()
|
||
check('不再转圈', spinning === 0, `${spinning} 个转圈`)
|
||
check('给「继续运行」', await page.getByRole('button', { name: '继续运行' }).count() === 1)
|
||
check('给「重新运行本轮」', await page.getByRole('button', { name: '重新运行本轮' }).count() === 1)
|
||
check('输入框可以接着问', await page.getByRole('button', { name: '停止这一轮' }).count() === 0)
|
||
await page.screenshot({ path: `${SHOTS}/restart-${theme}.png` })
|
||
|
||
await page.getByRole('button', { name: '继续运行' }).click()
|
||
check('继续运行：从断点续上并跑完', await page.waitForFunction(
|
||
() => window.__chat.getState().byConversation.c0resti?.at(-1)?.phase === 'done', null, { timeout: 8000 })
|
||
.then(() => true, () => false), (await chatState(page)).byConversation.c0resti.at(-1).phase)
|
||
check('续跑发给的是这一轮自己的运行', log.continues.at(-1) === 'run-restart', log.continues.join(','))
|
||
const cont = log.turnStarts.filter((b) => b.mode === 'continue').at(-1)
|
||
check('继续运行的请求：带着这一轮（retry_of、mode=continue），不带图', cont?.conversation_id === 'c0resti'
|
||
&& cont?.retry_of === t.serverId && !cont?.graph, JSON.stringify(cont))
|
||
check('续上之后答案出来了', await shows(page, '接着跑完了：三条记录。'))
|
||
check('续上之后补救按钮收起', await page.getByRole('button', { name: '继续运行' }).count() === 0)
|
||
await page.waitForTimeout(300)
|
||
await page.screenshot({ path: `${SHOTS}/restart-continued-${theme}.png` })
|
||
|
||
// 强杀：连结束标记都没有，连接直接断。重连上重启后的后端，照样收成中断
|
||
await goto(page, 'c0killn')
|
||
await send(page, '一个跑着跑着服务被强杀的问题')
|
||
check('连接直接断了：重连之后收成中断', await page.waitForFunction(
|
||
() => window.__chat.getState().byConversation.c0killn?.at(-1)?.phase === 'suspended', null, { timeout: 8000 })
|
||
.then(() => true, () => false), (await chatState(page)).byConversation.c0killn?.at(-1)?.phase)
|
||
check('强杀之后同样给「继续运行」', await page.getByRole('button', { name: '继续运行' }).count() === 1)
|
||
check('没有运行时报错', errors.length === 0, errors[0] ?? '')
|
||
await ctx.close()
|
||
})
|
||
|
||
await section('resume', '隔了几天再接着跑：计时接着上次的走', async () => {
|
||
const { page, ctx, errors } = await open(theme)
|
||
await goto(page, 'c0suspo')
|
||
check('从库里恢复的中断轮次认成中断', await shows(page, '服务重启，本轮已中断'))
|
||
const card = page.locator('[data-remedy="suspended"]')
|
||
await card.getByRole('button', { name: '继续运行' }).click()
|
||
const clock = page.locator('[data-turn="to1"] [title="已运行"]')
|
||
await clock.waitFor({ timeout: 6000 }).catch(() => {})
|
||
const text = await clock.innerText().catch(() => '')
|
||
check('计时从 42 秒接着走，不是从三天前算起', /^00:4\d/.test(text), text)
|
||
await page.screenshot({ path: `${SHOTS}/resumed-${theme}.png` })
|
||
await page.getByRole('button', { name: '停止这一轮' }).click()
|
||
check('没有运行时报错', errors.length === 0, errors[0] ?? '')
|
||
await ctx.close()
|
||
})
|
||
|
||
await section('history', '历史：已取消、续跑被拒、没查库、放宽步数', async () => {
|
||
const { page, ctx, errors } = await open(theme)
|
||
await goto(page, 'c0histf')
|
||
await shows(page, '步数用满的那一轮')
|
||
await page.waitForTimeout(800)
|
||
const turns = (await chatState(page)).byConversation.c0histf
|
||
check('老数据里的「已取消」认成已取消', turns[0].phase === 'cancelled', turns[0].phase)
|
||
const cancelledCard = page.locator('[data-remedy="cancelled"]')
|
||
check('已取消的那一轮没有「继续运行」', await cancelledCard.count() === 1
|
||
&& !(await cancelledCard.getByRole('button', { name: '继续运行' }).count()))
|
||
check('失败那一轮按运行状态给「继续运行」', turns[1].runStatus === 'failed'
|
||
&& await page.locator('[data-remedy="failed"]').getByRole('button', { name: '继续运行' }).count() === 1)
|
||
const failedCard = await page.locator('[data-turn="tf2"]').innerText()
|
||
// 说法和记录页同一份（lib/explain）：它把超时归成「等待超时」，原话收进技术细节
|
||
check('失败原因说人话、不露 Python 类名', /超时/.test(failedCard) && !failedCard.includes('KeyError'),
|
||
failedCard.split('\n').slice(0, 6).join(' / '))
|
||
// 人话只翻一遍：翻好的「操作超时：查询超时…」再翻一遍，标题和原因会各写一次「操作超时」
|
||
check('报错标题和原因不重复', !failedCard.includes('操作超时：'), failedCard.split('\n').slice(2, 6).join(' / '))
|
||
const rejected = page.locator('[data-turn="tf5"] + * [data-remedy="failed"]')
|
||
check('人驳回的失败：不给「继续运行」（原样续上只会再被驳回）', await rejected.count() === 1
|
||
&& !(await rejected.getByRole('button', { name: '继续运行' }).count()), (await rejected.innerText().catch(() => '')).replace(/\n/g, ' / '))
|
||
check('没查库的老答案标出来了', await shows(page, '本条未查询数据库'))
|
||
// 流里的报错照 store 拆好的那份画（3C REQ-12）：以前交的是原话，流拿它再讲一遍，当时记下的
|
||
// 标题、原因、怎么办全被换掉
|
||
const authErr = page.locator('[data-turn="tf6"] [data-turn-error]')
|
||
const authText = (await authErr.innerText().catch(() => '')).replace(/\n/g, ' / ')
|
||
check('落库的失败照原样画：标题、原因、怎么办都是当时记下的那份，只出现一次',
|
||
authText.startsWith('当时记下的标题：这把密钥被拒了') && authText.includes('当时记下的原因')
|
||
&& authText.includes('当时记下的怎么办') && authText.split('当时记下的标题').length === 2, authText.slice(0, 120))
|
||
check('老失败没记 fix：按原话认出该去模型接入，给直达入口',
|
||
await authErr.locator('a[data-fix="settings"][href="/settings/providers"]').count() === 1)
|
||
const stepBtn = page.getByRole('button', { name: /放宽步数后重新运行（12 → 24 步）/ })
|
||
check('步数用满给出「放宽步数后重新运行」并写明目标值', await stepBtn.count() === 1)
|
||
|
||
await page.locator('[data-turn="tf1"]').scrollIntoViewIfNeeded()
|
||
await page.screenshot({ path: `${SHOTS}/history-${theme}.png` })
|
||
await page.locator('[data-remedy="failed"]').getByRole('button', { name: '继续运行' }).click()
|
||
check('续跑被拒时说人话', await shows(page, '无需继续运行'))
|
||
await page.waitForTimeout(400)
|
||
await page.screenshot({ path: `${SHOTS}/continue-refused-${theme}.png` })
|
||
|
||
const startsBefore = log.runStarts.length
|
||
await stepBtn.click()
|
||
check('放宽步数：原图直接重跑，步数调到 24', await page.waitForFunction(
|
||
(n) => window.__chat.getState().byConversation.c0histf.at(-1).phase === 'done'
|
||
&& window.__chat.getState().byConversation.c0histf.at(-1).attempts?.length === 1, startsBefore, { timeout: 8000 })
|
||
.then(() => true, () => false))
|
||
const asked = log.turnStarts.filter((b) => b.retry_of === 'tf4').at(-1)
|
||
check('放宽步数的请求：带着这一轮、方式和目标步数（mode=max_steps、max_steps=24），图由服务端改',
|
||
asked?.conversation_id === 'c0histf' && asked?.mode === 'max_steps' && asked?.max_steps === 24 && !asked?.graph, JSON.stringify(asked))
|
||
const started = log.runStarts.slice(startsBefore).at(-1)
|
||
check('发起的图里 agent 步数是 24', started?.graph?.nodes?.find((n) => n.type === 'agent')?.data?.config?.max_steps === 24,
|
||
JSON.stringify(started?.graph?.nodes?.map((n) => n.data?.config)))
|
||
// 口径账本的来源链（next-round 02 做法 9）：运行记下属于哪段对话、哪一轮（服务端核对后封存进运行来源）
|
||
const server = await page.evaluate(() => window.__chat.getState().byConversation.c0histf.at(-1).serverId)
|
||
check('发起时带上这一轮属于哪段对话、哪一轮', started?.origin_ref?.conversation_id === 'c0histf'
|
||
&& !!server && started?.origin_ref?.turn_id === server, JSON.stringify(started?.origin_ref))
|
||
check('上一次留档：第 1 次尝试 · 结论不可用', await shows(page, '第 1 次尝试'))
|
||
await page.screenshot({ path: `${SHOTS}/retried-${theme}.png` })
|
||
check('没有运行时报错', errors.length === 0, errors[0] ?? '')
|
||
await ctx.close()
|
||
})
|
||
|
||
await section('rejected', '被驳回：补救区和列表写「被驳回」，重新运行后「之前的尝试」不写「已手动停止」', async () => {
|
||
ctl.rejectConvs = true
|
||
try {
|
||
const { page, ctx, errors } = await open(theme)
|
||
await goto(page, 'c0rejr')
|
||
await shows(page, '这个月的退款有多少')
|
||
await page.waitForTimeout(500)
|
||
const turn0 = (await chatState(page)).byConversation.c0rejr?.[0]
|
||
check('刷新之后照样认成被驳回（相位 cancelled，带着驳回的那一句）', turn0?.phase === 'cancelled'
|
||
&& turn0?.rejected?.sub === '张工 驳回了「查数」：数字对不上', JSON.stringify({ phase: turn0?.phase, rejected: turn0?.rejected }))
|
||
const card = page.locator('[data-remedy="cancelled"]')
|
||
const cardText = (await card.innerText().catch(() => '')).replace(/\n/g, ' / ')
|
||
check('补救区说是审批人驳回、谁在哪一步驳回，不说「你已停止本轮」', cardText.includes('被审批人驳回')
|
||
&& cardText.includes('张工 驳回了「查数」：数字对不上') && !cardText.includes('你已停止本轮'), cardText)
|
||
check('被驳回的没有「继续运行」', !(await card.getByRole('button', { name: '继续运行' }).count()))
|
||
// 过程里被驳回的那一步（node.failed 带 reason=rejected）：停止的剪影、中性色、行尾写「被驳回」，
|
||
// 不是绿色的「完成」勾，也不是失败的红三角（总方案 X9、细案 5.3「各处都写被驳回」）
|
||
await page.locator('[data-turn]').first().getByRole('button', { name: '查看执行过程' }).click()
|
||
const stepRow = page.locator('[data-turn]').first().locator('[data-node-id="ag"][data-step-status]').first()
|
||
await stepRow.waitFor({ timeout: 4000 }).catch(() => {})
|
||
const look = await stepRow.evaluate((el) => {
|
||
const badge = el.querySelector('svg[data-status]')
|
||
const line = el.querySelector('button')
|
||
return { status: el.dataset.stepStatus, badge: badge?.getAttribute('data-status') ?? null,
|
||
shape: badge?.getAttribute('data-shape') ?? null, badgeColor: badge ? getComputedStyle(badge).color : null,
|
||
done: getComputedStyle(document.documentElement).getPropertyValue('--st-done').trim(),
|
||
lineColor: line ? getComputedStyle(line).color : null, text: el.innerText.replace(/\n/g, ' / ') }
|
||
}).catch((e) => ({ error: String(e) }))
|
||
check('过程里被驳回的那一步：状态 rejected，剪影是停止、不是完成的勾', look.status === 'rejected'
|
||
&& look.badge === 'rejected' && look.shape === 'stop', JSON.stringify(look))
|
||
const doneRgb = await page.evaluate(() => {
|
||
const probe = document.createElement('span')
|
||
probe.style.color = 'var(--st-done)'
|
||
document.body.appendChild(probe)
|
||
const c = getComputedStyle(probe).color
|
||
probe.remove()
|
||
return c
|
||
})
|
||
check('……图标不是绿色（完成的颜色）', !!look.badgeColor && look.badgeColor !== doneRgb, `${look.badgeColor} vs ${doneRgb}`)
|
||
check('……副标题写谁驳回、备注，行尾写「被驳回」，不写「失败」', (look.text ?? '').includes('张工 驳回：数字对不上')
|
||
&& (look.text ?? '').includes('被驳回') && !(look.text ?? '').includes('失败'), look.text)
|
||
if (theme === 'dark') await page.screenshot({ path: `${SHOTS}/rejected-step-${theme}.png` })
|
||
const aside = page.locator('aside[aria-label="对话列表"]')
|
||
const rowText = async (t) => (await aside.locator('div.group', { hasText: t }).innerText().catch(() => '')).replace(/\n/g, ' / ')
|
||
// 标题里不带「被驳回」：标出来的是状态，不是标题里的字
|
||
check('列表里打开过的：标「被驳回」', (await rowText('会话：退款那一问')).includes('被驳回'), await rowText('会话：退款那一问'))
|
||
check('列表里没打开过的：按 last_status 标「被驳回」，不标「失败」', (await rowText('没打开过的那一问')).includes('被驳回')
|
||
&& !(await rowText('没打开过的那一问')).includes('失败'), await rowText('没打开过的那一问'))
|
||
|
||
const startsBefore = log.runStarts.length
|
||
await card.getByRole('button', { name: '重新运行本轮' }).click()
|
||
check('重新运行：留下一次尝试', await page.waitForFunction(
|
||
() => window.__chat.getState().byConversation.c0rejr.at(-1).attempts?.length === 1, null, { timeout: 8000 })
|
||
.then(() => true, () => false))
|
||
check('按原图重新发起了运行', log.runStarts.length > startsBefore, `${log.runStarts.length - startsBefore} 次`)
|
||
const [attempt] = (await chatState(page)).byConversation.c0rejr.at(-1).attempts ?? []
|
||
check('那一次记成被驳回、写谁在哪一步驳回', attempt?.outcome === 'rejected'
|
||
&& attempt?.summary === '被驳回：张工 驳回了「查数」：数字对不上', JSON.stringify(attempt))
|
||
const history = page.locator('[aria-label="之前的尝试"]')
|
||
const historyText = (await history.innerText().catch(() => '')).replace(/\n/g, ' / ')
|
||
check('「之前的尝试」写「被驳回」，不写「已手动停止」', historyText.includes('第 1 次尝试')
|
||
&& historyText.includes('被驳回：张工 驳回了「查数」') && !historyText.includes('已手动停止'), historyText)
|
||
check('没有运行时报错', errors.length === 0, errors[0] ?? '')
|
||
await ctx.close()
|
||
} finally {
|
||
ctl.rejectConvs = false
|
||
}
|
||
})
|
||
|
||
await section('ledger', '口径账本：轮次底部「正在比对口径」→ 摘要、按口径正式出具、对账（写请求在浏览器层拦下）', async () => {
|
||
ctl.ledgerConvs = true
|
||
try {
|
||
const { page, ctx, errors } = await open(theme)
|
||
// 文档记着的运行要和成果的运行对得上（run-led），片段、证据图按这次运行答
|
||
await page.route(/\/api\/artifacts\//, (route) => route.fulfill({ json: { id: EVIDENCE.doc_artifact,
|
||
content: { ...EVIDENCE.doc, run_id: 'run-led' } } }))
|
||
await page.route(/\/api\/runs\/run-led\/evidence(\/.*)?(\?.*)?$/, (route) => {
|
||
const m = new URL(route.request().url()).pathname.match(/\/segments\/([^/]+)$/)
|
||
return m ? route.fulfill({ json: EVIDENCE.segments[decodeURIComponent(m[1])] ?? {} }) : route.fulfill({ json: EVIDENCE.graph })
|
||
})
|
||
// 账本：头两次答 pending（正在比对），之后比对过了：整题等价于受管口径「周报口径」v2，另有一条查询可以对账
|
||
const led = { gets: [], writes: [], issued: [] }
|
||
await page.route(/\/api\/runs\/run-led\/ledger$/, (route) => {
|
||
led.gets.push(Date.now())
|
||
if (led.gets.length <= 2) return route.fulfill({ json: { run_id: 'run-led', state: 'pending', note: '封存后比对，不在封存范围内' } })
|
||
return route.fulfill({ json: ledgerOf('run-led', { labels: WHOLE_LABELS, whole: WHOLE, actions: [ISSUE, RECONCILE],
|
||
issued: led.issued }) })
|
||
})
|
||
await page.route(/\/api\/ledger\//, (route) => {
|
||
const req = route.request()
|
||
const path = new URL(req.url()).pathname
|
||
let body = null
|
||
try { body = req.postDataJSON() } catch { body = null }
|
||
led.writes.push({ key: `${req.method()} ${path}`, body })
|
||
if (req.method() === 'POST' && path === '/api/ledger/answers/run-led/issue') {
|
||
led.issued = [{ run_id: 'run-iss-1', status: 'succeeded', end_reason: null, workflow_id: 'wf-weekly', version: 2,
|
||
tier: 'formal', tier_text: '完整出具', auto: false, created_at: iso(0) }]
|
||
return route.fulfill({ json: { run_id: 'run-iss-1', created: true, issued: led.issued[0] } })
|
||
}
|
||
// 对账：这一段只核对按钮发出的请求体，结果卡片在 check-ledger 里看
|
||
if (req.method() === 'POST' && path === '/api/ledger/answers/run-led/reconcile') {
|
||
return route.fulfill({ status: 409, json: { detail: '检查脚本：不真的发起对账', code: 'ledger_reconcile_refused' } })
|
||
}
|
||
return route.fulfill({ status: 409, json: { detail: '检查脚本不写库' } })
|
||
})
|
||
await goto(page, 'c0ledgr')
|
||
const turnBox = page.locator('[data-turn]').first()
|
||
const waiting = await page.locator('[data-ledger-turn="waiting"]').first().waitFor({ timeout: 8000 }).then(() => true, () => false)
|
||
check('交付了的轮次：比对还没出结果时写「正在比对口径」', waiting
|
||
&& (await page.locator('[data-ledger-turn="waiting"]').first().innerText()).includes('正在比对口径'))
|
||
const summary = page.locator('[data-ledger-summary]').first()
|
||
const shown = await summary.waitFor({ timeout: 10000 }).then(() => true, () => false)
|
||
check('按退避再取，比对出结果后写摘要（不靠实时流）', shown && led.gets.length === 3, `${led.gets.length} 次`)
|
||
const gaps = led.gets.slice(1).map((t, i) => t - led.gets[i])
|
||
check('……隔得越来越久（退避），不是每一拍都取', gaps.length === 2 && gaps[0] >= 700 && gaps[1] >= gaps[0] - 100,
|
||
gaps.join('ms, ') + 'ms')
|
||
const text = (await summary.innerText()).replace(/\s+/g, ' ')
|
||
check('摘要：几个数来自哪份口径', text.includes('3 个数来自口径「周报口径」v2（受管）'), text)
|
||
check('整题等价：写等价于哪份口径、怎么认的', (await page.locator('[data-ledger-whole]').first().innerText())
|
||
.includes('整题等价于口径「周报口径」v2（受管） · 等价已证明'))
|
||
await page.waitForTimeout(3500)
|
||
check('比对过了、没有要等的：不再取', led.gets.length === 3, `${led.gets.length} 次`)
|
||
const issueBtn = page.locator('[data-ledger-issue-button]').first()
|
||
check('按服务端的动作摆「按口径正式出具」，下面写为什么不自动', await issueBtn.count() === 1
|
||
&& (await page.locator('[data-ledger-auto-reason]').first().innerText()).includes('设置里关了自动按口径出具'))
|
||
check('有相关口径：摆「对账」按钮', (await page.locator('[data-ledger-turn-reconcile]').first().innerText()).includes('对账：「订单口径」v1（已发布）'))
|
||
check('数字上挂着来自口径的小标记', await turnBox.locator('[data-ev-caliber-of="s6"]').count() === 1)
|
||
|
||
await issueBtn.click()
|
||
const posted = await until(() => led.writes.some((w) => w.key === 'POST /api/ledger/answers/run-led/issue'), 5000)
|
||
const issueBody = led.writes.find((w) => w.key === 'POST /api/ledger/answers/run-led/issue')?.body
|
||
check('出具的请求只说按哪份口径（参数由服务端从比对里取）', posted
|
||
&& JSON.stringify(issueBody) === JSON.stringify({ workflow_id: 'wf-weekly', version: 2 }), JSON.stringify(issueBody))
|
||
const issued = page.locator('[data-ledger-issued="run-iss-1"]').first()
|
||
const issuedShown = await issued.waitFor({ timeout: 6000 }).then(() => true, () => false)
|
||
const issuedText = issuedShown ? (await issued.innerText()).replace(/\s+/g, ' ') : ''
|
||
check('出具的正式运行挂在轮次下面：写档位', issuedText.includes('按口径正式出具') && issuedText.includes('「周报口径」v2（受管）')
|
||
&& issuedText.includes('完整出具'), issuedText)
|
||
check('……点开去运行', await issued.locator('a[href="/runs/run-iss-1"]').count() === 1)
|
||
check('出具过了：不再摆出具按钮', await page.locator('[data-ledger-issue-button]').count() === 0)
|
||
|
||
await page.locator('[data-ledger-turn-reconcile]').first().click()
|
||
const dialog = page.locator('[data-ledger-reconcile-dialog]')
|
||
check('点「对账」打开对账：差在哪、核对参数', await dialog.waitFor({ timeout: 4000 }).then(() => true, () => false)
|
||
&& (await dialog.innerText()).includes('口径只保留状态为已支付的行') && (await dialog.innerText()).includes('参数按这次的 SQL 推断，请核对'))
|
||
await page.locator('[data-ledger-param-input="start"]').fill('2026-09-02')
|
||
await page.locator('[data-ledger-reconcile-launch]').click()
|
||
await until(() => led.writes.some((w) => w.key === 'POST /api/ledger/answers/run-led/reconcile'), 5000)
|
||
const recBody = led.writes.find((w) => w.key === 'POST /api/ledger/answers/run-led/reconcile')?.body
|
||
check('对账的请求：哪条查询、哪份口径、核对过的参数（confirm）', recBody?.query === 'q-orders' && recBody?.workflow_id === 'wf-orders'
|
||
&& recBody?.version === 1 && recBody?.confirm === true && recBody?.inputs?.start === '2026-09-02' && !('store' in (recBody?.inputs ?? {})),
|
||
JSON.stringify(recBody))
|
||
check('服务端拒了：照它的话写在对账框里', (await page.locator('[data-ledger-reconcile-problem]').innerText().catch(() => ''))
|
||
.includes('检查脚本：不真的发起对账'))
|
||
check('写请求只有出具、对账这两次（都在浏览器层拦下）', led.writes.length === 2, led.writes.map((w) => w.key).join(' | '))
|
||
await page.screenshot({ path: `${SHOTS}/ledger-${theme}.png` })
|
||
check('没有运行时报错', errors.length === 0, errors[0] ?? '')
|
||
await ctx.close()
|
||
} finally {
|
||
ctl.ledgerConvs = false
|
||
}
|
||
})
|
||
|
||
await section('recall', '口径账本：答案身上的召回标记写在轮次底部（已按 vN 核对，不影响；依据有变；有待办时链到待办）', async () => {
|
||
// 召回在原答案的运行上追加 ledger.recalled、ledger.basis_changed（封存后），账本接口的 recall 把它们交出来（02 做法 7）。
|
||
// 以前只在「过程」里点开才看得到，使用者又进不了记录页。没比对的答案（state 不是 matched）照样可能被标依据有变
|
||
ctl.ledgerConvs = true
|
||
try {
|
||
// 比对过的那份另带一次做完的对账（另有两个人说过说明有用）：轮次里卡片上方没有「差在哪」，卡里写一份，反馈问的就是它
|
||
const reconciled = [{ ...reconciledOf('succeeded'), feedback: { mine: null, useful: 2, unclear: 0 } }]
|
||
for (const [name, LED] of [['matched', { ...ledgerOf('run-led', { labels: WHOLE_LABELS, whole: WHOLE, actions: [], reconciled }),
|
||
recall: RECALL }],
|
||
['ineligible', { run_id: 'run-led', state: 'ineligible', reason: '这种来源的运行不比对', note: '封存后比对，不在封存范围内',
|
||
recall: { checked: [], basis: [{ kind: 'structure', items: 1, new_errors: 0, at: iso(5) }], tasks: [] } }]]) {
|
||
const { page, ctx, errors } = await open(theme)
|
||
await page.route(/\/api\/artifacts\//, (route) => route.fulfill({ json: { id: EVIDENCE.doc_artifact,
|
||
content: { ...EVIDENCE.doc, run_id: 'run-led' } } }))
|
||
await page.route(/\/api\/runs\/run-led\/evidence(\/.*)?(\?.*)?$/, (route) => {
|
||
const m = new URL(route.request().url()).pathname.match(/\/segments\/([^/]+)$/)
|
||
return m ? route.fulfill({ json: EVIDENCE.segments[decodeURIComponent(m[1])] ?? {} }) : route.fulfill({ json: EVIDENCE.graph })
|
||
})
|
||
await page.route(/\/api\/runs\/run-led\/ledger$/, (route) => route.fulfill({ json: LED }))
|
||
await page.route(/\/api\/runs\/run-rec-1\/evidence\/metrics\/reconcile\/[^/]+$/, (route) =>
|
||
route.fulfill({ json: metricDetail('run-rec-1', new URL(route.request().url()).pathname.split('/').at(-1)) }))
|
||
await goto(page, 'c0ledgr')
|
||
const marks = page.locator('[data-ledger-turn] [data-ledger-recall]').first()
|
||
const shown = await marks.waitFor({ timeout: 8000 }).then(() => true, () => false)
|
||
const text = shown ? (await marks.innerText()).replace(/\s+/g, ' ') : ''
|
||
if (name === 'matched') {
|
||
check('轮次底部写「已按口径「周报口径」v3 核对，不影响」（核对过两次只写最近那一版）', shown
|
||
&& text.includes('已按口径「周报口径」v3 核对，不影响') && !text.includes('v2 核对'), text)
|
||
check('……依据有变：项数相加，写按新目录新出现几条错误级问题', text.includes('依据的数据目录有变（3 项）')
|
||
&& text.includes('按新目录检查新出现 1 条错误级问题'), text)
|
||
const link = marks.locator('[data-ledger-recall-task="tk-recall-1"] a')
|
||
check('……有说的是这个答案的召回待办：写数变了、链到待办', text.includes('口径升到 v2，这个答案用到的数变了')
|
||
&& await link.getAttribute('href').catch(() => null) === '/apps/inbox/tk-recall-1', text)
|
||
check('……和摘要在同一张账本卡里', await page.locator('[data-ledger-turn="matched"] [data-ledger-recall]').count() === 1)
|
||
const card = page.locator('[data-ledger-turn] [data-ledger-reconciled="run-rec-1"]')
|
||
const where = (await card.locator('[data-ledger-reconciled-where]').innerText().catch(() => '')).replace(/\s+/g, ' ')
|
||
check('轮次里做完的对账卡：写「差在哪」，下面问说明有没有帮助、写各几个人', where.includes('口径只保留状态为已支付的行')
|
||
&& (await card.locator('[data-ledger-feedback]').innerText().catch(() => '')).includes('2 人觉得有用')
|
||
&& await card.locator('[data-ledger-feedback-give]').count() === 2, where)
|
||
await page.screenshot({ path: `${SHOTS}/ledger-recall-${theme}.png` })
|
||
} else {
|
||
check('不比对的答案：轮次底部只写依据有变（表结构）', shown && text.includes('依据的表结构有变（1 项）')
|
||
&& await page.locator('[data-ledger-turn="recall"]').count() === 1 && await page.locator('[data-ledger-summary]').count() === 0, text)
|
||
}
|
||
check(`没有运行时报错（${name}）`, errors.length === 0, errors[0] ?? '')
|
||
await ctx.close()
|
||
}
|
||
} finally {
|
||
ctl.ledgerConvs = false
|
||
}
|
||
})
|
||
|
||
await section('fixfirst', '要改的东西在工具库里：先去改参数定义，「接着跑」退成次要', async () => {
|
||
ctl.fixConvs = true
|
||
try {
|
||
for (const [cid, tid, variant] of [['c0fixy', 'ty1', ''], ['c0fixz', 'tfz1', '（后端现在的说法）']]) {
|
||
const { page, ctx, errors } = await open(theme)
|
||
await goto(page, cid)
|
||
const card = page.locator('[data-remedy="failed"]')
|
||
await card.waitFor({ timeout: 6000 }).catch(() => {})
|
||
const fix = card.locator('[data-remedy-fix="tools"]')
|
||
const go = card.getByRole('button', { name: '继续运行' })
|
||
check(`主按钮是「去改参数定义」，直达那个工具的编辑框${variant}`,
|
||
(await fix.innerText().catch(() => '')).includes('去改参数定义')
|
||
&& /\bbtn-primary\b/.test(await fix.getAttribute('class').catch(() => '') ?? '')
|
||
&& await fix.getAttribute('href').catch(() => null) === '/tools/custom?edit=lookup_order',
|
||
await fix.getAttribute('href').catch(() => '没有这个按钮') ?? '')
|
||
check(`「继续运行」还在，但不是实心的，排在去改的后面${variant}`,
|
||
await go.count() === 1 && !/\bbtn-primary\b/.test(await go.getAttribute('class').catch(() => '') ?? '')
|
||
&& await page.evaluate(() => {
|
||
const c = document.querySelector('[data-remedy="failed"]')
|
||
const f = c?.querySelector('[data-remedy-fix]')
|
||
const g = [...(c?.querySelectorAll('button') ?? [])].find((b) => b.textContent?.includes('继续运行'))
|
||
return !!f && !!g && !!(f.compareDocumentPosition(g) & Node.DOCUMENT_POSITION_FOLLOWING)
|
||
}))
|
||
check(`没有别的实心按钮抢主位${variant}`, await card.locator('.btn-primary').count() === 1)
|
||
const streamFix = page.locator(`[data-turn="${tid}"] [data-turn-error] a[data-fix="tools"]`)
|
||
check(`流里的报错也直达那个工具${variant}`, await streamFix.getAttribute('href').catch(() => null) === '/tools/custom?edit=lookup_order'
|
||
&& (await streamFix.innerText().catch(() => '')).includes('去改参数定义'))
|
||
await page.screenshot({ path: `${SHOTS}/fixfirst-${cid}-${theme}.png` })
|
||
check(`没有运行时报错${variant}`, errors.length === 0, errors[0] ?? '')
|
||
await ctx.close()
|
||
}
|
||
} finally {
|
||
ctl.fixConvs = false
|
||
}
|
||
})
|
||
|
||
await section('build', '建流程就失败：说人话、原文收进技术细节、给补救', async () => {
|
||
const { page, ctx, errors } = await open(theme)
|
||
await goto(page, 'c0buildj')
|
||
await send(page, '一个建流程就失败的问题')
|
||
check('说出发生了什么', await shows(page, '模型返回的工作流结构有误'))
|
||
check('说出怎么办', await shows(page, '请重试，或把需求描述得更具体'))
|
||
const card = await page.locator('[data-turn]').last().innerText()
|
||
check('原始异常不直接摆出来', !card.includes('validation errors'), card.split('\n').slice(0, 5).join(' / '))
|
||
check('原文在「技术细节」里', await page.locator('[data-turn] details', { hasText: '技术细节' }).count() > 0)
|
||
const remedy = page.locator('[data-remedy="failed"]')
|
||
check('没有运行就没有「继续运行」', !(await remedy.getByRole('button', { name: '继续运行' }).count()))
|
||
check('给「重试本轮」', await remedy.getByRole('button', { name: '重试本轮' }).count() === 1)
|
||
await remedy.getByRole('button', { name: '换个说法' }).click()
|
||
const box = page.getByRole('textbox', { name: '向数据提问' })
|
||
await page.waitForTimeout(150)
|
||
const filled = await box.inputValue()
|
||
const focused = await box.evaluate((el) => el === document.activeElement)
|
||
check('「换个说法」把原问题填回输入框并聚焦', filled === '一个建流程就失败的问题' && focused,
|
||
`${filled} · ${focused ? '已聚焦' : '没聚焦'}`)
|
||
await page.screenshot({ path: `${SHOTS}/build-failed-${theme}.png` })
|
||
check('没有运行时报错', errors.length === 0, errors[0] ?? '')
|
||
await ctx.close()
|
||
})
|
||
|
||
await section('catalogpatch', '说了一条数据事实：「建议更新数据目录」卡片，保存撞上 409 能重新载入', async () => {
|
||
const { page, ctx, errors } = await open(theme)
|
||
await goto(page, 'c0emptg')
|
||
ctl.patchConflict = true
|
||
const before = log.catalogPatches.length
|
||
await send(page, 'status=9 表示作废，统计时要排除')
|
||
const card = page.locator('[data-catalog-patch="orders"]')
|
||
check('回答下面有「建议更新数据目录」卡片', await card.waitFor({ timeout: 6000 }).then(() => true, () => false)
|
||
&& (await card.innerText()).includes('建议更新数据目录'))
|
||
check('逐项写改前、改后、理由', (await card.locator('[data-patch-before]').innerText()).includes('1=已支付')
|
||
&& (await card.locator('[data-patch-after]').innerText()).includes('9=作废')
|
||
&& (await card.locator('[data-patch-reason]').innerText()).includes('status=9'))
|
||
const t = (await chatState(page)).byConversation.c0emptg.at(-1)
|
||
check('提案原样记进这一轮的操作流，回答照常交付', t.ops.some((o) => o.op === 'catalog_patch') && t.phase === 'done' && t.noQuery === true)
|
||
await card.locator('[data-patch-save]').click()
|
||
await card.locator('[data-patch-conflict]').waitFor({ timeout: 3000 })
|
||
check('409：提示这张表刚被修改过', (await card.innerText()).includes('这张表刚被修改过'))
|
||
await card.locator('[data-patch-reload]').click()
|
||
await card.locator('[data-patch-save]').waitFor({ timeout: 3000 })
|
||
check('重新载入后改前按最新的目录写', (await card.locator('[data-patch-before]').innerText()).includes('2=已退款'))
|
||
await card.locator('[data-patch-save]').click()
|
||
await card.locator('[data-patch-done]').waitFor({ timeout: 3000 })
|
||
const sent = log.catalogPatches.slice(before)
|
||
check('先带提案的版本，重新载入后带新版本', sent[0]?.if_version === 2 && !!sent[1]?.preview && sent[2]?.if_version === 3,
|
||
JSON.stringify(sent))
|
||
check('保存成功写明第几版、给到数据目录的入口', (await card.locator('[data-patch-done]').innerText()).includes('第 4 版')
|
||
&& (await card.locator('[data-patch-open]').getAttribute('href')) === '/data/catalog/src-shop/orders')
|
||
check('没有运行时报错', errors.length === 0, errors[0] ?? '')
|
||
await ctx.close()
|
||
})
|
||
|
||
await section('askback', '助手反问：这一轮记成「助手反问」，问题进历史，输入框提示回答', async () => {
|
||
const { page, ctx, errors } = await open(theme)
|
||
await goto(page, 'c0emptg')
|
||
const patches = log.patches.length
|
||
await send(page, '先问清楚再查上周销售额')
|
||
const t = await page.waitForFunction(() => window.__chat.getState().byConversation.c0emptg?.at(-1)?.phase !== 'planning', null,
|
||
{ timeout: 6000 }).then(() => true, () => false)
|
||
const turn = (await chatState(page)).byConversation.c0emptg.at(-1)
|
||
check('只问不改的一轮收尾，不再一直停在「正在理解需求」', t && turn.phase === 'done', turn.phase)
|
||
check('这一轮记成「助手反问」，问题作为这一轮的内容', turn.asked === '销售额按下单日还是支付日算？'
|
||
&& turn.output?.answer === turn.asked, JSON.stringify({ asked: turn.asked, output: turn.output }))
|
||
const card = page.locator('[data-turn]').last()
|
||
check('轮次头写「助手反问」，问题摆出来', (await card.innerText()).includes('助手反问')
|
||
&& (await card.innerText()).includes('销售额按下单日还是支付日算？'), (await card.innerText()).split('\n').slice(0, 6).join(' / '))
|
||
check('不标「本条未查询数据库」（这不是回答）', !(await card.innerText()).includes('本条未查询数据库'))
|
||
const box = page.getByRole('textbox', { name: '向数据提问' })
|
||
check('输入框提示回答助手的问题', (await box.getAttribute('placeholder')).includes('回答助手的问题'))
|
||
// 那一行由服务端写（答案列是助手问的那句话，meta 标着反问）：页面不 PATCH，照 end 带来的那一行摆
|
||
check('页面不写这一轮的结局；服务端写的那一行 meta 标着反问', log.patches.slice(patches).length === 0
|
||
&& !!turn.meta?.asked, JSON.stringify(turn.meta?.asked))
|
||
check('没有运行时报错', errors.length === 0, errors[0] ?? '')
|
||
await ctx.close()
|
||
|
||
// 刷新之后：库里的这一轮照样是「助手反问」，不当成没查库的回答
|
||
ctl.askBack = true
|
||
const again = await open(theme)
|
||
await goto(again.page, 'c0askbk')
|
||
const restored = again.page.locator('[data-turn]').last()
|
||
await restored.waitFor({ timeout: 6000 })
|
||
check('刷新后照样标「助手反问」', (await restored.innerText()).includes('助手反问'), (await restored.innerText()).split('\n').slice(0, 4).join(' / '))
|
||
check('……不标「本条未查询数据库」', !(await restored.innerText()).includes('本条未查询数据库'))
|
||
check('……输入框提示回答助手的问题', (await again.page.getByRole('textbox', { name: '向数据提问' }).getAttribute('placeholder'))
|
||
.includes('回答助手的问题'))
|
||
ctl.askBack = false
|
||
await again.ctx.close()
|
||
})
|
||
|
||
await section('selfcheck', '服务端已经自查修正过还没过：不再整轮重新生成', async () => {
|
||
const { page, ctx, errors } = await open(theme)
|
||
await goto(page, 'c0emptg')
|
||
const gens = log.gens.length
|
||
await send(page, '自查没修好的问题')
|
||
const done = await page.waitForFunction(() => window.__chat.getState().byConversation.c0emptg?.at(-1)?.phase === 'error', null,
|
||
{ timeout: 6000 }).then(() => true, () => false)
|
||
await page.waitForTimeout(500)
|
||
const turn = (await chatState(page)).byConversation.c0emptg.at(-1)
|
||
check('只生成了一次（同一轮的自查修正已经做过，不再整轮重来）', log.gens.length - gens === 1, `${log.gens.length - gens} 次`)
|
||
check('这一轮照实说没过、说明已经修正过', done && turn.failure?.reason?.includes('已按检查结果修正过'), JSON.stringify(turn.failure))
|
||
check('没有运行时报错', errors.length === 0, errors[0] ?? '')
|
||
await ctx.close()
|
||
})
|
||
|
||
await section('emptyretry', '助手只写了思考、一个操作都没给：照服务端的 next=retry 自动重来一次，还不行再照实报', async () => {
|
||
const { page, ctx, errors } = await open(theme)
|
||
await goto(page, 'c0emptg')
|
||
let gens = log.gens.length
|
||
await send(page, '只写了思考：有几家门店')
|
||
const done = await page.waitForFunction(() => window.__chat.getState().byConversation.c0emptg?.at(-1)?.phase === 'done',
|
||
null, { timeout: 6000 }).then(() => true, () => false)
|
||
await page.waitForTimeout(300)
|
||
const turn = (await chatState(page)).byConversation.c0emptg.at(-1)
|
||
check('第一次没给操作：自动重来一次，第二次照常回答', done && log.gens.length - gens === 2
|
||
&& String(turn.output?.answer ?? '').includes('3 家'), `${log.gens.length - gens} 次生成 · ${JSON.stringify(turn.output)}`)
|
||
check('重来时交回去的话带着原因（同服务端 retry_instruction 的从零重建措辞）',
|
||
log.gens.at(-1)?.instruction === '只写了思考：有几家门店\n\n（上一次没有成功：模型没有给出任何修改，也没有说明原因。请针对这个原因调整，重新给出完整的工作流。）',
|
||
log.gens.at(-1)?.instruction)
|
||
check('两次建图的用量都记进这一轮', Math.abs((turn.usage?.copilot?.cost_usd ?? 0) - 0.002) < 1e-9, JSON.stringify(turn.usage))
|
||
gens = log.gens.length
|
||
await send(page, '两次都只写了思考：有几家门店')
|
||
const failed = await page.waitForFunction(() => window.__chat.getState().byConversation.c0emptg?.at(-1)?.phase === 'error',
|
||
null, { timeout: 6000 }).then(() => true, () => false)
|
||
await page.waitForTimeout(500)
|
||
const last = (await chatState(page)).byConversation.c0emptg.at(-1)
|
||
check('重来还不行：照实报「助手本轮未完成」，不再试第三次', failed && log.gens.length - gens === 2
|
||
&& JSON.stringify(last.failure ?? {}).includes('助手本轮未完成'), `${log.gens.length - gens} 次生成 · ${JSON.stringify(last.failure)}`)
|
||
check('没有运行时报错', errors.length === 0, errors[0] ?? '')
|
||
await ctx.close()
|
||
})
|
||
|
||
await section('decide', '新服务端：照 final.next 和复核的 retry_instruction 走，这一轮的花费 = 助手 + 运行 + 复核', async () => {
|
||
const { page, ctx, errors } = await open(theme)
|
||
await goto(page, 'c0emptg')
|
||
const gens = log.gens.length
|
||
const runs = log.runStarts.length
|
||
const reviews = log.reviews.length
|
||
await send(page, '照服务端的决定：有几家门店')
|
||
const done = await page.waitForFunction(() => {
|
||
const t = window.__chat.getState().byConversation.c0emptg?.at(-1)
|
||
return t?.phase === 'done' && t?.run?.id?.startsWith('run-decide')
|
||
}, null, { timeout: 8000 }).then(() => true, () => false)
|
||
await page.waitForTimeout(500)
|
||
const turn = (await chatState(page)).byConversation.c0emptg.at(-1)
|
||
check('复核判不可用：带着原图重来一次', done && log.gens.length - gens === 2 && log.runStarts.length - runs === 2,
|
||
`${log.gens.length - gens} 次生成、${log.runStarts.length - runs} 次运行`)
|
||
check('重来时交回去的就是服务端给的那句话', log.gens.at(-1)?.instruction === DECIDE_RETRY, log.gens.at(-1)?.instruction)
|
||
check('第二次复核带着这一轮试过的次数', log.reviews.slice(reviews).map((r) => r.attempt).join(',') === '0,1',
|
||
JSON.stringify(log.reviews.slice(reviews)))
|
||
// 服务端写的那一行（end 带来的）
|
||
const last = turn.meta?.usage
|
||
check('meta.usage 记下助手、复核、换下来的运行和这一次运行', !!last && Math.abs(last.copilot?.cost_usd - 0.024) < 1e-9
|
||
&& Math.abs(last.review?.cost_usd - 0.003) < 1e-9 && Math.abs(last.earlier?.cost_usd - 0.02) < 1e-9
|
||
&& Math.abs(last.run?.cost_usd - 0.02) < 1e-9, JSON.stringify(last))
|
||
const tag = page.locator(`[data-turn="${turn.id}"] [data-turn-cost]`)
|
||
const tagText = await tag.innerText().catch(() => '')
|
||
const title = await tag.getAttribute('title').catch(() => '')
|
||
check('头部写这一轮的花费，悬停分得清助手、运行、复核', tagText.includes('$0.067') && (title ?? '').includes('助手建图')
|
||
&& (title ?? '').includes('复核'), `${tagText} | ${title}`)
|
||
|
||
// 复核慢过等待上限：答案先交出去，复核这一次的用量晚到了照记（这一轮的花费照实是助手 + 运行 + 复核，复核不落下）
|
||
ctl.lateReview = 900
|
||
// 服务端最多等复核这么久（检查里调短）
|
||
ctl.reviewWaitMs = 200
|
||
const patchesBefore = log.patches.length
|
||
await send(page, '照服务端的决定：复核晚到的那一轮')
|
||
const delivered = await page.waitForFunction(() => {
|
||
const t = window.__chat.getState().byConversation.c0emptg?.at(-1)
|
||
return t?.phase === 'done' && t?.question?.includes('复核晚到')
|
||
}, null, { timeout: 8000 }).then(() => true, () => false)
|
||
const early = (await chatState(page)).byConversation.c0emptg.at(-1)
|
||
check('复核没等到：答案先交出去，这时还没有复核的用量', delivered && !early.usage?.review, JSON.stringify(early.usage))
|
||
await page.waitForTimeout(1200)
|
||
const lateTurn = (await chatState(page)).byConversation.c0emptg.at(-1)
|
||
const lateMeta = lateTurn.meta?.usage
|
||
check('页面不写这一轮的结局', log.patches.slice(patchesBefore).length === 0)
|
||
check('复核晚到：它的用量照记进这一轮（服务端并进那一行的 meta.usage）', Math.abs((lateTurn.usage?.review?.cost_usd ?? 0) - 0.004) < 1e-9
|
||
&& Math.abs((lateMeta?.review?.cost_usd ?? 0) - 0.004) < 1e-9 && Math.abs((lateMeta?.copilot?.cost_usd ?? 0) - 0.012) < 1e-9,
|
||
JSON.stringify(lateMeta))
|
||
ctl.lateReview = 0
|
||
ctl.reviewWaitMs = 25_000
|
||
|
||
// 只建不跑：autorun 没写 false，next 是 ready，前端照 next 走、不发起运行
|
||
const before = log.runStarts.length
|
||
await send(page, '照服务端的决定，只建不跑')
|
||
const ready = await page.waitForFunction(() => window.__chat.getState().byConversation.c0emptg?.at(-1)?.phase === 'ready',
|
||
null, { timeout: 6000 }).then(() => true, () => false)
|
||
await page.waitForTimeout(300)
|
||
check('next=ready：图摆着等运行，不发起', ready && log.runStarts.length === before, `${log.runStarts.length - before} 次运行`)
|
||
check('没有运行时报错', errors.length === 0, errors[0] ?? '')
|
||
await ctx.close()
|
||
})
|
||
|
||
await section('makeapp', '做成应用：带着这次运行进蓝图，不再跳画布发布', async () => {
|
||
const { page, ctx, errors } = await open(theme)
|
||
const made = []
|
||
const extracted = []
|
||
await page.route('**/api/copilot/from-run**', (route) => { extracted.push(route.request().url()); return route.abort() })
|
||
await page.route(/\/api\/(blueprints|templates)(\/|\?|$)/, (route) => {
|
||
const r = route.request()
|
||
const path = new URL(r.url()).pathname
|
||
if (path === '/api/templates') return route.fulfill({ json: { items: [] } })
|
||
if (path === '/api/blueprints' && r.method() === 'POST') {
|
||
made.push(r.postDataJSON())
|
||
return route.fulfill({ status: 201, json: {
|
||
id: 'bp_from_run', name: 'B 的问题', owner: null, scope: 'shared', status: 'drafting', template_id: 'freeform',
|
||
template_version: 1, source: { kind: 'run', id: 'run-b1' }, conversation_id: null, workflow_id: null, app_id: null,
|
||
last_trial_run_id: null, graph_hash: 'x', joints: {}, copied_from: null, issues: [], created_at: iso(1), updated_at: iso(1),
|
||
spec: { format: 'agentlab.blueprint/1', purpose: { value: 'B 的问题', source: 'assistant', status: 'proposed' },
|
||
template: { id: 'freeform', version: 1 }, data: [], cadence: { value: null, source: 'assistant', status: 'missing' },
|
||
people: [], fields: [], writeback: null, access: { use: [], observe: [], manage: [] }, asks: [] },
|
||
} })
|
||
}
|
||
return route.fulfill({ status: 404, json: { detail: '检查脚本没有这个蓝图' } })
|
||
})
|
||
await goto(page, 'c0idleb')
|
||
const button = page.locator('[data-make-app]').last()
|
||
check('回答完成、有运行的轮次有「做成应用」', await button.waitFor({ timeout: 6000 }).then(() => true, () => false))
|
||
check('关登录（等同管理员）：建共享蓝图', await button.getAttribute('data-make-app') === 'shared')
|
||
await button.click()
|
||
check('建运行起点的蓝图', await until(() => made.length === 1) && made[0].source?.kind === 'run' && made[0].source?.id === 'run-b1'
|
||
&& made[0].scope === 'shared', JSON.stringify(made[0]))
|
||
check('进蓝图页，不去画布', await page.waitForURL(/\/blueprints\/bp_from_run$/, { timeout: 5000 }).then(() => true, () => false)
|
||
&& !page.url().includes('/studio/'), page.url())
|
||
check('不再提取工作流（from-run）', extracted.length === 0, extracted.join(' | '))
|
||
check('没有运行时报错', errors.length === 0, errors[0] ?? '')
|
||
await ctx.close()
|
||
})
|
||
|
||
await section('approval', '停在审批上：审批卡立刻出现，不等 4 秒轮询', async () => {
|
||
const { page, ctx, errors } = await open(theme)
|
||
ctl.approvals = []
|
||
ctl.pendingApproval = {
|
||
id: 'ap-1', run_id: 'run-wait', node_id: 'ag', mode: 'approve', title: '确认一下再查',
|
||
payload: { title: '确认一下再查' }, status: 'pending', response: {}, created_at: new Date().toISOString(),
|
||
workflow_name: '问数据', node_label: '查数', run_status: 'interrupted', run_class: 'exploratory',
|
||
}
|
||
ctl.interruptedAt = 0
|
||
await goto(page, 'c0waitk')
|
||
await send(page, '一个要审批的问题')
|
||
const seen = await page.locator('[data-approval-slot] button', { hasText: /通过|放行|批准/ }).first()
|
||
.waitFor({ timeout: 6000 }).then(() => true, () => false)
|
||
const lag = ctl.interruptedAt ? Date.now() - ctl.interruptedAt : -1
|
||
// 全局轮询是 4 秒一次；中断时立刻去拿的话，一秒内就该出来
|
||
check('审批卡在收到中断后马上出现（不等 4 秒轮询）', seen && lag >= 0 && lag < 1500, `${lag} ms`)
|
||
await page.waitForFunction(() => window.__chat.getState().byConversation.c0waitk?.at(-1)?.phase === 'waiting',
|
||
null, { timeout: 3000 }).catch(() => {})
|
||
const t = (await chatState(page)).byConversation.c0waitk.at(-1)
|
||
check('这一轮是等待审批，不算占着会话', t.phase === 'waiting'
|
||
&& await page.getByRole('button', { name: '停止这一轮' }).count() === 0, t.phase)
|
||
await page.screenshot({ path: `${SHOTS}/waiting-${theme}.png` })
|
||
ctl.approvals = []
|
||
check('没有运行时报错', errors.length === 0, errors[0] ?? '')
|
||
await ctx.close()
|
||
})
|
||
|
||
await section('reconnect', '实时流断了：按收到的最后一个 seq 重连，补齐、不重复（一轮在服务端照跑）', async () => {
|
||
const { page, ctx, errors } = await open(theme)
|
||
await goto(page, 'c0emptg')
|
||
ctl.opGapMs = 120
|
||
const streamsBefore = log.streams.length
|
||
const patchesBefore = log.patches.length
|
||
await send(page, '断线重连：查一下订单和明细')
|
||
await page.waitForFunction(() => (window.__chat.getState().byConversation.c0emptg?.at(-1)?.ops.length ?? 0) >= 2,
|
||
null, { timeout: 6000 }).catch(() => {})
|
||
// 接下来两次连接都直接断掉（网络抖了两下）：这一轮在服务端照跑，页面退避着重连
|
||
ctl.dropStreams = 2
|
||
const ran = await page.waitForFunction(() => {
|
||
const t = window.__chat.getState().byConversation.c0emptg?.at(-1)
|
||
return !!t?.run?.id && t.phase === 'running'
|
||
}, null, { timeout: 12000 }).then(() => true, () => false)
|
||
ctl.opGapMs = 10
|
||
await page.waitForTimeout(300)
|
||
const t = (await chatState(page)).byConversation.c0emptg.at(-1)
|
||
const T = srv.turns.get(t.serverId)
|
||
const afters = log.streams.slice(streamsBefore).filter((x) => x.turn === t.serverId).map((x) => x.after)
|
||
const resumedAt = afters.findIndex((a, i) => i >= 2 && a > 0 && afters[i - 1] === a && afters[i - 2] === a)
|
||
check('断了之后接着连：每次都带上收到的最后一个 seq（断掉的那两次和接上的那一次是同一个）', ran && resumedAt > 0,
|
||
afters.join(','))
|
||
check('从不往回要：重连的 seq 只增不减', afters.every((a, i) => i === 0 || a >= afters[i - 1]), afters.join(','))
|
||
check('补齐：服务端这一轮记下的事件，页面一条不少地收到了', !!T && t.turnSeq === T.events.length,
|
||
`${t.turnSeq} / ${T?.events.length}`)
|
||
// 心跳只落换了阶段的那一拍，同一阶段接着的不带 seq（服务端 chat/log.beat）：重连时服务端会补最近的一拍，页面接着的那一拍
|
||
// 顶替上一拍，不另起一条。按 seq 补齐的「一条只收一份」比的是心跳之外的操作
|
||
const ops = T?.events.filter((e) => e.type === 'op' && e.data?.op !== 'heartbeat').length ?? -1
|
||
const mine = t.ops.filter((o) => o.op !== 'heartbeat').length
|
||
check('不重复：助手的操作一条只收一份', mine === ops && t.ops.filter((o) => o.op === 'add_node').length === 3,
|
||
`${mine} / ${ops}`)
|
||
check('……补过的心跳不叠成好几条', t.ops.every((o, i) => o.op !== 'heartbeat' || t.ops[i + 1]?.op !== 'heartbeat'),
|
||
t.ops.map((o) => o.op).join(','))
|
||
check('画布边收边长到收尾：图有 3 个节点，运行接上了', t.graph?.nodes?.length === 3 && !!t.run?.id)
|
||
check('页面不写这一轮的结局（不 PATCH）', log.patches.slice(patchesBefore).length === 0)
|
||
await page.getByRole('button', { name: '停止这一轮' }).click().catch(() => {})
|
||
check('没有运行时报错', errors.length === 0, errors[0] ?? '')
|
||
await ctx.close()
|
||
})
|
||
|
||
await section('replay', '刷新、关掉再打开：服务端那一轮照跑，页面从这一次尝试的第一条接着播', async () => {
|
||
resetDb()
|
||
ctl.replayConvs = true
|
||
ctl.liveDetail = true
|
||
ctl.replayBuildGo = false
|
||
ctl.replayRunGo = false
|
||
try {
|
||
const patchesBefore = log.patches.length
|
||
let { page, ctx, errors } = await open(theme)
|
||
await goto(page, 'c0replay')
|
||
const startsBefore = log.turnStarts.length
|
||
await send(page, '刷新后接着看：各门店的订单数')
|
||
await page.waitForFunction(() => (window.__chat.getState().byConversation.c0replay?.at(-1)?.ops
|
||
.filter((o) => o.op === 'add_node').length ?? 0) === 2, null, { timeout: 6000 }).catch(() => {})
|
||
const serverId = (await chatState(page)).byConversation.c0replay.at(-1).serverId
|
||
|
||
// 刷新：库里那一行是 running、带着服务端的任务（建图中）
|
||
await page.reload()
|
||
await ready(page, () => !!window.__chat, '问数据页的 store（window.__chat）')
|
||
const live = await page.waitForFunction((id) => {
|
||
const t = window.__chat.getState().byConversation.c0replay?.find((x) => x.serverId === id)
|
||
return !!t && (t.phase === 'planning' || t.phase === 'building') && t.ops.filter((o) => o.op === 'add_node').length === 2
|
||
}, serverId, { timeout: 6000 }).then(() => true, () => false)
|
||
const t1 = (await chatState(page)).byConversation.c0replay?.find((x) => x.serverId === serverId)
|
||
check('刷新之后：这一轮还在服务端建图，照那个样子摆（不当成断了）', live, `${t1?.phase} · ${t1?.ops?.length} 条操作`)
|
||
check('……从这一次尝试的第一条回放，已经建出的两步一步不少、一步不重',
|
||
t1?.ops.filter((o) => o.op === 'add_node').length === 2 && t1?.ops.filter((o) => o.op === 'plan').length === 1)
|
||
check('……建图的过程照实时的样子摆着（不是要点开才取的历史轮次）', !t1?.restored
|
||
&& await page.locator('[data-turn]').last().locator('button', { hasText: '规划' }).count() > 0)
|
||
check('……刷新没有再发起一次（只发起过一次）', log.turnStarts.length - startsBefore === 1,
|
||
`${log.turnStarts.length - startsBefore} 次`)
|
||
check('……回放从这一次尝试开始（after 是 attempt_seq 之前）', log.streams.some((x) => x.turn === serverId && x.after === 0))
|
||
|
||
ctl.replayBuildGo = true
|
||
const running = await page.waitForFunction((id) => {
|
||
const t = window.__chat.getState().byConversation.c0replay?.find((x) => x.serverId === id)
|
||
return t?.phase === 'running' && t.run?.id === 'run-replay' && t.graph?.nodes?.length === 3
|
||
}, serverId, { timeout: 6000 }).then(() => true, () => false)
|
||
check('接着播：建完、服务端发起了运行，画布长全', running)
|
||
await page.screenshot({ path: `${SHOTS}/replay-running-${theme}.png` })
|
||
check('没有运行时报错（刷新后）', errors.length === 0, errors[0] ?? '')
|
||
await ctx.close()
|
||
|
||
// 关掉再打开：运行还在跑（停在闸上）
|
||
;({ page, ctx, errors } = await open(theme))
|
||
await goto(page, 'c0replay')
|
||
const reopened = await page.waitForFunction((id) => {
|
||
const t = window.__chat.getState().byConversation.c0replay?.find((x) => x.serverId === id)
|
||
return t?.phase === 'running' && t.run?.id === 'run-replay' && t.ops.filter((o) => o.op === 'add_node').length === 3
|
||
}, serverId, { timeout: 6000 }).then(() => true, () => false)
|
||
check('关掉再打开：照样接着看（运行中，建图的过程还在）', reopened,
|
||
JSON.stringify((await chatState(page)).byConversation.c0replay?.map((x) => x.phase)))
|
||
check('……运行的过程也接回来了', await page.waitForFunction((id) =>
|
||
(window.__chat.getState().byConversation.c0replay?.find((x) => x.serverId === id)?.events.length ?? 0) >= 2,
|
||
serverId, { timeout: 4000 }).then(() => true, () => false))
|
||
ctl.replayRunGo = true
|
||
const done = await page.waitForFunction((id) =>
|
||
window.__chat.getState().byConversation.c0replay?.find((x) => x.serverId === id)?.phase === 'done',
|
||
serverId, { timeout: 8000 }).then(() => true, () => false)
|
||
check('跑完、复核完：答案出来了', done && await shows(page, '刷新之后接着跑完了：两家门店。'))
|
||
check('页面一个结局都没写（不 PATCH）', log.patches.slice(patchesBefore).length === 0)
|
||
await page.screenshot({ path: `${SHOTS}/replay-done-${theme}.png` })
|
||
check('没有运行时报错（重新打开后）', errors.length === 0, errors[0] ?? '')
|
||
await ctx.close()
|
||
} finally {
|
||
ctl.replayConvs = false
|
||
ctl.liveDetail = false
|
||
ctl.replayBuildGo = true
|
||
ctl.replayRunGo = true
|
||
}
|
||
})
|
||
|
||
await section('stopreq', '停止：建图中也停得下，停止请求发给服务端的这一轮', async () => {
|
||
const { page, ctx, errors } = await open(theme)
|
||
await goto(page, 'c0emptg')
|
||
ctl.stopBuildGo = false
|
||
const patchesBefore = log.patches.length
|
||
const startsBefore = log.turnStarts.length
|
||
const runsBefore = log.runStarts.length
|
||
await send(page, '停止测试：各门店的订单数')
|
||
await page.waitForFunction(() => (window.__chat.getState().byConversation.c0emptg?.at(-1)?.ops.length ?? 0) >= 2,
|
||
null, { timeout: 6000 }).catch(() => {})
|
||
const asked = log.turnStarts.slice(startsBefore).at(-1)
|
||
check('发起的请求：这段对话、这个问题，不限数据源时不带 datasource_ids', asked?.conversation_id === 'c0emptg'
|
||
&& asked?.question === '停止测试：各门店的订单数' && !asked?.retry_of && !('datasource_ids' in (asked ?? {})), JSON.stringify(asked))
|
||
const t0 = (await chatState(page)).byConversation.c0emptg.at(-1)
|
||
await page.getByRole('button', { name: '停止这一轮' }).click()
|
||
await until(() => log.turnCancels.includes(t0.serverId), 3000)
|
||
check('停止请求发给服务端的这一轮（POST /api/chat/turns/{id}/cancel）', log.turnCancels.at(-1) === t0.serverId,
|
||
`${log.turnCancels.at(-1)} / ${t0.serverId}`)
|
||
await page.waitForTimeout(300)
|
||
const t = (await chatState(page)).byConversation.c0emptg.at(-1)
|
||
check('这一轮收成已取消，不画红', t.phase === 'cancelled' && !t.failure, t.phase)
|
||
ctl.stopBuildGo = true
|
||
await page.waitForTimeout(400)
|
||
const T = srv.turns.get(t0.serverId)
|
||
check('服务端那一轮也停了：不再往下建、不发起运行', T?.job.status === 'cancelled' && log.runStarts.length === runsBefore
|
||
&& T.events.filter((e) => e.type === 'op' && e.data.op === 'add_node').length === 2, `${T?.job.status} · ${log.runStarts.length - runsBefore} 次运行`)
|
||
check('建图中停止：补救区说「你在搭建工作流时停止了本轮」', await shows(page, '你在搭建工作流时停止了本轮'))
|
||
check('页面不写这一轮的结局（不 PATCH）', log.patches.slice(patchesBefore).length === 0)
|
||
check('没有运行时报错', errors.length === 0, errors[0] ?? '')
|
||
await ctx.close()
|
||
})
|
||
|
||
await section('userrole', '使用者（user 角色）：不给「继续运行」（服务端照老规矩不让使用者续跑），重试、重新运行本轮照常', async () => {
|
||
const { page, ctx, errors } = await open(theme)
|
||
// 开了登录、认出来是使用者、管理员开了问数据
|
||
await page.route('**/api/auth/me', (route) => route.fulfill({ json: { mode: 'local', username: 'uma', display_name: 'uma',
|
||
role: 'user', groups: [], verified: true, user_can_chat: true } }))
|
||
const continuesBefore = log.continues.length
|
||
await goto(page, 'c0suspo')
|
||
await shows(page, '服务重启，本轮已中断')
|
||
const card = page.locator('[data-remedy="suspended"]')
|
||
await card.waitFor({ timeout: 6000 }).catch(() => {})
|
||
const cardText = (await card.innerText().catch(() => '')).replace(/\n/g, ' / ')
|
||
check('中断的一轮：使用者看不到「继续运行」，说明里也不叫它去点', await card.getByRole('button', { name: '继续运行' }).count() === 0
|
||
&& !cardText.includes('继续运行') && cardText.includes('重新运行本轮'), cardText)
|
||
check('……「重新运行本轮」照常给', await card.getByRole('button', { name: '重新运行本轮' }).count() === 1)
|
||
await goto(page, 'c0histf')
|
||
await shows(page, '步数用满的那一轮')
|
||
await page.waitForTimeout(600)
|
||
const failed = page.locator('[data-turn="tf2"] + * [data-remedy="failed"]')
|
||
check('跑挂了的一轮：使用者也看不到「继续运行」，「重试本轮」照常给',
|
||
await page.getByRole('button', { name: '继续运行' }).count() === 0
|
||
&& await failed.getByRole('button', { name: /重试本轮|重新问一次/ }).count() === 1,
|
||
(await failed.innerText().catch(() => '')).replace(/\n/g, ' / '))
|
||
check('一次续跑都没发', log.continues.length === continuesBefore)
|
||
check('没有运行时报错', errors.length === 0, errors[0] ?? '')
|
||
await ctx.close()
|
||
})
|
||
|
||
await section('twotabs', '两个标签页同时打开停在半路的一轮：服务端只接着收尾一次，两边都接着看同一轮', async () => {
|
||
const a = await open(theme)
|
||
const b = await open(theme)
|
||
ctl.reviewDelay['run-late'] = 1200
|
||
const startsBefore = log.turnStarts.length
|
||
const postsBefore = log.patches.length
|
||
await Promise.all([goto(a.page, 'c0lateq'), goto(b.page, 'c0lateq')])
|
||
const done = (page) => page.waitForFunction(() => window.__chat.getState().byConversation.c0lateq?.[0]?.phase === 'done',
|
||
null, { timeout: 8000 }).then(() => true, () => false)
|
||
const [okA, okB] = await Promise.all([done(a.page), done(b.page)])
|
||
check('两边都收尾了，答案一样', okA && okB && await shows(a.page, '补交付的答案：三条记录。') && await shows(b.page, '补交付的答案：三条记录。'))
|
||
const resumed = log.turnStarts.slice(startsBefore).filter((x) => x.mode === 'resume' && x.retry_of === 'tq1')
|
||
check('服务端只接着收尾了一次（后到的那个 409，改成接着看）', resumed.length === 1, `${resumed.length} 次`)
|
||
const phases = await Promise.all([a.page, b.page].map((p) => p.evaluate(() => window.__chat.getState().byConversation.c0lateq?.[0]?.phase)))
|
||
check('后到的那个不说「无法确认本轮的当前状态」', !phases.includes('unknown'), phases.join(','))
|
||
check('两边都不写库（不 PATCH）', log.patches.length === postsBefore)
|
||
delete ctl.reviewDelay['run-late']
|
||
check('没有运行时报错', a.errors.length === 0 && b.errors.length === 0, a.errors[0] ?? b.errors[0] ?? '')
|
||
await a.ctx.close()
|
||
await b.ctx.close()
|
||
})
|
||
|
||
await section('steps', '历史轮次的执行过程：点开、收起', async () => {
|
||
const { page, ctx, errors } = await open(theme)
|
||
await goto(page, 'c0idleb')
|
||
await shows(page, 'B 的答案')
|
||
const toggle = page.getByRole('button', { name: /执行过程/ }).first()
|
||
check('默认只显示问题和答案', (await toggle.innerText()).includes('查看执行过程'))
|
||
await toggle.click()
|
||
check('点开后能收起', await page.getByRole('button', { name: '收起执行过程' })
|
||
.waitFor({ timeout: 4000 }).then(() => true, () => false))
|
||
await page.getByRole('button', { name: '收起执行过程' }).click()
|
||
check('收起之后按钮回到「查看执行过程」', await page.getByRole('button', { name: '查看执行过程' }).count() === 1)
|
||
check('没有运行时报错', errors.length === 0, errors[0] ?? '')
|
||
await ctx.close()
|
||
})
|
||
|
||
await section('writing', '模型出字：只给计数和思考的最后一句，不进 events', async () => {
|
||
for (const reducedMotion of ['no-preference', 'reduce']) {
|
||
const { page, ctx, errors } = await open(theme, { reducedMotion })
|
||
ctl.thinkGo = ctl.writingGo = false
|
||
ctl.writingAt = 0
|
||
await goto(page, 'c0writel')
|
||
await send(page, '一个正在写的问题')
|
||
const head = page.locator('[data-turn]').last()
|
||
if (reducedMotion === 'no-preference') {
|
||
const said = (text) => head.getByText(text, { exact: true }).first()
|
||
.waitFor({ timeout: 8000 }).then(() => true, () => false)
|
||
check('思考时头部写出最后一句完整的话', await said('正在思考：先看看 factory 里有哪些表'))
|
||
await page.waitForTimeout(400)
|
||
check('还在写的半句不上头部', !(await head.getByText('然后按月汇总出勤').count()))
|
||
await page.screenshot({ path: `${SHOTS}/thinking-${theme}.png` })
|
||
ctl.thinkGo = true
|
||
check('半句写完了才换上它', await said('正在思考：然后按月汇总出勤率，再和上个月比'))
|
||
}
|
||
ctl.thinkGo = true
|
||
ctl.writingGo = true
|
||
await until(() => ctl.writingAt > 0, 8000)
|
||
if (reducedMotion === 'reduce') {
|
||
// 关掉动效时一秒写回一次：半秒时还没出现，一秒多一点就有了
|
||
await page.waitForTimeout(Math.max(0, ctl.writingAt + 500 - Date.now()))
|
||
check('关掉动效时计数不逐帧跳（半秒时还没写回）', !(await head.getByText('已生成').count()))
|
||
}
|
||
const shown = await head.getByText('正在撰写 · 已生成 1,234 字').first()
|
||
.waitFor({ timeout: 3000 }).then(() => Date.now() - ctl.writingAt, () => -1)
|
||
check(reducedMotion === 'reduce' ? '关掉动效时一秒左右写回计数' : '出字时头部写「正在撰写 · 已生成 1,234 字」',
|
||
shown >= 0 && (reducedMotion === 'reduce' ? shown >= 800 : shown < 900), `${shown} ms`)
|
||
const t = (await chatState(page)).byConversation.c0writel.at(-1)
|
||
check('token 和思考增量都没进 events', !t.events.some((e) => e.type.startsWith('llm.token') || e.type === 'llm.thinking.delta'),
|
||
t.events.map((e) => e.type).join(','))
|
||
check('答案原文没有提前露出来', !(await head.innerText()).includes('字字字'))
|
||
await page.screenshot({ path: `${SHOTS}/writing-${reducedMotion === 'reduce' ? 'reduced-' : ''}${theme}.png` })
|
||
await page.getByRole('button', { name: '停止这一轮' }).click()
|
||
await page.waitForTimeout(200)
|
||
check('停下之后不再写「正在撰写」', !(await head.getByText('正在撰写').count()))
|
||
check('没有运行时报错', errors.length === 0, errors[0] ?? '')
|
||
await ctx.close()
|
||
}
|
||
})
|
||
|
||
await section('meta', '可信度信息刷新之后还在', async () => {
|
||
const { page, ctx, errors } = await open(theme)
|
||
await goto(page, 'c0metam')
|
||
await shows(page, '一线 96.2%')
|
||
await page.waitForTimeout(600)
|
||
const card = page.locator('[data-turn="tm1"]')
|
||
check('出具档位还在', await card.getByText('降档出具').count() > 0)
|
||
check('探索运行的标注还在', await card.getByText('探索运行 · 不进正式归档').count() > 0)
|
||
check('降档的原因还在', await card.getByText('叙述模板渲染为空').count() > 0)
|
||
check('耗时还在', (await card.innerText()).includes('12.3 s'))
|
||
check('这些都是从落库的 meta 读回来的，没有再去问运行', !log.runGets.includes('run-meta'), log.runGets.join(','))
|
||
await page.screenshot({ path: `${SHOTS}/meta-${theme}.png` })
|
||
check('没有运行时报错', errors.length === 0, errors[0] ?? '')
|
||
await ctx.close()
|
||
})
|
||
|
||
await section('list', '会话列表：分组、相对时间、完整标题、筛选、没打开过的也标状态', async () => {
|
||
resetDb()
|
||
const { page, ctx, errors } = await open(theme)
|
||
await goto(page, 'c0idleb')
|
||
const aside = page.locator('aside[aria-label="对话列表"]')
|
||
await aside.getByText('会话 B').first().waitFor({ timeout: 6000 })
|
||
const heads = await aside.locator('section').evaluateAll((els) => els.map((e) => e.getAttribute('aria-label')))
|
||
check('按今天 / 昨天 / 近 7 天 / 更早分组', heads.join(',') === '今天,昨天,近 7 天,更早', heads.join(','))
|
||
const rowA = aside.locator('div.group', { hasText: '会话 A' })
|
||
check('写相对时间', /分钟前|刚刚/.test(await rowA.innerText()), (await rowA.innerText()).replace(/\n/g, ' / '))
|
||
const tip = await aside.locator('div.group', { hasText: '会话 F' }).locator('button').first().getAttribute('title')
|
||
check('悬停能看到完整标题', !!tip && tip.startsWith('会话 F：取消过、失败过、不可用的历史'), tip ?? '')
|
||
// 这几条这次都没打开过，状态只能来自列表接口的 last_status
|
||
const rowText = async (t) => (await aside.locator('div.group', { hasText: t }).innerText()).replace(/\n/g, ' / ')
|
||
check('没打开过的：停在审批上的标「待审批」', (await rowText('会话 K')).includes('待审批'), await rowText('会话 K'))
|
||
check('没打开过的：失败的标「失败」', (await rowText('会话 J')).includes('失败'), await rowText('会话 J'))
|
||
check('没打开过的：服务重启挂起的标「已中断」', (await rowText('会话 I')).includes('已中断'), await rowText('会话 I'))
|
||
check('没打开过的：在跑的标「运行中」', (await rowText('会话 L')).includes('运行中'), await rowText('会话 L'))
|
||
check('正常完成的不标', !/待审批|失败|已中断|运行中/.test(await rowText('会话 M')), await rowText('会话 M'))
|
||
const rowL = aside.locator('div.group', { hasText: '会话 L' })
|
||
await rowL.hover()
|
||
check('列表说它在跑的，也不能删', await rowL.getByRole('button', { name: /删除/ }).isDisabled())
|
||
const box = aside.getByRole('searchbox', { name: '按标题或提问内容搜索对话' })
|
||
await box.fill('长会话')
|
||
const hits = await aside.locator('div.group').allInnerTexts()
|
||
check('筛选只留下匹配的对话', hits.length === 1 && hits[0].includes('会话 C'), hits.map((h) => h.split('\n')[0]).join(' | '))
|
||
await page.screenshot({ path: `${SHOTS}/list-filter-${theme}.png` })
|
||
await box.press('Escape')
|
||
check('Esc 清掉筛选', (await aside.locator('div.group').count()) === Object.keys(CONVS).length)
|
||
await box.fill('没有这样的对话')
|
||
check('没有匹配时说一声', await aside.getByText('没有标题或问题中包含「没有这样的对话」的对话').count() === 1)
|
||
await box.fill('')
|
||
await page.screenshot({ path: `${SHOTS}/list-status-${theme}.png` })
|
||
check('没有运行时报错', errors.length === 0, errors[0] ?? '')
|
||
await ctx.close()
|
||
})
|
||
|
||
await section('filter', '筛选框：有筛选词时不跟着列表变短一起消失', async () => {
|
||
resetDb()
|
||
ctl.few = true
|
||
const { page, ctx, errors } = await open(theme)
|
||
await goto(page, 'c0busya')
|
||
const aside = page.locator('aside[aria-label="对话列表"]')
|
||
const box = aside.getByRole('searchbox', { name: '按标题或提问内容搜索对话' })
|
||
check('8 个对话时有筛选框', await box.waitFor({ timeout: 6000 }).then(() => true, () => false))
|
||
await box.fill('会话 B')
|
||
const row = aside.locator('div.group', { hasText: '会话 B' })
|
||
await row.hover()
|
||
await row.getByRole('button', { name: /删除/ }).click()
|
||
// 老版本删除前还要确认一次
|
||
if (await page.getByRole('dialog').waitFor({ timeout: 800 }).then(() => true, () => false)) {
|
||
await page.getByRole('dialog').getByRole('button', { name: '删除' }).click()
|
||
}
|
||
await until(() => db.archived.has('c0idleb'))
|
||
await page.waitForTimeout(300)
|
||
check('删到 7 个：筛选词还在，筛选框也还在', await box.isVisible().catch(() => false)
|
||
&& (await box.inputValue().catch(() => '')) === '会话 B')
|
||
await page.screenshot({ path: `${SHOTS}/filter-shrunk-${theme}.png` })
|
||
// 老版本框已经没了，清不掉：不在这里卡 30 秒，下面几条照样判红
|
||
await box.fill('', { timeout: 2000 }).catch(() => {})
|
||
await page.waitForTimeout(200)
|
||
check('清掉筛选词：剩下 7 个都回来了', (await aside.locator('div.group').count()) === 7,
|
||
String(await aside.locator('div.group').count()))
|
||
check('没有筛选词、不够 8 个时，筛选框收起', !(await box.isVisible().catch(() => false)))
|
||
check('没有运行时报错', errors.length === 0, errors[0] ?? '')
|
||
ctl.few = false
|
||
resetDb()
|
||
await ctx.close()
|
||
})
|
||
|
||
await section('delete', '删除：先进回收站，可以撤销', async () => {
|
||
resetDb()
|
||
const { page, ctx, errors } = await open(theme)
|
||
await goto(page, 'c0idleb')
|
||
await shows(page, 'B 的答案')
|
||
const aside = page.locator('aside')
|
||
const trashBtn = aside.getByRole('button', { name: /^打开回收站/ })
|
||
check('回收站入口写着里面有几个', /2/.test(await trashBtn.innerText().catch(() => '')),
|
||
await trashBtn.innerText().catch(() => '没有入口'))
|
||
const row = page.locator('aside[aria-label="对话列表"] div.group', { hasText: '会话 B' })
|
||
await row.hover()
|
||
await row.getByRole('button', { name: /删除/ }).click()
|
||
// 回收站能找回来：删除不再先弹一次确认，撤销和回收站就是后悔药
|
||
const asked = await page.getByRole('dialog').waitFor({ timeout: 800 }).then(() => true, () => false)
|
||
check('删除不再先弹确认', !asked)
|
||
if (asked) await page.getByRole('dialog').getByRole('button', { name: '删除' }).click()
|
||
await until(() => log.archived.some(([id, b]) => id === 'c0idleb' && b.archived === true))
|
||
check('删除是归档，不是真删', log.archived.some(([id, b]) => id === 'c0idleb' && b.archived === true)
|
||
&& !log.deleted.includes('c0idleb'))
|
||
const undo = page.getByRole('button', { name: '撤销' }).first()
|
||
check('给了撤销', await undo.waitFor({ timeout: 4000 }).then(() => true, () => false))
|
||
check('提示里说去了回收站', await page.getByText(/已移到回收站/).count() > 0)
|
||
check('回收站的数跟着变成 3', /3/.test(await trashBtn.innerText().catch(() => '')), await trashBtn.innerText().catch(() => ''))
|
||
await undo.click().catch(() => {})
|
||
await until(() => log.archived.some(([id, b]) => id === 'c0idleb' && b.archived === false))
|
||
await page.waitForTimeout(200)
|
||
check('撤销放回去了', log.archived.some(([id, b]) => id === 'c0idleb' && b.archived === false)
|
||
&& await page.locator('aside[aria-label="对话列表"]').getByText('会话 B').count() > 0)
|
||
check('没有运行时报错', errors.length === 0, errors[0] ?? '')
|
||
await ctx.close()
|
||
})
|
||
|
||
await section('trash', '回收站：预览、恢复、彻底删除（要确认）', async () => {
|
||
resetDb()
|
||
const deletedBefore = log.deleted.length
|
||
const { page, ctx, errors } = await open(theme)
|
||
await goto(page, 'c0idleb')
|
||
await shows(page, 'B 的答案')
|
||
await page.locator('aside').getByRole('button', { name: /^打开回收站/ }).click()
|
||
const bin = page.locator('aside[aria-label="回收站"]')
|
||
check('打开回收站', await bin.waitFor({ timeout: 3000 }).then(() => true, () => false))
|
||
const rows = bin.locator('div.group')
|
||
check('列出回收站里的对话', (await rows.count()) === 2, (await rows.allInnerTexts()).map((t) => t.split('\n')[0]).join(' | '))
|
||
check('顶行和页头一样高（48px）', await bin.evaluate((el) => Math.round(el.firstElementChild.getBoundingClientRect().height)) === 48)
|
||
await page.screenshot({ path: `${SHOTS}/trash-${theme}.png` })
|
||
|
||
// 点开看一眼：内容照常显示，但不能接着问
|
||
await rows.filter({ hasText: '删掉的对话甲' }).locator('button').first().click()
|
||
await page.waitForURL(/\/chat\/c0trash1$/, { timeout: 5000 }).catch(() => {})
|
||
check('能打开回收站里的对话看内容', await shows(page, '回收站里的答案'))
|
||
check('说清它在回收站里', await shows(page, '该对话在回收站中'))
|
||
await page.getByRole('textbox', { name: '向数据提问' }).fill('回收站里还能问吗')
|
||
check('在回收站里不能接着问', await page.getByRole('button', { name: '发送' }).isDisabled())
|
||
await page.waitForTimeout(800) // 轮次有入场动画，等它落定再截
|
||
await page.screenshot({ path: `${SHOTS}/trash-preview-${theme}.png` })
|
||
await page.locator('[data-trash-banner]').getByRole('button', { name: '恢复' }).click()
|
||
await until(() => log.archived.some(([id, b]) => id === 'c0trash1' && b.archived === false))
|
||
await page.waitForTimeout(300)
|
||
check('恢复：放回列表，提示条收起，可以接着问', !(await page.locator('[data-trash-banner]').count())
|
||
&& !(await page.getByRole('button', { name: '发送' }).isDisabled()))
|
||
check('恢复之后回到对话列表，它在里面', await page.locator('aside[aria-label="对话列表"]')
|
||
.getByText('删掉的对话甲').count() > 0)
|
||
|
||
// 彻底删除：要确认，确认后才发 DELETE
|
||
await page.locator('aside').getByRole('button', { name: /^打开回收站/ }).click()
|
||
const left = bin.locator('div.group', { hasText: '删掉的对话乙' })
|
||
await left.hover()
|
||
await left.getByRole('button', { name: /彻底删除/ }).click()
|
||
const dialog = page.getByRole('dialog')
|
||
check('彻底删除前先确认', await dialog.waitFor({ timeout: 3000 }).then(() => true, () => false))
|
||
check('确认框说清无法恢复', (await dialog.innerText().catch(() => '')).includes('无法恢复'))
|
||
await page.waitForTimeout(300)
|
||
await page.screenshot({ path: `${SHOTS}/trash-purge-${theme}.png` })
|
||
await dialog.getByRole('button', { name: '彻底删除' }).click()
|
||
await until(() => log.deleted.length > deletedBefore)
|
||
check('确认后才真删', log.deleted.slice(deletedBefore).join() === 'c0trash2', log.deleted.slice(deletedBefore).join())
|
||
check('回收站空了就说一声', await shows(page, '回收站为空'))
|
||
await bin.getByRole('button', { name: /返回/ }).click()
|
||
check('返回对话列表', await page.locator('aside[aria-label="对话列表"]').waitFor({ timeout: 3000 }).then(() => true, () => false))
|
||
check('没有运行时报错', errors.length === 0, errors[0] ?? '')
|
||
resetDb()
|
||
await ctx.close()
|
||
})
|
||
|
||
await section('readonly', '回收站里只能看：不发起运行，正在跑的不给彻底删除', async () => {
|
||
resetDb()
|
||
ctl.trashMore = true
|
||
ctl.approvals = [{ id: 'ap-bin', run_id: 'run-trash-wait', node_id: 'ag', mode: 'approve', title: '确认一下再查',
|
||
payload: {}, status: 'pending', response: {}, created_at: new Date().toISOString() }]
|
||
const { page, ctx, errors } = await open(theme)
|
||
const counts = () => ({ starts: log.runStarts.length, continues: log.continues.length, gens: log.gens.length, posts: log.posts.length })
|
||
const before = counts()
|
||
await goto(page, 'c0trash3')
|
||
await page.locator('[data-trash-banner]').waitFor({ timeout: 6000 }).catch(() => {})
|
||
// 停在审批上的那一轮要核对完、成了「等待审批」，四种补救才都摆出来
|
||
await page.waitForFunction(() => window.__chat.getState().byConversation.c0trash3?.find((t) => t.id === 'ty1')?.phase === 'waiting',
|
||
null, { timeout: 6000 }).catch(() => {})
|
||
await page.waitForTimeout(400)
|
||
const STARTERS = /继续运行|重试本轮|重新运行本轮|重新问一次|不限数据源重试|放宽步数|^\s*运行\s*$/
|
||
// 补救动作排在轮次卡片外面（AssistantStream 的动作槽），整页找
|
||
const starters = page.locator('button', { hasText: STARTERS })
|
||
const labels = await starters.evaluateAll((bs) => bs.map((b) => `${b.textContent.trim()}${b.disabled ? '' : '（可点）'}`))
|
||
const has = (re) => labels.some((l) => re.test(l))
|
||
check('会发起运行的按钮照样摆着（恢复之后能做什么看得到）',
|
||
has(/^运行$/) && has(/继续运行/) && has(/重新问一次|重试本轮/) && has(/放宽步数/), labels.join(' | '))
|
||
check('……但一个都点不了', labels.length > 0 && labels.every((l) => !l.endsWith('（可点）')), labels.join(' | '))
|
||
const why = await starters.evaluateAll((bs) => [...new Set(bs.map((b) => b.title))])
|
||
check('按钮上说清为什么点不了', why.length === 1 && /回收站/.test(why[0]), why.join(' | '))
|
||
check('换个说法照常能用（它不发起运行）',
|
||
await page.locator('[data-remedy]').getByRole('button', { name: '换个说法' }).first().isEnabled().catch(() => false))
|
||
check('查看执行过程照常能用', await page.getByRole('button', { name: '查看执行过程' }).first().isEnabled().catch(() => false))
|
||
check('停在审批上的：不摆能点「批准」的审批卡，说清恢复之后再处理',
|
||
!(await page.getByRole('button', { name: '批准' }).count())
|
||
&& (await page.locator('[data-pending-approval]').innerText().catch(() => '')).includes('恢复对话后'))
|
||
// 按钮之外再绕一道：绕过禁用直接点、直接调 store，也一次运行都发不出去
|
||
await starters.evaluateAll((bs) => bs.forEach((b) => b.click()))
|
||
await page.evaluate(() => {
|
||
const s = window.__chat.getState()
|
||
s.retryTurn('c0trash3', 'ty3')
|
||
s.retryTurn('c0trash3', 'ty2', { rerun: true })
|
||
s.runNow('c0trash3', 'ty0')
|
||
void s.continueTurn('c0trash3', 'ty2')
|
||
s.ask('c0trash3', '回收站里偷偷问一句')
|
||
})
|
||
await page.waitForTimeout(900)
|
||
const after = counts()
|
||
check('一次运行都没发起：不建图、不开轮、不 POST /runs、不继续运行',
|
||
JSON.stringify(after) === JSON.stringify(before), `${JSON.stringify(before)} → ${JSON.stringify(after)}`)
|
||
await page.locator('[data-turn="ty2"]').scrollIntoViewIfNeeded().catch(() => {})
|
||
await page.waitForTimeout(300)
|
||
await page.screenshot({ path: `${SHOTS}/trash-readonly-${theme}.png` })
|
||
|
||
// 删掉之后又跑起来的：和列表里一样不给彻底删除
|
||
await page.locator('aside').getByRole('button', { name: /^打开回收站/ }).click()
|
||
const bin = page.locator('aside[aria-label="回收站"]')
|
||
const liveRow = bin.locator('div.group', { hasText: '删掉的对话丁' })
|
||
await liveRow.waitFor({ timeout: 3000 }).catch(() => {})
|
||
check('回收站里标出它在运行', (await liveRow.innerText().catch(() => '')).includes('运行中'),
|
||
(await liveRow.innerText().catch(() => '')).replace(/\n/g, ' / '))
|
||
await liveRow.hover()
|
||
const rowPurge = liveRow.getByRole('button', { name: /彻底删除/ })
|
||
check('正在跑的：行上的彻底删除点不了，说清为什么', await rowPurge.isDisabled().catch(() => false)
|
||
&& (await rowPurge.getAttribute('title')) === '该对话正在运行，请先停止再删除', await rowPurge.getAttribute('title').catch(() => ''))
|
||
await liveRow.locator('button').first().click()
|
||
const live = await page.waitForFunction(() => window.__chat.getState().byConversation.c0trash4?.[0]?.phase === 'running',
|
||
null, { timeout: 6000 }).then(() => true, () => false)
|
||
check('打开它：核对发现运行还活着，接了回去', live, (await chatState(page)).byConversation.c0trash4?.[0]?.phase)
|
||
const bannerPurge = page.locator('[data-trash-banner]').getByRole('button', { name: '彻底删除' })
|
||
check('提示条上的彻底删除也点不了', await bannerPurge.isDisabled().catch(() => false))
|
||
const stop = page.getByRole('button', { name: '停止这一轮' })
|
||
check('在回收站里也停得下来', await stop.count() === 1)
|
||
await page.waitForTimeout(300)
|
||
await page.screenshot({ path: `${SHOTS}/trash-running-${theme}.png` })
|
||
|
||
// 清空：只删没在跑的；在跑的留下，确认框里先说
|
||
const deletedBefore = log.deleted.length
|
||
await bin.getByRole('button', { name: '清空回收站' }).click()
|
||
const dialog = page.getByRole('dialog')
|
||
await dialog.waitFor({ timeout: 3000 }).catch(() => {})
|
||
const said = await dialog.innerText().catch(() => '')
|
||
check('清空前说清在跑的这次不删', said.includes('删掉的对话丁') && said.includes('正在运行'), said.replace(/\n/g, ' / '))
|
||
await dialog.getByRole('button', { name: '全部删除' }).click().catch(() => {})
|
||
await until(() => log.deleted.length - deletedBefore >= 3, 4000)
|
||
await page.waitForTimeout(300)
|
||
const gone = log.deleted.slice(deletedBefore)
|
||
check('清空：没在跑的都删了，在跑的没删', ['c0trash1', 'c0trash2', 'c0trash3'].every((id) => gone.includes(id))
|
||
&& !gone.includes('c0trash4'), gone.join(','))
|
||
check('在跑的那条事件流没被关掉、运行没被丢下', ctl.wsClosed['run-trash-live'] === false
|
||
&& (await chatState(page)).byConversation.c0trash4?.[0]?.phase === 'running')
|
||
check('它还在回收站里', await bin.locator('div.group', { hasText: '删掉的对话丁' }).count() === 1)
|
||
|
||
// 停下来之后就能删了
|
||
await stop.click().catch(() => {})
|
||
await until(() => log.cancels.includes('run-trash-live'), 3000)
|
||
await page.waitForTimeout(200)
|
||
check('停下之后：彻底删除可以点了', await bannerPurge.isEnabled().catch(() => false))
|
||
await bannerPurge.click().catch(() => {})
|
||
await dialog.waitFor({ timeout: 3000 }).catch(() => {})
|
||
await dialog.getByRole('button', { name: '彻底删除' }).click().catch(() => {})
|
||
await until(() => log.deleted.slice(deletedBefore).includes('c0trash4'), 3000)
|
||
check('确认之后删掉了', log.deleted.slice(deletedBefore).includes('c0trash4'))
|
||
check('没有运行时报错', errors.length === 0, errors[0] ?? '')
|
||
ctl.trashMore = false
|
||
ctl.approvals = []
|
||
resetDb()
|
||
await ctx.close()
|
||
})
|
||
|
||
await section('recheck', '恢复的半路轮次：核对撞上 500 有出路、补交付不计实时、终态落库', async () => {
|
||
const { page, ctx, errors } = await open(theme)
|
||
ctl.run500 = true
|
||
await goto(page, 'c0unkp')
|
||
// 先等这一轮取回来：取回来之前「不在核对」是空洞地成立的
|
||
const settled = await page.waitForFunction(() => {
|
||
const t = window.__chat.getState().byConversation.c0unkp?.[0]
|
||
return !!t && t.hydrated !== 'pending' && t.phase !== 'checking'
|
||
}, null, { timeout: 5000 }).then(() => true, () => false)
|
||
check('核对撞上 500：不再一直停在「正在核对」', settled, (await chatState(page)).byConversation.c0unkp?.[0]?.phase)
|
||
check('说清没核对上', await shows(page, '无法确认本轮的当前状态'))
|
||
const recheck = page.locator('[data-remedy]').getByRole('button', { name: '重新核对' })
|
||
check('给「重新核对」', await recheck.count() === 1)
|
||
check('不给会另起运行的按钮', !(await page.getByRole('button', { name: /重试本轮|重新运行本轮|继续运行/ }).count()))
|
||
check('这时不转圈', !(await page.locator('[data-turn="tp1"] .animate-spin').count()))
|
||
await page.screenshot({ path: `${SHOTS}/recheck-failed-${theme}.png` })
|
||
ctl.run500 = false
|
||
await recheck.click().catch(() => {})
|
||
check('重新核对：按运行的真实状态收尾', await page.waitForFunction(
|
||
() => window.__chat.getState().byConversation.c0unkp?.[0]?.phase === 'cancelled', null, { timeout: 5000 })
|
||
.then(() => true, () => false), (await chatState(page)).byConversation.c0unkp?.[0]?.phase)
|
||
ctl.run500 = true
|
||
|
||
// 三天前跑完、交付没赶上：补交付（等复核）期间不按实时运行计时
|
||
ctl.reviewDelay['run-late'] = 1800
|
||
await goto(page, 'c0lateq')
|
||
await page.waitForFunction(() => window.__chat.getState().byConversation.c0lateq?.[0]?.hydrated === 'done', null, { timeout: 5000 }).catch(() => {})
|
||
await page.waitForTimeout(300)
|
||
const mid = (await chatState(page)).byConversation.c0lateq?.[0]
|
||
check('补交付期间不算「运行中」', mid?.phase !== 'running' && mid?.phase !== 'done', mid?.phase)
|
||
const clock = await page.locator('[data-turn="tq1"] [title="已运行"]').allInnerTexts()
|
||
check('补交付期间头部不走实时计时（不会从三天前算起）', clock.length === 0, clock.join(','))
|
||
await page.screenshot({ path: `${SHOTS}/late-delivery-${theme}.png` })
|
||
check('补交付完成', await shows(page, '补交付的答案：三条记录。', 5000))
|
||
delete ctl.reviewDelay['run-late']
|
||
|
||
// 停在半路的几轮（这一波之前页面推的老轮次，服务端没有它们的任务）：页面不再自己收尾、不写库，请服务端接着等各自的
|
||
// 那次运行、照常收尾（POST /api/chat/turns mode=resume），结局照服务端写的那一行摆
|
||
ctl.approvals = [{ id: 'ap-t', run_id: 'run-t-wait', node_id: 'ag', mode: 'approve', title: '确认一下',
|
||
payload: {}, status: 'pending', response: {}, created_at: new Date().toISOString() }]
|
||
const patchesBefore = log.patches.length
|
||
const resumesBefore = log.turnStarts.length
|
||
await goto(page, 'c0termr')
|
||
const phaseOf = async (id) => (await chatState(page)).byConversation.c0termr?.find((t) => t.id === id)
|
||
for (let i = 0; i < 80; i++) {
|
||
const turns = (await chatState(page)).byConversation.c0termr ?? []
|
||
if (['tr1', 'tr2', 'tr3', 'tr4', 'tr5', 'tr6'].every((id) => ['cancelled', 'error', 'suspended', 'waiting']
|
||
.includes(turns.find((t) => t.id === id)?.phase))) break
|
||
await page.waitForTimeout(100)
|
||
}
|
||
const resumed = log.turnStarts.slice(resumesBefore).filter((b) => b.mode === 'resume').map((b) => b.retry_of)
|
||
check('停在半路的几轮都交给服务端接着收尾（mode=resume，带着各自这一轮）',
|
||
['tr1', 'tr2', 'tr3', 'tr4', 'tr5', 'tr6'].every((id) => resumed.includes(id))
|
||
&& log.turnStarts.slice(resumesBefore).every((b) => b.conversation_id === 'c0termr'), resumed.join(','))
|
||
check('页面不写这几轮的结局（不 PATCH 轮次）', log.patches.slice(patchesBefore).length === 0,
|
||
JSON.stringify(log.patches.slice(patchesBefore).map((p) => p.turn)))
|
||
const tr = Object.fromEntries(await Promise.all(['tr1', 'tr2', 'tr3', 'tr4', 'tr5', 'tr6'].map(async (id) => [id, await phaseOf(id)])))
|
||
check('在别处取消的：收成已取消', tr.tr1?.phase === 'cancelled', tr.tr1?.phase)
|
||
check('在别处失败的：收成失败，原因说人话', tr.tr2?.phase === 'error' && /鉴权/.test(tr.tr2?.error ?? ''), tr.tr2?.error)
|
||
check('服务重启挂起的：收成中断', tr.tr3?.phase === 'suspended', tr.tr3?.phase)
|
||
check('运行被删掉的：收成「上次未完成」', tr.tr4?.phase === 'error' && tr.tr4?.failure?.title === '上次未完成',
|
||
JSON.stringify(tr.tr4?.failure))
|
||
check('……后端现在的说法（鉴权失败（401））：同样认成鉴权失败', tr.tr6?.phase === 'error'
|
||
&& (tr.tr6?.error ?? '').startsWith('模型鉴权失败'), tr.tr6?.error)
|
||
check('还在等审批的：照旧等人（它还没结束）', tr.tr5?.phase === 'waiting', tr.tr5?.phase)
|
||
ctl.approvals = []
|
||
check('没有运行时报错', errors.length === 0, errors[0] ?? '')
|
||
await ctx.close()
|
||
})
|
||
|
||
await section('clipped', '长答案没取全：记进 meta、刷新后补全，2000 字按码点数', async () => {
|
||
const { page, ctx, errors } = await open(theme)
|
||
const patchesBefore = log.patches.length
|
||
await goto(page, 'c0clips')
|
||
await send(page, '一份取不到运行的长答案')
|
||
await page.waitForFunction(() => window.__chat.getState().byConversation.c0clips?.at(-1)?.phase === 'done', null, { timeout: 10000 }).catch(() => {})
|
||
const t = (await chatState(page)).byConversation.c0clips.at(-1)
|
||
// 一轮在服务端跑：交付读的是库里的完整成果，页面 GET 运行失败也不影响，新轮次不会再有「没取全」
|
||
check('服务端按库里的完整成果交付：答案完整、不标「没取全」', !t.clipped && String(t.output?.answer ?? '').length === FULL.length,
|
||
`${t.clipped} · ${String(t.output?.answer ?? '').length} / ${FULL.length}`)
|
||
check('页面不写这一轮的结局（不 PATCH 轮次）', log.patches.slice(patchesBefore).length === 0)
|
||
|
||
const runGetsBefore = log.runGets.length
|
||
await goto(page, 'c0clipt')
|
||
await shows(page, '多个成果键的长答案')
|
||
await until(() => log.runGets.slice(runGetsBefore).includes('run-multi'), 4000)
|
||
check('meta 说没取全：一进来就去补，不等滚进视口', log.runGets.slice(runGetsBefore).includes('run-multi'),
|
||
log.runGets.slice(runGetsBefore).join(','))
|
||
await page.waitForFunction(() => String(window.__chat.getState().byConversation.c0clipt?.[0]?.output?.answer ?? '').includes('【完整结尾】'),
|
||
null, { timeout: 4000 }).catch(() => {})
|
||
const tt1 = (await chatState(page)).byConversation.c0clipt[0]
|
||
check('多个成果键的长答案补全了', String(tt1.output?.answer ?? '').includes('【完整结尾】') && !tt1.clipped,
|
||
`${String(tt1.output?.answer ?? '').length} 字 · ${tt1.clipped}`)
|
||
await page.waitForTimeout(300)
|
||
const lost = await page.locator('[data-remedy]', { hasText: '截断为 2000 字' }).count()
|
||
check('恰好 2000 个码点（UTF-16 不是 2000）也认得出被截断、补不回来', lost === 2, `${lost} 处`)
|
||
await page.locator('[data-turn="tt2"]').scrollIntoViewIfNeeded().catch(() => {})
|
||
await page.screenshot({ path: `${SHOTS}/clipped-${theme}.png` })
|
||
check('没有运行时报错', errors.length === 0, errors[0] ?? '')
|
||
await ctx.close()
|
||
})
|
||
|
||
await section('launch', '发起那一下没到服务端：报错分得清标题和原因，补救说真话', async () => {
|
||
// 一轮在服务端跑之后，发起运行是服务端的事，「发起运行的请求没到后端」不会再发生；会断的是发起这一轮的那一下
|
||
//（POST /api/chat/turns）。这时服务端那边还没有这一轮，补救就是原样再问一次
|
||
const { page, ctx, errors } = await open(theme)
|
||
ctl.startFail = true
|
||
ctl.launchFail = false
|
||
await goto(page, 'c0launu')
|
||
await send(page, '一个启动失败的问题')
|
||
const alert = page.locator('[data-turn] [role="alert"]').first()
|
||
await alert.waitFor({ timeout: 8000 }).catch(() => {})
|
||
const title = await alert.locator('.font-medium').first().innerText().catch(() => '')
|
||
check('标题只是「无法连接服务」，原因不糊进粗体', title === '无法连接服务', title)
|
||
const text = await alert.innerText().catch(() => '')
|
||
check('不许诺页面上不存在的自动重试', !text.includes('自动重试'), text.replace(/\n/g, ' / '))
|
||
const retry = page.locator('[data-remedy="failed"]').getByRole('button', { name: '重试本轮' })
|
||
check('给「重试本轮」', await retry.count() === 1)
|
||
check('说出怎么办：连上之后点「重试本轮」', text.includes('重试本轮'), text.replace(/\n/g, ' / '))
|
||
await page.screenshot({ path: `${SHOTS}/launch-failed-${theme}.png` })
|
||
ctl.startFail = false
|
||
const gensBefore = log.gens.length
|
||
const startsBefore = log.turnStarts.length
|
||
await retry.click().catch(() => {})
|
||
await until(() => log.gens.length > gensBefore, 5000)
|
||
const again = log.turnStarts.slice(startsBefore).at(-1)
|
||
check('重试：服务端还没有这一轮，原样再问一次（同一个问题，不带 retry_of）', again?.question === '一个启动失败的问题'
|
||
&& !again?.retry_of && again?.conversation_id === 'c0launu', JSON.stringify(again))
|
||
check('这一次照常建图', log.gens.length > gensBefore, `${log.gens.length - gensBefore} 次建图`)
|
||
await page.getByRole('button', { name: '停止这一轮' }).click().catch(() => {})
|
||
ctl.launchFail = true
|
||
check('没有运行时报错', errors.length === 0, errors[0] ?? '')
|
||
await ctx.close()
|
||
})
|
||
|
||
await section('toolmissing', '发起就被拒（绑定的工具在本机不存在）：报错里给直达入口，不给接着跑', async () => {
|
||
const { page, ctx, errors } = await open(theme)
|
||
await goto(page, 'c0launu')
|
||
await send(page, '一个工具不在的问题')
|
||
const alert = page.locator('[data-turn] [data-turn-error]').first()
|
||
await alert.waitFor({ timeout: 8000 }).catch(() => {})
|
||
const text = await alert.innerText().catch(() => '')
|
||
check('标题点名那个工具', text.includes('绑定的工具「db_query__nope」不存在'), text.replace(/\n/g, ' / '))
|
||
check('说清这次运行没有发起、怎么办', text.includes('本次运行未启动') && text.includes('「数据」页接入'), text.replace(/\n/g, ' / '))
|
||
const fix = alert.locator('a[data-fix="data"]')
|
||
check('报错里有直达入口「前往「数据」页接入」（和画布上发起被拒同一套）', await fix.count() === 1
|
||
&& (await fix.innerText().catch(() => '')).includes('前往「数据」页接入'))
|
||
check('不给「继续运行」（原样继续运行还是缺这个工具）', await page.getByRole('button', { name: '继续运行' }).count() === 0)
|
||
// 这时还没有运行 id，流里的报错块不会再拿原话认一遍：入口得在 store 里就记进这一轮的 failure，
|
||
// 刷新之后（meta 里的 failure）照样有
|
||
const stored = await page.evaluate(() => (window.__chat.getState().byConversation.c0launu ?? []).at(-1)?.failure ?? null)
|
||
check('入口记进这一轮的 failure（fix / fixTo，不可继续运行）', stored?.fix === 'tools' && stored?.fixTo === '/data'
|
||
&& stored?.continuable === false, JSON.stringify(stored))
|
||
await fix.click().catch(() => {})
|
||
await page.waitForURL((u) => u.pathname.startsWith('/data'), { timeout: 5000 }).catch(() => {})
|
||
check('点了去数据页', new URL(page.url()).pathname.startsWith('/data'), page.url())
|
||
check('没有运行时报错', errors.length === 0, errors[0] ?? '')
|
||
await ctx.close()
|
||
})
|
||
|
||
await section('guard', '放宽步数只在对症时给：停滞、预算收的尾不给，跟随默认 100 步的也不给', async () => {
|
||
resetDb()
|
||
ctl.guardConvs = true
|
||
try {
|
||
const stepBtn = (page) => page.getByRole('button', { name: /放宽步数后重新运行/ })
|
||
// 新后端：默认 100 步、硬上限 100
|
||
const fresh = { run: { agent_max_steps: 100 }, limits: { max_agent_steps: 100 } }
|
||
const { page, ctx, errors } = await open(theme, { settings: fresh })
|
||
await goto(page, 'c0gstall')
|
||
await shows(page, '只查了一部分的结论')
|
||
check('连续几步没进展收的尾：不给「放宽步数」（加步数只会原样再撞一次）', await stepBtn(page).count() === 0)
|
||
await goto(page, 'c0gsteps')
|
||
await shows(page, '只查了一部分的结论')
|
||
const eight = stepBtn(page)
|
||
check('节点写死 8 步、步数用满：给「放宽步数」，从 8 放到 16', await eight.count() === 1
|
||
&& /8 → 16 步/.test(await eight.innerText()), await eight.innerText().catch(() => '（没有按钮）'))
|
||
await goto(page, 'c0gdflt')
|
||
await shows(page, '只查了一部分的结论')
|
||
check('节点上的 12 跟随默认 100 步、已经到硬上限：不许一个放不宽的数', await stepBtn(page).count() === 0)
|
||
check('没有运行时报错', errors.length === 0, errors[0] ?? '')
|
||
await ctx.close()
|
||
} finally {
|
||
ctl.guardConvs = false
|
||
}
|
||
})
|
||
|
||
await section('rerun', '重跑服务重启挂起的那一轮：旧运行顺手取消', async () => {
|
||
const { page, ctx, errors } = await open(theme)
|
||
await goto(page, 'c0suspo')
|
||
await shows(page, '服务重启，本轮已中断')
|
||
const before = log.cancels.length
|
||
const startsBefore = log.runStarts.length
|
||
await page.locator('[data-remedy="suspended"]').getByRole('button', { name: '重新运行本轮' }).click()
|
||
await until(() => log.runStarts.length > startsBefore, 5000)
|
||
await until(() => log.cancels.slice(before).includes('run-susp'), 2000)
|
||
check('挂起的旧运行被取消，不会一直挂在记录里', log.cancels.slice(before).includes('run-susp'), log.cancels.slice(before).join(','))
|
||
const rerun = log.turnStarts.filter((b) => b.retry_of === 'to1').at(-1)
|
||
check('重新运行本轮的请求：带着这一轮、mode=rerun', rerun?.mode === 'rerun' && rerun?.conversation_id === 'c0suspo',
|
||
JSON.stringify(rerun))
|
||
check('新的一次照常发起', log.runStarts.length > startsBefore)
|
||
await page.getByRole('button', { name: '停止这一轮' }).click().catch(() => {})
|
||
check('没有运行时报错', errors.length === 0, errors[0] ?? '')
|
||
await ctx.close()
|
||
})
|
||
|
||
await section('stale', '断在建图阶段的老轮次：页面不替它写库；重试本轮交给服务端', async () => {
|
||
// 这一波之前页面推的老轮次停在 running、又没有运行：建图那条流随页面、随上次的进程一起没了。以前页面按「多久没动静」
|
||
// 替它写回 error；现在服务端启动时整理它们（chat/job.startup），页面只照实摆「上次未完成」，不写库。
|
||
// 服务端还在建图的一轮带着任务（ConversationTurn.job），照它在跑的样子接着播（见「刷新后接着看」那一段）
|
||
const { page, ctx, errors } = await open(theme)
|
||
const patchesBefore = log.patches.length
|
||
await goto(page, 'c0buildx')
|
||
await shows(page, '重试之后二十分钟没动静的问题')
|
||
await page.waitForTimeout(600)
|
||
const turns = (await chatState(page)).byConversation.c0buildx ?? []
|
||
check('页面不替老轮次写库（不 PATCH）', log.patches.slice(patchesBefore).length === 0,
|
||
JSON.stringify(log.patches.slice(patchesBefore).map((p) => p.turn)))
|
||
check('断在建图阶段的老轮次：照实说上次未完成', turns.every((t) => t.phase === 'error' && t.failure?.title === '上次未完成'),
|
||
turns.map((t) => `${t.id}:${t.phase}`).join(','))
|
||
|
||
// 重试一次：交给服务端（带着这一轮、mode=retry），这一次的运行挂在这一轮上
|
||
const startsBefore = log.runStarts.length
|
||
const turnStarts = log.turnStarts.length
|
||
await page.locator('[data-remedy="failed"]').last().getByRole('button', { name: '重试本轮' }).click()
|
||
await until(() => log.runStarts.length > startsBefore, 6000)
|
||
const body = log.turnStarts.slice(turnStarts).at(-1)
|
||
check('重试本轮：请求带着这一轮和方式（retry_of、mode=retry）', body?.retry_of === 'tx8' && body?.mode === 'retry'
|
||
&& body?.conversation_id === 'c0buildx', JSON.stringify(body))
|
||
const newRun = log.runStarts.at(-1)?.id
|
||
await page.waitForFunction((id) => window.__chat.getState().byConversation.c0buildx?.find((t) => t.id === 'tx8')?.run?.id === id,
|
||
newRun, { timeout: 4000 }).catch(() => {})
|
||
const tx8 = (await chatState(page)).byConversation.c0buildx.find((t) => t.id === 'tx8')
|
||
check('服务端发起的这次运行挂在这一轮上，照常在跑', !!newRun && tx8?.run?.id === newRun && tx8?.phase === 'running',
|
||
`${tx8?.run?.id} · ${tx8?.phase}`)
|
||
check('页面照旧不写库', log.patches.slice(patchesBefore).length === 0)
|
||
await page.getByRole('button', { name: '停止这一轮' }).click().catch(() => {})
|
||
check('没有运行时报错', errors.length === 0, errors[0] ?? '')
|
||
await ctx.close()
|
||
})
|
||
|
||
await section('scope', '限定数据源：点一下只查这个库', async () => {
|
||
const { page, ctx, errors } = await open(theme)
|
||
await goto(page, 'c0emptg')
|
||
await shows(page, '问你的数据')
|
||
const chip = page.locator('[data-scope-chip="ds-plant"]')
|
||
check('首屏的数据源标签可以点', await chip.waitFor({ timeout: 5000 }).then(() => true, () => false))
|
||
check('默认不限', (await chip.getAttribute('aria-pressed')) === 'false')
|
||
await chip.click()
|
||
check('点一下就限定成它', (await chip.getAttribute('aria-pressed')) === 'true')
|
||
await page.waitForTimeout(300) // 颜色有过渡，等它换完再截
|
||
await page.screenshot({ path: `${SHOTS}/scope-hero-${theme}.png` })
|
||
const gensBefore = log.gens.length
|
||
await send(page, '只查这个库的问题')
|
||
await until(() => log.gens.length > gensBefore, 5000)
|
||
check('请求里带上 datasource_ids', JSON.stringify(log.gens.at(-1)?.datasource_ids) === '["ds-plant"]',
|
||
JSON.stringify(log.gens.at(-1)?.datasource_ids))
|
||
check('这一轮上写着只查了哪个库', await shows(page, '只查 factory'))
|
||
await page.getByRole('button', { name: '停止这一轮' }).click().catch(() => {})
|
||
await page.waitForTimeout(200)
|
||
|
||
// 在对话里：输入框左边的范围选择
|
||
await goto(page, 'c0scopv')
|
||
await shows(page, '上一问的答案')
|
||
const picker = page.getByRole('button', { name: /数据源范围/ })
|
||
check('输入框旁有数据源范围', await picker.count() === 1)
|
||
check('默认写「全部数据源」', (await picker.innerText().catch(() => '')).includes('全部数据源'), await picker.innerText().catch(() => ''))
|
||
await picker.click()
|
||
await page.getByRole('checkbox', { name: 'shop' }).check()
|
||
check('选了之后按钮写「只查 shop」', (await picker.innerText().catch(() => '')).includes('只查 shop'), await picker.innerText().catch(() => ''))
|
||
await page.screenshot({ path: `${SHOTS}/scope-picker-${theme}.png` })
|
||
await page.keyboard.press('Escape')
|
||
const before2 = log.gens.length
|
||
await send(page, '在对话里限定的问题')
|
||
await until(() => log.gens.length > before2, 5000)
|
||
check('对话里限定的也带上', JSON.stringify(log.gens.at(-1)?.datasource_ids) === '["ds-mig"]',
|
||
JSON.stringify(log.gens.at(-1)?.datasource_ids))
|
||
await page.getByRole('button', { name: '停止这一轮' }).click().catch(() => {})
|
||
await page.waitForTimeout(200)
|
||
|
||
// 认的是机读码，不是那句话的开头：后端换了说法，补救照样是放开范围（3C REQ-20）
|
||
ctl.scope400 = 'reworded'
|
||
await send(page, '限定的库换了说法')
|
||
await shows(page, '这一轮圈定的库一个也找不到了')
|
||
check('换了说法的 400：凭机读码认出来，补救是「不限数据源重试」',
|
||
await page.locator('[data-remedy="failed"]').last().getByRole('button', { name: '不限数据源重试' }).count() === 1)
|
||
await page.waitForTimeout(200)
|
||
|
||
// 限定的库被删了、停用了：后端 400，把原因原样说出来
|
||
ctl.scope400 = true
|
||
await send(page, '限定的库已经不在了')
|
||
check('库不在了：说清原因', await shows(page, '限定的数据源都已不可用'))
|
||
await page.screenshot({ path: `${SHOTS}/scope-400-${theme}.png` })
|
||
// 带着原范围的「重试本轮」只会再撞一次 400：补救得是放开范围
|
||
const loose = page.locator('[data-remedy="failed"]').last().getByRole('button', { name: '不限数据源重试' })
|
||
check('库不在了：补救是「不限数据源重试」', await loose.count() === 1)
|
||
const before3 = log.gens.length
|
||
await loose.click().catch(() => {})
|
||
await until(() => log.gens.length > before3, 5000)
|
||
check('不限数据源重试：请求里不再带 datasource_ids', log.gens.length > before3 && !log.gens.at(-1)?.datasource_ids?.length,
|
||
JSON.stringify(log.gens.at(-1)?.datasource_ids))
|
||
check('输入框旁的范围也一起放开', (await picker.innerText().catch(() => '')).includes('全部数据源'),
|
||
await picker.innerText().catch(() => ''))
|
||
await page.getByRole('button', { name: '停止这一轮' }).click().catch(() => {})
|
||
ctl.scope400 = false
|
||
|
||
// 库列表没取回来：从上一问接过来的范围照样会发出去，所以照样得摆在眼前
|
||
ctl.sourcesFail = true
|
||
await goto(page, 'c0scopw')
|
||
await shows(page, '上一问的答案 W')
|
||
const picker2 = page.getByRole('button', { name: /数据源范围/ })
|
||
await picker2.waitFor({ timeout: 3000 }).catch(() => {})
|
||
check('库列表没取回来：范围照样摆在输入框旁', (await picker2.innerText().catch(() => '')).includes('只查 factory'),
|
||
await picker2.innerText().catch(() => '没有范围按钮'))
|
||
await picker2.click().catch(() => {})
|
||
const stray = page.locator('[data-scope-stray]')
|
||
check('范围里的库照样列出来，注明状态获取失败', (await stray.innerText().catch(() => '')).includes('状态获取失败')
|
||
&& await stray.getByRole('checkbox').isChecked().catch(() => false), await stray.innerText().catch(() => ''))
|
||
await page.waitForTimeout(250)
|
||
await page.screenshot({ path: `${SHOTS}/scope-unconfirmed-${theme}.png` })
|
||
await stray.getByRole('checkbox').uncheck().catch(() => {})
|
||
check('取消勾选之后它还留在原处，按钮写回「全部数据源」', await stray.count() === 1
|
||
&& (await picker2.innerText().catch(() => '')).includes('全部数据源'), await picker2.innerText().catch(() => ''))
|
||
await stray.getByRole('checkbox').check().catch(() => {})
|
||
await page.keyboard.press('Escape')
|
||
const before4 = log.gens.length
|
||
await send(page, '库列表没取回来时的问题')
|
||
await until(() => log.gens.length > before4, 5000)
|
||
check('发出去的范围和看得见的一致', JSON.stringify(log.gens.at(-1)?.datasource_ids) === '["ds-plant"]',
|
||
JSON.stringify(log.gens.at(-1)?.datasource_ids))
|
||
await page.getByRole('button', { name: '停止这一轮' }).click().catch(() => {})
|
||
ctl.sourcesFail = false
|
||
check('没有运行时报错', errors.length === 0, errors[0] ?? '')
|
||
await ctx.close()
|
||
})
|
||
|
||
await section('opencanvas', '在画布里打开：画布有没保存的改动时，先问再建，整条路只问一遍', async () => {
|
||
const { page, ctx, errors } = await open(theme)
|
||
const creates = []
|
||
await page.route(/\/api\/workflows(\?.*)?$/, (route) => {
|
||
if (route.request().method() !== 'POST') return route.fallback()
|
||
creates.push(route.request().postDataJSON())
|
||
return route.fulfill({ status: 201, json: { id: 'wf-chat-open', name: '问数据：B 的问题', description: '', graph: GRAPH,
|
||
tags: [], version: 1, is_template: false, status: 'draft', published_version: null, published_by: null, run_count: 0,
|
||
created_at: iso(0), updated_at: iso(0) } })
|
||
})
|
||
await page.route(/\/api\/workflows\/wf-chat-open(\/.*)?(\?.*)?$/, (route) => {
|
||
if (route.request().method() !== 'GET') return route.abort()
|
||
const url = new URL(route.request().url())
|
||
if (url.pathname.endsWith('/versions')) return route.fulfill({ json: [] })
|
||
return route.fulfill({ json: { id: 'wf-chat-open', name: '问数据：B 的问题', description: '', graph: GRAPH, tags: [], version: 1,
|
||
is_template: false, status: 'draft', published_version: null, published_by: null, run_count: 0,
|
||
created_at: iso(0), updated_at: iso(0) } })
|
||
})
|
||
await goto(page, 'c0idleb')
|
||
await shows(page, 'B 的答案')
|
||
// 画布（store 里的那张）有没保存的改动：离开编排页它也还在
|
||
await page.evaluate(() => window.__studio.setState({ dirty: true,
|
||
workflow: { ...(window.__studio.getState().workflow ?? {}), id: 'wf-dirty', name: '没存的那张' } }))
|
||
await page.evaluate(() => {
|
||
window.__asks = 0
|
||
new MutationObserver(() => {
|
||
const d = [...document.querySelectorAll('[role=dialog]')].some((e) => e.textContent?.includes('未保存的改动'))
|
||
if (d && !window.__askOpen) window.__asks += 1
|
||
window.__askOpen = d
|
||
}).observe(document.body, { childList: true, subtree: true })
|
||
})
|
||
const openBtn = page.getByRole('button', { name: '在画布里打开' }).first()
|
||
const ask = page.getByRole('dialog').filter({ hasText: '未保存的改动' })
|
||
await openBtn.click()
|
||
const asked = await ask.waitFor({ timeout: 3000 }).then(() => true, () => false)
|
||
check('先问要不要放弃画布上的改动，还没建工作流', asked && creates.length === 0, `${asked ? '问了' : '没问'} · 建了 ${creates.length} 个`)
|
||
await ask.getByRole('button', { name: '取消' }).click().catch(() => {})
|
||
await page.waitForTimeout(300)
|
||
check('说了取消：不建，留在问数据页', creates.length === 0 && new URL(page.url()).pathname.startsWith('/chat'))
|
||
await openBtn.click()
|
||
await ask.waitFor({ timeout: 3000 }).catch(() => {})
|
||
await ask.getByRole('button', { name: '放弃并切换' }).click().catch(() => {})
|
||
await page.waitForURL((u) => u.pathname === '/studio/wf-chat-open', { timeout: 6000 }).catch(() => {})
|
||
await page.waitForTimeout(400)
|
||
const asks = await page.evaluate(() => window.__asks)
|
||
check('说了放弃：建一个、跳过去，整条路只问这一遍', creates.length === 1 && asks === 2
|
||
&& new URL(page.url()).pathname === '/studio/wf-chat-open', `建了 ${creates.length} 个 · 一共问了 ${asks} 遍（取消那次算一遍）`)
|
||
check('建的工作流带上来源：哪次运行、哪段对话、哪一轮（服务端核对）', JSON.stringify(creates[0]?.origin)
|
||
=== JSON.stringify({ run_id: 'run-b1', conversation_id: 'c0idleb', turn_id: 'tb1' }), JSON.stringify(creates[0]?.origin))
|
||
check('没有运行时报错', errors.length === 0, errors[0] ?? '')
|
||
await ctx.close()
|
||
})
|
||
|
||
await section('fresh', '数据源列表：旧了就在进页面时重取，别处加的库不用整页刷新', async () => {
|
||
const { page, ctx, errors } = await open(theme)
|
||
await goto(page, 'c0emptg')
|
||
await page.locator('[data-scope-chip="ds-plant"]').waitFor({ timeout: 5000 }).catch(() => {})
|
||
check('先看到的是启动时取回来的两个库', await page.locator('[data-scope-chip]').count() === 2)
|
||
// 别处加了一个库；这一页的列表已经是半分钟以前的了
|
||
ctl.extraSource = true
|
||
await page.evaluate(async () => {
|
||
const url = performance.getEntriesByType('resource').map((e) => e.name).find((n) => n.includes('/src/store/catalog.ts'))
|
||
const { useCatalog } = await import(url ?? '/src/store/catalog.ts')
|
||
useCatalog.setState((s) => ({ loadedAt: { ...s.loadedAt, datasources: Date.now() - 120_000 } }))
|
||
})
|
||
const getsBefore = log.sourceGets
|
||
// 站内跳走再回来：页面重新挂载，不整页刷新
|
||
await page.locator('nav[aria-label="主导航"] a[href="/runs"]').click()
|
||
await page.waitForURL((u) => u.pathname.startsWith('/runs'), { timeout: 5000 }).catch(() => {})
|
||
await page.locator('nav[aria-label="主导航"] a[href^="/chat"]').first().click()
|
||
await page.waitForURL((u) => u.pathname.startsWith('/chat'), { timeout: 5000 }).catch(() => {})
|
||
check('回到问数据页：重新取了一次库列表', await until(() => log.sourceGets > getsBefore, 5000),
|
||
`${log.sourceGets - getsBefore} 次`)
|
||
check('别处刚加的库出现在首屏的标签里', await page.locator('[data-scope-chip="ds-new"]').waitFor({ timeout: 5000 })
|
||
.then(() => true, () => false))
|
||
ctl.extraSource = false
|
||
check('没有运行时报错', errors.length === 0, errors[0] ?? '')
|
||
await ctx.close()
|
||
})
|
||
|
||
await section('composer', '画布助手：「在现有工作流上改 | 从头生成」分段控件', async () => {
|
||
const { page, ctx, errors } = await open(theme)
|
||
const WF = {
|
||
id: 'wf-chat-check', name: '__chat_check_composer__', description: '检查脚本伪造的工作流', graph: GRAPH,
|
||
tags: [], version: 1, is_template: false, status: 'draft', published_version: null, published_by: null,
|
||
run_count: 0, created_at: iso(60), updated_at: iso(60),
|
||
}
|
||
await page.route(/\/api\/workflows(\?.*)?$/, (route) => (route.request().method() === 'GET'
|
||
? route.fulfill({ json: [WF] }) : route.fallback()))
|
||
await page.route(/\/api\/workflows\/wf-chat-check(\/.*)?(\?.*)?$/, (route) => {
|
||
const url = new URL(route.request().url())
|
||
if (route.request().method() !== 'GET') return route.abort()
|
||
if (url.pathname.endsWith('/versions')) return route.fulfill({ json: [] })
|
||
return route.fulfill({ json: WF })
|
||
})
|
||
await page.goto(`${WEB}/studio/wf-chat-check`)
|
||
await ready(page, () => window.__studio?.getState().workflow?.id === 'wf-chat-check', '画布上伪造的工作流（window.__studio）')
|
||
.catch(() => {})
|
||
const group = page.getByRole('radiogroup', { name: '生成方式' })
|
||
check('有「生成方式」分段控件', await group.waitFor({ timeout: 6000 }).then(() => true, () => false))
|
||
const base = group.getByRole('radio', { name: '在现有工作流上改' })
|
||
const fresh = group.getByRole('radio', { name: '从头生成' })
|
||
check('默认「在现有工作流上改」', (await base.getAttribute('aria-checked')) === 'true'
|
||
&& (await fresh.getAttribute('aria-checked')) === 'false')
|
||
const size = await fresh.evaluate((el) => parseFloat(getComputedStyle(el).fontSize)).catch(() => 0)
|
||
check('字不小于 11px', size >= 11, `${size}px`)
|
||
// 从头生成只是不带画布上的图：后端照样垫上之前几轮对话和上一轮的图（studio-m2 终验）。
|
||
// 提示不能说成「整个重新生成」、像是什么都不记得
|
||
const freshTip = await fresh.getAttribute('title') ?? ''
|
||
check('「从头生成」的提示说清之前的对话仍会参考', /忽略画布上的现有工作流/.test(freshTip) && /之前.*对话/.test(freshTip)
|
||
&& /参考/.test(freshTip), freshTip)
|
||
await base.focus()
|
||
await page.keyboard.press('ArrowRight')
|
||
check('方向键切到「从头生成」', (await fresh.getAttribute('aria-checked')) === 'true'
|
||
&& await fresh.evaluate((el) => el === document.activeElement))
|
||
await page.waitForTimeout(300) // 底色有过渡，等它换完再截
|
||
await page.screenshot({ path: `${SHOTS}/composer-fresh-${theme}.png` })
|
||
const gensBefore = log.gens.length
|
||
const box = page.getByRole('textbox', { name: '描述要生成或修改的工作流' })
|
||
await box.fill('重新画一张三步的图')
|
||
await box.press('Enter')
|
||
await until(() => log.gens.length > gensBefore, 5000)
|
||
check('从头生成：不带现有的图', log.gens.at(-1)?.base_graph === null, JSON.stringify(log.gens.at(-1)?.base_graph)?.slice(0, 60))
|
||
await page.waitForFunction(() => !window.__studio.getState().copilot.active, null, { timeout: 8000 }).catch(() => {})
|
||
await base.click()
|
||
await box.fill('在最后加一步人工审批')
|
||
await box.press('Enter')
|
||
await until(() => log.gens.length > gensBefore + 1, 5000)
|
||
check('在现有工作流上改：带上现有的工作流', !!log.gens.at(-1)?.base_graph?.nodes?.length)
|
||
|
||
// 助手出错：出错的那一轮自己有报错块（怎么办、原文都在），输入框上面不再重复一条
|
||
await page.waitForFunction(() => !window.__studio.getState().copilot.active, null, { timeout: 8000 }).catch(() => {})
|
||
await box.fill('这一句会让建流程就失败')
|
||
await box.press('Enter')
|
||
const card = page.locator('[data-turn-error]', { hasText: '模型返回的工作流结构有误' })
|
||
const shown = await card.first().waitFor({ timeout: 6000 }).then(() => true, () => false)
|
||
const said = await card.first().innerText().catch(() => '')
|
||
check('助手出错：那一轮写出发生了什么和怎么办', shown && said.includes('请重试，或把需求描述得更具体'),
|
||
said.replace(/\n/g, ' / '))
|
||
check('同一句报错只说一遍（输入框上面不再摆一条）', await card.count() === 1
|
||
&& await page.locator('[data-copilot-error]').count() === 0,
|
||
`${await card.count()} 块 · 错误条 ${await page.locator('[data-copilot-error]').count()} 条`)
|
||
await page.waitForTimeout(300)
|
||
await page.screenshot({ path: `${SHOTS}/composer-error-${theme}.png` })
|
||
check('没有运行时报错', errors.length === 0, errors[0] ?? '')
|
||
await ctx.close()
|
||
})
|
||
}
|
||
|
||
await browser.close()
|
||
console.log(failed ? `\n✗ ${failed} 项未通过` : '\n✓ 问数据全部通过')
|
||
process.exit(failed ? 1 : 0)
|