mirror of
https://github.com/ThearchyHelios/AgentLab.git
synced 2026-11-15 03:31:58 +00:00
第三波前两批：六个「看着成功其实没干活」的平台漏洞堵上，记录页能回放，管理页与画布收尾
起因是一次探索运行：Copilot 改节点时把工具绑定整份抹掉，agent 没有工具只能写「假设调用」， 校验节点的自动修复编出 total_count=0 让 schema 通过，协作成员把工具调用标记当文字输出了四轮， 轮数用完后节点把原话当成功结果交出去——运行标成已完成，数据库一次都没查。 平台可信度（先写失败测试） - Copilot 修改节点改为合并：没写的字段保留，null 才删除；协作成员按名字合并。前后端同一套规则， 工具集合变小时自查报警、改图回执列出变化 - 新校验规则：提示词点名了某个工具而节点没有绑定它，判错误，Copilot 自查会交回去修 - 校验修复只许改格式：冒出原文没有的值就判修复无效；修复这次模型调用计入用量 - 识别模型把工具调用写成文字的情况（DSML、<tool_call> 等标记）：纠正一次，仍然如此就判失败并说明原因 - 协作团队用完轮数不再按成功交付：默认失败，可选降档，点出一次都没派到的成员 - 审批恢复后，同一节点里已经执行过的工具调用复用结果，不再重复执行 后端 - 查询时限交给数据库执行（Oracle call_timeout、PG statement_timeout、MySQL max_execution_time、 SQLite progress handler），可按数据源配置；超时报错与实际耗时一致 - 等审批的运行可以真正取消；会话列表带最后一轮状态；测连接结果持久化；探查结构可只看不存； 自定义工具可草稿试跑、写坏的参数定义保存时拒绝；MCP 服务改动后作废旧工具缓存； 检索命中带贡献分解；重建索引在后台跑并报进度；报错统一收进 core/errors 前端 - 公共层：管理页共用件、失败说明（lib/explain）、健康状态进公共库；目录按表各自加载并有超时， 一张表卡住不再拖垮整站；上传显示真实字节进度；离线横幅只播报不变的句子 - 画布：卡片只说游标所在的那一次执行；等审批时可放弃运行；团队收尾判定与未完成说清楚； 回放用量取那一刻的值 - 编排页：Copilot 合并语义与工具变化回执；「用完轮数时」选项；正式运行期间锁住所有改图入口 - 回答流 / 问数据：新失败类型说人话并给下一步；画布与右栏互相定位；数据源范围、回收站、重新核对 - 记录页：「航迹」页签复用画布时间轴，可拖动回放、列出关键时刻；列表行带执行 / 等人 / 失败缩略条 - 外壳：统一的离开前确认（data router），导航、⌘K、快捷键、后退都会先问 验证（沙箱）：tsc 干净，pytest 838 passed（两遍），14 个检查脚本中 13 个全绿； check-ui 余 1 项旧断言（按旧日期格式找运行行），3c 更新。
This commit is contained in:
ThearchyHelios
+414
-9
@@ -19,7 +19,11 @@
|
||||
// - 定位 / 去审批取景半路被拉远，落地卡在 compact 档（错误正文被藏）；
|
||||
// - 「只看执行路径」带进下一次运行；正式运行横幅盖住「恢复」；
|
||||
// - 同一刻的等人时长几处写几个数；只是点一下画布就暂停跟随；
|
||||
// - Esc 在右栏里按也清掉整次运行；泳道、节点格、坞高只能用鼠标。
|
||||
// - Esc 在右栏里按也清掉整次运行；泳道、节点格、坞高只能用鼠标；
|
||||
// - 运行中点节点，属性面板整块盖住右栏的运行视图（应当先请右栏滚到它的步骤）；
|
||||
// - 等审批时只能批不能放弃；回放时回边的 ×N、面板的用量拿终值冒充那一刻；
|
||||
// - Copilot 搭图时新节点落在视口外看不见；双击空白处只会缩放；1024 宽打开只看得见两张半卡；
|
||||
// - Copilot 删掉的节点一声不响就没了，看不出删的是哪个、原来在哪。
|
||||
import { chromium } from '../frontend/node_modules/playwright-core/index.mjs'
|
||||
|
||||
const WEB = process.env.AGENTLAB_WEB ?? 'http://localhost:5273'
|
||||
@@ -187,6 +191,13 @@ async function openStudio({ id = FX_ID, theme = 'dark', ...extra } = {}) {
|
||||
page.on('pageerror', (e) => errors.push('pageerror: ' + e.message))
|
||||
page.on('console', (m) => { if (m.type() === 'error' && !m.text().includes('404') && !m.text().includes('409')) errors.push(m.text()) })
|
||||
await fakeBackend(page)
|
||||
// 主题：设置接口里存的偏好会在启动时盖过本地缓存，不改它的话「暗色」那一遍画出来是亮的
|
||||
await page.route('**/api/settings', async (route) => {
|
||||
if (route.request().method() !== 'GET') return route.continue()
|
||||
const res = await route.fetch()
|
||||
const body = await res.json().catch(() => ({}))
|
||||
await route.fulfill({ response: res, json: { ...body, ui: { ...(body.ui ?? {}), theme } } })
|
||||
})
|
||||
await page.goto(`${WEB}/studio/${id}`, { waitUntil: 'networkidle' })
|
||||
await page.waitForFunction((wid) => window.__studio?.getState().workflow?.id === wid, id, { timeout: 15000 })
|
||||
await page.waitForTimeout(700)
|
||||
@@ -207,6 +218,18 @@ const seedRun = (page, extra = {}) => page.evaluate((x) => {
|
||||
}, extra)
|
||||
|
||||
const count = (page, sel) => page.locator(sel).count()
|
||||
/** 轮询到条件成立或超时。取景有动画、热更新可能插一脚，固定等几百毫秒会偶发失败 */
|
||||
async function until(fn, { timeout = 3000, every = 100 } = {}) {
|
||||
const end = Date.now() + timeout
|
||||
for (;;) {
|
||||
const v = await fn().catch(() => false)
|
||||
if (v || Date.now() > end) return v
|
||||
await new Promise((r) => setTimeout(r, every))
|
||||
}
|
||||
}
|
||||
/** 开发服务器热更新会把页面整个重载（别的代理恰好在改代码），store 里灌的运行就没了 */
|
||||
const markBoot = (page) => page.evaluate(() => { window.__fxBoot = true })
|
||||
const reloaded = (page) => page.evaluate(() => !window.__fxBoot).catch(() => true)
|
||||
const capsule = (page) => page.locator('.sf-capsule')
|
||||
/** 工具栏上的「运行」发起按钮（胶囊的名字也以「运行中」开头，按全名找） */
|
||||
const launcher = (page) => page.locator('[data-run-control]').getByRole('button', { name: /^运行$/ })
|
||||
@@ -330,6 +353,7 @@ console.log('=== 运行中：胶囊、光点、计时 ===')
|
||||
console.log('=== 等人：入边闸门，不流光点；去审批代替停止 ===')
|
||||
{
|
||||
const { browser, page } = await openStudio()
|
||||
await markBoot(page)
|
||||
await seedRun(page)
|
||||
await feed(page, [...PARALLEL, ...TEAM_DONE, ...TO_REVIEW])
|
||||
await page.waitForTimeout(400)
|
||||
@@ -351,12 +375,34 @@ console.log('=== 等人：入边闸门，不流光点；去审批代替停止 ==
|
||||
check('等人时全图 0 个光点', await count(page, '.edge-packet') === 0, `${await count(page, '.edge-packet')} 个`)
|
||||
check('小地图上等人的节点放大了一圈', await count(page, '.sf-mm-ring.sf-mm-waiting') === 1)
|
||||
|
||||
// 去审批：属性面板盖在右栏上时先让开，画布取景到等人的节点
|
||||
await page.evaluate(() => window.__studio.setState({ focusRequest: null, selectedId: 'gate' }))
|
||||
// 去审批：属性面板盖在右栏上时先让开，画布取景到等人的节点。右栏（助手面板）接
|
||||
// agentlab:goto-approval 事件切到运行层、滚到审批卡；它接了（preventDefault）胶囊就不再自己找
|
||||
await page.evaluate(() => {
|
||||
window.__goto = []
|
||||
window.addEventListener('agentlab:goto-approval', (e) => { window.__goto.push(e.detail); e.preventDefault() })
|
||||
window.__studio.setState({ focusRequest: null, selectedId: 'gate' })
|
||||
// 胶囊自己兜底时在整页里找审批卡（document.querySelector）；右栏接手时只在自己的
|
||||
// 滚动区里找。数一下前者，就知道胶囊有没有在右栏接手之后还去找
|
||||
window.__slotLookups = 0
|
||||
const find = Document.prototype.querySelector
|
||||
window.__restoreQuery = () => { Document.prototype.querySelector = find }
|
||||
Document.prototype.querySelector = function (sel) {
|
||||
if (String(sel).includes('data-approval-slot')) window.__slotLookups += 1
|
||||
return find.call(this, sel)
|
||||
}
|
||||
})
|
||||
await capsule(page).getByRole('button', { name: /去审批/ }).click()
|
||||
await page.waitForTimeout(200)
|
||||
const went = await page.evaluate(() => ({ sel: window.__studio.getState().selectedId, focus: window.__studio.getState().focusRequest?.id }))
|
||||
check('去审批：属性面板让开、取景到等人的节点', went.sel === null && went.focus === 'review', JSON.stringify(went))
|
||||
const goto = await page.evaluate(() => window.__goto)
|
||||
check('去审批：广播给右栏（带运行和节点）', goto.length === 1 && goto[0].nodeId === 'review' && goto[0].runId === 'fxrun000001',
|
||||
JSON.stringify(goto))
|
||||
// 描边、滚动都不能拿来判断——右栏接手时自己也描同一圈、滚到同一张卡
|
||||
await page.waitForTimeout(600)
|
||||
const lookups = await page.evaluate(() => { window.__restoreQuery(); return window.__slotLookups })
|
||||
check('右栏接手之后胶囊不再自己找卡片、不报「还没出现」', lookups === 0
|
||||
&& !(await page.locator('[aria-live]').allInnerTexts()).join('').includes('审批卡还没出现'), `整页找了 ${lookups} 次`)
|
||||
|
||||
// F：在需要处理的队列里逐个定位
|
||||
await page.evaluate(() => window.__studio.setState({ focusRequest: null }))
|
||||
@@ -365,6 +411,32 @@ console.log('=== 等人：入边闸门，不流光点；去审批代替停止 ==
|
||||
await page.waitForTimeout(200)
|
||||
const focus = await page.evaluate(() => window.__studio.getState().focusRequest?.id)
|
||||
check('F 定位到等人的节点', focus === 'review', String(focus))
|
||||
|
||||
// 放弃这次运行：等审批时的次级动作。关掉的审批找不回来，先问一句；确认后才调 /cancel
|
||||
let cancels = 0
|
||||
await page.route(/\/api\/runs\/fxrun000001\/cancel$/, (route) => {
|
||||
cancels += 1
|
||||
return route.fulfill({ status: 200, contentType: 'application/json', body: JSON.stringify({ ok: true, status: 'cancelled' }) })
|
||||
})
|
||||
const abandon = capsule(page).getByRole('button', { name: '放弃这次运行' })
|
||||
check('等审批时有「放弃这次运行」', await abandon.count() === 1)
|
||||
await abandon.click()
|
||||
const ask = page.getByRole('dialog').filter({ hasText: '放弃这次运行？' })
|
||||
await ask.waitFor({ timeout: 3000 }).catch(() => {})
|
||||
const askText = (await ask.innerText().catch(() => '')).replace(/\s+/g, ' ')
|
||||
check('放弃之前先问一句，写清后果', askText.includes('审批卡一并关闭') && cancels === 0, askText.slice(0, 80))
|
||||
await ask.getByRole('button', { name: '放弃这次运行' }).click().catch(() => {})
|
||||
await until(async () => cancels === 1, { timeout: 2000 })
|
||||
check('确认后才调 /cancel', cancels === 1, String(cancels))
|
||||
await feed(page, [ev(19, 'run.cancelled', null, {
|
||||
timing: { wall_ms: 20000, active_ms: 14000, wait_ms: 6000 }, actor: '检查员', message: '放弃了这次运行，1 条待审批一并关闭',
|
||||
}, 20)])
|
||||
await page.waitForTimeout(300)
|
||||
// 胶囊不见了多半是别的代理改代码触发了整页热更新（灌进 store 的运行没了）：报出来，别卡 30 秒崩掉
|
||||
const opened = await capsule(page).locator('.sf-cap-main').click({ timeout: 5000 }).then(() => true).catch(() => false)
|
||||
const line = opened ? await page.locator('.sf-hud-line').innerText().catch(() => '') : ''
|
||||
check('取消之后说清谁放弃的、关了什么', line.includes('检查员 放弃了这次运行，1 条待审批一并关闭'),
|
||||
opened ? line : `胶囊不见了${await reloaded(page) ? '：页面被热更新重载过，重跑确认' : ''}`)
|
||||
await browser.close()
|
||||
}
|
||||
|
||||
@@ -691,6 +763,17 @@ console.log('=== 正式运行：入口看 published_version；运行期间画布
|
||||
const formal = page.getByRole('button', { name: /正式运行 v2/ })
|
||||
check('草稿状态下正式运行入口还在（看 published_version）', await formal.count() === 1)
|
||||
check('画布有未保存改动时也能点', await formal.isEnabled())
|
||||
// 运行类别的记号：正式运行实心圆点、探索运行空心；盾牌只留给发布等级，不在发起入口上混用
|
||||
const marks = (sel) => page.evaluate((sel) => {
|
||||
const root = document.querySelector(sel)
|
||||
return {
|
||||
formal: root?.querySelectorAll('svg.sf-dot[data-class="formal"]').length ?? 0,
|
||||
exploratory: root?.querySelectorAll('svg.sf-dot[data-class="exploratory"]').length ?? 0,
|
||||
shield: root?.querySelectorAll('.lucide-shield-check, svg[class*="shield"]').length ?? 0,
|
||||
}
|
||||
}, sel)
|
||||
const bar = await marks('[data-run-control]')
|
||||
check('正式运行入口是实心圆点，没有盾牌', bar.formal === 1 && bar.shield === 0, JSON.stringify(bar))
|
||||
await formal.click()
|
||||
await page.waitForSelector('.sf-pop', { timeout: 5000 })
|
||||
await page.waitForTimeout(400)
|
||||
@@ -699,6 +782,16 @@ console.log('=== 正式运行：入口看 published_version；运行期间画布
|
||||
check('表单字段取已发布版本的（topic），不取画布的（q）',
|
||||
await page.locator('.sf-pop textarea#run-field-topic').count() === 1
|
||||
&& await page.locator('.sf-pop textarea#run-field-q').count() === 0)
|
||||
const formalPop = await marks('.sf-pop')
|
||||
check('正式运行浮层的标题和确认按钮都是实心圆点', formalPop.formal === 2 && formalPop.exploratory === 0 && formalPop.shield === 0,
|
||||
JSON.stringify(formalPop))
|
||||
await page.keyboard.press('Escape')
|
||||
await page.waitForTimeout(150)
|
||||
await launcher(page).click()
|
||||
await page.waitForSelector('.sf-pop', { timeout: 5000 })
|
||||
const explorePop = await marks('.sf-pop')
|
||||
check('探索运行浮层是空心圆点', explorePop.exploratory === 1 && explorePop.formal === 0 && explorePop.shield === 0,
|
||||
JSON.stringify(explorePop))
|
||||
await page.keyboard.press('Escape')
|
||||
|
||||
await page.evaluate(() => window.__studio.setState({ dirty: false }))
|
||||
@@ -782,13 +875,30 @@ console.log('=== 打开宽图：以入口为左锚，缩放不低于可读下限
|
||||
const vp = () => page.evaluate(() => document.querySelector('.react-flow__viewport').style.transform)
|
||||
const t0 = Date.now() / 1000
|
||||
const wev = (seq, type, node_id, dt) => ({ seq, type, node_id, data: {}, ts: t0 + dt })
|
||||
await markBoot(page)
|
||||
await seedRun(page, { workflow_id: WIDE_ID })
|
||||
await feed(page, [wev(1, 'run.started', null, 0), wev(2, 'node.started', 'n0', 0.1), wev(3, 'node.finished', 'n0', 0.2)])
|
||||
// 航迹坞升起来、画布变矮之后再量：坞出现时保持中心的那一下也会挪视口
|
||||
await page.waitForSelector('.tl', { timeout: 3000 }).catch(() => {})
|
||||
await page.waitForTimeout(300)
|
||||
check('n9 起初在视口外', !(await inView('n9')))
|
||||
const vpFrom = await vp()
|
||||
await feed(page, [wev(4, 'node.started', 'n9', 0.3)])
|
||||
await page.waitForTimeout(700)
|
||||
check('跟随执行：视口外的节点开跑，镜头跟过去', await inView('n9'))
|
||||
// 等视口真的动了、并且停在 n9 上：取景有 320ms 的动画，固定等一段时间会偶发落空
|
||||
const followed = await until(async () => (await vp()) !== vpFrom && await inView('n9'))
|
||||
const why = followed ? '' : JSON.stringify({
|
||||
reloaded: await reloaded(page),
|
||||
phase: await page.evaluate(() => window.__studio?.getState().runPhase).catch(() => null),
|
||||
follow: await page.evaluate(() => window.__studio?.getState().follow).catch(() => null),
|
||||
moved: (await vp()) !== vpFrom,
|
||||
})
|
||||
check('跟随执行：视口外的节点开跑，镜头跟过去', !!followed, why)
|
||||
// 等这一趟飞完，再让下一个远处的节点开跑，趁镜头飞到一半时人抓住画布拖一下。
|
||||
// React Flow 把被打断的那段动画的收尾推迟一拍才报，报到时人的拖动已经开始——
|
||||
// 以前它把人的起点吃掉，拖完跟随照样不暂停
|
||||
await until(async () => { const a = await vp(); await new Promise((r) => setTimeout(r, 120)); return a === await vp() })
|
||||
await feed(page, [wev(5, 'node.finished', 'n9', 0.4), wev(6, 'node.started', 'n1', 0.5)])
|
||||
await page.waitForTimeout(110)
|
||||
const pane = await page.locator('.react-flow__pane').boundingBox()
|
||||
await page.mouse.move(pane.x + 300, pane.y + 60)
|
||||
await page.mouse.down()
|
||||
@@ -796,14 +906,21 @@ console.log('=== 打开宽图：以入口为左锚，缩放不低于可读下限
|
||||
await page.mouse.up()
|
||||
await page.waitForTimeout(200)
|
||||
const paused = await vp()
|
||||
check('手动拖过画布：显示「已暂停跟随」', (await page.locator('.sf-follow').innerText()).includes('已暂停跟随'))
|
||||
await feed(page, [wev(5, 'node.finished', 'n9', 0.4), wev(6, 'node.started', 'n2', 0.5)])
|
||||
check('镜头飞到一半被人拖住：显示「已暂停跟随」',
|
||||
(await page.locator('.sf-follow').innerText({ timeout: 3000 }).catch(() => '')).includes('已暂停跟随'))
|
||||
await feed(page, [wev(7, 'node.finished', 'n1', 0.6), wev(8, 'node.started', 'n2', 0.7)])
|
||||
await page.waitForTimeout(700)
|
||||
check('暂停期间不再抢镜头', await vp() === paused)
|
||||
// 点名取景：已暂停也照做（是人要看它）
|
||||
await page.evaluate(() => window.__studio.getState().focusNode('n11'))
|
||||
await page.waitForTimeout(700)
|
||||
check('focusNode：画布取景到点名的节点', await inView('n11'))
|
||||
check('focusNode：画布取景到点名的节点', !!(await until(() => inView('n11'))))
|
||||
// 画布尺寸恰好在取景半路变了（航迹坞收起、升起）：保持中心那一下会打断动画，
|
||||
// 打断之后要重新飞到目标，不能停在半路
|
||||
await page.evaluate(() => window.__studio.getState().focusNode('n0'))
|
||||
await page.waitForTimeout(90)
|
||||
await page.locator('.tl').getByRole('button', { name: '收起航迹' }).click().catch(() => {})
|
||||
check('取景半路画布变高：照样落到目标上', !!(await until(() => inView('n0'))))
|
||||
await page.locator('.tl').getByRole('button', { name: '展开' }).click().catch(() => {})
|
||||
await page.evaluate(() => window.__studio.getState().clearRun())
|
||||
await page.waitForTimeout(200)
|
||||
|
||||
@@ -820,6 +937,294 @@ console.log('=== 打开宽图：以入口为左锚，缩放不低于可读下限
|
||||
await browser.close()
|
||||
}
|
||||
|
||||
// ---------------------------------------------------------------- 运行中点节点
|
||||
|
||||
console.log('=== 运行中点节点：先看它的步骤，属性面板是次级入口 ===')
|
||||
{
|
||||
const { browser, page, errors } = await openStudio()
|
||||
await seedRun(page)
|
||||
await feed(page, PARALLEL_NOW)
|
||||
await page.waitForTimeout(400)
|
||||
await page.evaluate(() => {
|
||||
window.__reveal = []
|
||||
window.addEventListener('agentlab:reveal-step', (e) => window.__reveal.push(e.detail))
|
||||
})
|
||||
const sheet = () => count(page, '[data-inspector-sheet]')
|
||||
const head = (id) => page.locator(`.react-flow__node[data-id="${id}"] .nc-head`)
|
||||
await head('start').click()
|
||||
await page.waitForTimeout(250)
|
||||
const picked = await page.evaluate(() => ({ sel: window.__studio.getState().selectedId, reveal: window.__reveal }))
|
||||
check('点跑过的节点：选中它', picked.sel === 'start', String(picked.sel))
|
||||
check('属性面板不自动盖上右栏', await sheet() === 0)
|
||||
check('请右栏滚到它的步骤（agentlab:reveal-step）', picked.reveal.length === 1 && picked.reveal[0].nodeId === 'start'
|
||||
&& picked.reveal[0].runId === 'fxrun000001', JSON.stringify(picked.reveal))
|
||||
const peek = page.locator('.sf-peek')
|
||||
check('画布上给出「…的配置」入口', await peek.count() === 1 && (await peek.innerText()).includes('问题'), await peek.innerText().catch(() => ''))
|
||||
await peek.click()
|
||||
await page.waitForTimeout(250)
|
||||
check('点入口才打开属性面板', await sheet() === 1 && await count(page, '.sf-peek') === 0)
|
||||
check('打开之后底部留着回到运行的那一条', await count(page, '[data-inspector-sheet] [data-run-phase]') === 1)
|
||||
// 没跑过的节点在右栏里没有步骤可看：照常打开属性面板
|
||||
await page.locator('.react-flow__pane').click({ position: { x: 20, y: 300 } })
|
||||
await page.waitForTimeout(150)
|
||||
await page.evaluate(() => { window.__reveal = [] })
|
||||
await head('gate').click()
|
||||
await page.waitForTimeout(250)
|
||||
check('点没跑过的节点：照常打开属性面板，不请右栏滚动', await sheet() === 1
|
||||
&& await page.evaluate(() => window.__reveal.length) === 0)
|
||||
// 双击跑过的节点：直接看配置
|
||||
await page.locator('.react-flow__pane').click({ position: { x: 20, y: 300 } })
|
||||
await page.waitForTimeout(150)
|
||||
await head('team').dblclick()
|
||||
await page.waitForTimeout(250)
|
||||
check('双击节点直接打开属性面板', await sheet() === 1 && await page.evaluate(() => window.__studio.getState().selectedId) === 'team')
|
||||
await page.keyboard.press('Escape')
|
||||
await page.waitForTimeout(150)
|
||||
check('Esc 收起属性面板、取消选中，运行还在', await sheet() === 0
|
||||
&& await page.evaluate(() => window.__studio.getState().runPhase) === 'running')
|
||||
// 先看了步骤、运行随后结束：节点还选中着、面板没开。再点它（选中没变）也得打开面板
|
||||
await head('start').click()
|
||||
await page.waitForTimeout(200)
|
||||
await page.evaluate(() => window.__studio.getState().clearRun())
|
||||
await page.waitForTimeout(200)
|
||||
const kept = await page.evaluate(() => window.__studio.getState().selectedId)
|
||||
check('运行清掉之后仍选中的节点留着「…的配置」入口', kept !== 'start' || await count(page, '.sf-peek') === 1, String(kept))
|
||||
await head('start').click()
|
||||
await page.waitForTimeout(250)
|
||||
check('没有运行时再点这个节点：打开属性面板', await sheet() === 1)
|
||||
check('没有运行时报错', errors.length === 0, errors.slice(0, 2).join(' | '))
|
||||
await browser.close()
|
||||
}
|
||||
|
||||
// ---------------------------------------------------------------- 回放：回边圈数、用量跟着游标
|
||||
|
||||
console.log('=== 回放：回边的 ×N 和面板的用量都是那一刻的 ===')
|
||||
{
|
||||
const { browser, page, errors } = await openStudio()
|
||||
await seedRun(page)
|
||||
// 两次模型调用，一次在第 5 秒、一次在第 12 秒；终态的后端累计和它们对得上
|
||||
const usage = [
|
||||
ev(30, 'llm.end', 'team', { agent: '检索员', model: 'm', input_tokens: 600, output_tokens: 200, cost_usd: 0.002 }, 5),
|
||||
ev(31, 'llm.end', 'team', { agent: '分析员', model: 'm', input_tokens: 400, output_tokens: 200, cost_usd: 0.001 }, 12),
|
||||
]
|
||||
const finished = TO_LOOP.map((e) => (e.type === 'run.finished'
|
||||
? { ...e, data: { ...e.data, usage: { input_tokens: 1000, output_tokens: 400, cost_usd: 0.003 } } } : e))
|
||||
const all = [...PARALLEL, ...usage, ...TEAM_DONE, ...finished].sort((a, b) => a.ts - b.ts)
|
||||
.map((e, i) => ({ ...e, seq: i + 1, replay: true }))
|
||||
await feed(page, all)
|
||||
await page.waitForTimeout(400)
|
||||
const badge = () => page.locator('.sf-loop-badge').innerText().catch(() => '')
|
||||
check('跑完：回边写兜回去几次（↺ ×1）', (await badge()).replace(/\s+/g, ' ') === '↺ ×1', await badge())
|
||||
const at = (t) => page.evaluate((ms) => window.__studio.getState().setReplayAt(ms), (BASE + t) * 1000)
|
||||
// 返工的第一轮还在跑：还没兜回去过
|
||||
await at(18)
|
||||
await page.waitForTimeout(250)
|
||||
check('回放到第一轮：回边还没兜回去，不写次数', (await badge()).trim() === '↺', await badge())
|
||||
// 协作团队跑到一半：只有第一次模型调用
|
||||
await at(8)
|
||||
await page.waitForTimeout(250)
|
||||
await capsule(page).locator('.sf-cap-main').click()
|
||||
const usageOf = async () => (await page.locator('.sf-metric', { hasText: '用量' }).innerText().catch(() => '')).replace(/\s+/g, ' ')
|
||||
const mid = await usageOf()
|
||||
check('回放：面板的用量是那一刻的（800 tok），不是终值', /800 tok/.test(mid) && !/1\.4k/.test(mid), mid)
|
||||
const side = await page.locator('.react-flow__node[data-id="team"] .nc-tele-side').innerText().catch(() => '')
|
||||
check('回放：卡片上的 token 也是那一刻的', /800 tok/.test(side), side)
|
||||
await page.keyboard.press('Escape')
|
||||
await page.evaluate(() => window.__studio.getState().setReplayAt(null))
|
||||
await page.waitForTimeout(250)
|
||||
await capsule(page).locator('.sf-cap-main').click()
|
||||
const live = await usageOf()
|
||||
check('回到实时：用量是整次运行的（1.4k tok）', /1\.4k tok/.test(live), live)
|
||||
check('没有运行时报错', errors.length === 0, errors.slice(0, 2).join(' | '))
|
||||
await browser.close()
|
||||
}
|
||||
|
||||
// ---------------------------------------------------------------- Copilot 搭图时的镜头
|
||||
|
||||
console.log('=== Copilot 搭图：新节点在视口外时镜头跟过去，人一动手这一轮就不跟 ===')
|
||||
for (const reducedMotion of ['no-preference', 'reduce']) {
|
||||
const { browser, page, errors } = await openStudio({ id: WIDE_ID, reducedMotion })
|
||||
const inView = (id) => page.evaluate((nid) => {
|
||||
const pane = document.querySelector('.react-flow').getBoundingClientRect()
|
||||
const r = document.querySelector(`.react-flow__node[data-id="${nid}"]`).getBoundingClientRect()
|
||||
return r.left >= pane.left && r.right <= pane.right && r.top >= pane.top && r.bottom <= pane.bottom
|
||||
}, id)
|
||||
const vp = () => page.evaluate(() => document.querySelector('.react-flow__viewport').style.transform)
|
||||
const building = (on, cursor) => page.evaluate(({ on, cursor }) => {
|
||||
const st = window.__studio
|
||||
st.setState({ copilot: { ...st.getState().copilot, active: on }, copilotCursor: cursor })
|
||||
}, { on, cursor })
|
||||
const label = reducedMotion === 'reduce' ? '（减少动效）' : ''
|
||||
check(`n11 起初在视口外${label}`, !(await inView('n11')))
|
||||
await building(true, 'n11')
|
||||
if (reducedMotion === 'reduce') {
|
||||
// 减少动效时不做动画：下一帧就到位
|
||||
await page.waitForTimeout(120)
|
||||
check('减少动效：镜头直接到位，不飞', await inView('n11'))
|
||||
} else {
|
||||
check('助手刚放下的节点在视口外：镜头跟过去', !!(await until(() => inView('n11'))))
|
||||
const z0 = await page.evaluate(() => document.querySelector('.react-flow').style.getPropertyValue('--zoom'))
|
||||
// 人拖了一下画布：这一轮不再跟
|
||||
const pane = await page.locator('.react-flow__pane').boundingBox()
|
||||
await page.mouse.move(pane.x + 200, pane.y + 200)
|
||||
await page.mouse.down()
|
||||
await page.mouse.move(pane.x + 380, pane.y + 240, { steps: 5 })
|
||||
await page.mouse.up()
|
||||
await page.waitForTimeout(250)
|
||||
const held = await vp()
|
||||
await building(true, 'n0')
|
||||
await page.waitForTimeout(700)
|
||||
check('人动过镜头之后，这一轮不再抢', await vp() === held)
|
||||
// 下一轮重新跟
|
||||
await building(false, null)
|
||||
await building(true, 'n0')
|
||||
check('下一轮又跟上', !!(await until(() => inView('n0'))))
|
||||
const z1 = await page.evaluate(() => document.querySelector('.react-flow').style.getPropertyValue('--zoom'))
|
||||
check('跟镜头只平移，不改缩放', z0 === z1, `${z0} → ${z1}`)
|
||||
}
|
||||
await building(false, null)
|
||||
check('没有运行时报错', errors.length === 0, errors.slice(0, 2).join(' | '))
|
||||
await browser.close()
|
||||
}
|
||||
|
||||
// ---------------------------------------------------------------- Copilot 删掉的节点：原地虚线框
|
||||
|
||||
console.log('=== Copilot 这一轮删掉的节点：原地留一道虚线框，停一下再淡出 ===')
|
||||
/** 照 store 的真实顺序走一轮：开始 → 节点没了 → 这一轮落定（diff 写着删了谁）→ 结束 */
|
||||
const removeInTurn = (page, id) => page.evaluate((id) => {
|
||||
const st = window.__studio
|
||||
const turn = { id: 'fx-turn', instruction: '去掉返工那一步', ops: [], phase: 'running', explanation: '', error: '' }
|
||||
st.setState({ copilot: { ...st.getState().copilot, active: true } })
|
||||
st.setState({ copilotTurns: [...st.getState().copilotTurns, turn] })
|
||||
st.setState({
|
||||
nodes: st.getState().nodes.filter((n) => n.id !== id),
|
||||
edges: st.getState().edges.filter((e) => e.source !== id && e.target !== id),
|
||||
})
|
||||
st.setState({ copilotTurns: st.getState().copilotTurns.map((t) => (t.id === turn.id ? {
|
||||
...t, phase: 'done', outcome: 'applied',
|
||||
diff: { added: [], removed: [id], changed: [], edgesAdded: 0, edgesRemoved: 2, total: 3 },
|
||||
} : t)) })
|
||||
st.setState({ copilot: { ...st.getState().copilot, active: false } })
|
||||
}, id)
|
||||
for (const reducedMotion of ['no-preference', 'reduce']) {
|
||||
const { browser, page, errors } = await openStudio({ reducedMotion })
|
||||
const was = await page.locator('.react-flow__node[data-id="fix"]').boundingBox()
|
||||
await removeInTurn(page, 'fix')
|
||||
await page.waitForTimeout(400)
|
||||
const ghost = page.locator('.sf-ghost[data-ghost="fix"]')
|
||||
if (reducedMotion === 'reduce') {
|
||||
check('减少动效：不画虚线框', await ghost.count() === 0)
|
||||
} else {
|
||||
const box = await ghost.boundingBox()
|
||||
const tag = await ghost.innerText().catch(() => '')
|
||||
check('删掉的节点原地留一道虚线框，写着删的是谁', !!box && tag.includes('已删除 · 返工'), tag)
|
||||
check('框就在它原来的位置、一样大', !!box && !!was && Math.abs(box.x - was.x) < 4 && Math.abs(box.y - was.y) < 4
|
||||
&& Math.abs(box.width - was.width) < 4, JSON.stringify({ was, box }))
|
||||
const tagPx = await ghost.locator('.sf-ghost-tag').evaluate((el) => el.getBoundingClientRect().height)
|
||||
check('标签在屏幕上读得清', tagPx >= 14, `${tagPx.toFixed(1)}px`)
|
||||
check('虚线框不挡操作', await ghost.evaluate((el) => getComputedStyle(el).pointerEvents) === 'none')
|
||||
await page.screenshot({ path: '/tmp/agentlab-canvas-fx-ghost.png' })
|
||||
check('停一下就淡出、收掉', !!(await until(async () => await ghost.count() === 0, { timeout: 4000 })))
|
||||
}
|
||||
check('没有运行时报错', errors.length === 0, errors.slice(0, 2).join(' | '))
|
||||
await browser.close()
|
||||
}
|
||||
// 窄画布打开就是精简档（缩放 0.4 上下）：标签的字照样补回 11px，不跟着缩成 7px
|
||||
{
|
||||
const { browser, page, errors } = await openStudio({ viewport: { width: 1024, height: 768 } })
|
||||
await page.waitForTimeout(400)
|
||||
const zoom = Number(await page.evaluate(() => document.querySelector('.react-flow').style.getPropertyValue('--zoom')))
|
||||
// 质量门在前三列里，这个缩放下在视口内：截图看得到
|
||||
await removeInTurn(page, 'gate')
|
||||
await page.waitForTimeout(400)
|
||||
const tag = await page.locator('.sf-ghost[data-ghost="gate"] .sf-ghost-tag').evaluate((el) => ({
|
||||
h: el.getBoundingClientRect().height, font: parseFloat(getComputedStyle(el).fontSize),
|
||||
clipped: el.scrollWidth > el.clientWidth + 1,
|
||||
})).catch(() => null)
|
||||
const px = tag ? tag.font * zoom : 0
|
||||
check(`精简档（缩放 ${zoom.toFixed(2)}）：虚线框的标签照样读得清（屏幕上 ≥ 11px、高 ≥ 14px）`,
|
||||
zoom < 0.5 && !!tag && px >= 10.9 && tag.h >= 14 && !tag.clipped,
|
||||
tag ? `字 ${px.toFixed(1)}px · 高 ${tag.h.toFixed(1)}px${tag.clipped ? ' · 被截断' : ''}` : '没画虚线框')
|
||||
await page.screenshot({ path: '/tmp/agentlab-canvas-fx-ghost-compact.png' })
|
||||
check('没有运行时报错', errors.length === 0, errors.slice(0, 2).join(' | '))
|
||||
await browser.close()
|
||||
}
|
||||
|
||||
// ---------------------------------------------------------------- 双击快速添加
|
||||
|
||||
console.log('=== 双击空白处：在光标处快速添加节点 ===')
|
||||
{
|
||||
const { browser, page, errors } = await openStudio()
|
||||
const spot = await emptySpot(page)
|
||||
const zoomOf = () => page.evaluate(() => document.querySelector('.react-flow').style.getPropertyValue('--zoom'))
|
||||
const z0 = await zoomOf()
|
||||
const n0 = await page.evaluate(() => window.__studio.getState().nodes.length)
|
||||
await page.mouse.dblclick(spot.x, spot.y)
|
||||
await page.waitForTimeout(250)
|
||||
const quick = page.locator('.sf-quick')
|
||||
check('双击空白处打开快速添加', await quick.count() === 1)
|
||||
check('双击不再缩放画布', await zoomOf() === z0, `${z0} → ${await zoomOf()}`)
|
||||
check('焦点直接在搜索框里', await page.evaluate(() => !!document.activeElement?.closest('.sf-quick input')))
|
||||
const panel = await quick.boundingBox()
|
||||
check('面板贴着双击处', !!panel && Math.abs(panel.x - spot.x) < 260 && Math.abs(panel.y - spot.y) < 320,
|
||||
panel ? `${Math.round(panel.x)},${Math.round(panel.y)} / ${Math.round(spot.x)},${Math.round(spot.y)}` : '没有面板')
|
||||
// 只打一个字：名字里有它的排在只是说明里提到它的前面（协作节点的说明里有「分派」）
|
||||
await page.keyboard.type('分')
|
||||
await page.waitForTimeout(100)
|
||||
const first = await quick.locator('[role="option"][aria-selected="true"]').innerText().catch(() => '')
|
||||
check('按名字搜，第一个就是要的', first.includes('条件分支'), first)
|
||||
await page.keyboard.press('Enter')
|
||||
await page.waitForTimeout(300)
|
||||
const added = await page.evaluate(() => {
|
||||
const st = window.__studio.getState()
|
||||
const n = st.nodes[st.nodes.length - 1]
|
||||
const el = document.querySelector(`.react-flow__node[data-id="${n.id}"]`)?.getBoundingClientRect()
|
||||
return { total: st.nodes.length, type: n.data.nodeType, cx: el ? el.left + el.width / 2 : null, cy: el ? el.top + el.height / 2 : null }
|
||||
})
|
||||
check('回车放下一个条件分支', added.total === n0 + 1 && added.type === 'branch', JSON.stringify(added))
|
||||
check('落在双击处（卡片中心对着光标）', added.cx != null && Math.abs(added.cx - spot.x) < 40 && Math.abs(added.cy - spot.y) < 40,
|
||||
JSON.stringify({ spot, at: [added.cx, added.cy] }))
|
||||
check('放下之后面板收起', await quick.count() === 0)
|
||||
await page.evaluate(() => window.__studio.getState().select(null))
|
||||
const spot2 = await emptySpot(page)
|
||||
await page.mouse.dblclick(spot2.x, spot2.y)
|
||||
await page.waitForTimeout(200)
|
||||
await page.keyboard.press('Escape')
|
||||
await page.waitForTimeout(150)
|
||||
check('Esc 收起、不加节点', await quick.count() === 0
|
||||
&& await page.evaluate(() => window.__studio.getState().nodes.length) === n0 + 1)
|
||||
check('没有运行时报错', errors.length === 0, errors.slice(0, 2).join(' | '))
|
||||
await browser.close()
|
||||
}
|
||||
|
||||
// ---------------------------------------------------------------- 窄画布的打开取景
|
||||
|
||||
console.log('=== 1024 宽打开：可读缩放下放不下三列时退到精简档，多看几列 ===')
|
||||
for (const theme of ['dark', 'light']) {
|
||||
const { browser, page } = await openStudio({ theme, viewport: { width: 1024, height: 768 } })
|
||||
await page.waitForTimeout(500)
|
||||
// 两遍真的是两套主题：不然「亮暗都看过」只是同一张图看了两次
|
||||
const shown = await page.evaluate(() => ({
|
||||
attr: document.documentElement.dataset.theme ?? null,
|
||||
scheme: getComputedStyle(document.documentElement).colorScheme,
|
||||
}))
|
||||
check(`${theme}: 画出来的确实是这一套主题`, shown.attr === theme && shown.scheme.includes(theme), JSON.stringify(shown))
|
||||
const view = await page.evaluate(() => {
|
||||
const rf = document.querySelector('.react-flow')
|
||||
const pane = rf.getBoundingClientRect()
|
||||
const nodes = [...document.querySelectorAll('.react-flow__node')]
|
||||
const whole = nodes.filter((n) => {
|
||||
const r = n.getBoundingClientRect()
|
||||
return r.left >= pane.left && r.right <= pane.right && r.top >= pane.top && r.bottom <= pane.bottom
|
||||
}).length
|
||||
return { zoom: Number(rf.style.getPropertyValue('--zoom')), lod: rf.dataset.lod, whole, total: nodes.length, w: Math.round(pane.width) }
|
||||
})
|
||||
check(`${theme}: 1024 宽打开至少看得全四张卡`, view.whole >= 4, JSON.stringify(view))
|
||||
check(`${theme}: 缩放不低于精简档下限 0.4，档位是精简卡`, view.zoom >= 0.4 - 1e-3 && view.lod === 'compact', JSON.stringify(view))
|
||||
await page.screenshot({ path: `/tmp/agentlab-canvas-fx-open-1024-${theme}.png` })
|
||||
await browser.close()
|
||||
}
|
||||
|
||||
// ---------------------------------------------------------------- 关掉动效
|
||||
|
||||
console.log('=== 系统关了动效 ===')
|
||||
|
||||
+760
-45
@@ -13,9 +13,19 @@
|
||||
// 3. 服务重启：stream.end status=interrupted 且没有待审批，这一轮收成
|
||||
// 「服务重启，这一轮中断了」，给接着跑，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. 已取消的轮次没有「接着跑」，续跑被拒时说人话；
|
||||
// 5. 已取消的、人驳回这类原样续上还会失败的轮次没有「接着跑」，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。
|
||||
// 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：核对撞上 500 有出路、补交付不走实时计时、查到的终态落库；
|
||||
// 9. 长答案没取全要记进 meta，刷新后补全；2000 字按码点数，不按 UTF-16；
|
||||
// 10. 发起运行失败：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；断在建图阶段按最后一次落库算有没有动静，发起运行时 meta 跟上新运行；
|
||||
// 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「在现有图上改 | 从头生成」分段控件；助手出错时同一句报错只说一遍。
|
||||
//
|
||||
// 所有写请求都被拦在浏览器里，不碰任何库。
|
||||
import { chromium } from '../frontend/node_modules/playwright-core/index.mjs'
|
||||
@@ -45,12 +55,17 @@ const check = (name, cond, detail = '') => {
|
||||
// ---------------------------------------------------------------- 伪造的数据
|
||||
|
||||
const iso = (minsAgo) => new Date(Date.now() - minsAgo * 60_000).toISOString()
|
||||
const conv = (id, title, turns, minsAgo) => ({
|
||||
const YESTERDAY_NOON = Math.round((Date.now() - new Date(new Date().setHours(12, 0, 0, 0) - 86_400_000).getTime()) / 60_000)
|
||||
const conv = (id, title, turns, minsAgo, extra = {}) => ({
|
||||
id, title, kind: 'chat', workflow_id: null, archived: false,
|
||||
created_at: iso(minsAgo + 5), last_active_at: iso(minsAgo), 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,
|
||||
@@ -70,16 +85,40 @@ const CONVS = {
|
||||
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),
|
||||
histF: conv('c0histf', '会话 F：取消过、失败过、不可用的历史', 4, 60 * 30),
|
||||
// 「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),
|
||||
restI: conv('c0resti', '会话 I：跑到一半服务重启', 0, 60 * 24 * 12),
|
||||
buildJ: conv('c0buildj', '会话 J：建流程就失败', 0, 60 * 24 * 13),
|
||||
waitK: conv('c0waitk', '会话 K：停在审批上', 0, 60 * 24 * 14),
|
||||
writeL: conv('c0writel', '会话 L：模型正在出字', 0, 60 * 24 * 15),
|
||||
// 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),
|
||||
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', 5, 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),
|
||||
}
|
||||
/** 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' }),
|
||||
}
|
||||
const DETAIL = {
|
||||
c0busya: [],
|
||||
@@ -93,6 +132,8 @@ const DETAIL = {
|
||||
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' }),
|
||||
// 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
|
||||
@@ -101,8 +142,8 @@ const DETAIL = {
|
||||
review: {
|
||||
verdict: 'annotated', note: 'agent 用满了 12 步还没给出结论。', answer: null, retry: true, severity: 'broken',
|
||||
signals: [{ kind: 'step_limit', detail: 'agent 用满了 12 步', severity: 'broken' }],
|
||||
meta: { v: 1, runId: 'run-broken', runClass: 'exploratory', runStatus: 'succeeded', queries: 3, ms: 41000 },
|
||||
},
|
||||
meta: { v: 1, runId: 'run-broken', runClass: 'exploratory', runStatus: 'succeeded', queries: 3, ms: 41000 },
|
||||
}),
|
||||
],
|
||||
c0emptg: [],
|
||||
@@ -115,37 +156,129 @@ const DETAIL = {
|
||||
// 三天前服务重启打断、跑了 42 秒的那一轮
|
||||
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),
|
||||
review: { meta: { v: 1, runId: 'run-susp', runStatus: 'interrupted', outcome: 'suspended', ms: 42000 } },
|
||||
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 review.meta 里：刷新后得从这里读回来，
|
||||
// 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',
|
||||
review: { 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' }),
|
||||
],
|
||||
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 用满了 12 步还没给出结论。', answer: null, retry: true, severity: 'broken',
|
||||
signals: [{ kind: 'step_limit', detail: 'agent 用满了 12 步', 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) })],
|
||||
}
|
||||
const RUNS = {
|
||||
'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-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 },
|
||||
'run-t-failed': { status: 'failed', output: {}, error: '鉴权没通过（401）：对方拒绝了这把密钥' },
|
||||
'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 },
|
||||
}
|
||||
const FULL = '这是完整答案的开头。' + '正文'.repeat(1920) + '【完整结尾】' // 3,850 字左右
|
||||
const SOURCES = [
|
||||
{ id: 'ds-mig', name: 'shop', kind: 'mysql', description: '' },
|
||||
{ id: 'ds-bi', name: 'bi', kind: 'oracle', description: '示例 BI 库：采购、物料、生产、设备' },
|
||||
{ id: 'ds-plant', name: 'factory', kind: 'oracle', description: '示例工厂库：采购、物料、生产、设备' },
|
||||
]
|
||||
const TABLES = {
|
||||
'ds-mig': ['__drizzle_migrations', '_migrations', 'admin_audit_log', 'app_log', 'test_sales', 'user', 'orders'],
|
||||
'ds-bi': ['ANALYTICS.test_sales', 'ANALYTICS.job_etl_audit', 'ANALYTICS.v_demo_table'],
|
||||
'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: [] }
|
||||
const ctl = { slowMs: 0, failE: true, wsClosed: {}, approvals: [], thinkGo: false, writingGo: false, writingAt: 0 }
|
||||
const log = {
|
||||
patches: [], cancels: [], continues: [], runStarts: [], archived: [], posts: [], approvalGets: [], runGets: [],
|
||||
deleted: [], gens: [],
|
||||
}
|
||||
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,
|
||||
/** 回收站里多摆两个（TRASH_MORE） */
|
||||
trashMore: false,
|
||||
/** 数据源列表取不回来 */
|
||||
sourcesFail: 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) : [])]
|
||||
.filter((c) => !db.purged.has(c.id))
|
||||
.map((c) => ({ ...c, archived: db.archived.has(c.id) }))
|
||||
let runSeq = 0
|
||||
|
||||
async function fakeApi(route) {
|
||||
@@ -156,18 +289,36 @@ async function fakeApi(route) {
|
||||
const json = (body, status = 200) => route.fulfill({ status, json: body })
|
||||
const body = () => { try { return req.postDataJSON() } catch { return {} } }
|
||||
|
||||
if (path === '/conversations' && method === 'GET') return json(Object.values(CONVS))
|
||||
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 = Object.values(CONVS).find((c) => c.id === id)
|
||||
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)
|
||||
return json({ ...known, turns: DETAIL[id] ?? [] })
|
||||
const turns = typeof DETAIL[id] === 'function' ? DETAIL[id]() : DETAIL[id]
|
||||
return json({ ...(known ?? conv('c0new', '新对话', 0, 0)), turns: 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: '' })
|
||||
}
|
||||
if (m && method === 'PATCH') { log.archived.push([m[1], body()]); return json({ ...CONVS[m[1]], ...body() }) }
|
||||
m = path.match(/^\/conversations\/([^/]+)\/turns$/)
|
||||
if (m && method === 'POST') {
|
||||
log.posts.push(m[1])
|
||||
@@ -177,7 +328,12 @@ async function fakeApi(route) {
|
||||
if (m && method === 'PATCH') { log.patches.push({ conv: m[1], turn: m[2], body: body() }); return json(turn(m[2], '', body())) }
|
||||
|
||||
if (path === '/copilot/generate-stream' && method === 'POST') {
|
||||
if (String(body().instruction ?? '').includes('建流程就失败')) {
|
||||
const b = body()
|
||||
log.gens.push(b)
|
||||
if (ctl.scope400 && b.datasource_ids?.length) {
|
||||
return json({ detail: '限定的数据源都不在了：可能已经被删掉或停用。去掉限定再问，或者到「数据」页确认它还在、而且是启用的' }, 400)
|
||||
}
|
||||
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 },
|
||||
@@ -199,12 +355,16 @@ async function fakeApi(route) {
|
||||
body: ops.map((o) => `data: ${JSON.stringify(o)}\n\n`).join('') })
|
||||
}
|
||||
if (path === '/copilot/review' && method === 'POST') {
|
||||
const wait = ctl.reviewDelay[body().run_id]
|
||||
if (wait) await new Promise((r) => setTimeout(r, wait))
|
||||
return json({ verdict: 'ok', note: '', answer: null, retry: false, severity: '', signals: [] })
|
||||
}
|
||||
if (path === '/runs' && method === 'POST') {
|
||||
const b = body()
|
||||
const conversation = String(b.input?.question ?? '')
|
||||
const id = conversation.includes('长答案') ? 'run-long'
|
||||
// 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 route.abort('connectionrefused')
|
||||
const id = conversation.includes('长答案') ? (conversation.includes('取不到运行') ? 'run-noget' : 'run-long')
|
||||
: conversation.includes('审批') ? 'run-wait'
|
||||
: conversation.includes('重启') ? 'run-restart'
|
||||
: conversation.includes('正在写') ? 'run-writing'
|
||||
@@ -226,6 +386,9 @@ async function fakeApi(route) {
|
||||
const id = m[1]
|
||||
log.runGets.push(id)
|
||||
if (id === 'run-long') return json({ id, status: 'succeeded', run_class: 'exploratory', output: { answer: FULL }, usage: {} })
|
||||
if (id === 'run-noget' || (id === 'run-unk' && ctl.run500)) return json({ detail: '后端出错了，详情看服务日志' }, 500)
|
||||
if (id === 'run-unk') return json({ id, status: 'cancelled', run_class: 'exploratory', output: {}, usage: {} })
|
||||
if (id.startsWith('run-404')) return json({ detail: '运行不存在' }, 404)
|
||||
if (RUNS[id]) return json({ id, run_class: 'exploratory', usage: {}, ...RUNS[id] })
|
||||
if (id.startsWith('run-c') || id === 'run-b1') return json({ id, status: 'succeeded', run_class: 'exploratory', output: {}, usage: {} })
|
||||
return json({ id, status: 'interrupted', run_class: 'exploratory', output: {}, usage: {} })
|
||||
@@ -240,7 +403,9 @@ async function fakeApi(route) {
|
||||
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') return json(SOURCES)
|
||||
if (path === '/datasources' && method === 'GET') {
|
||||
return ctl.sourcesFail ? json({ detail: '数据源列表暂时取不到' }, 500) : json(SOURCES)
|
||||
}
|
||||
m = path.match(/^\/datasources\/([^/]+)\/schema$/)
|
||||
if (m) return json({ tables: TABLES[m[1]] ?? [] })
|
||||
|
||||
@@ -251,6 +416,17 @@ async function fakeApi(route) {
|
||||
/** 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-long') {
|
||||
return {
|
||||
events: [
|
||||
@@ -315,7 +491,7 @@ function script(runId, after = 0) {
|
||||
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 bi 里有哪些表。' }),
|
||||
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。' }),
|
||||
@@ -414,8 +590,8 @@ for (const theme of THEMES) {
|
||||
const hints = await page.locator('button[title="填进输入框，改一改再发"]').allInnerTexts()
|
||||
check('有例句', hints.length > 0, hints.join(' | '))
|
||||
check('例句里没有系统表',
|
||||
hints.every((h) => !/__drizzle|_migrations|test_|etl_audit|_log\b/.test(h)), hints.join(' | '))
|
||||
check('描述里的业务词造了句', hints.some((h) => /bi：/.test(h)), hints.join(' | '))
|
||||
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()
|
||||
@@ -550,11 +726,16 @@ for (const theme of THEMES) {
|
||||
check('落库的也是完整答案', saved.length > 0 && saved.every((p) => p.body.answer.length === FULL.length),
|
||||
saved.map((p) => p.body.answer.length).join(','))
|
||||
// 落库是排队异步发的，交付之后再等它一下
|
||||
await until(() => log.patches.slice(patchesBefore).some((p) => p.body.review?.meta))
|
||||
const metas = log.patches.slice(patchesBefore).map((p) => p.body.review?.meta).filter(Boolean)
|
||||
await until(() => log.patches.slice(patchesBefore).some((p) => p.body.meta))
|
||||
const metas = log.patches.slice(patchesBefore).map((p) => p.body.meta).filter(Boolean)
|
||||
const meta = metas.at(-1)
|
||||
check('可信度信息跟着落库（运行、类别、查库次数）', meta?.runId === 'run-long' && meta?.runClass === 'exploratory',
|
||||
JSON.stringify(metas))
|
||||
// 后端有了单独的 meta 列：review 只放复核结论，不再夹带
|
||||
check('meta 单独写，不再塞进 review', log.patches.slice(patchesBefore)
|
||||
.every((p) => !p.body.review || !('meta' in p.body.review)),
|
||||
JSON.stringify(log.patches.slice(patchesBefore).map((p) => Object.keys(p.body.review ?? {}))))
|
||||
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】'))
|
||||
check('没有运行时报错', errors.length === 0, errors[0] ?? '')
|
||||
@@ -569,8 +750,9 @@ for (const theme of THEMES) {
|
||||
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)
|
||||
const spinning = await page.locator('main .animate-spin, [role="main"] .animate-spin').count()
|
||||
+ await page.locator('.animate-spin').count()
|
||||
// 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「接着跑」', await page.getByRole('button', { name: '接着跑' }).count() === 1)
|
||||
check('给「重跑这一轮」', await page.getByRole('button', { name: '重跑这一轮' }).count() === 1)
|
||||
@@ -627,10 +809,14 @@ for (const theme of THEMES) {
|
||||
check('失败那一轮按运行状态给「接着跑」', turns[1].runStatus === 'failed'
|
||||
&& await page.locator('[data-remedy="failed"]').getByRole('button', { name: '接着跑' }).count() === 1)
|
||||
const failedCard = await page.locator('[data-turn="tf2"]').innerText()
|
||||
check('失败原因说人话、不露 Python 类名', failedCard.includes('查询超时') && !failedCard.includes('KeyError'),
|
||||
failedCard.split('\n').slice(0, 4).join(' / '))
|
||||
// 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「接着跑」（原样续上只会再被驳回）', await rejected.count() === 1
|
||||
&& !(await rejected.getByRole('button', { name: '接着跑' }).count()), (await rejected.innerText().catch(() => '')).replace(/\n/g, ' / '))
|
||||
check('没查库的老答案标出来了', await shows(page, '这一条没有查库'))
|
||||
const stepBtn = page.getByRole('button', { name: /放宽步数重跑（12 → 24 步）/ })
|
||||
check('步数用满给出「放宽步数重跑」并写明目标值', await stepBtn.count() === 1)
|
||||
@@ -734,7 +920,7 @@ for (const theme of THEMES) {
|
||||
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 bi 里有哪些表'))
|
||||
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` })
|
||||
@@ -782,7 +968,8 @@ for (const theme of THEMES) {
|
||||
await ctx.close()
|
||||
}
|
||||
|
||||
if (section('list', '会话列表：分组、相对时间、完整标题、筛选')) {
|
||||
if (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')) {
|
||||
resetDb()
|
||||
const { page, ctx, errors } = await open(theme)
|
||||
await goto(page, 'c0idleb')
|
||||
const aside = page.locator('aside[aria-label="对话列表"]')
|
||||
@@ -793,6 +980,16 @@ for (const theme of THEMES) {
|
||||
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: '按标题或问过的话找对话' })
|
||||
await box.fill('长会话')
|
||||
const hits = await aside.locator('div.group').allInnerTexts()
|
||||
@@ -802,24 +999,69 @@ for (const theme of THEMES) {
|
||||
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('没有标题或问题里带「没有这样的对话」的对话').count() === 1)
|
||||
await box.fill('')
|
||||
await page.screenshot({ path: `${SHOTS}/list-status-${theme}.png` })
|
||||
check('没有运行时报错', errors.length === 0, errors[0] ?? '')
|
||||
await ctx.close()
|
||||
}
|
||||
|
||||
if (section('delete', '删除：确认，且可以撤销')) {
|
||||
if (section('filter', '筛选框：有筛选词时不跟着列表变短一起消失')) {
|
||||
resetDb()
|
||||
ctl.few = true
|
||||
const { page, ctx, errors } = await open(theme)
|
||||
await goto(page, 'c0busya')
|
||||
const aside = page.locator('aside[aria-label="对话列表"]')
|
||||
const box = aside.getByRole('searchbox', { name: '按标题或问过的话找对话' })
|
||||
check('8 个对话时有筛选框', await box.waitFor({ timeout: 6000 }).then(() => true, () => false))
|
||||
await box.fill('会话 B')
|
||||
const row = aside.locator('div.group', { hasText: '会话 B' })
|
||||
await row.hover()
|
||||
await row.getByRole('button', { name: /删除/ }).click()
|
||||
// 老版本删除前还要确认一次
|
||||
if (await page.getByRole('dialog').waitFor({ timeout: 800 }).then(() => true, () => false)) {
|
||||
await page.getByRole('dialog').getByRole('button', { name: '删除' }).click()
|
||||
}
|
||||
await until(() => db.archived.has('c0idleb'))
|
||||
await page.waitForTimeout(300)
|
||||
check('删到 7 个：筛选词还在，筛选框也还在', await box.isVisible().catch(() => false)
|
||||
&& (await box.inputValue().catch(() => '')) === '会话 B')
|
||||
await page.screenshot({ path: `${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()
|
||||
}
|
||||
|
||||
if (section('delete', '删除：先进回收站，可以撤销')) {
|
||||
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()
|
||||
check('删除前先确认', await page.getByRole('dialog').waitFor({ timeout: 3000 }).then(() => true, () => false))
|
||||
await page.getByRole('dialog').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))
|
||||
check('给了撤销', await page.getByRole('button', { name: '撤销' }).first()
|
||||
.waitFor({ timeout: 4000 }).then(() => true, () => false))
|
||||
await page.getByRole('button', { name: '撤销' }).first().click()
|
||||
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)
|
||||
@@ -827,6 +1069,479 @@ for (const theme of THEMES) {
|
||||
check('没有运行时报错', errors.length === 0, errors[0] ?? '')
|
||||
await ctx.close()
|
||||
}
|
||||
|
||||
if (section('trash', '回收站：预览、恢复、彻底删除（要确认）')) {
|
||||
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, '这个对话在回收站里'))
|
||||
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('确认框说清找不回来', (await dialog.innerText().catch(() => '')).includes('找不回来'))
|
||||
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, '回收站是空的'))
|
||||
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()
|
||||
}
|
||||
|
||||
if (section('readonly', '回收站里只能看：不发起运行，正在跑的不给彻底删除')) {
|
||||
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 = /接着跑|重试这一轮|重跑这一轮|重新问一次|不限数据源重试|放宽步数|跑一下/
|
||||
// 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(/跑一下/) && has(/接着跑/) && has(/重新问一次|重试这一轮/) && 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('看执行过程照常能用', await page.getByRole('button', { name: '看执行过程' }).first().isEnabled().catch(() => false))
|
||||
check('停在审批上的：不摆能点「通过」的审批卡，说清恢复之后再处理',
|
||||
!(await page.getByRole('button', { name: '通过' }).count())
|
||||
&& (await page.locator('[data-pending-approval]').innerText().catch(() => '')).includes('恢复这个对话之后'))
|
||||
// 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、不接着跑',
|
||||
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-${theme}.png` })
|
||||
|
||||
// 删掉之后又跑起来的：和列表里一样不给彻底删除
|
||||
await page.locator('aside').getByRole('button', { name: /^打开回收站/ }).click()
|
||||
const bin = page.locator('aside[aria-label="回收站"]')
|
||||
const liveRow = bin.locator('div.group', { hasText: '删掉的对话丁' })
|
||||
await liveRow.waitFor({ timeout: 3000 }).catch(() => {})
|
||||
check('回收站里标出它在运行', (await liveRow.innerText().catch(() => '')).includes('运行中'),
|
||||
(await liveRow.innerText().catch(() => '')).replace(/\n/g, ' / '))
|
||||
await liveRow.hover()
|
||||
const rowPurge = liveRow.getByRole('button', { name: /彻底删除/ })
|
||||
check('正在跑的：行上的彻底删除点不了，说清为什么', await rowPurge.isDisabled().catch(() => false)
|
||||
&& (await rowPurge.getAttribute('title')) === '这个对话正在运行，先停下来再删', await rowPurge.getAttribute('title').catch(() => ''))
|
||||
await liveRow.locator('button').first().click()
|
||||
const live = await page.waitForFunction(() => window.__chat.getState().byConversation.c0trash4?.[0]?.phase === 'running',
|
||||
null, { timeout: 6000 }).then(() => true, () => false)
|
||||
check('打开它：核对发现运行还活着，接了回去', live, (await chatState(page)).byConversation.c0trash4?.[0]?.phase)
|
||||
const bannerPurge = page.locator('[data-trash-banner]').getByRole('button', { name: '彻底删除' })
|
||||
check('提示条上的彻底删除也点不了', await bannerPurge.isDisabled().catch(() => false))
|
||||
const stop = page.getByRole('button', { name: '停止这一轮' })
|
||||
check('在回收站里也停得下来', await stop.count() === 1)
|
||||
await page.waitForTimeout(300)
|
||||
await page.screenshot({ path: `${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()
|
||||
}
|
||||
|
||||
if (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')) {
|
||||
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, '没核对上这一轮现在的状态'))
|
||||
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: /重试这一轮|重跑这一轮|接着跑/ }).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：各自的终态查到之后要落库，下次进来不再核对
|
||||
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
|
||||
await goto(page, 'c0termr')
|
||||
await until(() => ['tr1', 'tr2', 'tr3', 'tr4'].every((id) =>
|
||||
log.patches.slice(patchesBefore).some((p) => p.turn === id && p.body.status === 'error')), 6000)
|
||||
await page.waitForTimeout(300)
|
||||
const saved = (id) => log.patches.slice(patchesBefore).filter((p) => p.turn === id).at(-1)?.body
|
||||
check('在别处取消的：落库成已取消', saved('tr1')?.status === 'error' && saved('tr1')?.error === '已取消'
|
||||
&& saved('tr1')?.meta?.outcome === 'cancelled', JSON.stringify(saved('tr1')))
|
||||
check('在别处失败的：落库成失败，原因跟着写', saved('tr2')?.status === 'error' && /鉴权/.test(saved('tr2')?.error ?? ''),
|
||||
JSON.stringify(saved('tr2'))?.slice(0, 160))
|
||||
check('服务重启挂起的：落库成中断', saved('tr3')?.status === 'error' && saved('tr3')?.meta?.outcome === 'suspended',
|
||||
JSON.stringify(saved('tr3'))?.slice(0, 160))
|
||||
check('运行被删掉的：落库成「上次没有跑完」', saved('tr4')?.status === 'error' && saved('tr4')?.error === '上次没有跑完',
|
||||
JSON.stringify(saved('tr4'))?.slice(0, 160))
|
||||
check('还在等审批的不落库（它还没结束）', !saved('tr5'), JSON.stringify(saved('tr5')))
|
||||
ctl.approvals = []
|
||||
check('没有运行时报错', errors.length === 0, errors[0] ?? '')
|
||||
await ctx.close()
|
||||
}
|
||||
|
||||
if (section('clipped', '长答案没取全：记进 meta、刷新后补全，2000 字按码点数')) {
|
||||
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)
|
||||
check('GET 运行失败：先交截断的那份并标出来', t.clipped === 'partial', t.clipped)
|
||||
await until(() => log.patches.slice(patchesBefore).some((p) => p.body.meta?.clipped === 'partial'))
|
||||
check('「没取全」写进 meta，刷新后还认得出来', log.patches.slice(patchesBefore).some((p) => p.body.meta?.clipped === 'partial'),
|
||||
JSON.stringify(log.patches.slice(patchesBefore).map((p) => p.body.meta?.clipped)))
|
||||
|
||||
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 · ${tt1.clipped}`)
|
||||
await page.waitForTimeout(300)
|
||||
const lost = await page.locator('[data-remedy]', { hasText: '截在了 2000 字' }).count()
|
||||
check('恰好 2000 个码点（UTF-16 不是 2000）也认得出被截断、补不回来', lost === 2, `${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()
|
||||
}
|
||||
|
||||
if (section('launch', '发起运行失败：报错分得清标题和原因，补救说真话')) {
|
||||
const { page, ctx, errors } = await open(theme)
|
||||
ctl.launchFail = true
|
||||
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「连不上后端服务」，原因不糊进粗体', title === '连不上后端服务', 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 rerun = page.locator('[data-remedy="failed"]').getByRole('button', { name: '重跑这一轮' })
|
||||
check('流程已经搭好：给「重跑这一轮」，不让它重新搭', await rerun.count() === 1)
|
||||
check('说出怎么办：连上之后点「重跑这一轮」', text.includes('重跑这一轮'), text.replace(/\n/g, ' / '))
|
||||
await page.screenshot({ path: `${SHOTS}/launch-failed-${theme}.png` })
|
||||
ctl.launchFail = false
|
||||
const gensBefore = log.gens.length
|
||||
const startsBefore = log.runStarts.length
|
||||
await rerun.click().catch(() => {})
|
||||
await until(() => log.runStarts.length > startsBefore, 5000)
|
||||
check('重跑：原图直接发起，不再去问 Copilot', log.runStarts.length > startsBefore && log.gens.length === gensBefore,
|
||||
`${log.runStarts.length - startsBefore} 次发起 · ${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()
|
||||
}
|
||||
|
||||
if (section('rerun', '重跑服务重启挂起的那一轮：旧运行顺手取消')) {
|
||||
const { page, ctx, errors } = await open(theme)
|
||||
await goto(page, 'c0suspo')
|
||||
await shows(page, '服务重启，这一轮中断了')
|
||||
const before = log.cancels.length
|
||||
const startsBefore = log.runStarts.length
|
||||
await page.locator('[data-remedy="suspended"]').getByRole('button', { name: '重跑这一轮' }).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(','))
|
||||
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()
|
||||
}
|
||||
|
||||
if (section('stale', '断在建图阶段：按最后一次落库算有没有动静；发起运行时 meta 跟上新运行')) {
|
||||
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 until(() => log.patches.slice(patchesBefore).some((p) => p.turn === 'tx8'), 3000)
|
||||
await page.waitForTimeout(400)
|
||||
const mine = () => log.patches.slice(patchesBefore)
|
||||
check('两天前问的、两分钟前刚重试过：不当成断了写回 error',
|
||||
!mine().some((p) => p.turn === 'tx7' && p.body.status === 'error'), JSON.stringify(mine().filter((p) => p.turn === 'tx7')))
|
||||
check('重试之后二十分钟没动静的：认定断了，写回 error',
|
||||
mine().some((p) => p.turn === 'tx8' && p.body.status === 'error'), JSON.stringify(mine().filter((p) => p.turn === 'tx8')).slice(0, 200))
|
||||
|
||||
// 重试一次：搭好图时连 meta 一起写（有动静的凭据），发起运行时 meta.runId 换成新运行
|
||||
const startsBefore = log.runStarts.length
|
||||
const retryAt = log.patches.length
|
||||
await page.locator('[data-remedy="failed"]').getByRole('button', { name: '重试这一轮' }).click()
|
||||
await until(() => log.runStarts.length > startsBefore, 6000)
|
||||
const newRun = log.runStarts.at(-1)?.id
|
||||
await until(() => log.patches.slice(retryAt).some((p) => p.turn === 'tx8' && p.body.run_id === newRun), 3000)
|
||||
const ps = log.patches.slice(retryAt).filter((p) => p.turn === 'tx8').map((p) => p.body)
|
||||
const graphSave = ps.find((b) => b.graph)
|
||||
check('搭好图时连 meta 一起写，带上此刻', !!graphSave?.meta && Math.abs(Date.now() - (graphSave.meta.at ?? 0)) < 30_000,
|
||||
JSON.stringify(graphSave?.meta ?? null)?.slice(0, 120))
|
||||
const started = ps.find((b) => b.run_id === newRun)
|
||||
check('发起运行：run_id 列和 meta.runId 一起换成新运行，状态是 running',
|
||||
!!newRun && started?.meta?.runId === newRun && started?.status === 'running', JSON.stringify(started ?? null)?.slice(0, 160))
|
||||
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()
|
||||
}
|
||||
|
||||
if (section('scope', '限定数据源：点一下只查这个库')) {
|
||||
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-${theme}.png` })
|
||||
await page.keyboard.press('Escape')
|
||||
const before2 = log.gens.length
|
||||
await send(page, '在对话里限定的问题')
|
||||
await until(() => log.gens.length > before2, 5000)
|
||||
check('对话里限定的也带上', JSON.stringify(log.gens.at(-1)?.datasource_ids) === '["ds-mig"]',
|
||||
JSON.stringify(log.gens.at(-1)?.datasource_ids))
|
||||
await page.getByRole('button', { name: '停止这一轮' }).click().catch(() => {})
|
||||
await page.waitForTimeout(200)
|
||||
|
||||
// 限定的库被删了、停用了：后端 400，把原因原样说出来
|
||||
ctl.scope400 = true
|
||||
await send(page, '限定的库已经不在了')
|
||||
check('库不在了：说清原因', await shows(page, '限定的数据源都不在了'))
|
||||
await page.screenshot({ path: `${SHOTS}/scope-400-${theme}.png` })
|
||||
// 带着原范围的「重试这一轮」只会再撞一次 400：补救得是放开范围
|
||||
const loose = page.locator('[data-remedy="failed"]').getByRole('button', { name: '不限数据源重试' })
|
||||
check('库不在了：补救是「不限数据源重试」', await loose.count() === 1)
|
||||
const before3 = log.gens.length
|
||||
await loose.click().catch(() => {})
|
||||
await until(() => log.gens.length > before3, 5000)
|
||||
check('不限数据源重试：请求里不再带 datasource_ids', log.gens.length > before3 && !log.gens.at(-1)?.datasource_ids?.length,
|
||||
JSON.stringify(log.gens.at(-1)?.datasource_ids))
|
||||
check('输入框旁的范围也一起放开', (await picker.innerText().catch(() => '')).includes('全部数据源'),
|
||||
await picker.innerText().catch(() => ''))
|
||||
await page.getByRole('button', { name: '停止这一轮' }).click().catch(() => {})
|
||||
ctl.scope400 = false
|
||||
|
||||
// 库列表没取回来：从上一问接过来的范围照样会发出去，所以照样得摆在眼前
|
||||
ctl.sourcesFail = true
|
||||
await goto(page, 'c0scopw')
|
||||
await shows(page, '上一问的答案 W')
|
||||
const picker2 = page.getByRole('button', { name: /数据源范围/ })
|
||||
await picker2.waitFor({ timeout: 3000 }).catch(() => {})
|
||||
check('库列表没取回来：范围照样摆在输入框旁', (await picker2.innerText().catch(() => '')).includes('只查 factory'),
|
||||
await picker2.innerText().catch(() => '没有范围按钮'))
|
||||
await picker2.click().catch(() => {})
|
||||
const stray = page.locator('[data-scope-stray]')
|
||||
check('范围里的库照样列出来，注明没取到状态', (await stray.innerText().catch(() => '')).includes('没取到状态')
|
||||
&& await stray.getByRole('checkbox').isChecked().catch(() => false), await stray.innerText().catch(() => ''))
|
||||
await page.waitForTimeout(250)
|
||||
await page.screenshot({ path: `${SHOTS}/scope-unconfirmed-${theme}.png` })
|
||||
await stray.getByRole('checkbox').uncheck().catch(() => {})
|
||||
check('取消勾选之后它还留在原处，按钮写回「全部数据源」', await stray.count() === 1
|
||||
&& (await picker2.innerText().catch(() => '')).includes('全部数据源'), await picker2.innerText().catch(() => ''))
|
||||
await stray.getByRole('checkbox').check().catch(() => {})
|
||||
await page.keyboard.press('Escape')
|
||||
const before4 = log.gens.length
|
||||
await send(page, '库列表没取回来时的问题')
|
||||
await until(() => log.gens.length > before4, 5000)
|
||||
check('发出去的范围和看得见的一致', JSON.stringify(log.gens.at(-1)?.datasource_ids) === '["ds-plant"]',
|
||||
JSON.stringify(log.gens.at(-1)?.datasource_ids))
|
||||
await page.getByRole('button', { name: '停止这一轮' }).click().catch(() => {})
|
||||
ctl.sourcesFail = false
|
||||
check('没有运行时报错', errors.length === 0, errors[0] ?? '')
|
||||
await ctx.close()
|
||||
}
|
||||
|
||||
if (section('composer', '画布助手：「在现有图上改 | 从头生成」分段控件')) {
|
||||
const { page, ctx, errors } = await open(theme)
|
||||
const WF = {
|
||||
id: 'wf-chat-check', name: '__chat_check_composer__', description: '检查脚本伪造的工作流', graph: GRAPH,
|
||||
tags: [], version: 1, is_template: false, status: 'draft', published_version: null, published_by: null,
|
||||
run_count: 0, created_at: iso(60), updated_at: iso(60),
|
||||
}
|
||||
await page.route(/\/api\/workflows(\?.*)?$/, (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 page.waitForFunction(() => window.__studio?.getState().workflow?.id === 'wf-chat-check', null, { timeout: 15000 })
|
||||
.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: '在现有图上改' })
|
||||
const fresh = group.getByRole('radio', { name: '从头生成' })
|
||||
check('默认「在现有图上改」', (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, `${size}px`)
|
||||
await base.focus()
|
||||
await page.keyboard.press('ArrowRight')
|
||||
check('方向键切到「从头生成」', (await fresh.getAttribute('aria-checked')) === 'true'
|
||||
&& await fresh.evaluate((el) => el === document.activeElement))
|
||||
await page.waitForTimeout(300) // 底色有过渡，等它换完再截
|
||||
await page.screenshot({ path: `${SHOTS}/composer-fresh-${theme}.png` })
|
||||
const gensBefore = log.gens.length
|
||||
const box = page.getByRole('textbox', { name: '告诉助手要画什么流程' })
|
||||
await box.fill('重新画一张三步的图')
|
||||
await box.press('Enter')
|
||||
await until(() => log.gens.length > gensBefore, 5000)
|
||||
check('从头生成：不带现有的图', log.gens.at(-1)?.base_graph === null, JSON.stringify(log.gens.at(-1)?.base_graph)?.slice(0, 60))
|
||||
await page.waitForFunction(() => !window.__studio.getState().copilot.active, null, { timeout: 8000 }).catch(() => {})
|
||||
await base.click()
|
||||
await box.fill('在最后加一步人工审批')
|
||||
await box.press('Enter')
|
||||
await until(() => log.gens.length > gensBefore + 1, 5000)
|
||||
check('在现有图上改：带上现有的图', !!log.gens.at(-1)?.base_graph?.nodes?.length)
|
||||
|
||||
// 助手出错：出错的那一轮自己有报错块（怎么办、原文都在），输入框上面不再重复一条
|
||||
await page.waitForFunction(() => !window.__studio.getState().copilot.active, null, { timeout: 8000 }).catch(() => {})
|
||||
await box.fill('这一句会让建流程就失败')
|
||||
await box.press('Enter')
|
||||
const card = page.locator('[data-turn-error]', { hasText: '模型交回的工作流结构不对' })
|
||||
const shown = await card.first().waitFor({ timeout: 6000 }).then(() => true, () => false)
|
||||
const said = await card.first().innerText().catch(() => '')
|
||||
check('助手出错：那一轮写出发生了什么和怎么办', shown && said.includes('再试一次，或者把需求说得更具体些'),
|
||||
said.replace(/\n/g, ' / '))
|
||||
check('同一句报错只说一遍（输入框上面不再摆一条）', await card.count() === 1
|
||||
&& await page.locator('[data-copilot-error]').count() === 0,
|
||||
`${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-${theme}.png` })
|
||||
check('没有运行时报错', errors.length === 0, errors[0] ?? '')
|
||||
await ctx.close()
|
||||
}
|
||||
}
|
||||
|
||||
await browser.close()
|
||||
|
||||
@@ -592,6 +592,14 @@ console.log('\n=== 长运行：按轮折叠（runs-5）===')
|
||||
check('相邻重复行并成一行', rows.length === 2, rows.map((r) => r.title).join(' | '))
|
||||
check('并后的行写出次数、中位和最慢', rows[0].meta === '×3 · 中位 12 ms · 最慢 63 ms', rows[0].meta)
|
||||
check('警告并完还是警告', rows[1].level === 'warn' && rows[1].repeat?.count === 2)
|
||||
// 出了状况的行，副标题说的是为什么：两次修复凑出来的值不一样，并成一行就只剩第一次的
|
||||
const why = mod.compactSteps([
|
||||
{ id: 'r1', seq: 1, kind: 'llm', title: '让模型修复格式', level: 'warn', status: 'done', sub: '出现了原文没有的值（total_count=0）' },
|
||||
{ id: 'r2', seq: 2, kind: 'llm', title: '让模型修复格式', level: 'warn', status: 'done', sub: '出现了原文没有的值（region=华东）' },
|
||||
{ id: 'r3', seq: 3, kind: 'llm', title: '让模型修复格式', level: 'warn', status: 'done', sub: '出现了原文没有的值（region=华东）' },
|
||||
])
|
||||
check('原因不同的警告不并成一行，原因相同的照并', why.length === 2 && why[1].repeat?.count === 2,
|
||||
why.map((r) => `${r.sub} ${r.meta ?? ''}`).join(' | '))
|
||||
|
||||
// 恢复后的重放（resumed）不算新的一轮；循环体真的又跑一次才算
|
||||
const replay = mod.decodeRun([
|
||||
@@ -654,6 +662,180 @@ console.log('\n=== Copilot：心跳穿插、回话、少了一步、报错 ===')
|
||||
check('结局里的报错分开带着 hint 和 detail', eo.kind === 'error' && eo.error?.hint === 'h' && eo.error?.raw === 'd')
|
||||
}
|
||||
|
||||
console.log('\n=== 协作团队用完轮数：最后那次判定不是新的一轮（REQ-3A-3、NI-5）===')
|
||||
{
|
||||
const team = (evts) => mod.decodeRun(evts).find((s) => s.nodeId === 'team')
|
||||
const routes = (evts) => flatten(mod.decodeRun(evts)).filter((s) => s.kind === 'branch' && s.nodeId === 'team')
|
||||
|
||||
// 以前 closing 判定被写成「第 3 轮：调度者在想下一步…」→「第 3 轮：结束协作」，done=false 时
|
||||
// 读起来像团队正常收尾了；泳道还多出一列空的第 3 轮
|
||||
const failRoutes = routes(synthetic.exhaustedTeam('fail'))
|
||||
const closing = failRoutes[failRoutes.length - 1]
|
||||
check('判定没完成：说「轮数用完 · 调度者判定：未完成」并带理由',
|
||||
closing?.title === '轮数用完 · 调度者判定：未完成（还没有查到任何订单数据）', closing?.title)
|
||||
check('判定没完成的那一行是提醒色', closing?.level === 'warn', closing?.level)
|
||||
check('不再把判定写成「第 3 轮」', !failRoutes.some((s) => s.title.includes('第 3 轮')),
|
||||
failRoutes.map((s) => s.title).join(' | '))
|
||||
const judged = routes(synthetic.exhaustedTeam('judged'))
|
||||
check('判定完成：说「轮数用完 · 调度者判定：已完成」',
|
||||
judged[judged.length - 1]?.title === '轮数用完 · 调度者判定：已完成', judged[judged.length - 1]?.title)
|
||||
const live = routes(synthetic.exhaustedTeam('closing'))
|
||||
const pending = live[live.length - 1]
|
||||
check('还在判定时有一行进行中，不说「第 3 轮」',
|
||||
pending?.status === 'running' && pending.title === '轮数用完 · 调度者在做最后判定…', `${pending?.status} ${pending?.title}`)
|
||||
|
||||
for (const mode of ['fail', 'judged', 'degrade']) {
|
||||
const t = team(synthetic.exhaustedTeam(mode))?.team
|
||||
check(`${mode}：泳道只有派过活的 2 轮，判定不加列`, t?.rounds.length === 2, `${t?.rounds.length} 轮`)
|
||||
}
|
||||
// reduceTeam 是画布协作矩阵和右栏共用的那一份：直接喂事件也不能多出一列
|
||||
let direct
|
||||
for (const e of synthetic.exhaustedTeam('fail')) {
|
||||
const next = mod.reduceTeam(direct, e)
|
||||
if (next && e.node_id === 'team') direct = next
|
||||
}
|
||||
check('reduceTeam 直接喂也只有 2 轮', direct?.rounds.length === 2, `${direct?.rounds.length}`)
|
||||
check('reduceTeam 记下了判定结论', direct?.verdict?.closing === true && direct.verdict.done === false
|
||||
&& direct.verdict.reason === '还没有查到任何订单数据', JSON.stringify(direct?.verdict))
|
||||
// 还在判定：泳道据此写「调度者在判定」，而不是空着或多一列
|
||||
let judging
|
||||
for (const e of synthetic.exhaustedTeam('closing')) {
|
||||
const next = mod.reduceTeam(judging, e)
|
||||
if (next && e.node_id === 'team') judging = next
|
||||
}
|
||||
check('判定进行中：记下在判定、还没有结论，也不加列', judging?.verdict?.closing === true && judging.verdict.done === undefined
|
||||
&& judging.rounds.length === 2, JSON.stringify(judging?.verdict))
|
||||
const judgedTeam = team(synthetic.exhaustedTeam('judged'))?.team
|
||||
check('判定完成时团队算收尾了', judgedTeam?.finished === true && judgedTeam.verdict?.done === true)
|
||||
|
||||
// 成员把工具调用写成文字：这一步是失败，不是一个写着失败原因的「完成」格子
|
||||
const r0 = direct?.rounds[0]?.members.find((m) => m.agent === '取数员')
|
||||
check('失败的成员格子是 failed', r0?.status === 'failed', r0?.status)
|
||||
check('失败的成员带着原因', !!r0?.error?.includes('没有真正调用工具'), r0?.error)
|
||||
const r1 = direct?.rounds[1]?.members.find((m) => m.agent === '分析员')
|
||||
check('同一轮做完的成员照常是 done', r1?.status === 'done', r1?.status)
|
||||
const memberRow = flatten(mod.decodeRun(synthetic.exhaustedTeam('fail')))
|
||||
.find((s) => s.title.startsWith('取数员：') && s.status === 'failed')
|
||||
check('失败的成员那一行也画成失败', !!memberRow && memberRow.level === 'error', memberRow?.status)
|
||||
check('失败的成员那一行说清为什么', !!memberRow?.sub?.includes('没有真正调用工具'), memberRow?.sub)
|
||||
check('一轮里有人失败时不算「省下」', direct?.savedMs === 0, `${direct?.savedMs}`)
|
||||
|
||||
// 判失败：矩阵和右栏都要说「用完 N 轮未完成」，点出从没派到的成员
|
||||
check('判失败：团队的结局是 failed', direct?.verdict?.outcome === 'failed', JSON.stringify(direct?.verdict))
|
||||
check('判失败：认出轮数', direct?.verdict?.rounds === 2, `${direct?.verdict?.rounds}`)
|
||||
check('判失败：点出从没派到的成员', direct?.verdict?.never?.join(',') === '汇总员', direct?.verdict?.never?.join(','))
|
||||
|
||||
// 降档交付：节点「完成」了，但不能是一个安静的勾
|
||||
const degraded = mod.decodeRun(synthetic.exhaustedTeam('degrade'))
|
||||
const node = degraded.find((s) => s.nodeId === 'team')
|
||||
check('降档：团队节点行是提醒色', node?.status === 'done' && node.level === 'warn', `${node?.status} ${node?.level}`)
|
||||
check('降档：团队节点行说清是降档交付', !!node?.sub?.includes('用完 2 轮') && node.sub.includes('降档'), node?.sub)
|
||||
check('降档：泳道的结局是 degraded，点出没派到的成员',
|
||||
node?.team?.verdict?.outcome === 'degraded' && node.team.verdict.never?.join(',') === '汇总员',
|
||||
JSON.stringify(node?.team?.verdict))
|
||||
const note = flatten(degraded).find((s) => s.code === 'team_exhausted')
|
||||
check('降档那条提醒说人话', note?.title === '协作团队用完 2 轮仍未完成 · 按降档交付', note?.title)
|
||||
check('降档那条提醒给出下一步', !!note?.next && note.fix === 'canvas', `${note?.next} / ${note?.fix}`)
|
||||
// 认不出轮数（老后端、措辞变了）时照实说「用完了轮数」，不拿「?」凑一个数
|
||||
const vague = flatten(mod.decodeRun([
|
||||
{ seq: 1, type: 'node.started', node_id: 'team', ts: 1, data: { node_type: 'supervisor', label: '团队' } },
|
||||
{ seq: 2, type: 'log', node_id: 'team', ts: 2, data: { level: 'warn', code: 'team_exhausted', message: '协作团队没做完，按降档交付' } },
|
||||
{ seq: 3, type: 'node.finished', node_id: 'team', ts: 3, data: { duration_ms: 5, preview: { text: 'x', exhausted: true } } },
|
||||
]))
|
||||
const vagueNote = vague.find((s) => s.code === 'team_exhausted')
|
||||
const vagueNode = vague.find((s) => s.kind === 'node' && s.nodeId === 'team')
|
||||
check('认不出轮数时不写「?」', !vagueNote?.title.includes('?') && !vagueNode?.sub?.includes('?')
|
||||
&& !!vagueNote?.title.includes('用完了轮数'), `${vagueNote?.title} / ${vagueNode?.sub}`)
|
||||
}
|
||||
|
||||
console.log('\n=== 模型把工具调用写成了文字（NI-4）===')
|
||||
{
|
||||
const all = flatten(mod.decodeRun(synthetic.markupRun()))
|
||||
const warn = all.find((s) => s.code === 'tool_markup_leak')
|
||||
check('提醒行说人话，不贴原始标记', warn?.title === '模型把工具调用写成了文字，没有真正执行'
|
||||
&& !warn.title.includes('DSML'), warn?.title)
|
||||
check('说清已经提醒它重试', !!warn?.sub?.includes('重试一次'), warn?.sub)
|
||||
check('给出下一步：去画布绑定工具', !!warn?.next?.includes('绑定') && warn.fix === 'canvas', `${warn?.next} / ${warn?.fix}`)
|
||||
check('原始标记留在展开区给排查', !!warn?.detail?.includes('DSML'))
|
||||
const team = flatten(mod.decodeRun(synthetic.exhaustedTeam('fail'))).find((s) => s.code === 'tool_markup_leak')
|
||||
check('成员写成文字时说是哪个成员', team?.title === '取数员把工具调用写成了文字，没有真正执行', team?.title)
|
||||
const node = mod.decodeRun(synthetic.markupRun()).find((s) => s.nodeId === 'query')
|
||||
check('失败的节点照常是失败，原因留在详情', node?.status === 'failed' && !!node.detail?.includes('原始标记'))
|
||||
// 收尾轮那种：真调过工具，只是步数用完了还想接着查。节点照常完成，不能说成「没有真正执行」
|
||||
const settle = flatten(mod.decodeRun([
|
||||
{ seq: 1, type: 'node.started', node_id: 'q', ts: 1, data: { node_type: 'agent', label: '数据查询' } },
|
||||
{ seq: 2, type: 'log', node_id: 'q', ts: 2, data: { level: 'warn', code: 'tool_markup_leak',
|
||||
message: '模型把工具调用写成了文字（<tool_call>{"name": "db_query__shop"…），没有真正调用工具：步数用完后的收尾轮仍想调用工具，结论只基于之前查到的部分' } },
|
||||
{ seq: 3, type: 'node.finished', node_id: 'q', ts: 3, data: { duration_ms: 9, preview: { text: '华东 1,204 单' } } },
|
||||
])).find((s) => s.code === 'tool_markup_leak')
|
||||
check('收尾轮还想查：说步数用完，不说没有真正执行', !!settle?.title.includes('步数用完') && !settle.title.includes('没有真正执行'),
|
||||
settle?.title)
|
||||
check('收尾轮还想查：下一步是调大最多步数', !!settle?.next?.includes('最多步数'), settle?.next)
|
||||
}
|
||||
|
||||
console.log('\n=== 校验修复想凑数（NI-3）===')
|
||||
{
|
||||
const all = flatten(mod.decodeRun(synthetic.repairRun()))
|
||||
const repairs = all.filter((s) => s.code === 'repair')
|
||||
check('每次修复是一行模型调用', repairs.length === 2 && repairs.every((s) => s.kind === 'llm' && s.title === '让模型修复格式'),
|
||||
repairs.map((s) => s.title).join(' | '))
|
||||
check('修复调用带着耗时', repairs[0]?.meta === '1.3 s', repairs[0]?.meta)
|
||||
check('被作废的修复折进那一行，不另起孤行', !all.some((s) => s.code === 'repair_invented'),
|
||||
all.filter((s) => s.code === 'repair_invented').map((s) => s.title).join(','))
|
||||
check('作废的修复画成提醒，说出凑出来的值', repairs.every((s) => s.level === 'warn' && s.sub?.includes('total_count=0')),
|
||||
repairs.map((s) => `${s.level} ${s.sub}`).join(' | '))
|
||||
check('作废的修复给出下一步', !!repairs[0]?.next?.includes('上游'), repairs[0]?.next)
|
||||
check('第一次校验失败那条照常显示', all.some((s) => s.code === 'validate_retry' && s.title.includes('第 1 次校验失败')))
|
||||
// 老数据：没有 purpose=repair 的 llm.end，作废说明也不能丢
|
||||
const legacy = flatten(mod.decodeRun(synthetic.repairRun().filter((e) => e.type !== 'llm.end')))
|
||||
const orphan = legacy.find((s) => s.code === 'repair_invented')
|
||||
check('没有修复行可折时单独成一行说人话', !!orphan?.title.includes('修复被作废') && orphan.level === 'warn', orphan?.title)
|
||||
}
|
||||
|
||||
console.log('\n=== 工具时限（timeout_s / timed_out）===')
|
||||
{
|
||||
const live = flatten(mod.decodeRun(synthetic.timeoutRun('live'))).find((s) => s.kind === 'query')
|
||||
check('进行中的查询带着时限', live?.limitS === 30 && live.status === 'running', `${live?.limitS} ${live?.status}`)
|
||||
const done = flatten(mod.decodeRun(synthetic.timeoutRun('done'))).find((s) => s.kind === 'query')
|
||||
check('超时的查询是失败', done?.status === 'failed' && done.level === 'error', done?.status)
|
||||
check('超时的查询说清超了多少上限、已放弃等待', done?.sub === '超过 30 s 上限，已放弃等待', done?.sub)
|
||||
check('超时的原话留在结果里', !!done?.result?.includes('缩小范围'))
|
||||
}
|
||||
|
||||
console.log('\n=== 放弃等审批的运行 ===')
|
||||
{
|
||||
const steps = mod.decodeRun(synthetic.abandonedRun())
|
||||
const row = steps.find((s) => s.kind === 'lifecycle' && s.status === 'cancelled')
|
||||
check('取消那一行说清谁放弃的、一并关了什么', !!row?.sub?.includes('张工') && row.sub.includes('1 条待审批一并关闭'), row?.sub)
|
||||
check('放弃之后审批不再是待办', !flatten(steps).some((s) => s.status === 'waiting' || s.status === 'running'))
|
||||
const plain = mod.decodeRun(synthetic.cancelledRun()).find((s) => s.kind === 'lifecycle' && s.status === 'cancelled')
|
||||
check('老数据的取消行不硬编副标题', plain && !plain.sub, plain?.sub)
|
||||
}
|
||||
|
||||
console.log('\n=== 出具那一行记着档位 ===')
|
||||
{
|
||||
const is = mod.decodeRun(synthetic.mixedRun()).find((s) => s.kind === 'issuance')
|
||||
check('出具步骤带 tier，头部能据此提醒', is?.tier === 'degraded', is?.tier)
|
||||
}
|
||||
|
||||
console.log('\n=== Copilot 改图：工具绑定变化（NI-1）===')
|
||||
{
|
||||
const steps = mod.decodeCopilot(synthetic.COPILOT_TOOLS_DROPPED, { context: 'chat' })
|
||||
const change = steps.find((s) => s.code === 'tool_changes')
|
||||
check('工具绑定变化单独成一行', !!change, steps.map((s) => s.title).join(' | '))
|
||||
check('工具少了用提醒色，说清是哪个节点', change?.level === 'warn' && change.title.includes('数据查询'), change?.title)
|
||||
check('逐项列出前后', !!change?.detail?.includes('db_query__shop') && change.detail.includes('→'), change?.detail)
|
||||
const selfCheck = steps.find((s) => s.title.startsWith('自查通过'))
|
||||
check('自查通过但有工具被去掉：自查那一行是提醒，不是安静的通过', selfCheck?.level === 'warn'
|
||||
&& selfCheck.title.includes('工具'), `${selfCheck?.level} ${selfCheck?.title}`)
|
||||
const out = mod.copilotOutcome(synthetic.COPILOT_TOOLS_DROPPED)
|
||||
check('结局里带着工具绑定变化', out.toolChanges?.length === 1 && out.toolChanges[0].removed.length === 2,
|
||||
JSON.stringify(out.toolChanges))
|
||||
check('结局里带着「工具被去掉」的警告', out.dropped?.length === 1, JSON.stringify(out.dropped))
|
||||
const plain = mod.copilotOutcome(synthetic.COPILOT_STUCK)
|
||||
check('没有变化时两样都是空的', !plain.toolChanges?.length && !plain.dropped?.length)
|
||||
}
|
||||
|
||||
console.log('\n=== 术语 ===')
|
||||
{
|
||||
// 没起名的节点退到类型名，类型名跟全站同一张表
|
||||
|
||||
@@ -1,5 +1,5 @@
|
||||
// 几道"坏了就是整页事故"的护栏：输入法回车、未知节点类型、页面级错误边界、
|
||||
// 全局快捷键不抢输入框、未知地址不是白板。
|
||||
// 几道"坏了就是整页事故"的护栏：输入法回车、未知节点类型、页面级和外壳级的
|
||||
// 错误边界、全局快捷键不抢输入框、未知地址不是白板。
|
||||
//
|
||||
// 出事时的样子都很重：中文用户每选一个词就把半句话发去建图、跑图；
|
||||
// Copilot 编出一个不存在的节点类型，整站白屏、没保存的编辑一起丢；任何组件渲染
|
||||
@@ -90,6 +90,20 @@ await page.waitForURL(/\/chat/)
|
||||
check('换一页就恢复', await page.getByText('这一页出错了').waitFor({ state: 'detached', timeout: 8000 })
|
||||
.then(() => true, () => false))
|
||||
|
||||
console.log('\n=== 外壳自己出错：落在中文错误页，不是路由库的英文报错页 ===')
|
||||
// 导航、后台信号不在页面那一层错误边界里。路由换成 data router 之后，它们一出错，
|
||||
// 没人接的话 react-router 会画它自带的「Unexpected Application Error!」和一屏堆栈。
|
||||
// 导航上的「有运行在跑」要读画布的 trace：跑着却没有 trace，导航渲染就会抛
|
||||
await page.evaluate(() => {
|
||||
window.__savedTrace = window.__studio.getState().trace
|
||||
window.__studio.setState({ runPhase: 'running', trace: null })
|
||||
})
|
||||
check('给出中文的错误页', await shows('这一页出错了'))
|
||||
check('不是 react-router 的英文报错页', !(await page.locator('body').innerText()).includes('Unexpected Application Error'))
|
||||
await page.evaluate(() => window.__studio.setState({ runPhase: 'idle', trace: window.__savedTrace }))
|
||||
await page.getByRole('button', { name: '重试' }).first().click()
|
||||
check('点「重试」外壳就回来', await page.locator('nav[aria-label="主导航"]').waitFor({ timeout: 8000 }).then(() => true, () => false))
|
||||
|
||||
console.log('\n=== 全局快捷键：不抢输入框 ===')
|
||||
// 外壳在 window 的捕获阶段听键盘，比谁都先拿到。判错一次，问题框里打个问号就
|
||||
// 弹出一层面板、⌥3 直接把人带离写了一半的问题
|
||||
|
||||
+465
-56
@@ -41,15 +41,55 @@ try {
|
||||
|
||||
const browser = await chromium.launch({ executablePath: CHROME })
|
||||
|
||||
/**
|
||||
* 上传进度靠浏览器的 xhr.upload 事件，而 page.route 拦下的请求 Chrome 一个进度事件
|
||||
* 都不报（真请求会报，已实测）。所以在页面里替 XHR 报：发 /upload 时先报 40%，
|
||||
* 0.7 秒后报发完，再真正 send（被 page.route 接住）。发之前 abort 的，照浏览器的
|
||||
* 样子报一次 abort。只测界面怎么接这些事件，不碰后端
|
||||
*/
|
||||
function fakeUploadProgress() {
|
||||
const open = XMLHttpRequest.prototype.open
|
||||
const send = XMLHttpRequest.prototype.send
|
||||
const abort = XMLHttpRequest.prototype.abort
|
||||
XMLHttpRequest.prototype.open = function (method, url, ...rest) {
|
||||
this.__url = String(url)
|
||||
return open.call(this, method, url, ...rest)
|
||||
}
|
||||
XMLHttpRequest.prototype.abort = function () {
|
||||
if (this.__pending) {
|
||||
this.__pending = false
|
||||
this.dispatchEvent(new ProgressEvent('abort'))
|
||||
return
|
||||
}
|
||||
return abort.call(this)
|
||||
}
|
||||
XMLHttpRequest.prototype.send = function (body) {
|
||||
if (!/\/upload(\?|$)/.test(this.__url ?? '')) return send.call(this, body)
|
||||
const total = 2_000_000
|
||||
const up = this.upload
|
||||
this.__pending = true
|
||||
const fire = (type, loaded) => up.dispatchEvent(new ProgressEvent(type, { lengthComputable: true, loaded, total }))
|
||||
setTimeout(() => { if (this.__pending) fire('progress', 800_000) }, 50)
|
||||
setTimeout(() => {
|
||||
if (!this.__pending) return
|
||||
this.__pending = false
|
||||
fire('progress', total)
|
||||
fire('load', total)
|
||||
send.call(this, body)
|
||||
}, 700)
|
||||
}
|
||||
}
|
||||
|
||||
/**
|
||||
* 开一页：GET 放行，写请求交给 handlers（按「METHOD 路径正则」匹配），没配的一律
|
||||
* 拦成 503——既不写库，也顺带检验「写失败有反馈」。所有原生对话框都算失败。
|
||||
*/
|
||||
async function open(path, { handlers = [], theme = THEME, viewport = { width: 1280, height: 860 } } = {}) {
|
||||
async function open(path, { handlers = [], theme = THEME, viewport = { width: 1280, height: 860 }, uploadProgress = false } = {}) {
|
||||
const ctx = await browser.newContext({ viewport, colorScheme: theme, timezoneId: 'Asia/Shanghai' })
|
||||
await ctx.addInitScript((t) => {
|
||||
try { localStorage.setItem('agentlab.theme', t); localStorage.removeItem('agentlab.health') } catch { /* noop */ }
|
||||
}, theme)
|
||||
if (uploadProgress) await ctx.addInitScript(fakeUploadProgress)
|
||||
const page = await ctx.newPage()
|
||||
const sent = []
|
||||
const errors = []
|
||||
@@ -85,6 +125,15 @@ const goto = async (page, path) => {
|
||||
await page.waitForTimeout(300)
|
||||
}
|
||||
const json = (data, status = 200) => (route) => route.fulfill({ status, contentType: 'application/json', body: JSON.stringify(data) })
|
||||
/** 多久之前的服务器时间（后端列表里的 last_checked_at 就是这种带时区的 ISO） */
|
||||
const ago = (ms) => new Date(Date.now() - ms).toISOString()
|
||||
/** 等到条件成立或超时，返回最后一次的值 */
|
||||
const until = async (fn, ms = 8000, step = 200) => {
|
||||
const end = Date.now() + ms
|
||||
let v = await fn()
|
||||
while (!v && Date.now() < end) { await new Promise((r) => setTimeout(r, step)); v = await fn() }
|
||||
return v
|
||||
}
|
||||
const delayed = (ms, data, status = 200) => async (route) => { await new Promise((r) => setTimeout(r, ms)); return json(data, status)(route) }
|
||||
const shot = async (page, name) => { if (SHOTS) await page.screenshot({ path: `${SHOTS}/${name}.png` }) }
|
||||
const text = (page) => page.locator('main').innerText()
|
||||
@@ -135,8 +184,14 @@ for (const [path, title] of [['/tools', '工具'], ['/knowledge', '知识'], ['/
|
||||
console.log('\n=== 设置 · 模型接入 ===')
|
||||
{
|
||||
const p0 = providers[0]
|
||||
// 后端记着上次测连接的结果（last_check_*）：第一张卡 3 小时前测过、没通过，其余没测过
|
||||
const listed = providers.map((p, i) => (i === 0 ? {
|
||||
...p, last_checked_at: ago(3 * 3600_000), last_check_ok: false, last_latency_ms: null,
|
||||
last_error: '鉴权没通过（401）：对方拒绝了这把密钥',
|
||||
} : { ...p, last_checked_at: null, last_check_ok: null, last_latency_ms: null, last_error: null }))
|
||||
const { page, sent, natives, errors, close } = await open('/settings/providers', {
|
||||
handlers: [
|
||||
[/^GET \/providers$/, json(listed)],
|
||||
[/^POST \/providers\/[^/]+\/test$/, delayed(300, { ok: true, latency_ms: 128, model: p0?.default_model ?? 'm', reply: 'pong' })],
|
||||
[/^POST \/providers\/test$/, json({ ok: false, error: '测试没通过：连不上对方的服务', hint: '核对地址和端口，确认服务已经启动', detail: 'OpenAIConnectionError: Connection error.' })],
|
||||
[/^POST \/providers\/models$/, json({ ok: true, models: ['qwen-max', 'qwen-plus'], url: 'x' })],
|
||||
@@ -144,7 +199,14 @@ console.log('\n=== 设置 · 模型接入 ===')
|
||||
})
|
||||
if (p0) {
|
||||
const card = page.locator(`[data-provider="${p0.name}"]`)
|
||||
check('卡片上的状态点说的是连通，没测过写「未测试」', (await card.locator('[data-health]').getAttribute('data-health')) === 'idle')
|
||||
const first = await card.locator('[data-health]').innerText()
|
||||
check('卡片初值是后端记着的上次结果：连不上 · 3 小时前测（换了浏览器也在）', /连不上.*3 小时前测/.test(first), first)
|
||||
check('……原因也写在卡片上', (await card.innerText()).includes('对方拒绝了这把密钥'))
|
||||
const p1 = providers[1]
|
||||
if (p1) {
|
||||
check('后端没测过的卡片写「未测试」',
|
||||
(await page.locator(`[data-provider="${p1.name}"] [data-health]`).getAttribute('data-health')) === 'idle')
|
||||
}
|
||||
const disabled = providers.find((p) => !p.enabled)
|
||||
if (disabled) {
|
||||
check('停用单独标「已停用」，不再借圆点表达',
|
||||
@@ -155,7 +217,7 @@ console.log('\n=== 设置 · 模型接入 ===')
|
||||
check('测试中状态点在转、写已用时间', (await card.locator('[data-health]').getAttribute('data-health')) === 'checking')
|
||||
await page.waitForTimeout(600)
|
||||
const pill = await card.locator('[data-health]').innerText()
|
||||
check('测完留在卡片上，写明是什么时候测的：已连通 · 128 ms · 刚测过', /已连通.*128 ms.*刚测过/.test(pill), pill)
|
||||
check('本机刚测的比后端记的新，用本机的：已连通 · 128 ms · 刚测过', /已连通.*128 ms.*刚测过/.test(pill), pill)
|
||||
await shot(page, 'providers-tested')
|
||||
await page.getByRole('tab', { name: '运行环境' }).click()
|
||||
await page.waitForTimeout(300)
|
||||
@@ -214,27 +276,45 @@ console.log('\n=== 设置 · 模型接入 ===')
|
||||
|
||||
console.log('\n=== 设置 · 偏好 ===')
|
||||
{
|
||||
const settings = await get('/settings')
|
||||
// 设置读写都走假的一份：PUT 写进去，GET 读回来（主题的「已保存」是读回来核对过才写的）
|
||||
let stored = await get('/settings')
|
||||
const { page, sent, natives, close } = await open('/settings/prefs', {
|
||||
handlers: [[/^PUT \/settings$/, (route, { body }) => json({ ...settings, ...body.values })(route)]],
|
||||
handlers: [
|
||||
[/^GET \/settings$/, (route) => json(stored)(route)],
|
||||
[/^PUT \/settings$/, (route, { body }) => { stored = { ...stored, ...body.values }; return json(stored)(route) }],
|
||||
],
|
||||
})
|
||||
const themeRadio = (label) => page.getByRole('radiogroup', { name: '主题' }).getByRole('radio', { name: label })
|
||||
const checkedTheme = () => page.getByRole('radiogroup', { name: '主题' }).locator('[aria-checked="true"]').innerText()
|
||||
// 选一个和这次要截图的主题一致的值（浅色跑选「跟随系统」，上下文的 colorScheme
|
||||
// 是浅色）；它已经是选中的就先绕一下别的，点选中项不会发请求
|
||||
const [label, value] = THEME === 'dark' ? ['深色', 'dark'] : ['跟随系统', 'system']
|
||||
const radio = page.getByRole('radio', { name: label })
|
||||
if ((await radio.getAttribute('aria-checked')) === 'true') {
|
||||
await page.getByRole('radio', { name: '浅色' }).click()
|
||||
await page.waitForTimeout(300)
|
||||
if ((await themeRadio(label).getAttribute('aria-checked')) === 'true') {
|
||||
await themeRadio('浅色').click()
|
||||
await page.waitForTimeout(400)
|
||||
}
|
||||
await radio.click()
|
||||
await page.waitForTimeout(400)
|
||||
await themeRadio(label).click()
|
||||
await page.waitForTimeout(500)
|
||||
const put = sent.filter((s) => s.key === 'PUT /settings').at(-1)
|
||||
check('主题选中即保存，只 PUT ui 一组', !!put && Object.keys(put.body?.values ?? {}).join() === 'ui' && put.body.values.ui.theme === value,
|
||||
JSON.stringify(put?.body ?? {}).slice(0, 100))
|
||||
check('主题旁显示「已保存」', (await page.locator('[data-theme-save]').innerText()).includes('已保存'))
|
||||
check('主题旁显示「已保存」（读回来核对过）', (await page.locator('[data-theme-save]').innerText()).includes('已保存'))
|
||||
await page.waitForTimeout(1600)
|
||||
check('……约 1.6 秒后淡出，不一直挂着', await page.getByText('已保存', { exact: true }).count() === 0
|
||||
&& await page.locator('[data-theme-save="saved"]').count() === 0)
|
||||
|
||||
// REQ：从导航（或 ⌘K）换主题时偏好页开着，选项要跟着变，不能还显示旧值
|
||||
const resolvedBefore = await page.evaluate(() => {
|
||||
const a = document.documentElement.getAttribute('data-theme')
|
||||
return a === 'light' || a === 'dark' ? a : (matchMedia('(prefers-color-scheme: light)').matches ? 'light' : 'dark')
|
||||
})
|
||||
await page.getByRole('button', { name: /^切换到(浅色|深色)主题$/ }).click()
|
||||
await page.waitForTimeout(500)
|
||||
const synced = await checkedTheme()
|
||||
check('从导航换主题：偏好页的主题选项跟着变', synced.includes(resolvedBefore === 'dark' ? '浅色' : '深色'), synced)
|
||||
await themeRadio(label).click()
|
||||
await page.waitForTimeout(500)
|
||||
|
||||
check('默认知识库是下拉，选项来自已有知识库',
|
||||
await page.locator('#pref-collection').evaluate((el) => el.tagName) === 'SELECT'
|
||||
&& (await page.locator('#pref-collection option').allInnerTexts()).some((t) => t.startsWith(collections[0]?.collection ?? 'default')))
|
||||
@@ -252,23 +332,71 @@ console.log('\n=== 设置 · 偏好 ===')
|
||||
await page.waitForTimeout(200)
|
||||
check('选「留下」就还在偏好页', page.url().endsWith('/settings/prefs'), page.url())
|
||||
|
||||
const putsBefore = sent.filter((s) => s.key === 'PUT /settings').length
|
||||
await page.getByRole('button', { name: /保存设置/ }).click()
|
||||
await page.waitForTimeout(250)
|
||||
check('只改署名：存在本机就行，不发 PUT', sent.filter((s) => s.key === 'PUT /settings').length === putsBefore)
|
||||
check('保存后说「已保存」', (await page.locator('[data-prefs-bar]').innerText().catch(() => '')).includes('已保存'))
|
||||
check('署名写进本机', await page.evaluate(() => localStorage.getItem('agentlab_actor')) === '检查脚本')
|
||||
const initial = await page.locator('nav a[href="/settings/prefs"]').first().innerText().catch(() => '')
|
||||
check('……导航底部的署名首字立刻换成「检」（同一个标签页也收得到）', initial.trim() === '检', initial)
|
||||
|
||||
await page.locator('#pref-confirm-hint').locator('xpath=ancestor::label//input').click()
|
||||
await page.getByRole('button', { name: /保存设置/ }).click()
|
||||
await page.waitForTimeout(400)
|
||||
const put2 = sent.filter((s) => s.key === 'PUT /settings').at(-1)
|
||||
check('保存设置只 PUT run 一组', Object.keys(put2?.body?.values ?? {}).join() === 'run', JSON.stringify(put2?.body ?? {}).slice(0, 100))
|
||||
check('保存后说「已保存」', (await page.locator('[data-prefs-bar]').innerText().catch(() => '')).includes('已保存'))
|
||||
check('署名跟着保存写进本机', await page.evaluate(() => localStorage.getItem('agentlab_actor')) === '检查脚本')
|
||||
check('改了运行默认值：保存只 PUT run 一组', Object.keys(put2?.body?.values ?? {}).join() === 'run', JSON.stringify(put2?.body ?? {}).slice(0, 100))
|
||||
check('没有原生对话框', natives.length === 0, natives.join(' | '))
|
||||
await close()
|
||||
|
||||
// 保存失败要看得见（以前 pageerror 一条，界面毫无变化）
|
||||
const failing = await open('/settings/prefs')
|
||||
await failing.page.locator('#pref-actor').fill('x')
|
||||
await failing.page.locator('#pref-confirm-hint').locator('xpath=ancestor::label//input').click()
|
||||
await failing.page.getByRole('button', { name: /保存设置/ }).click()
|
||||
await failing.page.waitForTimeout(600)
|
||||
check('保存失败：条上写「没存上」', (await failing.page.locator('[data-prefs-bar]').innerText()).includes('没存上'))
|
||||
const bar = await failing.page.locator('[data-prefs-bar]').innerText()
|
||||
check('保存失败：条上写「没存上」', bar.includes('没存上'))
|
||||
check('……署名只在本机，已经存上了；条上只剩危险工具审批一项', bar.includes('有 1 项未保存') && bar.includes('危险工具审批') && !bar.includes('署名'), bar.replace(/\s+/g, ' '))
|
||||
check('保存失败不抛未捕获异常', failing.errors.length === 0, failing.errors[0] ?? '')
|
||||
await failing.close()
|
||||
|
||||
// 断线时在偏好页换的主题：先在本机生效，恢复后自动补存，而不是被服务端的旧值翻回去。
|
||||
// 断线要断得像真的：/health 也够不着，catalog 才会判断开、恢复时才算「重连」
|
||||
let offline = false
|
||||
let puts = 0
|
||||
let stored2 = await get('/settings')
|
||||
const cut = (route) => route.abort('internetdisconnected')
|
||||
const off = await open('/settings/prefs', {
|
||||
handlers: [
|
||||
[/^GET \/settings$/, (route) => (offline ? cut(route) : json(stored2)(route))],
|
||||
[/^PUT \/settings$/, (route, { body }) => {
|
||||
puts++
|
||||
if (offline) return cut(route)
|
||||
stored2 = { ...stored2, ...body.values }
|
||||
return json(stored2)(route)
|
||||
}],
|
||||
[/^GET \//, (route) => (offline ? cut(route) : route.continue())],
|
||||
],
|
||||
})
|
||||
const offRadio = (l) => off.page.getByRole('radiogroup', { name: '主题' }).getByRole('radio', { name: l })
|
||||
const target = (await offRadio('浅色').getAttribute('aria-checked')) === 'true' ? ['深色', 'dark'] : ['浅色', 'light']
|
||||
offline = true
|
||||
await offRadio(target[0]).click()
|
||||
await off.page.waitForTimeout(800)
|
||||
check('断线时换主题：说清先在本机生效、连上后自动存', (await off.page.locator('body').innerText()).includes('连上后端后自动存进设置'))
|
||||
check('……选项和页面都已经是新主题', (await offRadio(target[0]).getAttribute('aria-checked')) === 'true'
|
||||
&& await off.page.evaluate(() => document.documentElement.getAttribute('data-theme')) === target[1])
|
||||
const down = await until(() => off.page.locator('[data-offline-banner]').count().then((n) => n > 0), 6000)
|
||||
check('……整站也判了断开（横幅出来了）', !!down)
|
||||
const putsOffline = puts
|
||||
offline = false
|
||||
await until(() => stored2?.ui?.theme === target[1], 20000)
|
||||
await off.page.waitForTimeout(400)
|
||||
check('……后端恢复后自动补存这个主题', puts > putsOffline && stored2?.ui?.theme === target[1], `PUT ${puts} 次，存的是 ${stored2?.ui?.theme}`)
|
||||
check('……没有被服务端的旧值翻回去', await off.page.evaluate(() => document.documentElement.getAttribute('data-theme')) === target[1]
|
||||
&& (await offRadio(target[0]).getAttribute('aria-checked')) === 'true')
|
||||
await off.close()
|
||||
}
|
||||
|
||||
console.log('\n=== 数据 · 数据库 ===')
|
||||
@@ -276,22 +404,50 @@ console.log('\n=== 数据 · 数据库 ===')
|
||||
const dbs = sources.filter((s) => !/[\\/]uploads[\\/]tables[\\/]/.test(s.database ?? ''))
|
||||
const first = dbs.find((s) => s.table_count > 0) ?? dbs[0]
|
||||
const oracle = dbs.find((s) => s.kind === 'oracle')
|
||||
// 一个配置里的 schema 探不出对象的假库：只有这种库才给「换个 schema 探查」。
|
||||
// 顺带是个免密库（has_password:false），验证密码不再拦保存
|
||||
const noCheck = { last_checked_at: null, last_check_ok: null, last_latency_ms: null, last_error: null }
|
||||
// 假库一：配置里的 schema 探不出对象；后端记着 2 小时前测连接没通过。顺带是个免密库
|
||||
// （has_password:false），验证密码不再拦保存
|
||||
const empty = {
|
||||
id: 'check-manage-empty', name: 'zz_probe', kind: 'postgres', host: '10.0.0.9', port: null, database: 'mes',
|
||||
username: 'reader', options: { schema: 'ods' }, readonly: true, description: '检查脚本的假库', enabled: true,
|
||||
password_masked: '', has_password: false, table_count: 0, schema_synced_at: null,
|
||||
username: 'reader', options: { schema: 'ods', sslmode: 'disable' }, readonly: true, description: '检查脚本的假库', enabled: true,
|
||||
password_masked: '', has_password: false, table_count: 0, schema_synced_at: null, cached_schema: 'ods',
|
||||
schema_error: 'NoSuchTableError: ods', available_schemas: ['mes', 'ods', 'public'], tools: ['db_query__zz_probe'],
|
||||
last_checked_at: ago(2 * 3600_000), last_check_ok: false, last_latency_ms: null, last_error: '连不上：认证失败',
|
||||
}
|
||||
const probed = { ...empty, table_count: 3, schema_error: '', schema_synced_at: new Date().toISOString() }
|
||||
// 假库二：配置改成了 sales，缓存还是按 legacy 探的——助手看到的和配置对不上
|
||||
const drifted = {
|
||||
...empty, ...noCheck, id: 'check-manage-drift', name: 'zz_drift', options: { schema: 'sales' }, cached_schema: 'legacy',
|
||||
table_count: 2, schema_error: '', available_schemas: [], schema_synced_at: ago(3600_000), tools: ['db_query__zz_drift'],
|
||||
}
|
||||
let cur = empty
|
||||
let drift = drifted
|
||||
const { page, sent, natives, errors, close } = await open('/data/databases', {
|
||||
handlers: [
|
||||
[/^GET \/datasources$/, json([...sources, empty])],
|
||||
[/^GET \/datasources\/check-manage-empty\/schema$/, json({ tables: ['mes.work_order', 'mes.line', 'mes.shift'], summary: '', synced_at: probed.schema_synced_at })],
|
||||
[/^POST \/datasources\/check-manage-empty\/introspect$/, (route, { url }) =>
|
||||
json(url.searchParams.get('schema') === 'mes' ? probed : empty)(route)],
|
||||
[/^PATCH \/datasources\/check-manage-empty$/, (route, { body }) => json({ ...empty, ...body, options: body?.options ?? empty.options })(route)],
|
||||
[/^GET \/datasources$/, (route) => json([...sources, cur, drift])(route)],
|
||||
[/^GET \/datasources\/check-manage-empty\/schema$/, json({ tables: ['mes.work_order', 'mes.line', 'mes.shift'], summary: '', synced_at: ago(0) })],
|
||||
[/^GET \/datasources\/check-manage-drift\/schema$/, json({ tables: ['sales.orders', 'sales.sales_daily'], summary: '', synced_at: ago(0) })],
|
||||
[/^POST \/datasources\/check-manage-empty\/introspect$/, (route, { url }) => {
|
||||
const schema = url.searchParams.get('schema')
|
||||
if (url.searchParams.get('dry_run') === 'true') {
|
||||
return json(schema === 'mes'
|
||||
? { dry_run: true, schema: 'mes', table_count: 3, tables: ['mes.work_order', 'mes.line', 'mes.shift'], truncated: false, total: 3, schema_error: '', available_schemas: empty.available_schemas }
|
||||
: { dry_run: true, schema, table_count: 0, tables: [], truncated: false, total: 0, schema_error: '', available_schemas: empty.available_schemas })(route)
|
||||
}
|
||||
const hit = cur.options.schema === 'mes'
|
||||
cur = { ...cur, cached_schema: cur.options.schema ?? '', table_count: hit ? 3 : 0, schema_error: hit ? '' : empty.schema_error,
|
||||
schema_synced_at: hit ? ago(0) : cur.schema_synced_at }
|
||||
return json(cur)(route)
|
||||
}],
|
||||
[/^POST \/datasources\/check-manage-drift\/introspect$/, (route) => {
|
||||
drift = { ...drift, cached_schema: drift.options.schema, schema_synced_at: ago(0) }
|
||||
return json(drift)(route)
|
||||
}],
|
||||
// 后端的规矩：连接配置（地址、账号、库）改了，上次测连接的结果作废
|
||||
[/^PATCH \/datasources\/check-manage-empty$/, (route, { body }) => {
|
||||
const reconnect = ['host', 'port', 'database', 'username'].some((k) => k in (body ?? {}) && body[k] !== cur[k]) || 'password' in (body ?? {})
|
||||
cur = { ...cur, ...body, options: body?.options ?? cur.options, ...(reconnect ? noCheck : {}) }
|
||||
return json(cur)(route)
|
||||
}],
|
||||
[/^POST \/datasources\/[^/]+\/test$/, delayed(250, { ok: true, elapsed_ms: 42, url: 'x' })],
|
||||
[/^POST \/datasources\/test$/, json({ ok: false, error: '连不上：认证失败', hint: '核对用户名和密码', detail: 'ORA-01017' })],
|
||||
[/^POST \/datasources\/[^/]+\/introspect/, (route, { url }) => {
|
||||
@@ -332,7 +488,7 @@ console.log('\n=== 数据 · 数据库 ===')
|
||||
await rows.first().click()
|
||||
await page.waitForTimeout(800)
|
||||
const cols = await card.locator('[data-schema-browser] tbody tr').count()
|
||||
check('点表名懒加载出逐列信息', cols > 0, `${cols} 列`)
|
||||
check('点表名懒加载出逐列信息（后端给的结构化列）', cols > 0, `${cols} 列`)
|
||||
const scroller = card.locator('[data-schema-browser] .overflow-y-auto')
|
||||
await scroller.evaluate((el) => { el.scrollTop = 60 })
|
||||
const before = await scroller.evaluate((el) => el.scrollTop)
|
||||
@@ -345,8 +501,9 @@ console.log('\n=== 数据 · 数据库 ===')
|
||||
await shot(page, 'datasource-schema')
|
||||
}
|
||||
|
||||
// 好好的库不给「换个 schema 探查」：后端探查会落进缓存，换着探就把助手看到的结构换掉了
|
||||
check('结构正常的库不给「换个 schema 探查」', await card.locator('[data-probe-schema]').count() === 0)
|
||||
// 「换个 schema 看看」只看不存，好好的库也给；单文件的 SQLite 没有 schema 可换
|
||||
check(first.kind === 'sqlite' ? 'SQLite 不给「换个 schema 看看」' : '结构正常的库也能「换个 schema 看看」（只看不存，没有风险）',
|
||||
await card.locator('[data-probe-schema]').count() === (first.kind === 'sqlite' ? 0 : 1))
|
||||
|
||||
await card.getByRole('button', { name: `删除数据源 ${first.name}` }).click()
|
||||
const del = dialog(page)
|
||||
@@ -358,40 +515,50 @@ console.log('\n=== 数据 · 数据库 ===')
|
||||
}
|
||||
|
||||
{
|
||||
// 配置里的 schema 探不出东西：才给「换个 schema 探查」
|
||||
const ec = page.locator(`[data-source="${empty.name}"]`)
|
||||
check('探查失败的库给「换个 schema 探查」', await ec.locator('[data-probe-schema]').count() === 1)
|
||||
const pill0 = await ec.locator('[data-health]').innerText()
|
||||
check('卡片初值是后端记着的上次测连接：连不上 · 2 小时前测', /连不上.*2 小时前测/.test(pill0), pill0)
|
||||
check('……原因写在卡片上', (await ec.innerText()).includes('认证失败'))
|
||||
|
||||
// 换个 schema 看看：只看不存（dry_run），缓存和配置都不动
|
||||
check('探查失败的库给「换个 schema 看看」', await ec.locator('[data-probe-schema]').count() === 1)
|
||||
await ec.locator('[data-probe-schema]').click()
|
||||
const pd = dialog(page)
|
||||
const pdText = await pd.innerText()
|
||||
check('……先说清楚：探到的结构会换掉缓存，不写进配置就换回来', pdText.includes('换掉') && pdText.includes('换回「ods」'), pdText.slice(0, 120))
|
||||
check('……先说清楚：只看不存，缓存和配置都不动', pdText.includes('只看不存') && pdText.includes('缓存和配置都不动'), pdText.slice(0, 120))
|
||||
await pd.locator('input').fill('mes')
|
||||
await pd.getByRole('button', { name: '探查', exact: true }).click()
|
||||
await pd.getByRole('button', { name: '看看', exact: true }).click()
|
||||
await page.waitForTimeout(500)
|
||||
const probeReq = sent.filter((s) => s.key.endsWith('/introspect')).at(-1)
|
||||
check('……请求带上指定的 schema', !!probeReq?.url.includes('schema=mes'), probeReq?.url ?? '')
|
||||
check('……请求带上 schema 和 dry_run=true（不写缓存）', !!probeReq?.url.includes('schema=mes') && probeReq.url.includes('dry_run=true'), probeReq?.url ?? '')
|
||||
const ask = dialog(page)
|
||||
const askText = await ask.innerText().catch(() => '')
|
||||
check('……探到了再问要不要写进配置，并举出探到的表名', askText.includes('mes') && askText.includes('work_order'), askText.slice(0, 160))
|
||||
check('……「不改」说的是什么都不动，不再「换回去」', askText.includes('什么都不动') && !askText.includes('换回'), askText.slice(0, 200))
|
||||
await shot(page, 'datasource-probe-schema')
|
||||
const before = sent.filter((s) => s.key.endsWith('/introspect')).length
|
||||
await ask.getByRole('button', { name: '不改，换回 ods' }).click()
|
||||
const before = sent.length
|
||||
await ask.getByRole('button', { name: '不改', exact: true }).click()
|
||||
await page.waitForTimeout(500)
|
||||
const restore = sent.filter((s) => s.key.endsWith('/introspect')).slice(before)
|
||||
check('……不改：马上按配置里的 schema 重探一次，缓存和配置对得上',
|
||||
restore.length === 1 && !restore[0].url.includes('schema='), restore.map((r) => r.url).join(' | '))
|
||||
check('……不改就不发 PATCH', !sent.some((s) => s.key.startsWith('PATCH /datasources')))
|
||||
check('……卡片回到配置里那份（探查失败、0 个对象）', (await ec.innerText()).includes('上次探查失败'))
|
||||
check('……不改：一个请求都不再发（以前要再探一次把缓存换回来）', sent.length === before, sent.slice(before).map((r) => r.url).join(' | '))
|
||||
check('……卡片还是配置里那份（探查失败、0 个对象）', (await ec.innerText()).includes('上次探查失败'))
|
||||
|
||||
// 失败框里的候选 chip：探到了选「改」就写进配置
|
||||
await ec.getByRole('button', { name: '用 mes 探查' }).click()
|
||||
// 失败框里的候选 chip：探到了选「改」就写进配置，再按新配置真探一次
|
||||
await ec.getByRole('button', { name: '看看 mes' }).click()
|
||||
await page.waitForTimeout(500)
|
||||
await dialog(page).getByRole('button', { name: '改成 mes' }).click()
|
||||
await page.waitForTimeout(400)
|
||||
await page.waitForTimeout(600)
|
||||
const patch = sent.filter((s) => s.key === 'PATCH /datasources/check-manage-empty').at(-1)
|
||||
check('……选「改成 mes」：PATCH 写进 options.schema', patch?.body?.options?.schema === 'mes', JSON.stringify(patch?.body ?? {}))
|
||||
check('……选「改成 mes」：PATCH 写进 options.schema，别的连接参数原样带上',
|
||||
patch?.body?.options?.schema === 'mes' && patch?.body?.options?.sslmode === 'disable', JSON.stringify(patch?.body ?? {}))
|
||||
const real = sent.filter((s) => s.key.endsWith('check-manage-empty/introspect')).at(-1)
|
||||
check('……再按新配置真探一次（不带 dry_run）', !!real && !real.url.includes('dry_run') && !real.url.includes('schema='), real?.url ?? '')
|
||||
const ecText = await ec.innerText()
|
||||
check('……卡片换成新 schema 的结构', /3\s*个对象/.test(ecText) && !ecText.includes('上次探查失败'), ecText.replace(/\s+/g, ' ').slice(0, 160))
|
||||
|
||||
// 免密库：密码不拦保存
|
||||
// 本机刚测的比后端记的新；只改说明不动它，改了连接配置就作废
|
||||
await ec.getByRole('button', { name: /测连接/ }).click()
|
||||
await page.waitForTimeout(500)
|
||||
check('本机刚测过：卡片换成本机的结果', /已连通.*42 ms/.test(await ec.locator('[data-health]').innerText()))
|
||||
await ec.getByRole('button', { name: '编辑' }).click()
|
||||
const ed = dialog(page)
|
||||
await ed.getByLabel('说明', { exact: true }).fill('检查脚本的假库（改过说明）')
|
||||
@@ -407,6 +574,24 @@ console.log('\n=== 数据 · 数据库 ===')
|
||||
const saved = sent.filter((s) => s.key === 'PATCH /datasources/check-manage-empty').at(-1)
|
||||
check('……保存时不带 password（不动它）', !!saved && !('password' in (saved.body ?? {})) && saved.body?.description?.includes('改过说明'),
|
||||
JSON.stringify(saved?.body ?? {}).slice(0, 120))
|
||||
check('……只改说明：刚测的结果还作数', /已连通/.test(await ec.locator('[data-health]').innerText()))
|
||||
await ec.getByRole('button', { name: '编辑' }).click()
|
||||
await dialog(page).getByLabel(/主机/).fill('10.0.0.10')
|
||||
await dialog(page).getByRole('button', { name: '保存' }).click()
|
||||
await page.waitForTimeout(400)
|
||||
const after = await ec.locator('[data-health]').innerText()
|
||||
check('……换了主机：之前测的不再代表它，回到「未测试」', after.includes('未测试'), after)
|
||||
|
||||
// 缓存和配置对不上：常驻一行 warn，一键按配置重探
|
||||
const dc = page.locator(`[data-source="${drifted.name}"]`)
|
||||
const driftText = await dc.locator('[data-schema-drift]').innerText().catch(() => '')
|
||||
check('缓存按 legacy 探、配置写着 sales：卡片上常驻一行说对不上', driftText.includes('legacy') && driftText.includes('sales'), driftText)
|
||||
await shot(page, 'datasource-drift')
|
||||
await dc.getByRole('button', { name: '按配置重新探查' }).click()
|
||||
await page.waitForTimeout(500)
|
||||
const re = sent.filter((s) => s.key === 'POST /datasources/check-manage-drift/introspect').at(-1)
|
||||
check('……点「按配置重新探查」：按配置真探（不带 schema、不带 dry_run）', !!re && !re.url.includes('?'), re?.url ?? '')
|
||||
check('……探完那行 warn 就消失', await dc.locator('[data-schema-drift]').count() === 0)
|
||||
}
|
||||
|
||||
if (oracle) {
|
||||
@@ -463,11 +648,21 @@ console.log('\n=== 数据 · 表格 ===')
|
||||
] }],
|
||||
}
|
||||
const { page, sent, natives, close } = await open('/data/tables', {
|
||||
uploadProgress: true,
|
||||
handlers: [
|
||||
[/^POST \/datasources\/upload$/, delayed(400, fake, 201)],
|
||||
// 传上来的表只有一两张，卡片会自动摊开结构：假的那张也得有结构可取
|
||||
// 列注释里带两个空格：以前解析 detail 文本（按两个以上空格切列），「见附表」会被切丢
|
||||
[/^GET \/datasources\/check-manage-fake\/schema$/, (route, { url }) => json(url.searchParams.get('table')
|
||||
? { table: 'sales_demo', detail: '表 sales_demo\n 2026-01 REAL\n region TEXT\n Unnamed: 2 TEXT' }
|
||||
? {
|
||||
table: 'sales_demo', found: true, qualified: 'sales_demo', kind: 'table', comment: '按月的销售明细',
|
||||
detail: '表 sales_demo（按月的销售明细）\n 2026-01 REAL\n region TEXT 主键、非空、销售区域 见附表\n Unnamed: 2 TEXT',
|
||||
columns: [
|
||||
{ name: '2026-01', type: 'REAL', pk: false, not_null: false, comment: null },
|
||||
{ name: 'region', type: 'TEXT', pk: true, not_null: true, comment: '销售区域 见附表' },
|
||||
{ name: 'Unnamed: 2', type: 'TEXT', pk: false, not_null: false, comment: null },
|
||||
],
|
||||
}
|
||||
: { tables: ['sales_demo'], summary: '', synced_at: fake.source.schema_synced_at })(route)],
|
||||
],
|
||||
})
|
||||
@@ -476,8 +671,21 @@ console.log('\n=== 数据 · 表格 ===')
|
||||
await dlg.locator('input[type="file"]').setInputFiles({ name: 'sales_demo.csv', mimeType: 'text/csv', buffer: Buffer.from('2026-01,region\n1,east\n') })
|
||||
check('文件名自动变成数据源名', (await dlg.locator('input.mono').first().inputValue()) === 'sales_demo')
|
||||
await dlg.getByRole('button', { name: /导入/ }).click()
|
||||
await page.waitForTimeout(100)
|
||||
await page.waitForTimeout(250)
|
||||
check('导入中按钮写已用时间', /导入中/.test(await dlg.getByRole('button', { name: /导入中/ }).innerText().catch(() => '')))
|
||||
const sending = await dlg.locator('[data-upload-progress]').innerText().catch(() => '')
|
||||
check('上传中画真实的字节进度：已传多少 / 总共多少 · 百分比', sending.includes('已传') && sending.includes('40%')
|
||||
&& (await dlg.locator('[data-upload-progress] [role="progressbar"]').getAttribute('aria-valuenow').catch(() => '')) === '40',
|
||||
sending.replace(/\s+/g, ' '))
|
||||
check('……字节没发完之前可以取消', await dlg.getByRole('button', { name: '取消上传' }).count() === 1)
|
||||
await shot(page, 'table-uploading')
|
||||
await page.waitForTimeout(650)
|
||||
const processing = await dlg.locator('[data-upload-progress]').innerText().catch(() => '')
|
||||
check('……发完写「正在读表」，不把条拉满冒充完成', processing.includes('正在读表')
|
||||
&& await dlg.locator('[data-upload-progress] [role="progressbar"]').count() === 0, processing.replace(/\s+/g, ' '))
|
||||
const cancel = dlg.getByRole('button', { name: '取消', exact: true })
|
||||
check('……发完之后不能取消（后端已经在建表），悬停说明为什么',
|
||||
await cancel.isDisabled() && ((await cancel.getAttribute('title')) ?? '').includes('传完'))
|
||||
await page.waitForTimeout(700)
|
||||
const res = await dialog(page).innerText()
|
||||
check('导入完弹窗不关，停在列名和类型上', res.includes('region') && res.includes('TEXT'))
|
||||
@@ -490,7 +698,26 @@ console.log('\n=== 数据 · 表格 ===')
|
||||
await dialog(page).getByRole('button', { name: /表头不对/ }).click()
|
||||
await page.waitForTimeout(200)
|
||||
check('改行号重传：回到表单，行号 +1', (await dialog(page).locator('input[type="number"]').inputValue()) === '2')
|
||||
// 这次发到一半取消：后端收不全，什么都不会建
|
||||
await dialog(page).getByRole('button', { name: /导入/ }).click()
|
||||
await page.waitForTimeout(250)
|
||||
await dialog(page).getByRole('button', { name: '取消上传' }).click()
|
||||
await page.waitForTimeout(900)
|
||||
check('发到一半取消：说清什么都没建，弹窗留在表单上', (await page.locator('body').innerText()).includes('没传完就取消了')
|
||||
&& await dialog(page).getByRole('button', { name: /导入/ }).isEnabled())
|
||||
await dialog(page).getByRole('button', { name: '取消' }).click()
|
||||
await page.waitForTimeout(300)
|
||||
// 结构浏览器的列清单用后端给的结构化列，不再解析 detail 文本
|
||||
const tcard = page.locator('[data-source="sales_demo"]')
|
||||
await tcard.locator('[data-schema-browser] button[aria-expanded]').first().click()
|
||||
await page.waitForTimeout(500)
|
||||
const region = tcard.locator('[data-schema-browser] tbody tr', { hasText: 'region' })
|
||||
const cells = await region.locator('td').allInnerTexts()
|
||||
check('列清单：类型、主键、非空、说明各归各的格子', cells.length === 3 && cells[1] === 'TEXT'
|
||||
&& await region.locator('[aria-label="主键"]').count() === 1 && cells[2].includes('非空'), cells.join(' | '))
|
||||
check('……列注释里的两个空格不再把说明切丢（「见附表」还在）', cells[2]?.includes('销售区域') && cells[2]?.includes('见附表'), cells[2] ?? '')
|
||||
check('……表注释写在列表上方', (await tcard.locator('[data-schema-browser]').innerText()).includes('按月的销售明细'))
|
||||
await shot(page, 'table-columns')
|
||||
check('没有原生对话框', natives.length === 0, natives.join(' | '))
|
||||
check('上传只发了一次', sent.filter((s) => s.key === 'POST /datasources/upload').length === 1)
|
||||
await close()
|
||||
@@ -499,6 +726,7 @@ console.log('\n=== 数据 · 表格 ===')
|
||||
console.log('\n=== 知识库 ===')
|
||||
{
|
||||
const { page, sent, natives, errors, close } = await open('/knowledge/kb', {
|
||||
uploadProgress: true,
|
||||
handlers: [[/^POST \/kb\/upload$/, delayed(1200, {
|
||||
id: 'fake-doc', collection: 'default', title: 'note.md', source: 'note.md', mime: 'text/markdown',
|
||||
chunk_count: 0, status: 'processing', meta: { progress: { done: 1, total: 4 } },
|
||||
@@ -515,13 +743,34 @@ console.log('\n=== 知识库 ===')
|
||||
await page.locator('[data-alpha-off] button', { hasText: '重置' }).click()
|
||||
check('一键回到运行时值', Number(await page.locator('input[data-alpha]').inputValue()) === embedding.default_alpha)
|
||||
|
||||
await page.locator('input[data-kb-upload]').setInputFiles({ name: 'note.md', mimeType: 'text/markdown', buffer: Buffer.from('# hi') })
|
||||
await page.waitForTimeout(250)
|
||||
check('上传中列表顶上有占位行', await page.locator('[data-uploading]').count() === 1)
|
||||
// 一次选两份：一次传一个，第二份排队；排队的能取消
|
||||
await page.locator('input[data-kb-upload]').setInputFiles([
|
||||
{ name: 'note.md', mimeType: 'text/markdown', buffer: Buffer.from('# hi') },
|
||||
{ name: 'more.md', mimeType: 'text/markdown', buffer: Buffer.from('# more') },
|
||||
])
|
||||
await page.waitForTimeout(300)
|
||||
check('上传中列表顶上有占位行，一份一行', await page.locator('[data-uploading]').count() === 2)
|
||||
check('上传按钮写着上传中', (await page.getByRole('button', { name: /上传中/ }).count()) === 1)
|
||||
const first = page.locator('[data-uploading]', { hasText: 'note.md' })
|
||||
const sendingRow = await first.innerText()
|
||||
check('……在传的那份画真实的字节进度', sendingRow.includes('已传') && sendingRow.includes('40%')
|
||||
&& await first.locator('[role="progressbar"]').count() === 1, sendingRow.replace(/\s+/g, ' '))
|
||||
const second = page.locator('[data-uploading]', { hasText: 'more.md' })
|
||||
check('……第二份写「排队中」，还能取消', (await second.innerText()).includes('排队中')
|
||||
&& await second.getByRole('button', { name: '取消上传 more.md' }).count() === 1)
|
||||
await shot(page, 'kb-uploading')
|
||||
await page.waitForTimeout(1300)
|
||||
await second.getByRole('button', { name: '取消上传 more.md' }).click()
|
||||
await page.waitForTimeout(200)
|
||||
check('……排队的取消了立刻从列表拿掉', await page.locator('[data-uploading]', { hasText: 'more.md' }).count() === 0)
|
||||
await page.waitForTimeout(500)
|
||||
const processingRow = await first.innerText()
|
||||
check('……字节发完写「处理中」（后端在解析、切块），不把条拉满冒充完成',
|
||||
processingRow.includes('后端在解析、切块') && await first.locator('[role="progressbar"]').count() === 0,
|
||||
processingRow.replace(/\s+/g, ' '))
|
||||
check('……发完的那份不给取消（后端已经在切块）', await page.getByRole('button', { name: '取消上传 note.md' }).count() === 0)
|
||||
await page.waitForTimeout(1500)
|
||||
check('传完占位行消失', await page.locator('[data-uploading]').count() === 0)
|
||||
check('……取消的那份没有发出去', sent.filter((s) => s.key === 'POST /kb/upload').length === 1)
|
||||
|
||||
await page.locator('input[data-kb-upload]').setInputFiles({ name: 'data.csv', mimeType: 'text/csv', buffer: Buffer.from('a,b\n1,2') })
|
||||
await page.waitForTimeout(300)
|
||||
@@ -548,6 +797,104 @@ console.log('\n=== 知识库 ===')
|
||||
await close()
|
||||
}
|
||||
|
||||
console.log('\n=== 知识库 · 检索贡献条与重建进度 ===')
|
||||
{
|
||||
const hits = {
|
||||
query: '出勤', collection: 'default', degraded: [], alpha: embedding.default_alpha ?? 0.5,
|
||||
results: [
|
||||
{ chunk_id: 'c1', document_id: 'd1', title: 'hr_attendance.md', ordinal: 3, content: '出勤率 = 实际出勤 / 计划出勤',
|
||||
score: 0.42, signals: { vector: 0.6, keyword: 3.21 }, contrib: { vector: 0.3, keyword: 0.12 } },
|
||||
{ chunk_id: 'c2', document_id: 'd1', title: 'hr_attendance.md', ordinal: 4, content: '计划出勤按排班表算',
|
||||
score: 0.25, signals: { vector: 0.5 }, contrib: { vector: 0.25, keyword: 0 } },
|
||||
],
|
||||
}
|
||||
const job = (phase, chunks, memories, state = 'running') => ({
|
||||
id: 'check-job', state, phase, collection: 'default', total: 15, done: chunks + memories,
|
||||
chunks: { total: 12, done: chunks }, memories: { total: 3, done: memories },
|
||||
started_at: ago(4000), finished_at: state === 'running' ? null : ago(0), error: null, hint: null,
|
||||
result: state === 'done' ? { reindexed: 12, memories_reindexed: 3, embedder: 'openai:demo-embed' } : null,
|
||||
})
|
||||
// GET /kb/reindex：点重建之前没在跑（开发模式下进页那一眼会问两次）；之后依次是
|
||||
// 片段 → 倒排 → 记忆 → 做完
|
||||
const steps = [job('chunks', 6, 0), job('index', 12, 0), job('memories', 12, 2), job('done', 12, 3, 'done')]
|
||||
let started = false
|
||||
let polled = 0
|
||||
let rebuilt = false
|
||||
const { page, sent, natives, errors, close } = await open('/knowledge/kb', {
|
||||
handlers: [
|
||||
[/^GET \/kb\/search$/, json(hits)],
|
||||
[/^GET \/kb\/embedding$/, (route) => json({ ...embedding, fallback: false, has_semantics: true, unindexed_chunks: 0,
|
||||
stale_chunks: rebuilt ? 0 : 12, stale_memories: rebuilt ? 0 : 3 })(route)],
|
||||
[/^POST \/kb\/reindex$/, (route) => { started = true; return json(job('chunks', 0, 0), 202)(route) }],
|
||||
[/^GET \/kb\/reindex$/, (route) => {
|
||||
if (!started) return json({ state: 'idle' })(route)
|
||||
const out = steps[Math.min(polled++, steps.length - 1)]
|
||||
if (out.state === 'done') rebuilt = true
|
||||
return json(out)(route)
|
||||
}],
|
||||
],
|
||||
})
|
||||
await page.getByLabel('检索内容').fill('出勤')
|
||||
await page.getByRole('button', { name: '检索', exact: true }).click()
|
||||
await page.waitForTimeout(400)
|
||||
const hit = page.locator('[data-hit]').first()
|
||||
const contrib = await hit.locator('[data-contrib]').innerText().catch(() => '')
|
||||
check('命中拆成两路贡献：语义 +0.300、关键词 +0.120（两段之和就是总分）',
|
||||
/语义\s*\+0\.300/.test(contrib) && /关键词\s*\+0\.120/.test(contrib), contrib.replace(/\s+/g, ' '))
|
||||
check('……原始分另写成数：余弦、BM25', contrib.includes('余弦') && contrib.includes('0.600') && contrib.includes('3.21'))
|
||||
const segs = await hit.locator('span[title^="语义贡献"] > span').count()
|
||||
const segs2 = await page.locator('[data-hit]').nth(1).locator('span[title^="语义贡献"] > span').count()
|
||||
check('……贡献条两段；只靠语义命中的那条只有一段', segs === 2 && segs2 === 1, `${segs} / ${segs2}`)
|
||||
check('……结果上方有图例：总分 = 语义贡献 + 关键词贡献', (await page.locator('[data-contrib-legend]').innerText().catch(() => '')).includes('关键词贡献'))
|
||||
await shot(page, 'kb-search-contrib')
|
||||
|
||||
const rebuild = page.getByRole('button', { name: /重建索引（15 段）/ })
|
||||
check('向量对不上时给「重建索引（15 段）」', await rebuild.count() === 1,
|
||||
(await page.locator('[data-embedder]').innerText().catch(() => '')).replace(/\s+/g, ' ').slice(0, 160))
|
||||
await rebuild.click()
|
||||
const post = sent.find((s) => s.key === 'POST /kb/reindex')
|
||||
check('重建走后台：POST 带 background=true', !!post?.url.includes('background=true'), post?.url ?? '')
|
||||
const chunksText = await until(async () => {
|
||||
const t = await page.locator('[data-reindex="chunks"]').innerText().catch(() => '')
|
||||
return t.includes('6 / 12') ? t : ''
|
||||
}, 4000)
|
||||
check('进度是后端按批记的真实数：正在重算知识片段的向量 · 6 / 12 段', !!chunksText, chunksText)
|
||||
check('……进度条的值就是 done / total', (await page.locator('[data-reindex] [role="progressbar"]').getAttribute('aria-valuenow').catch(() => '')) === '6')
|
||||
check('……跑着的时候不能换向量模型', await page.getByRole('button', { name: '换一个' }).isDisabled())
|
||||
await shot(page, 'kb-reindex-progress')
|
||||
const indexText = await until(() => page.locator('[data-reindex="index"]').innerText().catch(() => ''), 4000)
|
||||
check('倒排那一段没有分批进度：写明，不去假装它在走', indexText.includes('没有分批进度'), indexText.replace(/\s+/g, ' '))
|
||||
const memText = await until(() => page.locator('[data-reindex="memories"]').innerText().catch(() => ''), 4000)
|
||||
check('……接着重算记忆：2 / 3 条', memText.includes('2 / 3 条'), memText.replace(/\s+/g, ' '))
|
||||
const doneToast = await until(async () => {
|
||||
const t = await page.locator('body').innerText()
|
||||
return t.includes('重建 12 段知识') ? t : ''
|
||||
}, 4000)
|
||||
check('做完了报一句：用哪个模型重建了多少', !!doneToast && doneToast.includes('openai:demo-embed') && doneToast.includes('3 条记忆'))
|
||||
check('……进度收起，重建按钮也跟着消失（不再有对不上的）',
|
||||
await page.locator('[data-reindex]').count() === 0 && await page.getByRole('button', { name: /重建索引/ }).count() === 0)
|
||||
check('没有原生对话框', natives.length === 0, natives.join(' | '))
|
||||
check('没有运行时报错', errors.length === 0, errors[0] ?? '')
|
||||
await close()
|
||||
|
||||
// 已经有一次在跑：回 409 就说明，并接着看它的进度，不起第二次
|
||||
let other = false
|
||||
const busy = await open('/knowledge/kb', {
|
||||
handlers: [
|
||||
[/^GET \/kb\/embedding$/, json({ ...embedding, fallback: false, has_semantics: true, unindexed_chunks: 0, stale_chunks: 12, stale_memories: 3 })],
|
||||
[/^POST \/kb\/reindex$/, (route) => { other = true; return json({ detail: '已经在重建索引了（全部集合，已完成 3/15），等它做完再点' }, 409)(route) }],
|
||||
// 进页那一眼还没在跑；点了之后别处起的那一次才露面
|
||||
[/^GET \/kb\/reindex$/, (route) => json(other ? { ...job('chunks', 3, 0), collection: null } : { state: 'idle' })(route)],
|
||||
],
|
||||
})
|
||||
await busy.page.getByRole('button', { name: /重建索引（15 段）/ }).click()
|
||||
await busy.page.waitForTimeout(600)
|
||||
check('已经在重建：回 409 时说明已经在跑', (await busy.page.locator('body').innerText()).includes('已经在重建索引了'))
|
||||
check('……接着显示那一次的进度', ((await busy.page.locator('[data-reindex]').innerText().catch(() => '')).includes('3 / 12')))
|
||||
check('……只发了一次 POST，不起第二次', busy.sent.filter((s) => s.key === 'POST /kb/reindex').length === 1)
|
||||
await busy.close()
|
||||
}
|
||||
|
||||
console.log('\n=== 长期记忆 ===')
|
||||
{
|
||||
const withRun = memories.find((m) => m.source?.kind === 'run' && m.source?.run_exists)
|
||||
@@ -607,6 +954,20 @@ console.log('\n=== 长期记忆 ===')
|
||||
await close()
|
||||
}
|
||||
|
||||
console.log('\n=== 目录没取回来：空态不冒充「还没有」 ===')
|
||||
for (const [path, pattern, label] of [
|
||||
['/settings/providers', /^GET \/providers$/, '模型接入'],
|
||||
['/knowledge/skills', /^GET \/skills$/, 'Skill'],
|
||||
]) {
|
||||
const { page, close } = await open(path, { handlers: [[pattern, json({ detail: '后端出错了' }, 500)]] })
|
||||
await page.waitForTimeout(600)
|
||||
const el = page.locator('main [data-empty-unknown]')
|
||||
check(`${label}那张表没取回来：不说「还没有」，说没取回来并给「重新读取」`,
|
||||
(await el.getAttribute('data-empty-unknown').catch(() => null)) === 'error' && (await el.innerText().catch(() => '')).includes('重新读取'))
|
||||
check('……收起「添加 / 新建」这类动作，免得照着建出重复的', await el.getByRole('button', { name: /添加接入|新建 Skill/ }).count() === 0)
|
||||
await close()
|
||||
}
|
||||
|
||||
console.log('\n=== Skill ===')
|
||||
{
|
||||
const { page, sent, natives, close } = await open('/knowledge/skills', {
|
||||
@@ -637,10 +998,23 @@ console.log('\n=== 工具 ===')
|
||||
[/^POST \/tools\/[^/]+\/run$/, (route, { body }) => {
|
||||
calls++
|
||||
return body?.confirm
|
||||
? json({ ok: true, result: 'done', duration_ms: 5 })(route)
|
||||
? json({ ok: true, result: 'done', duration_ms: 5, note: '参数名 cuont 不存在，已按唯一候选 count 执行' })(route)
|
||||
: json({ detail: '会在 playground 工作目录里写入「x.txt」，同名文件会被覆盖。在工具库里执行不经过审批，确认后才会执行。' }, 409)(route)
|
||||
}],
|
||||
[/^POST \/mcp\/refresh$/, (route) => route.abort()],
|
||||
[/^POST \/custom-tools\/test$/, (route, { body }) => json(body?.config?.url?.includes('fail')
|
||||
? { ok: false, error: '工具执行失败：HTTP 404', hint: '核对接口地址', detail: '', duration_ms: 12 }
|
||||
: { ok: true, result: { echo: body?.args ?? {} }, duration_ms: 34, note: '参数名 qurey 不存在，已按唯一候选 query 执行' })(route)],
|
||||
// 沙箱里没有 MCP 服务：两台假的，一台后端记着 5 分钟前连上过，一台老数据只有 status、没有测连接记录
|
||||
[/^GET \/mcp\/servers$/, json([
|
||||
{ id: 'check-mcp-ok', name: 'files_demo', transport: 'stdio', command: 'npx', args: ['-y', 'demo-mcp'], env: {}, url: null,
|
||||
enabled: true, status: 'ok', last_error: null, tools_cache: ['read', 'list'],
|
||||
last_checked_at: ago(5 * 60_000), last_check_ok: true, last_latency_ms: 88 },
|
||||
{ id: 'check-mcp-old', name: 'legacy_demo', transport: 'http', command: null, args: [], env: {}, url: 'https://example.com/mcp',
|
||||
enabled: true, status: 'error', last_error: '连不上', tools_cache: [],
|
||||
last_checked_at: null, last_check_ok: null, last_latency_ms: null },
|
||||
])],
|
||||
[/^POST \/mcp\/servers\/check-mcp-ok\/probe$/, delayed(200, { ok: true, tools: [{ name: 'read' }, { name: 'list' }, { name: 'write' }], elapsed_ms: 120 })],
|
||||
],
|
||||
})
|
||||
if (danger) {
|
||||
@@ -661,17 +1035,52 @@ console.log('\n=== 工具 ===')
|
||||
const runs = sent.filter((s) => /^POST \/tools\/.+\/run$/.test(s.key))
|
||||
check('确认后带 confirm:true 重发', runs.length === 2 && runs[1].body?.confirm === true, JSON.stringify(runs.map((r) => r.body?.confirm)))
|
||||
check('结果显示成功', (await page.locator('[data-tool-result]').getAttribute('data-tool-result')) === 'ok')
|
||||
const note = await page.locator('[data-tool-note]').innerText().catch(() => '')
|
||||
check('参数名被纠正过：成功结果下面用 warn 写出来，提醒先改对再抄进工作流', note.includes('cuont') && note.includes('改对'), note)
|
||||
}
|
||||
await page.getByRole('tab', { name: '自定义工具' }).click()
|
||||
await page.waitForTimeout(400)
|
||||
await page.getByRole('button', { name: /新建工具/ }).first().click()
|
||||
const params = await dialog(page).locator('textarea').nth(1).inputValue()
|
||||
const ed = dialog(page)
|
||||
const params = await ed.locator('textarea').nth(1).inputValue()
|
||||
check('新建工具的参数给了能跑的示例，不是 {}', params.includes('"query"'), params.slice(0, 60))
|
||||
check('编辑器里有试跑区（保存后可用）', (await dialog(page).locator('[data-tool-trial]').innerText()).includes('先保存'))
|
||||
await dialog(page).getByRole('button', { name: '取消' }).click()
|
||||
const trialBox = ed.locator('[data-tool-trial]')
|
||||
check('新建时不用先保存就能试跑', (await trialBox.innerText()).includes('不用先保存'))
|
||||
const tryBtn = trialBox.getByRole('button', { name: /试跑/ }).last()
|
||||
check('……URL 还没填时试跑不可点，悬停说缺什么', await tryBtn.isDisabled() && ((await tryBtn.getAttribute('title')) ?? '').includes('URL'))
|
||||
await ed.getByLabel(/^URL/).fill('https://api.example.com/search?q={{ query }}')
|
||||
await tryBtn.click()
|
||||
await page.waitForTimeout(400)
|
||||
const draft = sent.filter((s) => s.key === 'POST /custom-tools/test').at(-1)
|
||||
check('……试跑的是眼前这份没保存的配置（POST /custom-tools/test）',
|
||||
draft?.body?.kind === 'http' && draft?.body?.config?.url?.includes('api.example.com') && 'query' in (draft?.body?.parameters?.properties ?? {})
|
||||
&& 'args' in (draft?.body ?? {}), JSON.stringify(draft?.body ?? {}).slice(0, 160))
|
||||
check('……没有保存（没发 POST /custom-tools）', !sent.some((s) => s.key === 'POST /custom-tools'))
|
||||
const trialNote = await trialBox.locator('[data-tool-note]').innerText().catch(() => '')
|
||||
check('……试跑结果里参数名被纠正过也写出来', trialNote.includes('qurey'), trialNote)
|
||||
await shot(page, 'custom-tool-trial')
|
||||
await ed.getByLabel(/^URL/).fill('https://api.example.com/fail')
|
||||
check('……改了配置，刚才的结果标成旧的', (await trialBox.innerText()).includes('这是改之前的结果'))
|
||||
await tryBtn.click()
|
||||
await page.waitForTimeout(400)
|
||||
check('……失败：原因和怎么办写在试跑区', (await trialBox.innerText()).includes('HTTP 404') && (await trialBox.innerText()).includes('核对接口地址'))
|
||||
await ed.getByRole('button', { name: '取消' }).click()
|
||||
await page.waitForTimeout(200)
|
||||
const discard = dialog(page).getByRole('button', { name: '放弃修改' })
|
||||
if (await discard.count()) await discard.click()
|
||||
|
||||
await page.getByRole('tab', { name: 'MCP 接入' }).click()
|
||||
await page.waitForTimeout(400)
|
||||
const okCard = page.locator('article', { hasText: 'files_demo' })
|
||||
const okPill = await okCard.locator('[data-health]').innerText().catch(() => '')
|
||||
check('MCP 卡片初值是后端记着的上次探测：已连通 · 88 ms · 5 分钟前测', /已连通.*88 ms.*5 分钟前测/.test(okPill), okPill)
|
||||
check('……只有 status、没有测连接记录的老数据写「未测试」，不再拿 status 猜',
|
||||
(await page.locator('article', { hasText: 'legacy_demo' }).locator('[data-health]').getAttribute('data-health')) === 'idle')
|
||||
await okCard.getByRole('button', { name: /测试/ }).click()
|
||||
await page.waitForTimeout(500)
|
||||
const probed = await okCard.locator('[data-health]').innerText()
|
||||
check('……测一次：用后端量的耗时（120 ms），不是往返时间', /已连通.*120 ms.*刚测过/.test(probed), probed)
|
||||
await shot(page, 'mcp-health')
|
||||
await page.getByRole('button', { name: /重新加载/ }).click()
|
||||
await page.waitForTimeout(700)
|
||||
check('写操作失败有 error toast（以前一声不响）', await page.locator('[role="alert"][aria-live="assertive"] > *').count() > 0)
|
||||
|
||||
@@ -14,6 +14,9 @@
|
||||
// 这些推导状态和循环容器「两轮之间」也按那一刻算
|
||||
// - 协作矩阵的表头不和底部打架：并行的一轮交回一部分、全部交回等调度时都不说「串行」
|
||||
// - 去审批浮层键盘可达；出口密的分支节点标签不压别人的线；精简档读数够大
|
||||
// - 收场要说人话：协作团队用完轮数（判失败 / 降档交付）写清「用完 N 轮未完成」、点出
|
||||
// 一次都没被派到的成员，收尾判定不另算一轮；模型把工具调用写成文字的失败写原因，
|
||||
// 不是半句原话；工具超出后端声明的时限直说「已超出 Ns 上限」；校验靠修复才过要提一句
|
||||
//
|
||||
// 用法：AGENTLAB_WEB=http://localhost:5373 AGENTLAB_API=http://localhost:8100/api node scripts/check-run-states.mjs
|
||||
// 截图落在 RUN_STATES_SHOTS（默认 /tmp/agentlab-run-states）。
|
||||
@@ -25,8 +28,10 @@ const API = process.env.AGENTLAB_API ?? 'http://localhost:8000/api'
|
||||
const CHROME = process.env.CHROME_PATH
|
||||
?? '/Applications/Google Chrome.app/Contents/MacOS/Google Chrome'
|
||||
const SHOTS = process.env.RUN_STATES_SHOTS ?? '/tmp/agentlab-run-states'
|
||||
// 截图的像素密度：要看清卡片里 11px 的字时设成 2
|
||||
const DSF = Number(process.env.RUN_STATES_DSF ?? 1) || 1
|
||||
mkdirSync(SHOTS, { recursive: true })
|
||||
// 只跑其中几段（逗号分隔：themes, interact, team, replay, history, reduced, exits），调样式时省时间
|
||||
// 只跑其中几段（逗号分隔：themes, interact, team, endings, replay, history, reduced, exits），调样式时省时间
|
||||
const ONLY = (process.env.RUN_STATES_ONLY ?? '').split(',').filter(Boolean)
|
||||
const want = (part) => !ONLY.length || ONLY.includes(part)
|
||||
|
||||
@@ -237,6 +242,7 @@ async function open({ theme = 'dark', reduced = false, workflow = WORKFLOW } = {
|
||||
const browser = await chromium.launch({ executablePath: CHROME })
|
||||
const ctx = await browser.newContext({
|
||||
viewport: { width: 1440, height: 900 },
|
||||
deviceScaleFactor: DSF,
|
||||
colorScheme: theme,
|
||||
reducedMotion: reduced ? 'reduce' : 'no-preference',
|
||||
})
|
||||
@@ -372,6 +378,20 @@ async function shot(page, name) {
|
||||
return path
|
||||
}
|
||||
|
||||
/** 只截几张卡（外扩一圈）：看卡片里的字用，整张画布截图里它们太小 */
|
||||
async function cardShot(page, ids, name) {
|
||||
const boxes = (await Promise.all(ids.map((id) => page.locator(`.react-flow__node[data-id="${id}"]`).boundingBox())))
|
||||
.filter(Boolean)
|
||||
if (!boxes.length) return null
|
||||
const x = Math.min(...boxes.map((b) => b.x)) - 16
|
||||
const y = Math.min(...boxes.map((b) => b.y)) - 16
|
||||
const w = Math.max(...boxes.map((b) => b.x + b.width)) + 16 - x
|
||||
const h = Math.max(...boxes.map((b) => b.y + b.height)) + 16 - y
|
||||
const path = `${SHOTS}/${name}.png`
|
||||
await page.screenshot({ path, clip: { x: Math.max(0, x), y: Math.max(0, y), width: w, height: h } })
|
||||
return path
|
||||
}
|
||||
|
||||
const SCENARIOS = [
|
||||
{ key: 'live', events: LIVE, expect: { team: 'running', loop: 'running', body: 'running', join: 'queued', skip: 'skipped', gate: 'idle', in: 'done' } },
|
||||
{ key: 'waiting', events: WAITING, expect: { human: 'waiting', body: 'done', loop: 'done', team: 'done', out: 'idle' } },
|
||||
@@ -717,10 +737,32 @@ if (want('interact')) {
|
||||
const stampInfo = await stamp.evaluate((el) => ({
|
||||
tag: el.tagName, role: el.getAttribute('role'), tab: el.tabIndex, cursor: getComputedStyle(el).cursor, title: el.title,
|
||||
})).catch(() => null)
|
||||
// 右栏还没有可以滚过去的出具横幅：它是一枚章，不装作能点
|
||||
check('印章不是个点了没反应的按钮', !!stampInfo && stampInfo.tag !== 'BUTTON' && stampInfo.role === 'img'
|
||||
// 它是一枚章，不是按钮：不进 Tab 序列，悬停前没有手形
|
||||
check('印章不是个按钮', !!stampInfo && stampInfo.tag !== 'BUTTON' && stampInfo.role === 'img'
|
||||
&& stampInfo.tab < 0 && stampInfo.cursor !== 'pointer', JSON.stringify(stampInfo && { ...stampInfo, title: undefined }))
|
||||
check('印章悬停写出「校验没跑全」', (stampInfo?.title ?? '').includes('校验没跑全：指标集为空'), (stampInfo?.title ?? '').split('\n').join(' / '))
|
||||
// 右栏看得见时点它是去看完整判定的捷径：广播 agentlab:goto-issuance，右栏在对话层时自己切过去
|
||||
await page.evaluate(() => {
|
||||
window.__issuance = []
|
||||
window.addEventListener('agentlab:goto-issuance', (e) => window.__issuance.push(e.detail))
|
||||
})
|
||||
await stamp.hover()
|
||||
const cursor = await stamp.evaluate((el) => getComputedStyle(el).cursor)
|
||||
await stamp.click()
|
||||
const went = await page.evaluate(() => window.__issuance)
|
||||
check('右栏看得见：悬停成了链接，点它请右栏摆出出具横幅', cursor === 'pointer' && went.length === 1 && went[0].runId === RUN_ID,
|
||||
`${cursor} · ${JSON.stringify(went)}`)
|
||||
// 回指明细带口径（issuance.matched[].caliber）：同一个指标换过口径，看得出这次按的是哪一版
|
||||
await page.evaluate(() => {
|
||||
const st = window.__studio
|
||||
const out = st.getState().run.output
|
||||
st.setState({ run: { ...st.getState().run, output: { ...out, _issuance: { ...out._issuance,
|
||||
matched: [{ token: '12.4%', metric: '毛利率', caliber: '月度口径 @ v2' }, { token: '1,204', metric: '订单数' }] } } } })
|
||||
})
|
||||
await page.waitForTimeout(150)
|
||||
const matchedTitle = await stamp.getAttribute('title')
|
||||
check('印章悬停列出回指明细和口径', (matchedTitle ?? '').includes('12.4% → 毛利率（月度口径 @ v2）')
|
||||
&& (matchedTitle ?? '').includes('1,204 → 订单数'), (matchedTitle ?? '').split('\n').join(' / '))
|
||||
check('运行结束后端口恢复可连', await page.evaluate(() =>
|
||||
document.querySelector('.react-flow__node[data-id="gate"] .react-flow__handle')?.classList.contains('connectable')))
|
||||
// 探索运行进行中也不能连线：端口跟着 React Flow 的 connectable 走
|
||||
@@ -787,6 +829,371 @@ if (want('team')) {
|
||||
await browser.close()
|
||||
}
|
||||
|
||||
// ---------------------------------------------------------------- 收场：用完轮数、写成文字的工具调用、超时
|
||||
|
||||
if (want('endings')) {
|
||||
// 一张小图：协作团队两轮上限、三个成员（「定稿员」一次都派不到）；一个 agent 节点把
|
||||
// 工具调用写成了文字；一个工具节点查询超时；一个校验节点靠修复才过
|
||||
const graph = await (await fetch(`${API}/copilot/layout`, {
|
||||
method: 'POST', headers: { 'content-type': 'application/json' },
|
||||
body: JSON.stringify({ graph: {
|
||||
nodes: [
|
||||
{ id: 'in', type: 'input', data: { label: '入口', config: { fields: [{ name: 'week' }] } } },
|
||||
{ id: 'team', type: 'supervisor', data: { label: '复盘小组', config: {
|
||||
goal: '汇总本周订单异常', max_rounds: 2, max_parallel: 2,
|
||||
agents: [agent('检索员', '拉订单明细'), agent('分析员', '找异常原因'), agent('定稿员', '写结论')],
|
||||
} } },
|
||||
{ id: 'ask', type: 'agent', data: { label: '订单查询', config: { prompt: '查 orders 本周异常', tools: ['db_query__shop'] } } },
|
||||
{ id: 'q', type: 'tool', data: { label: '取订单', config: { tool: 'db_query__shop' } } },
|
||||
{ id: 'chk', type: 'validate', data: { label: '结构校验', config: { repair_with_llm: true, max_retries: 2 } } },
|
||||
{ id: 'out', type: 'output', data: { label: '周报', config: { fields: [{ name: 'report' }] } } },
|
||||
],
|
||||
edges: [
|
||||
{ source: 'in', target: 'team' }, { source: 'in', target: 'ask' }, { source: 'in', target: 'q' },
|
||||
{ source: 'q', target: 'chk' }, { source: 'team', target: 'out' }, { source: 'ask', target: 'out' },
|
||||
{ source: 'chk', target: 'out' },
|
||||
],
|
||||
} }),
|
||||
})).json()
|
||||
const workflow = { ...WORKFLOW, id: 'wcard-endings', name: '__endings__', graph }
|
||||
|
||||
const MARKUP = '模型输出了工具调用的原始标记，但没有真正调用工具。常见原因：节点没有绑定工具，或者模型、服务不支持工具调用。'
|
||||
const EXHAUSTED = '协作团队用完 2 轮仍未完成：还没有定稿。一次都没被派到的成员：定稿员。先看成员有没有绑定要用的工具，再调大「最多轮数」；也可以把「用完轮数时」改成降档交付'
|
||||
const START = [
|
||||
[0, 'run.started', null, { nodes: 6 }],
|
||||
[0.1, 'node.started', 'in', {}],
|
||||
[0.05, 'node.finished', 'in', { duration_ms: 3, preview: { week: '2026-W39' } }],
|
||||
]
|
||||
// 两轮都派了、调度者始终没说完成；第一轮的检索员把工具调用写成了文字，这一步记失败
|
||||
const ROUNDS = [
|
||||
[0.05, 'node.started', 'team', {}],
|
||||
[0.02, 'agent.route.start', 'team', { round: 0 }],
|
||||
[1.2, 'agent.route.end', 'team', { round: 0, duration_ms: 1200, agents: ['检索员'], parallel: 1, done: false, reason: '先拉订单明细' }],
|
||||
[0.02, 'agent.step.start', 'team', { agent: '检索员', instruction: '拉本周订单', round: 0, parallel: 1 }],
|
||||
[2.0, 'agent.step.end', 'team', { agent: '检索员', duration_ms: 2000, round: 0, parallel: 1, preview: '<tool_call>{"name": "db_query__shop"}</tool_call>',
|
||||
failed: true, error: '模型输出了工具调用的原始标记，但没有真正调用工具' }],
|
||||
[0.02, 'agent.route.start', 'team', { round: 1 }],
|
||||
[1.0, 'agent.route.end', 'team', { round: 1, duration_ms: 1000, agents: ['分析员'], parallel: 1, done: false, reason: '换人分析' }],
|
||||
[0.02, 'agent.step.start', 'team', { agent: '分析员', instruction: '找异常原因', round: 1, parallel: 1 }],
|
||||
[1.5, 'agent.step.end', 'team', { agent: '分析员', duration_ms: 1500, round: 1, parallel: 1, preview: '缺数据，无法判断' }],
|
||||
]
|
||||
// 轮数用完之后的收尾判定：只判定、不派活，round 是上限（从 0 数就是「第 3 轮」）
|
||||
const closing = (done, reason) => [
|
||||
[0.02, 'agent.route.start', 'team', { round: 2, closing: true }],
|
||||
[0.9, 'agent.route.end', 'team', { round: 2, closing: true, duration_ms: 900, agents: [], parallel: 0, done, reason }],
|
||||
[0.05, 'llm.end', 'team', { agent: '调度者', model: 'm', input_tokens: 900, output_tokens: 120, cost_usd: 0.001 }],
|
||||
]
|
||||
const FAIL = script([...START,
|
||||
[0.02, 'node.started', 'ask', {}],
|
||||
...ROUNDS, ...closing(false, '还没有定稿'),
|
||||
[0.02, 'log', 'ask', { level: 'warn', code: 'tool_markup_leak', message: '模型输出了工具调用的原始标记，但没有真正调用工具' }],
|
||||
[0.02, 'node.failed', 'ask', { duration_ms: 6800, error: MARKUP }],
|
||||
[0.05, 'node.failed', 'team', { duration_ms: 6800, error: EXHAUSTED }],
|
||||
[0.05, 'run.failed', null, { error: '「复盘小组」失败', node_id: 'team', timing: { wall_ms: 7000, active_ms: 7000, wait_ms: 0 } }],
|
||||
])
|
||||
const DEGRADE = script([...START,
|
||||
[0.02, 'node.started', 'ask', {}],
|
||||
// 第一次查询超出声明的 2s 时限被放弃，第二次查到了；收尾时模型还想要工具
|
||||
[0.05, 'tool.start', 'ask', { tool: 'db_query__shop', call_id: 'c1', timeout_s: 2 }],
|
||||
[3.0, 'tool.error', 'ask', { tool: 'db_query__shop', call_id: 'c1', duration_ms: 3000, timed_out: true, error: '查询超过 2s 没有返回，已放弃等待' }],
|
||||
[0.05, 'tool.start', 'ask', { tool: 'db_query__shop', call_id: 'c2', timeout_s: 2 }],
|
||||
[0.4, 'tool.end', 'ask', { tool: 'db_query__shop', call_id: 'c2', duration_ms: 400, preview: '{"row_count": 12}' }],
|
||||
[0.1, 'log', 'ask', { level: 'warn', code: 'tool_markup_leak', message: '收尾轮仍想调用工具：步数用完了' }],
|
||||
[0.05, 'llm.end', 'ask', { agent: '订单查询', model: 'm', input_tokens: 1200, output_tokens: 300, cost_usd: 0.002 }],
|
||||
[0.02, 'node.finished', 'ask', { duration_ms: 3700, preview: { text: '本周异常 12 单' } }],
|
||||
[0.02, 'node.started', 'q', {}],
|
||||
[0.02, 'tool.start', 'q', { tool: 'db_query__shop', call_id: 'c3', timeout_s: 30 }],
|
||||
[0.3, 'tool.end', 'q', { tool: 'db_query__shop', call_id: 'c3', duration_ms: 300, preview: '{"row_count": 12}' }],
|
||||
[0.02, 'node.finished', 'q', { duration_ms: 320, preview: { row_count: 12 } }],
|
||||
[0.02, 'node.started', 'chk', {}],
|
||||
[0.4, 'llm.end', 'chk', { model: 'm', purpose: 'repair', duration_ms: 400, input_tokens: 300, output_tokens: 40, cost_usd: 0.0005, calls: 1 }],
|
||||
[0.02, 'log', 'chk', { level: 'warn', code: 'repair_invented', message: '修复时出现了原文没有的值：total_count=0' }],
|
||||
[0.4, 'llm.end', 'chk', { model: 'm', purpose: 'repair', duration_ms: 400, input_tokens: 300, output_tokens: 40, cost_usd: 0.0005, calls: 1 }],
|
||||
[0.02, 'node.finished', 'chk', { duration_ms: 850, preview: { total_count: 12 } }],
|
||||
...ROUNDS, ...closing(false, '还没有定稿'),
|
||||
[0.02, 'log', 'team', { level: 'warn', code: 'team_exhausted', message: '协作团队用完 2 轮仍未完成：还没有定稿。一次都没被派到的成员：定稿员。按降档交付' }],
|
||||
[0.02, 'node.finished', 'team', { duration_ms: 6800, preview: { exhausted: true, exhausted_reason: '还没有定稿', never_dispatched: '[1 项]', text: '缺数据，无法判断' } }],
|
||||
[0.05, 'node.started', 'out', {}],
|
||||
[0.05, 'node.finished', 'out', { duration_ms: 4, preview: { report: '…' } }],
|
||||
[0.05, 'run.finished', null, { output: {}, usage: {}, duration_ms: 9000, timing: { wall_ms: 9000, active_ms: 9000, wait_ms: 0 } }],
|
||||
])
|
||||
// 进行中：收尾判定还没出结论；工具节点的查询已经超出声明的 2s
|
||||
const LIVE_JUDGING = script([...START,
|
||||
...ROUNDS,
|
||||
[0.02, 'agent.route.start', 'team', { round: 2, closing: true }],
|
||||
[0.02, 'node.started', 'q', {}],
|
||||
[0.02, 'tool.start', 'q', { tool: 'db_query__shop', call_id: 'c9', timeout_s: 2 }],
|
||||
[3.2, 'log', null, { level: 'info', message: '心跳' }],
|
||||
])
|
||||
// 收尾判定说完成了：正常交付，表头说清是轮数用完之后判的，不点名谁没被派到
|
||||
const CLOSED_OK = script([...START,
|
||||
...ROUNDS, ...closing(true, '材料够了，可以收尾'),
|
||||
[0.05, 'node.finished', 'team', { duration_ms: 6800, preview: { text: '结论：…' } }],
|
||||
[0.05, 'run.finished', null, { output: {}, usage: {}, duration_ms: 7000, timing: { wall_ms: 7000, active_ms: 7000, wait_ms: 0 } }],
|
||||
])
|
||||
const askLlm = (out) => [0.05, 'llm.end', 'ask', { agent: '订单查询', model: 'm', input_tokens: 600, output_tokens: out, cost_usd: 0.001 }]
|
||||
// 同一个节点执行两次：第一次的收场不能挂到第二次上。团队用完轮数判失败 → 调大轮数
|
||||
// 接着跑（后端给失败的节点标 resumed，但这是重新执行；前两轮在断点里，不重发）→ 派到
|
||||
// 定稿员、调度者判完成。agent 第一次收尾时还想调工具，第二次（循环体每一轮也是这样）干干净净
|
||||
const RERUN = script([...START,
|
||||
[0.02, 'node.started', 'ask', {}],
|
||||
[0.3, 'log', 'ask', { level: 'warn', code: 'tool_markup_leak',
|
||||
message: '模型把工具调用写成了文字（<tool_call>{"name": "db_query__shop"…），没有真正调用工具：步数用完后的收尾轮仍想调用工具，没能给出结论' }],
|
||||
askLlm(300),
|
||||
[0.02, 'node.finished', 'ask', { duration_ms: 400, preview: { text: '本周异常 12 单' } }],
|
||||
...ROUNDS, ...closing(false, '还没有定稿'),
|
||||
[0.05, 'node.failed', 'team', { duration_ms: 6800, error: EXHAUSTED }],
|
||||
[0.05, 'run.failed', null, { error: '「复盘小组」失败', node_id: 'team', timing: { wall_ms: 7000, active_ms: 7000, wait_ms: 0 } }],
|
||||
[2.0, 'run.resumed', null, { from: ['team'], message: '从「复盘小组」接着跑，前面 2 个节点的结果保留' }],
|
||||
[0.02, 'run.started', null, { nodes: 6, resumed: true }],
|
||||
[0.05, 'node.started', 'team', { resumed: true }],
|
||||
[0.02, 'agent.route.start', 'team', { round: 2 }],
|
||||
[0.8, 'agent.route.end', 'team', { round: 2, duration_ms: 800, agents: ['定稿员'], parallel: 1, done: false, reason: '派定稿员写结论' }],
|
||||
[0.02, 'agent.step.start', 'team', { agent: '定稿员', instruction: '写结论', round: 2, parallel: 1 }],
|
||||
[1.2, 'agent.step.end', 'team', { agent: '定稿员', duration_ms: 1200, round: 2, parallel: 1, preview: '结论：…' }],
|
||||
[0.02, 'agent.route.start', 'team', { round: 3 }],
|
||||
[0.6, 'agent.route.end', 'team', { round: 3, duration_ms: 600, agents: [], parallel: 0, done: true, reason: '已经定稿' }],
|
||||
[0.05, 'node.finished', 'team', { duration_ms: 2700, preview: { text: '结论：…' } }],
|
||||
[0.05, 'node.started', 'ask', {}],
|
||||
askLlm(240),
|
||||
[0.3, 'node.finished', 'ask', { duration_ms: 320, preview: { text: '本周异常 9 单' } }],
|
||||
[0.05, 'node.started', 'out', {}],
|
||||
[0.05, 'node.finished', 'out', { duration_ms: 4, preview: { report: '…' } }],
|
||||
[0.05, 'run.finished', null, { output: {}, usage: {}, duration_ms: 12000, timing: { wall_ms: 12000, active_ms: 10000, wait_ms: 0 } }],
|
||||
])
|
||||
// 写成文字、被提醒之后重答了：成果是重答的那一次，不是「收尾时仍想调用工具」那种不完整。
|
||||
// 团队里检索员提醒后重答了；分析员提醒后还这样，这一步记失败，不能算进「纠正过」
|
||||
const NUDGE = script([...START,
|
||||
[0.02, 'node.started', 'ask', {}],
|
||||
[0.3, 'log', 'ask', { level: 'warn', code: 'tool_markup_leak',
|
||||
message: '模型把工具调用写成了文字（<tool_call>{"name": "db_query__shop"…），没有真正调用工具，已提醒它重试一次' }],
|
||||
[0.02, 'tool.start', 'ask', { tool: 'db_query__shop', call_id: 'n1', timeout_s: 30 }],
|
||||
[0.3, 'tool.end', 'ask', { tool: 'db_query__shop', call_id: 'n1', duration_ms: 300, preview: '{"row_count": 12}' }],
|
||||
askLlm(280),
|
||||
[0.02, 'node.finished', 'ask', { duration_ms: 700, preview: { text: '本周异常 12 单' } }],
|
||||
[0.02, 'node.started', 'team', {}],
|
||||
[0.02, 'agent.route.start', 'team', { round: 0 }],
|
||||
[0.8, 'agent.route.end', 'team', { round: 0, duration_ms: 800, agents: ['检索员', '分析员'], parallel: 2, done: false, reason: '先拉明细、再找原因' }],
|
||||
[0.02, 'agent.step.start', 'team', { agent: '检索员', instruction: '拉本周订单', round: 0, parallel: 2 }],
|
||||
[0.01, 'agent.step.start', 'team', { agent: '分析员', instruction: '找异常原因', round: 0, parallel: 2 }],
|
||||
[0.4, 'log', 'team', { level: 'warn', code: 'tool_markup_leak',
|
||||
message: '检索员把工具调用写成了文字（<tool_call>{"name": "db_query__shop"…），没有真正调用工具，已提醒它重试一次' }],
|
||||
[0.2, 'log', 'team', { level: 'warn', code: 'tool_markup_leak',
|
||||
message: '分析员把工具调用写成了文字（<tool_call>{"name": "db_query__shop"…），没有真正调用工具，已提醒它重试一次' }],
|
||||
[0.6, 'agent.step.end', 'team', { agent: '检索员', duration_ms: 1200, round: 0, parallel: 2, preview: '拉到 12 单' }],
|
||||
[0.3, 'agent.step.end', 'team', { agent: '分析员', duration_ms: 1500, round: 0, parallel: 2, preview: '',
|
||||
failed: true, error: '模型输出了工具调用的原始标记，但没有真正调用工具，这一步什么都没查到' }],
|
||||
[0.02, 'agent.route.start', 'team', { round: 1 }],
|
||||
[0.6, 'agent.route.end', 'team', { round: 1, duration_ms: 600, agents: [], parallel: 0, done: true, reason: '明细够了' }],
|
||||
[0.05, 'node.finished', 'team', { duration_ms: 2500, preview: { text: '结论：…' } }],
|
||||
[0.05, 'run.finished', null, { output: {}, usage: {}, duration_ms: 3500, timing: { wall_ms: 3500, active_ms: 3500, wait_ms: 0 } }],
|
||||
])
|
||||
/** 改团队节点的配置（运行之后调过「最多轮数」、或者压根没配）。值给 null 就是删掉这一项 */
|
||||
const setTeamConfig = (page, patch) => page.evaluate((patch) => {
|
||||
const st = window.__studio
|
||||
const node = st.getState().nodes.find((n) => n.id === 'team')
|
||||
const config = { ...node.data.config, ...patch }
|
||||
for (const [k, v] of Object.entries(patch)) if (v === null) delete config[k]
|
||||
st.setState({ nodes: st.getState().nodes.map((n) => (n.id === 'team' ? { ...n, data: { ...n.data, config } } : n)) })
|
||||
}, patch)
|
||||
|
||||
const at = (page, id) => page.locator(`.react-flow__node[data-id="${id}"]`)
|
||||
const flat = (t) => (t ?? '').replace(/\s+/g, ' ').trim()
|
||||
const matrixOf = (page) => page.evaluate(() => {
|
||||
const m = document.querySelector('.react-flow__node[data-id="team"] .team-matrix')
|
||||
const right = m?.querySelector('.team-head-right')
|
||||
const row = (name) => [...(m?.querySelectorAll('.team-row') ?? [])].find((r) => r.textContent.includes(name))
|
||||
const member = (name) => {
|
||||
const r = row(name)
|
||||
return r ? { status: r.dataset.memberStatus, text: r.innerText.replace(/\s+/g, ' ').trim(), title: r.title } : null
|
||||
}
|
||||
return {
|
||||
head: (m?.querySelector('.team-head')?.textContent ?? '').replace(/\s+/g, ' ').trim(),
|
||||
headTitle: m?.querySelector('.team-head')?.getAttribute('title') ?? '',
|
||||
right: right?.textContent ?? '', tone: right?.classList.contains('is-err') ? 'err' : right?.classList.contains('is-warn') ? 'warn' : '',
|
||||
dispatch: (m?.querySelector('.team-row-dispatch')?.textContent ?? '').replace(/\s+/g, ' ').trim(),
|
||||
a: member('检索员'), b: member('分析员'), c: member('定稿员'),
|
||||
}
|
||||
})
|
||||
const note = (page, id) => page.evaluate((id) => {
|
||||
const el = document.querySelector(`.react-flow__node[data-id="${id}"] .nc-tele-note`)
|
||||
return el ? { text: el.textContent.trim(), title: el.title, warn: el.classList.contains('is-warn') } : null
|
||||
}, id)
|
||||
|
||||
for (const theme of ['dark', 'light']) {
|
||||
console.log(`\n=== ${theme === 'dark' ? '暗色' : '亮色'}：收场说人话（用完轮数、工具调用写成文字、超时、修复） ===`)
|
||||
const { browser, page, errors } = await open({ theme, workflow })
|
||||
await setLod(page, 'full')
|
||||
const base = Object.fromEntries(Object.entries(await cards(page)).map(([id, c]) => [id, c.h]))
|
||||
|
||||
// ---- 判失败：用完轮数（默认）
|
||||
await feed(page, FAIL(Date.now()))
|
||||
await page.waitForTimeout(900)
|
||||
let c = await cards(page)
|
||||
let m = await matrixOf(page)
|
||||
const teamNote = await note(page, 'team')
|
||||
check(`${theme}: 团队用完轮数判失败，卡片写「用完 2 轮仍未完成」`, c.team.state === 'failed'
|
||||
&& teamNote?.text === '协作团队用完 2 轮仍未完成', JSON.stringify(teamNote?.text))
|
||||
check(`${theme}: 悬停写清理由和怎么办`, (teamNote?.title ?? '').includes('还没有定稿') && (teamNote?.title ?? '').includes('怎么办'),
|
||||
flat(teamNote?.title))
|
||||
check(`${theme}: 收尾判定不另算一轮（表头仍是第 2 轮 / 上限 2）`, m.head.startsWith('第 2 轮 / 上限 2'), m.head)
|
||||
check(`${theme}: 表头说「用完 2 轮 · 未完成」，失败上红`, m.right === '用完 2 轮 · 未完成' && m.tone === 'err', `${m.right} · ${m.tone}`)
|
||||
check(`${theme}: 调度者那一行写「收尾判定」`, m.dispatch.includes('收尾判定'), m.dispatch)
|
||||
check(`${theme}: 一次都没被派到的成员点出来`, m.c?.status === 'never' && m.c.text.includes('未派到')
|
||||
&& m.headTitle.includes('一次都没被派到：定稿员'), JSON.stringify(m.c))
|
||||
check(`${theme}: 成员把工具调用写成文字：格子是失败，悬停写原因`, m.a?.status === 'failed'
|
||||
&& (m.a.title ?? '').includes('原因：模型输出了工具调用的原始标记'), JSON.stringify(m.a))
|
||||
const askNote = await note(page, 'ask')
|
||||
check(`${theme}: agent 把工具调用写成文字而失败：卡片写原因，不是半句原话`, c.ask.state === 'failed'
|
||||
&& askNote?.text === '模型没有真正调用工具' && (askNote?.title ?? '').includes('怎么办'), JSON.stringify(askNote))
|
||||
const grew = Object.entries(c).filter(([id, x]) => x.h !== base[id])
|
||||
check(`${theme}: 失败收场不撑高卡片`, grew.length === 0, grew.map(([id, x]) => `${id} ${base[id]}→${x.h}`).join(', '))
|
||||
shots.push(await shot(page, `${theme}-endings-failed`), await cardShot(page, ['team', 'ask'], `${theme}-endings-failed-cards`))
|
||||
|
||||
// ---- 降档交付、收尾时还想调工具、查询超时、修复才过
|
||||
await feed(page, DEGRADE(Date.now()))
|
||||
await page.waitForTimeout(900)
|
||||
c = await cards(page)
|
||||
m = await matrixOf(page)
|
||||
const degr = await note(page, 'team')
|
||||
const chip = await at(page, 'team').locator('.nc-head .nc-chip-warn', { hasText: '降档' }).count()
|
||||
check(`${theme}: 降档交付：卡片写「用完 2 轮未完成 · 降档交付」，琥珀色、带「降档」标`, c.team.state === 'done'
|
||||
&& degr?.text === '用完 2 轮未完成 · 降档交付' && degr.warn && chip === 1, JSON.stringify({ degr, chip }))
|
||||
check(`${theme}: 降档交付：表头琥珀、点出没被派到的成员`, m.right === '用完 2 轮 · 未完成' && m.tone === 'warn'
|
||||
&& m.c?.status === 'never', `${m.right} · ${m.tone} · ${m.c?.status}`)
|
||||
const askDone = await note(page, 'ask')
|
||||
check(`${theme}: 收尾时还想调工具：跑完了但琥珀色提一句`, c.ask.state === 'done' && askDone?.text === '收尾时仍想调用工具' && askDone.warn,
|
||||
JSON.stringify(askDone))
|
||||
const tools = await at(page, 'ask').locator('.nc-tools').evaluate((el) => ({ bad: el.classList.contains('is-bad'), title: el.title }))
|
||||
.catch(() => null)
|
||||
check(`${theme}: 超时的那次调用算失败，悬停写「超时（上限 2s）」`, !!tools?.bad && tools.title.includes('超时（上限 2s）'), JSON.stringify(tools))
|
||||
const chkNote = await note(page, 'chk')
|
||||
check(`${theme}: 校验靠修复才过：写「修复 2 次后通过」，有一次凑数被作废就上琥珀`, chkNote?.text === '修复 2 次后通过'
|
||||
&& chkNote.warn && chkNote.title.includes('total_count=0'), JSON.stringify(chkNote))
|
||||
shots.push(await shot(page, `${theme}-endings-degraded`), await cardShot(page, ['team', 'ask', 'q'], `${theme}-endings-degraded-cards`))
|
||||
// 校验节点在右边，常被右栏挡住：取景过去再截
|
||||
await page.evaluate(() => window.__studio.getState().focusNode('chk'))
|
||||
await page.waitForTimeout(600)
|
||||
shots.push(await cardShot(page, ['chk'], `${theme}-endings-degraded-validate`))
|
||||
|
||||
// 回放到收尾判定还没出结论的那一刻：不提前降档、不提前点名
|
||||
const judging = await page.evaluate(() => {
|
||||
const segs = window.__studio.getState().trace.nodes.team.segments.filter((s) => s.kind === 'dispatch')
|
||||
return segs[segs.length - 1].start + 300
|
||||
})
|
||||
await page.evaluate((t) => window.__studio.getState().setReplayAt(t), judging)
|
||||
await page.waitForTimeout(300)
|
||||
m = await matrixOf(page)
|
||||
const early = await at(page, 'team').locator('.nc-head .nc-chip-warn', { hasText: '降档' }).count()
|
||||
check(`${theme}: 回放到收尾判定中：表头写「调度者在判定」，不提前降档、不点名`, m.right.includes('调度者在判定') && early === 0
|
||||
&& m.c?.status !== 'never', `${m.right} · 降档标 ${early} · ${m.c?.status}`)
|
||||
await page.evaluate(() => window.__studio.getState().setReplayAt(null))
|
||||
|
||||
// ---- 进行中：收尾判定中；查询超出声明的时限
|
||||
await feed(page, LIVE_JUDGING(Date.now()))
|
||||
await page.waitForTimeout(600)
|
||||
m = await matrixOf(page)
|
||||
check(`${theme}: 进行中的收尾判定：表头写「轮数用完 · 调度者在判定」，不是新的一轮`,
|
||||
m.right === '轮数用完 · 调度者在判定' && m.head.startsWith('第 2 轮 / 上限 2') && m.dispatch.includes('判定'), `${m.head} | ${m.dispatch}`)
|
||||
const over = await at(page, 'q').locator('.nc-over').innerText().catch(() => '')
|
||||
const overTitle = await at(page, 'q').locator('.nc-tools').getAttribute('title').catch(() => '') ?? ''
|
||||
// 这句话整个落在卡片里，不被右边沿切掉
|
||||
const overFits = await page.evaluate(() => {
|
||||
const el = document.querySelector('.react-flow__node[data-id="q"] .nc-over')
|
||||
const card = el?.closest('.nc')
|
||||
if (!el || !card) return false
|
||||
return el.getBoundingClientRect().right <= card.getBoundingClientRect().right - 2
|
||||
})
|
||||
check(`${theme}: 查询超出声明的时限：直说「已超出 2s 上限」，工具名在悬停里`, over === '已超出 2s 上限' && overTitle.includes('db_query__shop ⋯ 进行中（上限 2s）') && overFits, `${over} · ${flat(overTitle)}${overFits ? '' : ' · 超出了卡片'}`)
|
||||
shots.push(await shot(page, `${theme}-endings-live`), await cardShot(page, ['team', 'q'], `${theme}-endings-live-cards`))
|
||||
await setLod(page, 'compact')
|
||||
const brief = await at(page, 'team').locator('.nc-lod-read').innerText()
|
||||
check(`${theme}: 精简档：收尾判定中写清是判定，不写「第 3 轮」`, brief === '轮数用完 · 收尾判定中', brief)
|
||||
// 收尾判定认事件里的 closing，不拿配置里的「最多轮数」比：没配（后端按缺省值跑）、
|
||||
// 运行之后调大了（接着跑之前、回放历史运行），照样认得出是在判定
|
||||
await setTeamConfig(page, { max_rounds: null })
|
||||
await page.waitForTimeout(150)
|
||||
const briefBare = await at(page, 'team').locator('.nc-lod-read').innerText()
|
||||
await setLod(page, 'full')
|
||||
m = await matrixOf(page)
|
||||
check(`${theme}: 没配最多轮数：照样认出收尾判定（精简档、表头、调度者那一行）`, briefBare === '轮数用完 · 收尾判定中'
|
||||
&& m.right === '轮数用完 · 调度者在判定' && m.dispatch.includes('判定'), `${briefBare} | ${m.right} | ${m.dispatch}`)
|
||||
await setTeamConfig(page, { max_rounds: 4 })
|
||||
await page.waitForTimeout(150)
|
||||
m = await matrixOf(page)
|
||||
check(`${theme}: 运行之后调大了最多轮数：照样认出收尾判定`, m.right === '轮数用完 · 调度者在判定'
|
||||
&& m.dispatch.includes('判定'), `${m.right} | ${m.dispatch}`)
|
||||
await setTeamConfig(page, { max_rounds: 2 })
|
||||
|
||||
// ---- 收尾判定说完成了
|
||||
await feed(page, CLOSED_OK(Date.now()))
|
||||
await page.waitForTimeout(600)
|
||||
m = await matrixOf(page)
|
||||
check(`${theme}: 收尾判定已完成：表头说清、安静、不点名`, m.right === '轮数用完 · 判定已完成' && m.tone === ''
|
||||
&& m.head.startsWith('第 2 轮 / 上限 2') && m.c?.status !== 'never', `${m.head} · ${m.tone} · ${m.c?.status}`)
|
||||
check(`${theme}: 悬停写完整判定`, m.headTitle.includes('轮数用完 · 调度者判定：已完成') && m.headTitle.includes('材料够了'), flat(m.headTitle))
|
||||
|
||||
// ---- 同一个节点执行第二次：上一次的收场不挂到这一次上。照真实流程先把最多轮数调大
|
||||
// 再接着跑：回放到第一次时「用完 2 轮」得按那一次的原话，不按现在的配置写成 4
|
||||
await setTeamConfig(page, { max_rounds: 4 })
|
||||
await feed(page, RERUN(Date.now()))
|
||||
await page.waitForTimeout(700)
|
||||
const quietEnding = async (label) => {
|
||||
const cs = await cards(page)
|
||||
const mx = await matrixOf(page)
|
||||
const never = await at(page, 'team').locator('[data-member-status="never"]').count()
|
||||
const chips = await at(page, 'team').locator('.nc-head .nc-chip-warn', { hasText: '降档' }).count()
|
||||
const tn = await note(page, 'team')
|
||||
const an = await note(page, 'ask')
|
||||
check(`${theme}: ${label}：团队表头不挂上一次的「判定未完成」、安静、不点名`, cs.team.state === 'done' && mx.tone === ''
|
||||
&& !mx.right.includes('未完成') && !mx.headTitle.includes('未完成') && never === 0 && chips === 0 && !tn?.warn,
|
||||
`${mx.right} · 语气「${mx.tone}」· 未派到 ${never} · 降档标 ${chips} · ${JSON.stringify(tn)}`)
|
||||
check(`${theme}: ${label}：agent 这一次干干净净，写产出量，不写上一次的「收尾时仍想调用工具」`, cs.ask.state === 'done'
|
||||
&& !!an && !an.warn && /tok/.test(an.text), JSON.stringify(an))
|
||||
}
|
||||
await quietEnding('调大轮数接着跑成功、agent 再执行一次')
|
||||
shots.push(await cardShot(page, ['team', 'ask'], `${theme}-endings-rerun-cards`))
|
||||
const execs = await page.evaluate(() => {
|
||||
const n = window.__studio.getState().trace.nodes
|
||||
const runs = (id) => n[id].segments.filter((s) => s.kind === 'run')
|
||||
return { teamRuns: runs('team').length, askRuns: runs('ask').length,
|
||||
first: runs('team')[0].end + 300, second: runs('ask').at(-1).end + 10 }
|
||||
})
|
||||
check(`${theme}: 夹具确实是两次执行（团队、agent 各两段）`, execs.teamRuns === 2 && execs.askRuns === 2, JSON.stringify(execs))
|
||||
await page.evaluate((t) => window.__studio.getState().setReplayAt(t), execs.second)
|
||||
await page.waitForTimeout(300)
|
||||
await quietEnding('回放到第二次执行里')
|
||||
// 回放到第一次失败的那一刻：说的是那一次的收场，不因为后来成功了就不说
|
||||
await page.evaluate((t) => window.__studio.getState().setReplayAt(t), execs.first)
|
||||
await page.waitForTimeout(300)
|
||||
m = await matrixOf(page)
|
||||
const firstTeam = await note(page, 'team')
|
||||
const firstAsk = await note(page, 'ask')
|
||||
check(`${theme}: 回放到第一次失败那一刻：团队写那一次的「用完 2 轮仍未完成」，表头红、点名定稿员`,
|
||||
firstTeam?.text === '协作团队用完 2 轮仍未完成' && m.right === '用完 2 轮 · 未完成' && m.tone === 'err' && m.c?.status === 'never',
|
||||
`${JSON.stringify(firstTeam?.text)} · ${m.right} · ${m.tone} · ${m.c?.status}`)
|
||||
check(`${theme}: 回放到第一次执行：agent 写那一次的「收尾时仍想调用工具」`, firstAsk?.text === '收尾时仍想调用工具' && firstAsk.warn,
|
||||
JSON.stringify(firstAsk))
|
||||
await page.evaluate(() => window.__studio.getState().setReplayAt(null))
|
||||
await setTeamConfig(page, { max_rounds: 2 })
|
||||
|
||||
// ---- 写成文字、被提醒之后重答了：不是「收尾时仍想调用工具」
|
||||
await feed(page, NUDGE(Date.now()))
|
||||
await page.waitForTimeout(700)
|
||||
const askNudge = await note(page, 'ask')
|
||||
check(`${theme}: agent 写成文字被提醒后重答了：写「纠正过一次工具调用」，不说成果可能不完整`,
|
||||
askNudge?.text === '纠正过一次工具调用' && askNudge.warn && askNudge.title.includes('提醒之后重答了')
|
||||
&& !askNudge.title.includes('可能不完整'), JSON.stringify(askNudge))
|
||||
const teamNudge = await note(page, 'team')
|
||||
check(`${theme}: 协作成员提醒后重答了：团队卡同样写「纠正过一次」，提醒后仍失败的那位不算进去`,
|
||||
teamNudge?.text === '纠正过一次工具调用' && teamNudge.warn && teamNudge.title.includes('检索员')
|
||||
&& !teamNudge.title.includes('分析员'), JSON.stringify(teamNudge))
|
||||
shots.push(await cardShot(page, ['team', 'ask'], `${theme}-endings-nudge-cards`))
|
||||
check('没有运行时报错', errors.length === 0, errors.slice(0, 3).join(' | '))
|
||||
await browser.close()
|
||||
}
|
||||
}
|
||||
|
||||
if (want('replay')) {
|
||||
console.log('\n=== 回放：卡片回到游标那一刻 ===')
|
||||
const { browser, page, errors } = await open({ theme: 'dark' })
|
||||
|
||||
+619
-4
@@ -357,9 +357,28 @@ if (authFail) {
|
||||
check('有「接着跑」', await page.locator('[data-run-banner=failed] [data-action=continue]').count() === 1)
|
||||
check('有「去模型接入」', await page.locator('[data-run-banner=failed] [data-action=settings]').count() === 1)
|
||||
check('有「复制错误」', await page.locator('[data-run-banner=failed] [data-action=copy-error]').count() === 1)
|
||||
const href = await page.locator('[data-run-banner=failed] [data-action=locate]').getAttribute('href')
|
||||
check('「在画布中定位」带 run 和 focus', href === `/studio/${authFail.workflow_id}?run=${authFail.id}&focus=${full.error_node_id}`, href ?? '')
|
||||
check('失败的运行不给「提取模板」', await page.locator('[data-action=extract]').count() === 0)
|
||||
// 横幅里的定位是排错的主路，详情头上的回放是看全程：两处都得够得着画布。失败的节点
|
||||
// 在工作流现在的图里已经没了的话（之后改过结构），对准它会落空：横幅改给「在航迹中看」，
|
||||
// 详情头照样回放、只是不带 focus。两种情形下面的航迹夹具各测一遍，这里按库里的实情测
|
||||
const authWf = (await getJson('/workflows')).find((w) => w.id === authFail.workflow_id)
|
||||
const stillThere = !!authWf?.graph?.nodes?.some((n) => n.id === full.error_node_id)
|
||||
const canvasBase = `/studio/${authFail.workflow_id}?run=${authFail.id}`
|
||||
const headLink = page.locator('[data-run-detail] > header [data-action=open-canvas]')
|
||||
const headHref = await headLink.getAttribute('href').catch(() => null)
|
||||
const headText = await headLink.innerText().catch(() => '')
|
||||
if (stillThere) {
|
||||
const href = await page.locator('[data-run-banner=failed] [data-action=locate]').getAttribute('href').catch(() => null)
|
||||
check('「在画布中定位」带 run 和 focus', href === `${canvasBase}&focus=${full.error_node_id}`, href ?? '')
|
||||
check('失败的运行详情头上也有「在画布中回放」，对准失败的节点',
|
||||
headText.includes('在画布中回放') && headHref === href, headHref ?? '没有这个按钮')
|
||||
} else {
|
||||
check('失败的节点在工作流现在的图里没了：不给落空的「在画布中定位」，改给「在航迹中看」',
|
||||
await page.locator('[data-run-banner=failed] [data-action=locate]').count() === 0
|
||||
&& await page.locator('[data-run-banner=failed] [data-action=locate-trace]').count() === 1)
|
||||
check('失败的运行详情头上照样「在画布中回放」，不带对不上的 focus',
|
||||
headText.includes('在画布中回放') && headHref === canvasBase, headHref ?? '没有这个按钮')
|
||||
}
|
||||
|
||||
// 接着跑：POST 伪造成功，实时流在上面已经伪造好
|
||||
fakes.set(`POST /api/runs/${authFail.id}/continue`, () => ({
|
||||
@@ -463,7 +482,16 @@ if (syntaxFail) {
|
||||
await page.locator('[data-run-banner=failed]').waitFor()
|
||||
check('表达式写错的失败不给「接着跑」', await page.locator('[data-run-banner=failed] [data-action=continue]').count() === 0)
|
||||
const locate = page.locator('[data-run-banner=failed] [data-action=locate]')
|
||||
check('「在画布中定位」成了主按钮', await locate.count() === 1 && /\bbtn-primary\b/.test(await locate.getAttribute('class') ?? ''))
|
||||
// 失败的节点在工作流现在的图里没了的话，定位落空，主路改成去航迹看（定位作主按钮的
|
||||
// 情形由后面「模型没有真正调用工具」的夹具测）
|
||||
const synNode = (await getJson(`/runs/${syntaxFail.id}`)).error_node_id
|
||||
const synWf = (await getJson('/workflows')).find((w) => w.id === syntaxFail.workflow_id)
|
||||
if (synWf?.graph?.nodes?.some((n) => n.id === synNode)) {
|
||||
check('「在画布中定位」成了主按钮', await locate.count() === 1 && /\bbtn-primary\b/.test(await locate.getAttribute('class') ?? ''))
|
||||
} else {
|
||||
check('失败的节点在工作流现在的图里没了：不给落空的定位，给「在航迹中看」',
|
||||
await locate.count() === 0 && await page.locator('[data-run-banner=failed] [data-action=locate-trace]').count() === 1)
|
||||
}
|
||||
} else {
|
||||
check('沙箱里有一条表达式写错的失败运行', false)
|
||||
}
|
||||
@@ -478,6 +506,8 @@ if (held) {
|
||||
await page.waitForTimeout(500)
|
||||
check('显示「已挂起」而不是「等待审批」', await page.locator('[data-run-detail]').getAttribute('data-run-code') === 'held')
|
||||
check('挂起横幅带「接着跑」', await page.locator('[data-run-banner=held] [data-action=resume]').count() === 1)
|
||||
// 挂起的没人会再去接：不放弃的话它在记录里永远挂着「可续跑」
|
||||
check('挂起横幅也能放弃这次运行', await page.locator('[data-run-banner=held] [data-action=abandon]').count() === 1)
|
||||
}
|
||||
|
||||
// ------------------------------------------------------------------ 封存凭证
|
||||
@@ -573,6 +603,16 @@ await page.waitForTimeout(700)
|
||||
const clockB = await page.locator('[data-telemetry=wall]').innerText()
|
||||
check('墙钟在走（mm:ss.s）', /^\d\d:\d\d\.\d$/.test(clockB) && clockA !== clockB, `${clockA} → ${clockB}`)
|
||||
check('状态格写着当前节点', (await page.locator('[data-run-telemetry]').innerText()).includes('当前「总结」'))
|
||||
// 在跑的运行看航迹：游标跟着现在走，「此刻」列出在跑的节点
|
||||
await page.locator('[data-run-detail] [role=tab][data-tab=trace]').click()
|
||||
await page.locator('[data-trace-readout]').waitFor({ timeout: 5000 }).catch(() => {})
|
||||
const liveAtA = await page.locator('[data-trace-readout] [data-readout=at]').innerText().catch(() => '')
|
||||
await page.waitForTimeout(500)
|
||||
const liveAtB = await page.locator('[data-trace-readout] [data-readout=at]').innerText().catch(() => '')
|
||||
const liveMode = await page.locator('[data-trace-readout]').getAttribute('data-mode').catch(() => '')
|
||||
check('航迹上游标是实时的，而且在走', liveMode === '实时' && liveAtA !== liveAtB, `${liveMode} ${liveAtA} → ${liveAtB}`.replace(/\s+/g, ' '))
|
||||
check('「此刻」列出在跑的「总结」', await page.locator('[data-trace-now] [data-now-node=b]').count() === 1)
|
||||
await page.locator('[data-run-detail] [role=tab][data-tab=stream]').click()
|
||||
await page.getByRole('button', { name: '更多操作' }).click()
|
||||
const del = page.locator('[data-menu=delete]')
|
||||
check('运行中「删除记录」不可点，并说明要先停止', await del.isDisabled() && (await del.innerText()).includes('先停止'))
|
||||
@@ -703,8 +743,583 @@ await page.getByRole('dialog').getByRole('button', { name: '删除记录' }).cli
|
||||
check('409 时把后端的话原样告诉人', await page.getByText('先停止它，再删除').first().waitFor({ timeout: 4000 }).then(() => true, () => false))
|
||||
check('409 时停在原地', new URL(page.url()).pathname === `/runs/${BUSY}`)
|
||||
|
||||
// ------------------------------------------------------------------ 航迹
|
||||
|
||||
console.log('\n=== 航迹：按时间摊开，拖到哪一刻读到哪一刻 ===')
|
||||
// 一次等过 5 分钟审批的运行：查询 → 审批 → 汇总。名字都是编的通用示例
|
||||
const wfHost = (await getJson('/workflows'))[0]
|
||||
const iso = (s) => new Date(s * 1000).toISOString()
|
||||
const TRACE = 'fake0trace00000000000000000000'
|
||||
const TRACE_OFF = 'fake0traceoff00000000000000000'
|
||||
const tt = Date.now() / 1000 - 3600
|
||||
const WAIT_S = 300
|
||||
const traceGraph = { nodes: [
|
||||
{ id: 'in', type: 'input', position: { x: 0, y: 0 }, data: { label: '输入', config: {} } },
|
||||
{ id: 'query', type: 'agent', position: { x: 200, y: 0 }, data: { label: '查询订单', config: { tools: ['db_query__shop'] } } },
|
||||
{ id: 'gate', type: 'human', position: { x: 400, y: 0 }, data: { label: '人工审批', config: {} } },
|
||||
{ id: 'sum', type: 'llm', position: { x: 600, y: 0 }, data: { label: '汇总', config: {} } },
|
||||
], edges: [{ source: 'in', target: 'query' }, { source: 'query', target: 'gate' }, { source: 'gate', target: 'sum' }] }
|
||||
const traceEvents = (usageIn = 2000) => [
|
||||
{ seq: 1, type: 'run.started', node_id: null, ts: tt, data: { nodes: 4 } },
|
||||
{ seq: 2, type: 'node.started', node_id: 'in', ts: tt + 0.1, data: { node_type: 'input', label: '输入' } },
|
||||
{ seq: 3, type: 'node.finished', node_id: 'in', ts: tt + 0.2, data: { duration_ms: 100 } },
|
||||
{ seq: 4, type: 'node.started', node_id: 'query', ts: tt + 0.3, data: { node_type: 'agent', label: '查询订单' } },
|
||||
{ seq: 5, type: 'tool.start', node_id: 'query', ts: tt + 0.5,
|
||||
data: { tool: 'db_query__shop', call_id: 'c1', args: { sql: 'select count(*) as n from orders' } } },
|
||||
{ seq: 6, type: 'tool.end', node_id: 'query', ts: tt + 2.5,
|
||||
data: { tool: 'db_query__shop', call_id: 'c1', duration_ms: 2000, preview: '{"columns":["n"],"rows":[[42]]}' } },
|
||||
{ seq: 7, type: 'llm.end', node_id: 'query', ts: tt + 3.0,
|
||||
data: { agent: '查询订单', model: 'demo-model', input_tokens: 1200, output_tokens: 300, cost_usd: 0.004 } },
|
||||
{ seq: 8, type: 'node.finished', node_id: 'query', ts: tt + 3.1, data: { duration_ms: 2800, preview: '共 42 单' } },
|
||||
{ seq: 9, type: 'node.started', node_id: 'gate', ts: tt + 3.2, data: { node_type: 'human', label: '人工审批' } },
|
||||
{ seq: 10, type: 'human.requested', node_id: 'gate', ts: tt + 3.3, data: { mode: 'approve', prompt: '放行吗？' } },
|
||||
{ seq: 11, type: 'run.interrupted', node_id: null, ts: tt + 3.4, data: {} },
|
||||
{ seq: 12, type: 'human.resolved', node_id: 'gate', ts: tt + 3.4 + WAIT_S, data: { response: { approved: true }, actor: '张工' } },
|
||||
{ seq: 13, type: 'run.resumed', node_id: null, ts: tt + 3.4 + WAIT_S, data: { actor: '张工' } },
|
||||
{ seq: 14, type: 'node.finished', node_id: 'gate', ts: tt + 3.5 + WAIT_S, data: { duration_ms: 1 } },
|
||||
{ seq: 15, type: 'node.started', node_id: 'sum', ts: tt + 3.6 + WAIT_S, data: { node_type: 'llm', label: '汇总' } },
|
||||
{ seq: 16, type: 'llm.end', node_id: 'sum', ts: tt + 6.0 + WAIT_S,
|
||||
data: { model: 'demo-model', input_tokens: 800, output_tokens: 500, cost_usd: 0.003 } },
|
||||
{ seq: 17, type: 'node.finished', node_id: 'sum', ts: tt + 6.1 + WAIT_S, data: { duration_ms: 2500 } },
|
||||
{ seq: 18, type: 'run.finished', node_id: null, ts: tt + 6.2 + WAIT_S, data: {
|
||||
output: { answer: '共 42 单' }, usage: { input_tokens: usageIn, output_tokens: 800, cost_usd: 0.007 },
|
||||
timing: { wall_ms: (6.2 + WAIT_S) * 1000, active_ms: 6200, wait_ms: WAIT_S * 1000 } } },
|
||||
]
|
||||
const traceRun = (id, usageIn = 2000) => ({
|
||||
...base, id, workflow_id: wfHost.id, workflow_name: wfHost.name, status: 'succeeded', run_class: 'exploratory',
|
||||
version: null, version_hash: null, manifest_hash: null, manifest_seq: null, error: null, error_node_id: null,
|
||||
input: { q: '上周的订单' }, output: { answer: '共 42 单' },
|
||||
usage: { input_tokens: usageIn, output_tokens: 800, cost_usd: 0.007, duration_ms: 6200,
|
||||
wall_ms: (6.2 + WAIT_S) * 1000, active_ms: 6200, wait_ms: WAIT_S * 1000 },
|
||||
created_at: iso(tt), started_at: iso(tt), finished_at: iso(tt + 6.2 + WAIT_S),
|
||||
})
|
||||
const ART_Q = 'a1'.repeat(32)
|
||||
const ART_QOUT = 'b2'.repeat(32)
|
||||
const ART_SUM = 'c3'.repeat(32)
|
||||
const traceArtifacts = [
|
||||
{ id: ART_Q, kind: 'query_snapshot', node_id: 'query', size: 2048, meta: {}, created_at: iso(tt + 2.5) },
|
||||
{ id: ART_QOUT, kind: 'node_output', node_id: 'query', size: 120, meta: { type: 'agent', attempt: 1 }, created_at: iso(tt + 3.1) },
|
||||
{ id: ART_SUM, kind: 'node_output', node_id: 'sum', size: 300, meta: { type: 'llm', attempt: 1 }, created_at: iso(tt + 6.1 + WAIT_S) },
|
||||
]
|
||||
for (const [id, usageIn] of [[TRACE, 2000], [TRACE_OFF, 5000]]) {
|
||||
fakes.set(`GET /api/runs/${id}`, () => ({ status: 200, json: traceRun(id, usageIn) }))
|
||||
fakes.set(`GET /api/runs/${id}/events`, () => ({ status: 200, json: traceEvents(usageIn) }))
|
||||
fakes.set(`GET /api/runs/${id}/graph`, () => ({ status: 200, json: { graph: traceGraph, workflow_id: wfHost.id, version: null } }))
|
||||
fakes.set(`GET /api/runs/${id}/artifacts`, () => ({ status: 200, json: traceArtifacts }))
|
||||
}
|
||||
fakes.set(`GET /api/artifacts/${ART_Q}`, () => ({ status: 200, json: { id: ART_Q, content: {
|
||||
tool: 'db_query__shop', args: { sql: 'select count(*) as n from orders' }, result: '{"columns":["n"],"rows":[[42]]}',
|
||||
} } }))
|
||||
// 工作流「现在」的图：这些运行之后改过结构——「汇总」删了，换成「通知」；「查询订单」还在。
|
||||
// 画布上还对得准的节点给「在画布中看这一步」，已经没了的不给（给了也对不到任何东西）
|
||||
const wfNow = (await getJson('/workflows')).map((w) => (w.id !== wfHost.id ? w : { ...w, graph: {
|
||||
nodes: [...traceGraph.nodes.filter((n) => n.id !== 'sum'),
|
||||
{ id: 'notify', type: 'output', position: { x: 600, y: 0 }, data: { label: '通知', config: {} } }],
|
||||
edges: [...traceGraph.edges.filter((e) => e.target !== 'sum'), { source: 'gate', target: 'notify' }],
|
||||
} }))
|
||||
fakes.set('GET /api/workflows', () => ({ status: 200, json: wfNow }))
|
||||
|
||||
const viewTab = (k) => page.locator(`[data-run-detail] [role=tab][data-tab=${k}]`)
|
||||
const readout = (k) => page.locator(`[data-trace-readout] [data-readout=${k}]`).innerText().catch(() => '')
|
||||
// 实时、回放、终态写在游标那一格的标签行，数值那一行只有时刻
|
||||
const mode = () => page.locator('[data-trace-readout]').getAttribute('data-mode').catch(() => '')
|
||||
await page.goto(`${WEB}/runs/${TRACE}`, { waitUntil: 'networkidle' })
|
||||
await page.locator('[data-run-detail]').waitFor()
|
||||
check('详情分「时间线 / 航迹 / 工件」三个页签',
|
||||
await viewTab('stream').count() === 1 && await viewTab('trace').count() === 1 && await viewTab('artifacts').count() === 1)
|
||||
check('默认停在时间线', await viewTab('stream').getAttribute('aria-selected').catch(() => null) === 'true')
|
||||
check('工件页签上写着有几件', (await viewTab('artifacts').innerText().catch(() => '')).includes('3'))
|
||||
check('读屏念的是「3 件」，不是「3 条」',
|
||||
await viewTab('artifacts').locator('[aria-label]').getAttribute('aria-label').catch(() => null) === '3 件',
|
||||
await viewTab('artifacts').locator('[aria-label]').getAttribute('aria-label').catch(() => '') ?? '')
|
||||
await viewTab('trace').click().catch(() => {})
|
||||
await page.locator('[data-run-trace] section[aria-label="航迹"]').waitFor({ timeout: 5000 }).catch(() => {})
|
||||
check('点「航迹」地址记下 ?view=trace', new URL(page.url()).searchParams.get('view') === 'trace')
|
||||
const lanes = await page.locator('[data-run-trace] [data-lane]').evaluateAll((els) => els.map((e) => e.getAttribute('data-lane')))
|
||||
check('每个节点一条泳道，按执行顺序排', lanes.join(',') === 'in,query,gate,sum', lanes.join(',') || '没有泳道')
|
||||
check('没拖游标时读的是终态：已完成、2.8k tok',
|
||||
(await readout('phase')).includes('已完成') && (await readout('tokens')).includes('2.8k'),
|
||||
`${await readout('phase')} / ${await readout('tokens')}`)
|
||||
check('没选节点时列出最耗时的节点（慢在哪）',
|
||||
(await page.locator('[data-trace-slowest] [data-slow-node]').first().getAttribute('data-slow-node').catch(() => null)) === 'query')
|
||||
|
||||
// 地址里带着时刻：从工件、分享的链接直接落到那一刻
|
||||
await page.goto(`${WEB}/runs/${TRACE}?view=trace&at=2500`, { waitUntil: 'networkidle' })
|
||||
await page.locator('[data-trace-readout]').waitFor({ timeout: 5000 }).catch(() => {})
|
||||
check('?at=2500 落到 T+00:02.5', (await readout('at')).includes('00:02.5'), await readout('at'))
|
||||
check('游标在时间轴上也在回放', (await page.locator('[data-run-trace] [role=slider]').getAttribute('aria-valuenow').catch(() => null)) === '2500')
|
||||
check('那一刻在跑的是「查询订单」', await page.locator('[data-trace-now] [data-now-node=query]').count() === 1)
|
||||
check('那一刻整次运行还没用 token（模型 3.0 s 才回）', /^0\b/.test((await readout('tokens')).trim()), await readout('tokens'))
|
||||
check('进来之后地址里的时刻摘掉（拖动时不再和它对不上）', !new URL(page.url()).searchParams.has('at'))
|
||||
|
||||
// 慢在哪：那一刻卡在哪个工具上（查询 0.5 s 开始、2.5 s 返回）
|
||||
await page.goto(`${WEB}/runs/${TRACE}?view=trace&at=2000`, { waitUntil: 'networkidle' })
|
||||
await page.locator('[data-trace-readout]').waitFor({ timeout: 5000 }).catch(() => {})
|
||||
await page.locator('[data-run-trace] [data-lane=query] .tl-label').click().catch(() => {})
|
||||
const callText = await page.locator('[data-trace-node=query] [data-node-field=call]').innerText().catch(() => '')
|
||||
check('那一刻在跑的节点，卡片写着正在调哪个工具、调了多久', /db_query__shop.*已 1\.5 s/.test(callText), callText || '没有这一行')
|
||||
check('点泳道名只是选中节点，游标不动', (await readout('at')).includes('00:02.0'), await readout('at'))
|
||||
|
||||
// 拖动：在坞的刻度行上按下、拖过去，读数跟着游标走
|
||||
const track = page.locator('[data-run-trace] .tl-scrub').first()
|
||||
const box = await track.boundingBox().catch(() => null)
|
||||
if (box) {
|
||||
await page.mouse.move(box.x + box.width * 0.05, box.y + box.height / 2)
|
||||
await page.mouse.down()
|
||||
await page.mouse.move(box.x + box.width * 0.15, box.y + box.height / 2, { steps: 4 })
|
||||
await page.waitForTimeout(120)
|
||||
const mid = `${await mode()} ${await readout('at')}`
|
||||
await page.mouse.move(box.x + box.width * 0.3, box.y + box.height / 2, { steps: 4 })
|
||||
await page.mouse.up()
|
||||
await page.waitForTimeout(150)
|
||||
const after = `${await mode()} ${await readout('at')}`
|
||||
check('拖动游标：读数跟着走，并标成回放', mid.includes('回放') && after.includes('回放') && mid !== after, `${mid} → ${after}`)
|
||||
} else {
|
||||
check('拖动游标：读数跟着走，并标成回放', false, '找不到坞的刻度行')
|
||||
}
|
||||
const slider = page.locator('[data-run-trace] [role=slider]')
|
||||
await slider.focus().catch(() => {})
|
||||
await page.keyboard.press('Home')
|
||||
await page.waitForTimeout(100)
|
||||
check('键盘 Home：游标回到开始', (await readout('at')).includes('00:00.0'), await readout('at'))
|
||||
await page.keyboard.press('End')
|
||||
await page.waitForTimeout(100)
|
||||
check('键盘 End：回到终态读数', await mode() === '终态', await mode())
|
||||
|
||||
// 关键时刻：长时间等人的空档被压缩了，拖游标很难正好停在那一下
|
||||
const moments = page.locator('[data-trace-moments] [data-moment]')
|
||||
const momentText = (await moments.allInnerTexts().catch(() => [])).map((t) => t.replace(/\s+/g, ' '))
|
||||
check('列出关键时刻：开始、停下等审批、审批已处理（等了多久）、结局',
|
||||
momentText.length === 4 && momentText[0].includes('开始运行') && momentText[1].includes('「人工审批」停下等审批')
|
||||
&& momentText[2].includes('审批已处理') && momentText[2].includes('5 分 00 秒') && momentText[3].includes('运行完成'),
|
||||
momentText.join(' | '))
|
||||
check('没拖游标时，当前落在结局那一条', await moments.nth(3).getAttribute('aria-current').catch(() => null) === 'step')
|
||||
await moments.nth(2).click().catch(() => {})
|
||||
await page.waitForTimeout(150)
|
||||
check('点「审批已处理」：游标跳到那一刻（T+05:03.4）', await mode() === '回放' && (await readout('at')).includes('05:03.4'),
|
||||
`${await mode()} ${await readout('at')}`)
|
||||
check('跳过去之后它成了当前那一条', await moments.nth(2).getAttribute('aria-current').catch(() => null) === 'step')
|
||||
check('坞上的游标也跟着过去', (await slider.getAttribute('aria-valuenow').catch(() => null)) === '303400',
|
||||
await slider.getAttribute('aria-valuenow').catch(() => '') ?? '')
|
||||
await moments.nth(3).click().catch(() => {})
|
||||
await page.waitForTimeout(150)
|
||||
check('点结局那一条回到终态', await mode() === '终态', await mode())
|
||||
|
||||
await page.goto(`${WEB}/runs/${TRACE}?view=trace&at=30000`, { waitUntil: 'networkidle' })
|
||||
await page.locator('[data-trace-readout]').waitFor({ timeout: 5000 }).catch(() => {})
|
||||
check('拖到等审批那一段：状态是等待审批', (await readout('phase')).includes('等待审批'), await readout('phase'))
|
||||
check('那一刻等审批的是「人工审批」', await page.locator('[data-trace-now] [data-now-node=gate]').count() === 1)
|
||||
check('用量只算到那一刻：1.5k', (await readout('tokens')).includes('1.5k'), await readout('tokens'))
|
||||
await page.locator('[data-run-trace] [data-lane=query]').click().catch(() => {})
|
||||
const nodeCard = page.locator('[data-trace-node=query]')
|
||||
check('点泳道看这个节点在那一刻的样子', await nodeCard.count() === 1)
|
||||
check('节点的用量、工具数取自那一刻',
|
||||
(await nodeCard.locator('[data-node-field=tokens]').innerText().catch(() => '')).includes('1.5k')
|
||||
&& (await nodeCard.locator('[data-node-field=tools]').innerText().catch(() => '')).includes('1'))
|
||||
check('节点卡能回画布看这一步',
|
||||
(await nodeCard.locator('[data-action=node-canvas]').getAttribute('href').catch(() => null)) === `/studio/${wfHost.id}?run=${TRACE}&focus=query`)
|
||||
await page.locator('[data-run-trace] [data-lane=sum] .tl-label').click().catch(() => {})
|
||||
const goneCard = page.locator('[data-trace-node=sum]')
|
||||
check('工作流现在的图里已经没有的节点：节点卡不给落空的「在画布中看这一步」，说明为什么',
|
||||
await goneCard.count() === 1 && await goneCard.locator('[data-action=node-canvas]').count() === 0
|
||||
&& await goneCard.locator('[data-node-gone]').count() === 1)
|
||||
await page.locator('[data-run-trace] [data-lane=query] .tl-label').click().catch(() => {})
|
||||
await nodeCard.locator('[data-action=node-stream]').click().catch(() => {})
|
||||
await page.waitForTimeout(300)
|
||||
check('「在时间线里看」切回时间线', await viewTab('stream').getAttribute('aria-selected').catch(() => null) === 'true')
|
||||
check('并且把那一步描出来', await page.locator('[data-run-detail] [data-node-id="query"][data-flash]').count() >= 1)
|
||||
|
||||
await page.goto(`${WEB}/runs/${TRACE_OFF}?view=trace&at=30000`, { waitUntil: 'networkidle' })
|
||||
await page.locator('[data-trace-readout]').waitFor({ timeout: 5000 }).catch(() => {})
|
||||
check('各次调用加起来对不上后端总数时，那一刻的用量写「—」，不拿终值冒充', (await readout('tokens')).trim().startsWith('—'), await readout('tokens'))
|
||||
|
||||
// 模型把工具调用写成了文字、一次都没真正查（引擎判失败）：结局是失败，而且说得出为什么
|
||||
console.log('\n=== 航迹：模型没有真正调用工具 ===')
|
||||
const LEAK = 'fake0leak000000000000000000000'
|
||||
const LEAK_ERROR = '模型输出了工具调用的原始标记，但没有真正调用工具，这一步一次都没查到数据。常见原因：节点没有绑定工具，或者模型、服务不支持工具调用。到画布里给这个节点绑定要用的工具；绑定了还这样，就换一个支持工具调用的模型'
|
||||
fakes.set(`GET /api/runs/${LEAK}`, () => ({ status: 200, json: {
|
||||
...traceRun(LEAK), status: 'failed', output: null, error: LEAK_ERROR, error_node_id: 'query',
|
||||
usage: { input_tokens: 900, output_tokens: 400, cost_usd: 0.002, duration_ms: 4200, wall_ms: 4200, active_ms: 4200, wait_ms: 0 },
|
||||
finished_at: iso(tt + 4.2),
|
||||
} }))
|
||||
fakes.set(`GET /api/runs/${LEAK}/events`, () => ({ status: 200, json: [
|
||||
...traceEvents().slice(0, 4),
|
||||
{ seq: 5, type: 'llm.end', node_id: 'query', ts: tt + 2.0,
|
||||
data: { agent: '查询订单', model: 'demo-model', input_tokens: 450, output_tokens: 200, cost_usd: 0.001 } },
|
||||
{ seq: 6, type: 'log', node_id: 'query', ts: tt + 2.1, data: {
|
||||
level: 'warn', code: 'tool_markup_leak', message: '模型把工具调用写成了文字（<｜｜DSML｜｜invoke name="db_query__shop"…），没有真正调用工具' } },
|
||||
{ seq: 7, type: 'llm.end', node_id: 'query', ts: tt + 4.0,
|
||||
data: { agent: '查询订单', model: 'demo-model', input_tokens: 450, output_tokens: 200, cost_usd: 0.001 } },
|
||||
{ seq: 8, type: 'node.failed', node_id: 'query', ts: tt + 4.1, data: { error: LEAK_ERROR, duration_ms: 3800 } },
|
||||
{ seq: 9, type: 'run.failed', node_id: null, ts: tt + 4.2, data: {
|
||||
error: LEAK_ERROR, node_id: 'query', timing: { wall_ms: 4200, active_ms: 4200, wait_ms: 0 } } },
|
||||
] }))
|
||||
fakes.set(`GET /api/runs/${LEAK}/graph`, () => ({ status: 200, json: { graph: traceGraph, workflow_id: wfHost.id, version: null } }))
|
||||
fakes.set(`GET /api/runs/${LEAK}/artifacts`, () => ({ status: 200, json: [] }))
|
||||
await page.goto(`${WEB}/runs/${LEAK}?view=trace`, { waitUntil: 'networkidle' })
|
||||
await page.locator('[data-trace-readout]').waitFor({ timeout: 5000 }).catch(() => {})
|
||||
const leakBanner = (await page.locator('[data-run-banner=failed]').innerText().catch(() => '')).replace(/\s+/g, ' ')
|
||||
check('失败横幅说清是模型没有真正调用工具', leakBanner.includes('没有真正调用工具'), leakBanner.slice(0, 60))
|
||||
const leakLocate = page.locator('[data-run-banner=failed] [data-action=locate]')
|
||||
check('失败的节点还在工作流现在的图里：「在画布中定位」对准它；接着跑过不去，它就是主按钮',
|
||||
await leakLocate.getAttribute('href').catch(() => null) === `/studio/${wfHost.id}?run=${LEAK}&focus=query`
|
||||
&& /\bbtn-primary\b/.test(await leakLocate.getAttribute('class').catch(() => '') ?? '')
|
||||
&& await page.locator('[data-run-banner=failed] [data-action=locate-trace]').count() === 0)
|
||||
const leakMoments = (await page.locator('[data-trace-moments] [data-moment]').allInnerTexts().catch(() => [])).map((t) => t.replace(/\s+/g, ' '))
|
||||
check('关键时刻里有那一下警告，结局写失败于哪个节点',
|
||||
leakMoments.some((t) => t.includes('「查询订单」把工具调用写成了文字'))
|
||||
&& (leakMoments.at(-1) ?? '').includes('失败于「查询订单」') && (leakMoments.at(-1) ?? '').includes('没有真正调用工具'),
|
||||
leakMoments.join(' | '))
|
||||
const leakCard = (await page.locator('[data-trace-node=query]').innerText().catch(() => '')).replace(/\s+/g, ' ')
|
||||
check('失败的节点默认选中，卡片上写着警告和原因',
|
||||
leakCard.includes('把工具调用写成了文字') && leakCard.includes('没有真正调用工具') && leakCard.includes('失败'),
|
||||
leakCard.slice(0, 80))
|
||||
await page.locator('[data-trace-moments] [data-moment="query:markup"]').click().catch(() => {})
|
||||
await page.waitForTimeout(150)
|
||||
check('跳到警告那一刻：节点还在跑，卡片上已经有警告',
|
||||
(await readout('at')).includes('00:02.1') && (await page.locator('[data-trace-node=query] [data-node-warn]').count()) === 1,
|
||||
await readout('at'))
|
||||
|
||||
// 失败在「汇总」上，而工作流之后把「汇总」删了：画布上对准它什么也对不到
|
||||
console.log('\n=== 失败的节点在工作流现在的图里已经没了 ===')
|
||||
const GONE = 'fake0gonefail00000000000000000'
|
||||
const GONE_ERROR = '节点「汇总」失败：模型服务出错了（HTTP 500）'
|
||||
fakes.set(`GET /api/runs/${GONE}`, () => ({ status: 200, json: {
|
||||
...traceRun(GONE), status: 'failed', output: null, error: GONE_ERROR, error_node_id: 'sum', finished_at: iso(tt + 6.2 + WAIT_S),
|
||||
} }))
|
||||
fakes.set(`GET /api/runs/${GONE}/events`, () => ({ status: 200, json: [
|
||||
...traceEvents().slice(0, 15),
|
||||
{ seq: 16, type: 'node.failed', node_id: 'sum', ts: tt + 6.0 + WAIT_S, data: { error: GONE_ERROR, duration_ms: 2400 } },
|
||||
{ seq: 17, type: 'run.failed', node_id: null, ts: tt + 6.1 + WAIT_S, data: {
|
||||
error: GONE_ERROR, node_id: 'sum', timing: { wall_ms: (6.1 + WAIT_S) * 1000, active_ms: 6100, wait_ms: WAIT_S * 1000 } } },
|
||||
] }))
|
||||
fakes.set(`GET /api/runs/${GONE}/graph`, () => ({ status: 200, json: { graph: traceGraph, workflow_id: wfHost.id, version: null } }))
|
||||
fakes.set(`GET /api/runs/${GONE}/artifacts`, () => ({ status: 200, json: [] }))
|
||||
await page.goto(`${WEB}/runs/${GONE}`, { waitUntil: 'networkidle' })
|
||||
await page.locator('[data-run-banner=failed]').waitFor({ timeout: 5000 }).catch(() => {})
|
||||
check('横幅不给落空的「在画布中定位」，改给「在航迹中看」',
|
||||
await page.locator('[data-run-banner=failed] [data-action=locate]').count() === 0
|
||||
&& await page.locator('[data-run-banner=failed] [data-action=locate-trace]').count() === 1)
|
||||
const goneHead = page.locator('[data-run-detail] > header [data-action=open-canvas]')
|
||||
check('详情头照样能在画布中回放，只是不带对不上的 focus',
|
||||
await goneHead.getAttribute('href').catch(() => null) === `/studio/${wfHost.id}?run=${GONE}`
|
||||
&& !/对准/.test(await goneHead.getAttribute('title').catch(() => '') ?? ''),
|
||||
await goneHead.getAttribute('href').catch(() => '没有这个按钮') ?? '')
|
||||
await page.locator('[data-run-banner=failed] [data-action=locate-trace]').click().catch(() => {})
|
||||
await page.locator('[data-trace-readout]').waitFor({ timeout: 5000 }).catch(() => {})
|
||||
check('「在航迹中看」：切到航迹，选中失败的那个节点，停在终态',
|
||||
await viewTab('trace').getAttribute('aria-selected').catch(() => null) === 'true'
|
||||
&& await page.locator('[data-trace-node=sum]').count() === 1 && await mode() === '终态')
|
||||
check('那张节点卡也说清画布上已经没有它', await page.locator('[data-trace-node=sum] [data-node-gone]').count() === 1)
|
||||
|
||||
// ------------------------------------------------------------------ 工件
|
||||
|
||||
console.log('\n=== 工件：按节点分组，就地打开 ===')
|
||||
await page.goto(`${WEB}/runs/${TRACE}?view=artifacts`, { waitUntil: 'networkidle' })
|
||||
await page.locator('[data-run-artifacts]').waitFor({ timeout: 5000 }).catch(() => {})
|
||||
const groups = await page.locator('[data-run-artifacts] [data-artifact-group]').evaluateAll((els) => els.map((e) => e.getAttribute('data-artifact-group')))
|
||||
check('按节点分组，按执行顺序', groups.join(',') === 'query,sum', groups.join(',') || '没有分组')
|
||||
const qText = await page.locator(`[data-artifact-id="${ART_Q}"]`).innerText().catch(() => '')
|
||||
check('每件写着类型和大小', qText.includes('查询快照') && qText.includes('2.0 KB'), qText.replace(/\s+/g, ' '))
|
||||
await page.locator(`[data-artifact-id="${ART_Q}"] [data-action=artifact-open]`).click().catch(() => {})
|
||||
const artDialog = page.getByRole('dialog')
|
||||
// 弹窗先出来、内容取回来之后才画表：等到解开的那一块再读
|
||||
await artDialog.locator('[data-artifact]').waitFor({ timeout: 5000 }).catch(() => {})
|
||||
const artText = await artDialog.innerText().catch(() => '')
|
||||
check('用证据查看器打开：带节点名，解开 SQL 和结果表',
|
||||
artText.includes('查询订单') && artText.includes('select count(*)') && artText.includes('42'), artText.replace(/\s+/g, ' ').slice(0, 90))
|
||||
await page.keyboard.press('Escape')
|
||||
await page.locator(`[data-artifact-id="${ART_Q}"] [data-action=artifact-moment]`).click().catch(() => {})
|
||||
await page.waitForTimeout(400)
|
||||
check('「在航迹里看这一刻」跳到航迹、定位到它产出的时刻',
|
||||
await viewTab('trace').getAttribute('aria-selected').catch(() => null) === 'true' && (await readout('at')).includes('00:02.5'),
|
||||
await readout('at'))
|
||||
const realWithArt = []
|
||||
for (const r of allRuns.slice(0, 40)) {
|
||||
const a = await getJson(`/runs/${r.id}/artifacts`)
|
||||
if (a.length) { realWithArt.push([r, a]); break }
|
||||
}
|
||||
if (realWithArt.length) {
|
||||
const [r, a] = realWithArt[0]
|
||||
await page.goto(`${WEB}/runs/${r.id}`, { waitUntil: 'networkidle' })
|
||||
await page.locator('[data-run-detail]').waitFor()
|
||||
await page.waitForTimeout(300)
|
||||
check('真实运行：工件页签的数和接口一致', (await viewTab('artifacts').innerText().catch(() => '')).includes(String(a.length)),
|
||||
`接口 ${a.length} 件`)
|
||||
}
|
||||
|
||||
// ------------------------------------------------------------------ 回画布
|
||||
|
||||
console.log('\n=== 在画布中回放 ===')
|
||||
await page.goto(`${WEB}/runs/${TRACE}`, { waitUntil: 'networkidle' })
|
||||
await page.locator('[data-run-detail]').waitFor()
|
||||
const replayLink = page.locator('[data-run-detail] > header [data-action=open-canvas]')
|
||||
check('跑完的运行，详情头是「在画布中回放」',
|
||||
(await replayLink.innerText().catch(() => '')).includes('在画布中回放')
|
||||
&& await replayLink.getAttribute('href').catch(() => null) === `/studio/${wfHost.id}?run=${TRACE}`,
|
||||
await replayLink.getAttribute('href').catch(() => '') ?? '')
|
||||
check('工作流之后改过结构：链接说明画布上是现在的图',
|
||||
await replayLink.getAttribute('data-graph-drift').catch(() => null) === '1'
|
||||
&& /改过/.test(await replayLink.getAttribute('title').catch(() => '') ?? ''))
|
||||
fakes.delete('GET /api/workflows')
|
||||
|
||||
// ------------------------------------------------------------------ 列表：缩略条、取消原因
|
||||
|
||||
console.log('\n=== 列表：每行一条缩略条 ===')
|
||||
const FAILED_ROW = 'fake0failedrow0000000000000000'
|
||||
const CANCEL_ROW = 'fake0cancelrow0000000000000000'
|
||||
const failedRow = { ...traceRun(FAILED_ROW), status: 'failed', error: '节点「汇总」失败：模型服务出错了（HTTP 500）', error_node_id: 'sum',
|
||||
usage: { wall_ms: 4000, active_ms: 4000, wait_ms: 0, duration_ms: 4000 }, finished_at: iso(tt + 4) }
|
||||
const cancelRow = { ...traceRun(CANCEL_ROW), status: 'cancelled', error: '用户取消：在等审批时放弃了这次运行',
|
||||
usage: { wall_ms: 90000, active_ms: 3000, wait_ms: 87000, duration_ms: 3000 }, finished_at: iso(tt + 90) }
|
||||
fakes.set('GET /api/runs', () => ({ status: 200, json: [traceRun(TRACE), failedRow, cancelRow, ...listTop] }))
|
||||
await page.goto(`${WEB}/runs`, { waitUntil: 'networkidle' })
|
||||
await rowsOf().first().waitFor()
|
||||
const shapeCount = await page.locator('[data-runs-list] [data-run-id] [data-run-shape]').count()
|
||||
check('每一行都有一条缩略条', shapeCount === await rowsOf().count() && shapeCount > 0, `${shapeCount} 条`)
|
||||
const waitFrac = await page.locator(`[data-run-id="${TRACE}"] [data-shape-seg=wait]`).evaluate(
|
||||
(el) => el.getBoundingClientRect().width / el.parentElement.getBoundingClientRect().width).catch(() => 0)
|
||||
check('等人占了九成八的墙钟，缩略条上等人那段也占大半', waitFrac > 0.9, waitFrac.toFixed(2))
|
||||
check('失败的那一行收在失败色上', await page.locator(`[data-run-id="${FAILED_ROW}"] [data-shape-end=failed]`).count() === 1)
|
||||
check('缩略条的 title 写着执行和等人',
|
||||
/执行 .+ · 等人 .+/.test(await page.locator(`[data-run-id="${TRACE}"] [data-run-shape]`).getAttribute('title').catch(() => '') ?? ''))
|
||||
const cancelText = (await page.locator(`[data-run-id="${CANCEL_ROW}"]`).innerText().catch(() => '')).replace(/\s+/g, ' ')
|
||||
check('已取消写原因，不重复「用户取消」', cancelText.includes('在等审批时放弃了这次运行') && !cancelText.includes('用户取消'), cancelText.slice(0, 80))
|
||||
fakes.delete('GET /api/runs')
|
||||
|
||||
console.log('\n=== 工作流筛选：未保存的工作流 ===')
|
||||
await page.goto(`${WEB}/runs`, { waitUntil: 'networkidle' })
|
||||
await rowsOf().first().waitFor()
|
||||
const wfOptions = await page.locator('[data-runs-workflow] option').evaluateAll((els) => els.map((e) => e.value))
|
||||
check('工作流下拉里有「未保存的工作流」', wfOptions.includes('__none__'))
|
||||
if (wfOptions.includes('__none__')) {
|
||||
await listAfter(() => page.locator('[data-runs-workflow]').selectOption('__none__'), (p) => p.get('workflow_id') === '__none__')
|
||||
check('地址记下 ?wf=__none__', new URL(page.url()).searchParams.get('wf') === '__none__')
|
||||
const names = await page.locator('[data-runs-list] [data-run-id] > div:first-of-type').evaluateAll((els) => els.map((e) => e.textContent ?? ''))
|
||||
check('筛出来的都是未保存的工作流', names.length > 0 && names.every((n) => n.includes('未保存的工作流')), `${names.length} 行`)
|
||||
await page.locator('[data-runs-workflow]').selectOption('')
|
||||
await settle()
|
||||
}
|
||||
|
||||
// ------------------------------------------------------------------ 放弃
|
||||
|
||||
if (pending.length) {
|
||||
console.log('\n=== 等审批的运行就地放弃 ===')
|
||||
const target = pending[0]
|
||||
const full = await getJson(`/runs/${target.run_id}`)
|
||||
const evs = await getJson(`/runs/${target.run_id}/events`)
|
||||
await page.goto(`${WEB}/runs/${target.run_id}`, { waitUntil: 'networkidle' })
|
||||
await page.locator('[data-run-detail][data-run-code=waiting]').waitFor({ timeout: 8000 }).catch(() => {})
|
||||
const abandon = page.locator('[data-run-banner=waiting] [data-action=abandon]')
|
||||
check('等审批的横幅上有「放弃这次运行」', await abandon.count() === 1)
|
||||
if (await abandon.count()) {
|
||||
let cancelCalls = 0
|
||||
fakes.set(`POST /api/runs/${target.run_id}/cancel`, () => { cancelCalls++; return { status: 200, json: { ok: true, status: 'cancelled' } } })
|
||||
await abandon.click()
|
||||
const ask = page.getByRole('dialog')
|
||||
await ask.waitFor({ timeout: 5000 }).catch(() => {})
|
||||
check('放弃前要确认，并写清待审批一并关闭', (await ask.innerText().catch(() => '')).includes('待审批'))
|
||||
const tc = Date.now() / 1000
|
||||
fakes.set(`GET /api/runs/${target.run_id}`, () => ({ status: 200, json: {
|
||||
...full, status: 'cancelled', error: '用户取消：在等审批时放弃了这次运行', finished_at: iso(tc),
|
||||
} }))
|
||||
fakes.set(`GET /api/runs/${target.run_id}/events`, () => ({ status: 200, json: [...evs, {
|
||||
seq: (evs.at(-1)?.seq ?? 0) + 1, type: 'run.cancelled', node_id: null, ts: tc,
|
||||
data: { actor: '张工', message: '放弃了这次运行，1 条待审批一并关闭', timing: { wall_ms: 1000, active_ms: 10, wait_ms: 990 } },
|
||||
}] }))
|
||||
fakes.set('GET /api/approvals', (url) => ({
|
||||
status: 200, json: pending.filter((a) => a.id !== target.id && (!url.searchParams.get('run_id') || a.run_id === url.searchParams.get('run_id'))),
|
||||
}))
|
||||
await ask.getByRole('button', { name: '放弃这次运行' }).click({ timeout: 5000 }).catch(() => {})
|
||||
const gone = await page.waitForFunction(
|
||||
() => document.querySelector('[data-run-detail]')?.getAttribute('data-run-code') === 'cancelled',
|
||||
null, { timeout: 6000 },
|
||||
).then(() => true, () => false)
|
||||
check('发的是 POST /cancel', cancelCalls === 1)
|
||||
check('状态变成已取消', gone)
|
||||
const status = (await page.locator('[data-run-telemetry]').innerText().catch(() => '')).replace(/\s+/g, ' ')
|
||||
check('状态格写着谁放弃的', status.includes('张工') && status.includes('放弃'), status.slice(0, 60))
|
||||
fakes.delete(`POST /api/runs/${target.run_id}/cancel`)
|
||||
fakes.delete(`GET /api/runs/${target.run_id}`)
|
||||
fakes.delete(`GET /api/runs/${target.run_id}/events`)
|
||||
fakes.delete('GET /api/approvals')
|
||||
}
|
||||
}
|
||||
|
||||
// ------------------------------------------------------------------ 等了几天、窄屏
|
||||
|
||||
console.log('\n=== 等了几天的审批、1024 宽 ===')
|
||||
// 停在审批上十天多：等人要写「10 天 05 小时」，而不是被截成「245 小时 3…」；游标写「T+10 天 05:30:07」，
|
||||
// 不写「T+245:30:07.3」（同一行的墙钟、等人写的是「10 天」，而且放不下）。等人的时长在涨，游标要走
|
||||
const STALE = 'fake0stalewait0000000000000000'
|
||||
const tw = Date.now() / 1000 - 10 * 86400 - 5.5 * 3600
|
||||
const longApproval = {
|
||||
id: 'fake0longappr0000000000000000', run_id: STALE, node_id: 'gate', mode: 'approve', title: '放行吗？',
|
||||
payload: { kind: 'human_node', node_id: 'gate', mode: 'approve', title: '放行吗？', message: '', schema: {}, context: {} },
|
||||
status: 'pending', response: {}, created_at: iso(tw + 3.3), resolved_by: null, resolved_at: null,
|
||||
workflow_id: wfHost.id, workflow_name: '订单日报', node_label: '人工审批', run_status: 'interrupted', run_class: 'exploratory',
|
||||
}
|
||||
fakes.set(`GET /api/runs/${STALE}`, () => ({ status: 200, json: {
|
||||
...traceRun(STALE), status: 'interrupted', output: null, finished_at: null, created_at: iso(tw), started_at: iso(tw),
|
||||
usage: { input_tokens: 1200, output_tokens: 300, cost_usd: 0.004 },
|
||||
} }))
|
||||
fakes.set(`GET /api/runs/${STALE}/events`, () => ({ status: 200, json: traceEvents().slice(0, 11).map((e) => ({ ...e, ts: e.ts - tt + tw })) }))
|
||||
fakes.set(`GET /api/runs/${STALE}/graph`, () => ({ status: 200, json: { graph: traceGraph, workflow_id: wfHost.id, version: null } }))
|
||||
fakes.set(`GET /api/runs/${STALE}/artifacts`, () => ({ status: 200, json: [] }))
|
||||
fakes.set('GET /api/approvals', (url) => ({
|
||||
status: 200, json: !url.searchParams.get('run_id') || url.searchParams.get('run_id') === STALE ? [longApproval] : [],
|
||||
}))
|
||||
const fits = (sel) => page.locator(sel).evaluateAll((els) => els.length > 0 && els.every((e) => e.scrollWidth <= e.clientWidth + 0.5))
|
||||
const DAY_AT = /^T\+10 天 05:3\d:\d\d$/
|
||||
// 1300 宽：详情区 836px，排成一行的话一格只剩 78px，「10 天 05 小时」放不下，得排两行
|
||||
for (const [w, h] of [[1440, 900], [1300, 860], [1024, 768]]) {
|
||||
await page.setViewportSize({ width: w, height: h })
|
||||
await page.goto(`${WEB}/runs/${STALE}?view=trace`, { waitUntil: 'networkidle' })
|
||||
await page.locator('[data-trace-readout]').waitFor({ timeout: 5000 }).catch(() => {})
|
||||
const waitRead = await readout('wait')
|
||||
check(`${w} 宽：等人写成跨天的跨度（10 天 05 小时），不被截断`,
|
||||
waitRead.startsWith('10 天 05 小时') && await fits('[data-trace-readout] [data-readout=wait]'), waitRead)
|
||||
const atRead = (await readout('at')).trim()
|
||||
check(`${w} 宽：游标写成跨天的时刻（T+10 天 05:30:xx），不被截断`,
|
||||
DAY_AT.test(atRead) && await fits('[data-trace-readout] [data-readout=at]'), atRead)
|
||||
}
|
||||
await page.setViewportSize({ width: 1440, height: 900 })
|
||||
await page.goto(`${WEB}/runs/${STALE}?view=trace`, { waitUntil: 'networkidle' })
|
||||
await page.locator('[data-trace-readout]').waitFor({ timeout: 5000 }).catch(() => {})
|
||||
check('「此刻」的等审批也写跨天', (await page.locator('[data-trace-now] [data-now-node=gate]').innerText().catch(() => '')).includes('10 天 05 小时'))
|
||||
const waitAtA = await readout('at')
|
||||
// 跨天的时刻只写到秒：隔一秒多再读
|
||||
await page.waitForTimeout(1100)
|
||||
const waitAtB = await readout('at')
|
||||
check('停在审批上时游标照样跟着现在走（和坞里同一拍）', await mode() === '实时' && waitAtA !== waitAtB, `${waitAtA} → ${waitAtB}`.replace(/\s+/g, ' '))
|
||||
|
||||
// 去审批卡：审批卡在时间线里。从航迹、工件、原始事件点过来，要切回时间线、描一下卡片，
|
||||
// 并且焦点进到卡片里——键盘和读屏用户才接得着。视图经地址切，提交比点击晚好几帧
|
||||
console.log('\n=== 去审批卡：从哪个视图点都能把焦点送进卡片 ===')
|
||||
for (const [from, label] of [['trace', '航迹'], ['artifacts', '工件'], ['raw', '原始事件'], ['stream', '时间线']]) {
|
||||
await page.goto(`${WEB}/runs/${STALE}${from === 'trace' || from === 'artifacts' ? `?view=${from}` : ''}`, { waitUntil: 'networkidle' })
|
||||
await page.locator('[data-run-banner=waiting] [data-action=jump-approval]').waitFor({ timeout: 5000 }).catch(() => {})
|
||||
if (from === 'raw') {
|
||||
await page.locator('[data-action=raw]').click().catch(() => {})
|
||||
await page.locator('[data-raw-events]').waitFor({ timeout: 3000 }).catch(() => {})
|
||||
}
|
||||
await page.locator('[data-run-banner=waiting] [data-action=jump-approval]').click().catch(() => {})
|
||||
const into = await page.waitForFunction(() => !!document.activeElement?.closest('[data-approval-card]'), null, { timeout: 3000 })
|
||||
.then(() => true, () => false)
|
||||
const ringed = await page.locator(`[data-approval-card="${longApproval.id}"].runs-focus-ring`).count()
|
||||
const back = await viewTab('stream').getAttribute('aria-selected').catch(() => null) === 'true'
|
||||
check(`从「${label}」点「去审批卡」：回到时间线，焦点进了审批卡，卡片描了一下边`, into && ringed === 1 && back,
|
||||
`时间线 ${back ? '是' : '否'} · 描边 ${ringed} · 焦点在 ${await page.evaluate(() => {
|
||||
const a = document.activeElement
|
||||
return `${a?.tagName}${a?.getAttribute('data-action') ? `[${a.getAttribute('data-action')}]` : ''}`
|
||||
})}`)
|
||||
}
|
||||
|
||||
// 批过的十天等待：「审批已处理」「运行完成」都在 T+10 天之后。关键时刻里时刻那一栏按最宽的
|
||||
// 一条排齐，不压到后面的徽标上；终态的游标放得下
|
||||
console.log('\n=== 跨天的关键时刻 ===')
|
||||
const LONGDONE = 'fake0longdone00000000000000000'
|
||||
const td = Date.now() / 1000 - 11 * 86400
|
||||
const LONG_S = 10 * 86400
|
||||
const longTiming = { wall_ms: (6.2 + LONG_S) * 1000, active_ms: 6200, wait_ms: LONG_S * 1000 }
|
||||
fakes.set(`GET /api/runs/${LONGDONE}`, () => ({ status: 200, json: {
|
||||
...traceRun(LONGDONE), created_at: iso(td), started_at: iso(td), finished_at: iso(td + 6.2 + LONG_S),
|
||||
usage: { input_tokens: 2000, output_tokens: 800, cost_usd: 0.007, duration_ms: 6200, ...longTiming },
|
||||
} }))
|
||||
fakes.set(`GET /api/runs/${LONGDONE}/events`, () => ({ status: 200, json: traceEvents().map((e) => ({
|
||||
...e,
|
||||
ts: e.ts - tt + td + (e.seq >= 12 ? LONG_S - WAIT_S : 0),
|
||||
...(e.type === 'run.finished' ? { data: { ...e.data, timing: longTiming } } : {}),
|
||||
})) }))
|
||||
fakes.set(`GET /api/runs/${LONGDONE}/graph`, () => ({ status: 200, json: { graph: traceGraph, workflow_id: wfHost.id, version: null } }))
|
||||
fakes.set(`GET /api/runs/${LONGDONE}/artifacts`, () => ({ status: 200, json: [] }))
|
||||
const momentsAligned = () => page.locator('[data-trace-moments] [data-moment]').evaluateAll((els) => {
|
||||
const cols = els.map((b) => {
|
||||
const at = b.querySelector('[data-moment-at]')
|
||||
return { at: at.getBoundingClientRect(), fit: at.scrollWidth <= at.clientWidth + 0.5, mark: at.nextElementSibling.getBoundingClientRect().left }
|
||||
})
|
||||
return cols.length > 0 && cols.every((c) => c.fit && c.at.right <= c.mark + 0.5 && Math.abs(c.mark - cols[0].mark) < 0.5)
|
||||
})
|
||||
for (const [w, h] of [[1440, 900], [1024, 768]]) {
|
||||
await page.setViewportSize({ width: w, height: h })
|
||||
await page.goto(`${WEB}/runs/${LONGDONE}?view=trace`, { waitUntil: 'networkidle' })
|
||||
await page.locator('[data-trace-moments]').waitFor({ timeout: 5000 }).catch(() => {})
|
||||
const texts = (await page.locator('[data-trace-moments] [data-moment]').allInnerTexts().catch(() => [])).map((t) => t.replace(/\s+/g, ' '))
|
||||
check(`${w} 宽：跨天的关键时刻写「T+10 天 00:00:03」，时刻一栏排齐、不压到徽标上`,
|
||||
texts.some((t) => t.includes('T+10 天 00:00:03') && t.includes('审批已处理') && t.includes('等了 10 天')) && await momentsAligned(),
|
||||
texts.join(' | '))
|
||||
check(`${w} 宽：终态的游标写「T+10 天 00:00:06」，放得下`,
|
||||
(await readout('at')).trim() === 'T+10 天 00:00:06' && await fits('[data-trace-readout] [data-readout=at]'), await readout('at'))
|
||||
}
|
||||
|
||||
// 屏幕矮、节点多：九个节点的失败运行，横幅占去一截。读数区先拿够它要的，默认选中的失败
|
||||
// 节点卡、关键时刻都得在首屏；坞拿剩下的，不矮过能拖到的最矮，泳道在坞里滚
|
||||
console.log('\n=== 屏幕矮、节点多：读数区不被航迹坞压住 ===')
|
||||
const FAN = 'fake0fanfail000000000000000000'
|
||||
const FAN_ERROR = '节点「查询订单」失败：模型服务出错了（HTTP 500）'
|
||||
const FAN_STEPS = ['check', 'score', 'route', 'publish', 'archive']
|
||||
const fanGraph = {
|
||||
nodes: [...traceGraph.nodes, ...FAN_STEPS.map((id, i) => ({
|
||||
id, type: 'llm', position: { x: 800 + i * 200, y: 0 }, data: { label: `步骤 ${i + 1}`, config: {} },
|
||||
}))],
|
||||
edges: [...traceGraph.edges, ...FAN_STEPS.map((id, i) => ({ source: i ? FAN_STEPS[i - 1] : 'sum', target: id }))],
|
||||
}
|
||||
fakes.set(`GET /api/runs/${FAN}`, () => ({ status: 200, json: {
|
||||
...traceRun(FAN), status: 'failed', output: null, error: FAN_ERROR, error_node_id: 'query',
|
||||
usage: { input_tokens: 0, output_tokens: 0, cost_usd: 0, duration_ms: 1200, wall_ms: 1200, active_ms: 1200, wait_ms: 0 },
|
||||
finished_at: iso(tt + 1.2),
|
||||
} }))
|
||||
fakes.set(`GET /api/runs/${FAN}/events`, () => ({ status: 200, json: [
|
||||
...traceEvents().slice(0, 4),
|
||||
{ seq: 5, type: 'node.failed', node_id: 'query', ts: tt + 1.1, data: { error: FAN_ERROR, duration_ms: 800 } },
|
||||
{ seq: 6, type: 'run.failed', node_id: null, ts: tt + 1.2, data: {
|
||||
error: FAN_ERROR, node_id: 'query', timing: { wall_ms: 1200, active_ms: 1200, wait_ms: 0 } } },
|
||||
] }))
|
||||
fakes.set(`GET /api/runs/${FAN}/graph`, () => ({ status: 200, json: { graph: fanGraph, workflow_id: wfHost.id, version: null } }))
|
||||
fakes.set(`GET /api/runs/${FAN}/artifacts`, () => ({ status: 200, json: [] }))
|
||||
const firstScreen = () => page.evaluate(() => {
|
||||
const read = document.querySelector('[data-trace-readout]')?.getBoundingClientRect()
|
||||
const whole = (sel) => {
|
||||
const b = document.querySelector(sel)?.getBoundingClientRect()
|
||||
return !!b && !!read && b.top >= read.top - 0.5 && b.bottom <= read.bottom + 1
|
||||
}
|
||||
return {
|
||||
card: whole('[data-trace-node=query]'), moments: whole('[data-trace-moments]'),
|
||||
dock: Math.round(document.querySelector('[data-run-trace] section[aria-label="航迹"]')?.getBoundingClientRect().height ?? 0),
|
||||
}
|
||||
})
|
||||
for (const [w, h] of [[1440, 900], [1180, 800], [1024, 768]]) {
|
||||
await page.setViewportSize({ width: w, height: h })
|
||||
await page.goto(`${WEB}/runs/${FAN}?view=trace`, { waitUntil: 'networkidle' })
|
||||
await page.locator('[data-trace-node=query]').waitFor({ timeout: 5000 }).catch(() => {})
|
||||
await page.waitForTimeout(150)
|
||||
const seen = await firstScreen()
|
||||
check(`${w}×${h}：失败节点卡、关键时刻整块都在首屏，坞不矮过 132px`, seen.card && seen.moments && seen.dock >= 132, JSON.stringify(seen))
|
||||
}
|
||||
// 再矮（或者横幅更高）就放不下了：坞不再让，读数区底边淡出，看得出下面还有
|
||||
await page.setViewportSize({ width: 1024, height: 640 })
|
||||
await page.waitForTimeout(300)
|
||||
check('1024×640：读数区放不下时底边淡出，提示能往下滚', await page.locator('[data-trace-readout][data-more]').count() === 1)
|
||||
|
||||
// 1024 宽：详情区只有 600px，读数一格不到 70px 的字宽
|
||||
await page.setViewportSize({ width: 1024, height: 768 })
|
||||
fakes.set('GET /api/runs', () => ({ status: 200, json: [traceRun(TRACE), ...listTop] }))
|
||||
await page.goto(`${WEB}/runs/${TRACE}`, { waitUntil: 'networkidle' })
|
||||
await page.locator('[data-run-telemetry]').waitFor({ timeout: 5000 }).catch(() => {})
|
||||
const cut = await page.locator('[data-run-telemetry] [data-telemetry]').evaluateAll((els) =>
|
||||
els.filter((e) => e.scrollWidth > e.clientWidth + 0.5).map((e) => `${e.getAttribute('data-telemetry')}「${e.textContent}」`))
|
||||
check('1024 宽：详情头的读数没有一格被截成省略号', cut.length === 0, cut.join('、') || '全部放得下')
|
||||
const footH = await page.locator(`[data-runs-list] [data-run-id="${TRACE}"] [data-run-duration]`).evaluate(
|
||||
(el) => el.parentElement.getBoundingClientRect().height).catch(() => 99)
|
||||
check('1024 宽：列表行底下那一行不折行（时长太长就截断，用量不被拆开）', footH <= 18, `${footH}px`)
|
||||
await page.setViewportSize({ width: 1440, height: 900 })
|
||||
fakes.delete('GET /api/runs')
|
||||
fakes.delete('GET /api/approvals')
|
||||
|
||||
console.log('\n=== 收尾 ===')
|
||||
check('只有伪造过的写请求', writes.every((w) => /\/continue$|\/cancel$|\/copilot\/from-run$|DELETE \/api\/runs\/fake0|^POST \/api\/runs$/.test(w)),
|
||||
// 提取模板之后会跳进画布：画布一打开就拿图去做校验、变量分析，这两个是带图的只读 POST
|
||||
// （同样被探针拦下，不落库）。赶上它们发出来之前就离开了画布的话就没有
|
||||
check('只有伪造过的写请求', writes.every((w) => /\/continue$|\/cancel$|\/copilot\/from-run$|DELETE \/api\/runs\/fake0|^POST \/api\/runs$|^POST \/api\/workflows\/(validate|variables)$/.test(w)),
|
||||
writes.join(', '))
|
||||
check('没有未捕获的运行时错误', errors.length === 0, errors[0] ?? '')
|
||||
|
||||
|
||||
+219
-21
@@ -82,6 +82,8 @@ async function open({ theme = 'dark', state }) {
|
||||
if (state.offline) return route.abort('connectionrefused')
|
||||
// 后端连得上但不回：/health 照常，其余请求一直挂着
|
||||
if (state.hang && url.pathname !== '/api/health') return
|
||||
// 只卡一张表：别的照常回
|
||||
if (state.hangPath === url.pathname) return
|
||||
if (r.method() !== 'GET') {
|
||||
state.writes.push({ method: r.method(), path: url.pathname, body: r.postData() })
|
||||
return route.fulfill({ json: {} })
|
||||
@@ -98,11 +100,19 @@ async function open({ theme = 'dark', state }) {
|
||||
const nav = (page) => page.locator('nav[aria-label="主导航"]')
|
||||
const waitNav = (page) => nav(page).waitFor({ timeout: 15000 })
|
||||
const isMac = process.platform === 'darwin'
|
||||
// 只跑其中几段：CHECK_ONLY=离开前确认,启动页 node scripts/check-shell.mjs（按段名包含匹配）
|
||||
const ONLY = process.env.CHECK_ONLY?.split(',').map((s) => s.trim()).filter(Boolean)
|
||||
const want = (name) => !ONLY?.length || ONLY.some((k) => name.includes(k))
|
||||
/** 这一段要跑就打出段标题、返回 true */
|
||||
const section = (name, title) => {
|
||||
if (!want(name)) return false
|
||||
console.log(title)
|
||||
return true
|
||||
}
|
||||
const MOD = isMac ? 'Meta' : 'Control'
|
||||
|
||||
// ---------------------------------------------------------------------------
|
||||
console.log('=== 导航 ===')
|
||||
{
|
||||
if (section('导航', '=== 导航 ===')) {
|
||||
const state = { approvals: PENDING, offline: false, writes: [] }
|
||||
const { ctx, page } = await open({ state })
|
||||
await page.goto(`${WEB}/chat`)
|
||||
@@ -187,8 +197,7 @@ console.log('=== 导航 ===')
|
||||
}
|
||||
|
||||
// ---------------------------------------------------------------------------
|
||||
console.log('\n=== 待审批徽标 ===')
|
||||
for (const theme of ['dark', 'light']) {
|
||||
if (section('待审批徽标', '\n=== 待审批徽标 ===')) for (const theme of ['dark', 'light']) {
|
||||
const state = { approvals: PENDING, offline: false, writes: [] }
|
||||
const { ctx, page } = await open({ theme, state })
|
||||
await page.goto(`${WEB}/chat`)
|
||||
@@ -225,8 +234,7 @@ for (const theme of ['dark', 'light']) {
|
||||
}
|
||||
|
||||
// ---------------------------------------------------------------------------
|
||||
console.log('\n=== 404 ===')
|
||||
{
|
||||
if (section('404', '\n=== 404 ===')) {
|
||||
const state = { approvals: [], offline: false, writes: [] }
|
||||
const { ctx, page } = await open({ state })
|
||||
await page.goto(`${WEB}/does-not-exist/at-all`)
|
||||
@@ -244,8 +252,7 @@ console.log('\n=== 404 ===')
|
||||
}
|
||||
|
||||
// ---------------------------------------------------------------------------
|
||||
console.log('\n=== 标签页标题、favicon、后台提醒 ===')
|
||||
{
|
||||
if (section('标签页标题', '\n=== 标签页标题、favicon、后台提醒 ===')) {
|
||||
const state = { approvals: [], offline: false, writes: [] }
|
||||
const { ctx, page } = await open({ state })
|
||||
await page.goto(`${WEB}/runs`)
|
||||
@@ -331,8 +338,7 @@ console.log('\n=== 标签页标题、favicon、后台提醒 ===')
|
||||
}
|
||||
|
||||
// ---------------------------------------------------------------------------
|
||||
console.log('\n=== ⌘K 命令面板 ===')
|
||||
{
|
||||
if (section('⌘K 命令面板', '\n=== ⌘K 命令面板 ===')) {
|
||||
const state = { approvals: PENDING, offline: false, writes: [] }
|
||||
const { ctx, page } = await open({ state })
|
||||
await page.goto(`${WEB}/runs`)
|
||||
@@ -374,6 +380,22 @@ console.log('\n=== ⌘K 命令面板 ===')
|
||||
check('点工作流进它的画布', await page.waitForURL(new RegExp(`/studio/${wf.id}`), { timeout: 5000 }).then(() => true, () => false), page.url())
|
||||
}
|
||||
|
||||
// 页面刚挂载时的自动聚焦（比如画布助手的输入框）会把焦点从开着的面板里抢走，
|
||||
// 之后打的字、按的 Esc 就都落到了面板底下
|
||||
await page.keyboard.press(`${MOD}+KeyK`)
|
||||
await dialog.waitFor({ timeout: 3000 })
|
||||
const stolen = await page.evaluate(() => {
|
||||
const el = document.querySelector('main textarea') ?? document.querySelector('main input, main button, nav a')
|
||||
el?.focus()
|
||||
return el?.tagName ?? null
|
||||
})
|
||||
check('面板开着时焦点被页面抢走，会回到搜索框', await dialog.getByRole('combobox').evaluate((el) => el === document.activeElement),
|
||||
`${stolen} → ${await page.evaluate(() => document.activeElement?.tagName)}`)
|
||||
await page.keyboard.press('Escape')
|
||||
check('……Esc 照样关掉面板', await dialog.waitFor({ state: 'detached', timeout: 3000 }).then(() => true, () => false))
|
||||
// 上面没过时面板还开着：⌘K 再按一下把它收掉，后面几项照常跑
|
||||
if (await dialog.count()) await page.keyboard.press(`${MOD}+KeyK`)
|
||||
|
||||
await page.keyboard.press(`${MOD}+KeyK`)
|
||||
await dialog.getByRole('combobox').fill('zzqq没有这个东西')
|
||||
check('没结果时说清楚', (await dialog.innerText()).includes('没有匹配'))
|
||||
@@ -403,8 +425,7 @@ console.log('\n=== ⌘K 命令面板 ===')
|
||||
}
|
||||
|
||||
// ---------------------------------------------------------------------------
|
||||
console.log('\n=== ? 快捷键说明 ===')
|
||||
{
|
||||
if (section('快捷键说明', '\n=== ? 快捷键说明 ===')) {
|
||||
const state = { approvals: [], offline: false, writes: [] }
|
||||
const { ctx, page } = await open({ state })
|
||||
await page.goto(`${WEB}/studio`)
|
||||
@@ -428,8 +449,117 @@ console.log('\n=== ? 快捷键说明 ===')
|
||||
}
|
||||
|
||||
// ---------------------------------------------------------------------------
|
||||
console.log('\n=== 离线横幅 ===')
|
||||
{
|
||||
if (section('离开前确认', '\n=== 离开前确认 ===')) {
|
||||
// 页面有没保存的改动时登记一道守卫（lib/leave）。以前设置页只在捕获阶段拦 <a> 的
|
||||
// 点击：⌘K、⌥ 数字直接调 navigate()，浏览器后退也不走 <a>，改到一半的表单就这么丢了。
|
||||
// 这里用 dev 构建挂出的 window.__leave 登记一道测试守卫：每问一次记一笔，答案由脚本给
|
||||
const state = { approvals: PENDING, offline: false, writes: [] }
|
||||
const { ctx, page } = await open({ state })
|
||||
await page.goto(`${WEB}/runs`)
|
||||
await waitNav(page)
|
||||
const hooked = await page.waitForFunction(() => !!window.__leave, null, { timeout: 5000 }).then(() => true, () => false)
|
||||
check('dev 构建挂出了 window.__leave', hooked)
|
||||
if (hooked) {
|
||||
await page.evaluate(() => {
|
||||
window.__asks = []
|
||||
window.__answer = null
|
||||
window.__unguard = window.__leave.register({
|
||||
confirm: (next) => new Promise((resolve) => {
|
||||
window.__asks.push(next ? next.pathname + next.search : null)
|
||||
window.__answer = resolve
|
||||
}),
|
||||
})
|
||||
})
|
||||
const asks = () => page.evaluate(() => window.__asks.length)
|
||||
const asked = (n) => page.waitForFunction((n) => window.__asks.length >= n, n, { timeout: 3000 }).then(() => true, () => false)
|
||||
const answer = (ok) => page.evaluate((ok) => { const r = window.__answer; window.__answer = null; r?.(ok) }, ok)
|
||||
const where = () => { const u = new URL(page.url()); return u.pathname + u.search }
|
||||
const settle = () => page.waitForTimeout(250)
|
||||
const dialog = page.getByRole('dialog', { name: '命令面板' })
|
||||
|
||||
await page.keyboard.press(`${MOD}+KeyK`)
|
||||
await dialog.getByRole('combobox').fill('设置')
|
||||
await page.keyboard.press('Enter')
|
||||
check('⌘K 跳页先问守卫，问的时候地址不动', await asked(1) && where().startsWith('/runs'), `${await asks()} 次 · ${where()}`)
|
||||
check('守卫拿得到要去哪', (await page.evaluate(() => window.__asks[0])) === '/settings', await page.evaluate(() => window.__asks[0]))
|
||||
await answer(false)
|
||||
await settle()
|
||||
check('选「留下」：还在原地，面板也关了', where().startsWith('/runs') && (await dialog.count()) === 0, where())
|
||||
|
||||
await page.evaluate(() => document.activeElement?.blur())
|
||||
await page.keyboard.press('Alt+Digit1')
|
||||
check('⌥1 切页先问守卫', await asked(2) && where().startsWith('/runs'), where())
|
||||
await answer(true)
|
||||
check('选「离开」：去到问数据', await page.waitForURL(/\/chat/, { timeout: 5000 }).then(() => true, () => false), where())
|
||||
|
||||
await nav(page).locator('a[href="/data"]').click()
|
||||
check('点左侧导航先问守卫', await asked(3))
|
||||
await answer(false)
|
||||
await settle()
|
||||
check('选「留下」：地址不变', where().startsWith('/chat'), where())
|
||||
|
||||
await page.evaluate(() => history.back())
|
||||
check('浏览器后退也先问守卫', await asked(4))
|
||||
await answer(false)
|
||||
await page.waitForTimeout(400)
|
||||
check('选「留下」：地址退回原处', where().startsWith('/chat'), where())
|
||||
await page.evaluate(() => history.back())
|
||||
await asked(5)
|
||||
await answer(true)
|
||||
check('后退时选「离开」：真的退回去了', await page.waitForURL(/\/runs/, { timeout: 5000 }).then(() => true, () => false), where())
|
||||
|
||||
const before = await asks()
|
||||
await page.locator('[data-approval-badge]').click()
|
||||
await page.waitForURL(/tab=approvals/, { timeout: 5000 }).catch(() => {})
|
||||
check('同一页里只换 ?tab 不问', (await asks()) === before && /\/runs\?tab=approvals/.test(where()), `${await asks() - before} 次 · ${where()}`)
|
||||
|
||||
// 先建东西再跳的命令：得先问，不然人点了「留下」，会话已经建出来了
|
||||
state.writes.length = 0
|
||||
await page.keyboard.press(`${MOD}+KeyK`)
|
||||
await dialog.getByRole('combobox').fill('新对话')
|
||||
await page.keyboard.press('Enter')
|
||||
check('⌘K「新对话」先问守卫', await asked(before + 1))
|
||||
check('……问的时候还没建会话', !state.writes.some((w) => w.method === 'POST' && w.path === '/api/conversations'),
|
||||
state.writes.map((w) => `${w.method} ${w.path}`).join(', '))
|
||||
await answer(false)
|
||||
await settle()
|
||||
check('选「留下」：不建会话、不跳', !state.writes.some((w) => w.method === 'POST' && w.path === '/api/conversations')
|
||||
&& where().startsWith('/runs'), where())
|
||||
|
||||
await page.keyboard.press(`${MOD}+KeyK`)
|
||||
await dialog.getByRole('combobox').fill('新建工作流')
|
||||
await page.keyboard.press('Enter')
|
||||
check('⌘K「新建工作流」先问守卫，再问名字', await asked(before + 2)
|
||||
&& (await page.getByRole('dialog', { name: '新建工作流' }).count()) === 0)
|
||||
await answer(false)
|
||||
await settle()
|
||||
check('选「留下」：不弹起名框', (await page.getByRole('dialog', { name: '新建工作流' }).count()) === 0 && where().startsWith('/runs'))
|
||||
|
||||
// 关页、刷新：守卫在时浏览器会问（它只给通用文案），撤了就不问
|
||||
const unloadAsks = []
|
||||
page.on('dialog', (d) => { unloadAsks.push(d.type()); void d.dismiss() })
|
||||
await page.close({ runBeforeUnload: true })
|
||||
await page.waitForTimeout(500)
|
||||
check('有守卫时关页，浏览器先问', unloadAsks.includes('beforeunload') && !page.isClosed(), unloadAsks.join(',') || '没问')
|
||||
|
||||
await page.evaluate(() => window.__unguard())
|
||||
const n = await asks()
|
||||
await page.evaluate(() => document.activeElement?.blur())
|
||||
await page.keyboard.press('Alt+Digit1')
|
||||
check('撤掉守卫后照常切页，不再问', await page.waitForURL(/\/chat/, { timeout: 5000 }).then(() => true, () => false)
|
||||
&& (await asks()) === n, where())
|
||||
unloadAsks.length = 0
|
||||
await page.close({ runBeforeUnload: true })
|
||||
// 页面这回真关了，只能在外面等
|
||||
await new Promise((r) => setTimeout(r, 500))
|
||||
check('撤掉守卫后关页不再问', unloadAsks.length === 0 && page.isClosed(), unloadAsks.join(','))
|
||||
}
|
||||
check('没有未捕获的运行时错误', page.errors.length === 0, page.errors[0] ?? '')
|
||||
await ctx.close()
|
||||
}
|
||||
|
||||
// ---------------------------------------------------------------------------
|
||||
if (section('离线横幅', '\n=== 离线横幅 ===')) {
|
||||
const state = { approvals: [], offline: false, writes: [] }
|
||||
const { ctx, page } = await open({ state })
|
||||
await page.goto(`${WEB}/chat`)
|
||||
@@ -440,7 +570,11 @@ console.log('\n=== 离线横幅 ===')
|
||||
const banner = page.locator('main').getByText('后端未连接')
|
||||
check('断网后出横幅', await banner.waitFor({ timeout: 15000 }).then(() => true, () => false))
|
||||
check('横幅是主区第一个孩子（toast 才会跟着下移）',
|
||||
await page.evaluate(() => document.querySelector('main')?.firstElementChild?.getAttribute('role') === 'alert'))
|
||||
// role=alert 只包不变的那几句（整块重念会把倒计时也念一遍），外层认 data-offline-banner
|
||||
await page.evaluate(() => document.querySelector('main')?.firstElementChild?.hasAttribute('data-offline-banner') === true))
|
||||
const alertText = await page.evaluate(() => document.querySelector('[data-offline-banner] [role="alert"]')?.textContent ?? null)
|
||||
check('横幅的播报区只有不变的那几句：倒计时不在里面', alertText != null && alertText.includes('后端未连接') && !/秒后/.test(alertText)
|
||||
&& /秒后自动重试/.test(await page.locator('[data-offline-banner]').innerText()), alertText ?? '没有 role=alert')
|
||||
check('遥测点变成离线', /离线/.test(await nav(page).locator('[data-telemetry]').innerText()))
|
||||
state.offline = false
|
||||
await page.getByRole('button', { name: '立即重试' }).first().click()
|
||||
@@ -463,8 +597,7 @@ const bootLiveRegions = (page) => page.evaluate(() => {
|
||||
})
|
||||
|
||||
// ---------------------------------------------------------------------------
|
||||
console.log('\n=== 启动页 ===')
|
||||
for (const theme of ['dark', 'light']) {
|
||||
if (section('启动页', '\n=== 启动页 ===')) for (const theme of ['dark', 'light']) {
|
||||
const state = { approvals: [], offline: true, writes: [] }
|
||||
const { ctx, page } = await open({ theme, state })
|
||||
await page.goto(`${WEB}/chat`)
|
||||
@@ -488,20 +621,27 @@ for (const theme of ['dark', 'light']) {
|
||||
await ctx.close()
|
||||
}
|
||||
|
||||
{
|
||||
if (want('启动页')) {
|
||||
// 后端连着、只有一张表报错：照常进，但要说出来，不然那张表对应的下拉就是空的，看着像"没有"
|
||||
const state = { approvals: [], offline: false, writes: [], broken: '/api/skills' }
|
||||
const { ctx, page } = await open({ state })
|
||||
await page.goto(`${WEB}/chat`)
|
||||
await waitNav(page)
|
||||
const warned = await page.getByText(/有 1 项没加载成功：Skill（500）/).waitFor({ timeout: 5000 }).then(() => true, () => false)
|
||||
const warn = page.getByText(/有 1 项没加载成功（Skill 500）/)
|
||||
const warned = await warn.waitFor({ timeout: 5000 }).then(() => true, () => false)
|
||||
check('部分请求报错：照常进页面，并说出哪一项没加载', warned)
|
||||
const tone = await nav(page).locator('[data-telemetry]').getAttribute('data-tone')
|
||||
check('遥测点记为「降级」（琥珀），不是一片绿', tone === '降级', tone)
|
||||
// 那张表一直是空的，提示就一直在：4 秒一闪就没，人还是会对着空列表去新建
|
||||
await page.waitForTimeout(4500)
|
||||
check('提示常驻，直到那张表取回来', await warn.isVisible())
|
||||
state.broken = null
|
||||
await page.locator('[role="status"] > div', { hasText: '没加载成功' }).getByRole('button', { name: '重试' }).click({ timeout: 5000 }).catch(() => {})
|
||||
check('重试取回来：提示撤掉', await warn.waitFor({ state: 'detached', timeout: 5000 }).then(() => true, () => false))
|
||||
await ctx.close()
|
||||
}
|
||||
|
||||
{
|
||||
if (want('启动页')) {
|
||||
// 后端连得上、请求却卡着不回：「先进去看看」之后不能是一片绿加一屏「还没有工作流」
|
||||
const state = { approvals: [], offline: false, writes: [], hang: true }
|
||||
const { ctx, page } = await open({ state })
|
||||
@@ -515,6 +655,9 @@ for (const theme of ['dark', 'light']) {
|
||||
await enter.click()
|
||||
await waitNav(page)
|
||||
const telemetry = nav(page).locator('[data-telemetry]')
|
||||
// 这一路唯一的 /health 探活和「先进去看看」同时出现，按钮点得快时它还在路上
|
||||
await page.waitForFunction(() => document.querySelector('[data-telemetry]')?.getAttribute('data-telemetry') !== 'checking', null, { timeout: 5000 })
|
||||
.catch(() => {})
|
||||
const tone = await telemetry.getAttribute('data-tone')
|
||||
check('进去以后遥测点不是「在线」，而是「加载中」', tone === '加载中', `${tone} / ${await telemetry.innerText()}`)
|
||||
const warn = page.getByText(/没取回来.*先别新建/)
|
||||
@@ -537,7 +680,62 @@ for (const theme of ['dark', 'light']) {
|
||||
await ctx.close()
|
||||
}
|
||||
|
||||
{
|
||||
if (want('启动页')) {
|
||||
// 只有一张表卡住（后端连着）：每个请求最多等 15 秒，到点记成出错。遥测点从「加载中」
|
||||
// 转成「降级」，提示也得跟着改口说「没加载成功」，而不是悄悄撤掉——那张表还是空的
|
||||
const state = { approvals: [], offline: false, writes: [], hangPath: '/api/tools' }
|
||||
const { ctx, page } = await open({ state })
|
||||
await page.goto(`${WEB}/chat`)
|
||||
// 第一次全量加载要等那一项超时才算落定，启动页到点自己放人进来
|
||||
const entered = await nav(page).waitFor({ timeout: 25000 }).then(() => true, () => false)
|
||||
check('卡住的那一项超时后，启动页自己放人进来', entered)
|
||||
const warn = page.getByText(/有 1 项没加载成功（工具 超时）.*先别新建/)
|
||||
check('超时的那一项：进门就说它没加载成功，劝先别新建', await warn.waitFor({ timeout: 5000 }).then(() => true, () => false),
|
||||
await warn.innerText().catch(() => '没有这条提示'))
|
||||
const telemetry = nav(page).locator('[data-telemetry]')
|
||||
check('遥测点记为「降级」', (await telemetry.getAttribute('data-tone')) === '降级', await telemetry.getAttribute('data-tone'))
|
||||
await telemetry.click()
|
||||
const row = page.getByRole('dialog', { name: '后端连接' }).locator('[data-check="tools"]')
|
||||
const rowText = (await row.innerText()).replace(/\s+/g, ' ')
|
||||
check('加载清单里那一项写「超时」，不写「连不上」', /超时/.test(rowText), rowText)
|
||||
await page.keyboard.press('Escape')
|
||||
await page.waitForTimeout(4500)
|
||||
check('提示常驻，不会自己消失', await warn.isVisible())
|
||||
state.hangPath = null
|
||||
await page.locator('[role="status"] > div', { hasText: '没加载成功' }).getByRole('button', { name: '重试' }).click({ timeout: 5000 }).catch(() => {})
|
||||
check('重试取回来：提示撤掉', await warn.waitFor({ state: 'detached', timeout: 5000 }).then(() => true, () => false))
|
||||
check('遥测点回到「在线」', await page.waitForFunction(() => document.querySelector('[data-telemetry]')?.getAttribute('data-tone') === '在线', null, { timeout: 8000 })
|
||||
.then(() => true, () => false), await telemetry.getAttribute('data-tone'))
|
||||
check('没有未捕获的运行时错误', page.errors.length === 0, page.errors[0] ?? '')
|
||||
await ctx.close()
|
||||
}
|
||||
|
||||
if (want('启动页')) {
|
||||
// 取回来过的表，后来的刷新卡住了：手上的列表是真的（失败不清空），遥测点说「加载中」
|
||||
// 就够了，不能再弹「列表可能不全，先别新建」吓人
|
||||
const state = { approvals: [], offline: false, writes: [] }
|
||||
const { ctx, page } = await open({ state })
|
||||
await page.goto(`${WEB}/chat`)
|
||||
await waitNav(page)
|
||||
const telemetry = nav(page).locator('[data-telemetry]')
|
||||
await page.waitForFunction(() => document.querySelector('[data-telemetry]')?.getAttribute('data-tone') === '在线', null, { timeout: 10000 }).catch(() => {})
|
||||
state.hangPath = '/api/tools'
|
||||
await telemetry.click()
|
||||
await page.getByRole('dialog', { name: '后端连接' }).getByRole('button', { name: '重新检测' }).click()
|
||||
await page.keyboard.press('Escape')
|
||||
await page.waitForTimeout(3800)
|
||||
check('刷新卡住：遥测点说「加载中」', (await telemetry.getAttribute('data-tone')) === '加载中', await telemetry.getAttribute('data-tone'))
|
||||
check('……但不弹「先别新建」：那张表之前取回来过', (await page.getByText(/先别新建/).count()) === 0)
|
||||
await telemetry.click()
|
||||
const panel = page.getByRole('dialog', { name: '后端连接' })
|
||||
check('遥测浮层照样写出哪一项还没回来，但不劝别新建', /工具还没回来/.test(await panel.innerText()) && !/先别急着新建/.test(await panel.innerText()),
|
||||
(await panel.innerText()).replace(/\n/g, ' '))
|
||||
await page.keyboard.press('Escape')
|
||||
check('没有未捕获的运行时错误', page.errors.length === 0, page.errors[0] ?? '')
|
||||
await ctx.close()
|
||||
}
|
||||
|
||||
if (want('启动页')) {
|
||||
// 断线时在导航上换了主题：重连后补存进设置，不能被服务端的旧值翻回去
|
||||
const state = { approvals: [], offline: true, writes: [] }
|
||||
const { ctx, page } = await open({ state })
|
||||
|
||||
@@ -939,6 +939,401 @@ console.log('\n=== 编号列、追问、CSV ===')
|
||||
await q.page.close()
|
||||
}
|
||||
|
||||
// ---------------------------------------------------------------- 第三波
|
||||
/** 量一个 CSS 颜色表达式算出来的 rgb，跟元素的实际颜色比 */
|
||||
const colorOf = (page, v) => page.evaluate((expr) => {
|
||||
const probe = document.createElement('span')
|
||||
probe.style.color = expr
|
||||
document.body.append(probe)
|
||||
const c = getComputedStyle(probe).color
|
||||
probe.remove()
|
||||
return c
|
||||
}, v)
|
||||
const count = (text, needle) => text.split(needle).length - 1
|
||||
|
||||
console.log('\n=== 协作团队用完轮数：泳道、头部、报错（NI-5、REQ-3A-3）===')
|
||||
for (const [w, dense, label] of [[1100, '0', '宽栏'], [380, '1', '窄栏']]) {
|
||||
const { page, errors } = await open(`syn=exhausted&dense=${dense}`, { w, h: 900, name: `exhausted-${dense === '1' ? 'dense' : 'wide'}` })
|
||||
const body = await page.locator('body').innerText()
|
||||
const lanes = await page.locator('[data-team-lanes]').innerText().catch(() => '')
|
||||
check(`${label}：泳道头说「用完 2 轮仍未完成」`, lanes.includes('用完 2 轮仍未完成'), lanes.split('\n')[0])
|
||||
check(`${label}：泳道点出从没派到的成员`, lanes.includes('汇总员') && lanes.includes('没派到'), lanes.replace(/\n/g, ' ').slice(0, 120))
|
||||
check(`${label}：最后那次判定不画成第 3 轮`, !/第\s?3\s?轮/.test(lanes), lanes.replace(/\n/g, ' ').slice(0, 160))
|
||||
const cell = await page.evaluate(() => {
|
||||
const b = document.querySelector('[data-team-lanes] button[aria-label^="取数员 第 1 轮"]')
|
||||
return b ? getComputedStyle(b).backgroundColor : ''
|
||||
})
|
||||
check(`${label}：失败的成员格子是失败色`, !!cell && cell === await colorOf(page, 'var(--st-failed)'), cell)
|
||||
check(`${label}：判定那一行说「轮数用完 · 调度者判定：未完成」`, body.includes('轮数用完 · 调度者判定：未完成'))
|
||||
// 报错用 lib/explain 讲清为什么、怎么办，不再原样贴后端那一整句
|
||||
const alert = await page.locator('[data-turn] [role="alert"]').first().innerText().catch(() => '')
|
||||
check(`${label}：报错标题说人话`, alert.startsWith('协作团队用完 2 轮仍未完成'), alert.split('\n')[0])
|
||||
check(`${label}：报错给出怎么办`, alert.includes('最多轮数'), alert.replace(/\n/g, ' ').slice(0, 160))
|
||||
check(`${label}：没有运行时报错`, errors.length === 0, errors.join(' | '))
|
||||
const overflow = await page.evaluate(() => document.documentElement.scrollWidth - document.documentElement.clientWidth)
|
||||
check(`${label}：不横向溢出`, overflow <= 0, `${overflow}px`)
|
||||
await page.close()
|
||||
}
|
||||
{
|
||||
const { page } = await open('syn=exhausted-degrade', { w: 1100, h: 900, name: 'exhausted-degrade' })
|
||||
const head = await page.locator('[data-turn] .sticky').first().innerText()
|
||||
check('降档交付：头部不是一个安静的「已完成」', head.includes('协作团队没做完'), head.replace(/\n/g, ' '))
|
||||
const lanes = await page.locator('[data-team-lanes]').innerText().catch(() => '')
|
||||
check('降档交付：泳道头说按降档交付', lanes.includes('按降档交付'), lanes.split('\n')[0])
|
||||
const body = await page.locator('body').innerText()
|
||||
check('降档交付：团队那一行说清是降档', body.includes('用完 2 轮仍未完成，按降档交付'))
|
||||
check('降档交付：提醒行给出下一步', body.includes('不能当结论用'))
|
||||
await page.close()
|
||||
const r = await open('syn=exhausted-closing&dense=1', { w: 380, h: 800, name: 'exhausted-closing' })
|
||||
const rb = await r.page.locator('body').innerText()
|
||||
check('最后那次判定进行中：写「调度者在做最后判定」', rb.includes('轮数用完 · 调度者在做最后判定'))
|
||||
check('最后那次判定进行中：不写「第 3 轮」', !rb.includes('第 3 轮'))
|
||||
const rl = await r.page.locator('[data-team-lanes]').innerText().catch(() => '')
|
||||
check('最后那次判定进行中：泳道也说在判定', rl.includes('轮数用完 · 调度者在判定'), rl.split('\n').slice(0, 2).join(' '))
|
||||
await r.page.close()
|
||||
}
|
||||
|
||||
console.log('\n=== 模型把工具调用写成文字、修复凑数、工具超时（NI-3/4）===')
|
||||
{
|
||||
const { page, errors } = await open('syn=markup&dense=1', { w: 380, h: 900, name: 'markup' })
|
||||
const body = await page.locator('body').innerText()
|
||||
const alert = await page.locator('[data-turn] [role="alert"]').first().innerText().catch(() => '')
|
||||
check('报错标题：模型没有真正调用工具', alert.startsWith('模型没有真正调用工具'), alert.split('\n')[0])
|
||||
check('报错说原因和怎么办', alert.includes('当成文字') && alert.includes('绑定'), alert.replace(/\n/g, ' '))
|
||||
check('原始标记不出现在正文里', !body.includes('DSML'))
|
||||
const row = await page.locator('[data-node-id="query"][data-step-status="failed"]').first().innerText().catch(() => '')
|
||||
check('失败的节点行说人话', row.includes('模型没有真正调用工具'), row.replace(/\n/g, ' '))
|
||||
const warn = page.locator('[data-step-code="tool_markup_leak"]').first()
|
||||
const warnText = await warn.innerText().catch(() => '')
|
||||
check('提醒行说人话', warnText.includes('模型把工具调用写成了文字，没有真正执行'), warnText.replace(/\n/g, ' '))
|
||||
check('下一步不用展开就看得到', await warn.locator('[data-step-next]').count() === 1
|
||||
&& (await warn.locator('[data-step-next]').innerText()).includes('绑定'))
|
||||
check('没有运行时报错', errors.length === 0, errors.join(' | '))
|
||||
await page.close()
|
||||
|
||||
const rp = await open('syn=repair', { w: 1100, h: 900, name: 'repair' })
|
||||
const rbody = await rp.page.locator('body').innerText()
|
||||
// 两次修复凑出的是同一个值：并成一行「×2」，原因照样写在行上；值不同时分行（见 check-decode）
|
||||
check('每次修复都算上、说出凑出来的值', rbody.includes('让模型修复格式') && /×2/.test(rbody) && rbody.includes('total_count=0'),
|
||||
rbody.split('\n').filter((l) => l.includes('修复') || l.includes('×')).join(' | '))
|
||||
const ralert = await rp.page.locator('[data-turn] [role="alert"]').first().innerText().catch(() => '')
|
||||
check('报错：校验修复被拒绝', ralert.startsWith('校验修复被拒绝'), ralert.split('\n')[0])
|
||||
await rp.page.close()
|
||||
|
||||
const tl = await open('syn=timeout-live&dense=1', { w: 380, h: 700, name: 'timeout-live' })
|
||||
const over = await tl.page.evaluate(() => {
|
||||
const row = document.querySelector('[aria-busy="true"][data-step-status="running"]')
|
||||
const clock = row?.querySelector('[data-over-limit]')
|
||||
return { text: clock?.textContent ?? '', color: clock ? getComputedStyle(clock).color : '' }
|
||||
})
|
||||
check('进行中的查询越过时限要说破', over.text.includes('已超出 30 s 上限'), over.text)
|
||||
check('越过时限的秒表是提醒色', over.color === await colorOf(tl.page, 'var(--st-waiting)'), over.color)
|
||||
await tl.page.close()
|
||||
const td = await open('syn=timeout', { w: 1100, h: 700 })
|
||||
check('超时的查询写明超了多少上限', (await td.page.locator('body').innerText()).includes('超过 30 s 上限，已放弃等待'))
|
||||
await td.page.close()
|
||||
}
|
||||
|
||||
console.log('\n=== 放弃、结构化的报错、恢复的轮次、出具横幅（REQ-1/5/6、REQ-3A-2）===')
|
||||
{
|
||||
const { page } = await open('syn=abandoned', { w: 1100, h: 600, name: 'abandoned' })
|
||||
const mark = await page.locator('[data-phase-mark]').allInnerTexts()
|
||||
check('放弃的运行：分段线说清谁放弃的、一并关了什么', mark.some((t) => t.includes('已取消') && t.includes('张工')
|
||||
&& t.includes('1 条待审批一并关闭')), mark.join(' | '))
|
||||
await page.close()
|
||||
|
||||
const s = await open('syn=structured', { w: 1100, h: 600, name: 'structured' })
|
||||
const alert = await s.page.locator('[data-turn] [role="alert"]').first().innerText()
|
||||
// 以前页面拆好的「操作超时」又被翻译一遍：「操作超时 / 操作超时：查询超时…」
|
||||
check('拆好的报错照原样画，标题不重复', count(alert, '操作超时') === 1 && alert.startsWith('操作超时'), alert.replace(/\n/g, ' '))
|
||||
check('原因、怎么办各一行', alert.includes('查询超时：数据库 90 秒没有返回') && alert.includes('缩小时间范围'))
|
||||
check('原文收进技术细节', !alert.includes('OperationalError') && alert.includes('技术细节'))
|
||||
await s.page.close()
|
||||
|
||||
const r = await open('syn=restored', { w: 1100, h: 600 })
|
||||
const head = await r.page.locator('[data-turn] .sticky').first().innerText()
|
||||
check('恢复的轮次：头部照样写几次查询', head.includes('3 次查询'), head.replace(/\n/g, ' '))
|
||||
await r.page.close()
|
||||
|
||||
const i = await open('syn=issued', { w: 1100, h: 900 })
|
||||
check('出具横幅带 data-issuance-banner（画布印章据此滚过来）', await i.page.locator('[data-issuance-banner]').count() === 1)
|
||||
const live = await i.page.locator('[data-turn] [role="status"][aria-live="polite"]').first().innerText()
|
||||
check('读屏也播报出具档位', live.includes('降档出具'), live)
|
||||
await i.page.close()
|
||||
|
||||
const c = await open('syn=tools-dropped&dense=1', { w: 380, h: 700, name: 'tools-dropped' })
|
||||
const cb = await c.page.locator('body').innerText()
|
||||
check('改图时工具被清空要单独说', cb.includes('「数据查询」的工具被清空了'))
|
||||
check('自查通过但工具被去掉，自查那行不是安静的通过', cb.includes('自查通过，但有 1 处工具被去掉了'))
|
||||
await c.page.close()
|
||||
}
|
||||
|
||||
console.log('\n=== 吸顶的轮次头贴着顶边（REQ-8）===')
|
||||
for (const [q, w, h, label] of [['syn=mixed', 1024, 768, '宽栏 1024'], ['syn=mixed', 1180, 800, '宽栏 1180'],
|
||||
['syn=mixed&dense=1', 380, 700, '窄栏']]) {
|
||||
const { page } = await open(q, { w, h })
|
||||
const gap = await page.evaluate(() => {
|
||||
const sc = document.querySelector('[data-stream-scroll]')
|
||||
sc.scrollTop = Math.min(sc.scrollHeight - sc.clientHeight, 260)
|
||||
sc.dispatchEvent(new Event('scroll'))
|
||||
const head = document.querySelector('[data-turn] .sticky')
|
||||
const s = sc.getBoundingClientRect()
|
||||
const hd = head.getBoundingClientRect()
|
||||
// 顶边往下 1px 那一行像素：落在头上才算没缝
|
||||
const hit = document.elementFromPoint(s.left + s.width / 2, s.top + 1)
|
||||
return { gap: Math.round(hd.top - s.top), onHead: !!hit && head.contains(hit), scrolled: sc.scrollTop }
|
||||
})
|
||||
check(`${label}：往下翻后轮次头贴着滚动区顶边，上面不露正文`, gap.scrolled > 0 && gap.gap === 0 && gap.onHead, JSON.stringify(gap))
|
||||
await page.close()
|
||||
}
|
||||
|
||||
console.log('\n=== 画布右栏：去审批、放弃、出具横幅、回执（REQ-1/3/4、REQ-3A-2）===')
|
||||
{
|
||||
const WF_ID = 'fx-stream3'
|
||||
const RUN = 'fxstream0003'
|
||||
const GRAPH = {
|
||||
nodes: [
|
||||
{ id: 'in', type: 'input', position: { x: 0, y: 0 }, data: { label: '问题', config: { fields: [{ name: 'q' }] } } },
|
||||
{ id: 'review', type: 'human', position: { x: 300, y: 0 }, data: { label: '主管审批', config: { mode: 'approve' } } },
|
||||
{ id: 'out', type: 'output', position: { x: 620, y: 0 }, data: { label: '成果', config: {} } },
|
||||
{ id: 'query', type: 'agent', position: { x: 300, y: 200 }, data: { label: '数据查询', config: { tools: [] } } },
|
||||
],
|
||||
edges: [{ id: 'e1', source: 'in', target: 'review' }, { id: 'e2', source: 'review', target: 'out', sourceHandle: 'approved' }],
|
||||
}
|
||||
const WF = { id: WF_ID, name: '右栏检查三', description: '', graph: GRAPH, tags: [], version: 1, status: 'draft',
|
||||
published_version: null, run_count: 0, created_at: '2026-09-26T00:00:00Z', updated_at: '2026-09-26T00:00:00Z' }
|
||||
const APPROVAL = { id: 'ap-3', run_id: RUN, node_id: 'review', mode: 'approve', title: '这份报表可以发出吗？',
|
||||
payload: { message: '华东 1,204 单' }, status: 'pending', response: {}, created_at: new Date(Date.now() - 60_000).toISOString(),
|
||||
workflow_name: '右栏检查三', node_label: '主管审批', run_class: 'exploratory' }
|
||||
const ctx = await browser.newContext({ viewport: { width: 1440, height: 900 }, colorScheme: 'dark' })
|
||||
await ctx.addInitScript(() => { try { localStorage.setItem('agentlab_actor', '张工') } catch { /* 隐私窗口 */ } })
|
||||
const page = await ctx.newPage()
|
||||
const errors = []
|
||||
page.on('pageerror', (e) => errors.push(e.message))
|
||||
const json = (route, body, status = 200) => route.fulfill({ status, contentType: 'application/json', body: JSON.stringify(body) })
|
||||
let pending = true
|
||||
let status = 'interrupted'
|
||||
const cancels = []
|
||||
await page.route(/\/api\/workflows(\?.*)?$/, async (route) => {
|
||||
if (route.request().method() !== 'GET') return json(route, { detail: '检查脚本不写库' }, 409)
|
||||
const real = await (await route.fetch()).json().catch(() => [])
|
||||
return json(route, [WF, ...(Array.isArray(real) ? real : [])])
|
||||
})
|
||||
await page.route(/\/api\/workflows\/fx-stream3(\/.*)?(\?.*)?$/, (route) =>
|
||||
route.request().method() === 'GET' ? json(route, WF) : json(route, { detail: '检查脚本不写库' }, 409))
|
||||
await page.route(/\/api\/conversations(\/.*)?(\?.*)?$/, (route) =>
|
||||
route.request().method() === 'GET' ? json(route, []) : json(route, { id: 'fx-conv3', kind: 'canvas', title: '', turns: [] }))
|
||||
await page.route(/\/api\/approvals(\/.*)?(\?.*)?$/, (route) =>
|
||||
route.request().method() !== 'GET' ? json(route, { detail: '检查脚本不写库' }, 409) : json(route, pending ? [APPROVAL] : []))
|
||||
await page.route(/\/api\/runs(\/.*)?(\?.*)?$/, (route) => {
|
||||
const req = route.request()
|
||||
if (req.method() === 'POST' && req.url().includes(`${RUN}/cancel`)) {
|
||||
cancels.push(req.headers()['x-actor'] ?? '')
|
||||
return json(route, { ok: true, status: 'cancelled' })
|
||||
}
|
||||
if (req.method() === 'GET' && req.url().includes(RUN)) {
|
||||
return json(route, { id: RUN, workflow_id: WF_ID, status, input: {}, output: {}, error: null, usage: {}, run_class: 'exploratory' })
|
||||
}
|
||||
return req.method() === 'GET' ? route.continue() : json(route, { detail: '检查脚本不写库' }, 409)
|
||||
})
|
||||
await page.goto(`${WEB}/studio/${WF_ID}`, { waitUntil: 'networkidle' })
|
||||
await page.waitForFunction((id) => window.__studio?.getState().workflow?.id === id, WF_ID, { timeout: 15000 })
|
||||
const now = Date.now() / 1000
|
||||
const ev = (seq, type, node_id, data = {}, t = 0) => ({ seq, type, node_id, data, ts: now - 30 + t })
|
||||
const WAIT = [
|
||||
ev(1, 'run.started', null, { nodes: 3 }, 0),
|
||||
ev(2, 'node.started', 'in', { node_type: 'input', label: '问题' }, 0.1),
|
||||
ev(3, 'node.finished', 'in', { duration_ms: 3, preview: { q: 'x' } }, 0.2),
|
||||
ev(4, 'node.started', 'review', { node_type: 'human', label: '主管审批' }, 0.3),
|
||||
ev(5, 'human.requested', 'review', { kind: 'human_node', node_id: 'review', mode: 'approve', title: '这份报表可以发出吗？' }, 0.4),
|
||||
ev(6, 'run.interrupted', 'review', { payload: { node_id: 'review', title: '这份报表可以发出吗？' } }, 0.5),
|
||||
]
|
||||
const feed = (list) => page.evaluate((l) => { const s = window.__studio.getState(); for (const e of l) s.applyEvent(e) }, list)
|
||||
await page.evaluate((run) => window.__studio.setState({ run: { id: run, workflow_id: 'fx-stream3', status: 'queued', input: {},
|
||||
output: {}, error: null, usage: {}, run_class: 'exploratory', version: null }, streaming: true, unsubscribe: () => {} }), RUN)
|
||||
await feed(WAIT)
|
||||
await page.waitForTimeout(600)
|
||||
const panel = page.locator('.sheet-in').first()
|
||||
check('等人：栏头「去审批」旁边有「放弃这次运行」', await panel.getByRole('button', { name: '放弃这次运行' }).count() === 1)
|
||||
|
||||
// 胶囊的「去审批」在右栏处于对话层时派发 agentlab:goto-approval：右栏得自己切过去、把焦点放进审批卡
|
||||
await panel.getByRole('button', { name: /助手/ }).first().click()
|
||||
await page.waitForTimeout(300)
|
||||
check('回到对话层后运行层收起', await page.locator('.sheet-in').count() === 0)
|
||||
// 右栏条和胶囊同一套说法：停在谁、等了多久
|
||||
const strip = await page.locator('[data-run-strip]').innerText().catch(() => '')
|
||||
check('对话层的运行条说停在哪个节点', strip.includes('主管审批'), strip.replace(/\n/g, ' '))
|
||||
if (SHOTS) {
|
||||
for (const theme of ['dark', 'light']) {
|
||||
await page.evaluate((t) => document.documentElement.setAttribute('data-theme', t), theme)
|
||||
await page.waitForTimeout(200)
|
||||
await page.locator('[data-assistant-panel]').screenshot({ path: `${SHOTS}/stream-panel-strip-${theme}.png` })
|
||||
}
|
||||
await page.evaluate(() => document.documentElement.setAttribute('data-theme', 'dark'))
|
||||
}
|
||||
// 右栏收起时它还挂着（只是藏起来）：那时不接手，让胶囊自己兜底，别切到一层没人看得见的运行层
|
||||
const unhandled = await page.evaluate((run) => {
|
||||
const aside = document.querySelector('[data-assistant-panel]').closest('aside')
|
||||
aside.style.visibility = 'hidden'
|
||||
const passed = window.dispatchEvent(new CustomEvent('agentlab:goto-approval', { detail: { runId: run }, cancelable: true }))
|
||||
aside.style.visibility = ''
|
||||
return passed
|
||||
}, RUN)
|
||||
await page.waitForTimeout(200)
|
||||
check('右栏收起时不接手 goto-approval', unhandled && await page.locator('.sheet-in').count() === 0)
|
||||
await page.evaluate((run) => window.dispatchEvent(new CustomEvent('agentlab:goto-approval', { detail: { runId: run, nodeId: 'review' } })), RUN)
|
||||
await page.waitForTimeout(900)
|
||||
const focusIn = await page.evaluate(() => !!document.activeElement?.closest('[data-approval]'))
|
||||
check('goto-approval：切到运行层', await page.locator('.sheet-in').count() === 1)
|
||||
check('goto-approval：焦点落在审批卡里', focusIn)
|
||||
check('goto-approval：审批卡描了一圈', await page.locator('[data-approval].sf-flash').count() === 1)
|
||||
if (SHOTS) await page.screenshot({ path: `${SHOTS}/stream-panel-goto-approval-dark.png` })
|
||||
|
||||
// 放弃：确认之后调 cancel，带着署名；事件流过来就收成「已取消」，说清是谁放弃的
|
||||
await page.locator('.sheet-in').first().getByRole('button', { name: '放弃这次运行' }).click()
|
||||
await page.waitForTimeout(300)
|
||||
const dialog = await page.locator('[role="dialog"]').innerText().catch(() => '')
|
||||
check('放弃前先确认，说清后果', dialog.includes('待审批') && dialog.includes('不能再接着跑'), dialog.replace(/\n/g, ' ').slice(0, 160))
|
||||
await page.locator('[role="dialog"]').getByRole('button', { name: '放弃这次运行' }).click()
|
||||
await page.waitForTimeout(400)
|
||||
check('放弃：调了 cancel，带着署名', cancels.length === 1 && decodeURIComponent(cancels[0]) === '张工', JSON.stringify(cancels))
|
||||
pending = false
|
||||
status = 'cancelled'
|
||||
await feed([ev(7, 'run.cancelled', null, { timing: { wall_ms: 30000, active_ms: 500, wait_ms: 29500 }, actor: '张工',
|
||||
message: '放弃了这次运行，1 条待审批一并关闭' }, 30), ev(8, 'stream.end', null, { status: 'cancelled' }, 30)])
|
||||
await page.waitForTimeout(500)
|
||||
const head = await page.locator('.sheet-in [data-turn] .sticky').first().innerText().catch(() => '')
|
||||
check('放弃：头部写已取消', head.includes('已取消'), head.replace(/\n/g, ' '))
|
||||
const marks = await page.locator('.sheet-in [data-phase-mark]').allInnerTexts()
|
||||
check('放弃：分段线写清谁放弃、一并关了什么', marks.some((t) => t.includes('张工') && t.includes('一并关闭')), marks.join(' | '))
|
||||
check('放弃之后不再给「去审批」和「放弃」', await page.locator('.sheet-in').getByRole('button', { name: /去审批|放弃这次运行/ }).count() === 0)
|
||||
if (SHOTS) {
|
||||
for (const theme of ['dark', 'light']) {
|
||||
await page.evaluate((t) => document.documentElement.setAttribute('data-theme', t), theme)
|
||||
await page.waitForTimeout(200)
|
||||
await page.screenshot({ path: `${SHOTS}/stream-panel-abandoned-${theme}.png` })
|
||||
}
|
||||
await page.evaluate(() => document.documentElement.setAttribute('data-theme', 'dark'))
|
||||
}
|
||||
|
||||
// 出具横幅：画布上的印章派发 agentlab:goto-issuance；右栏在对话层时先切过去再滚到横幅
|
||||
await page.evaluate((run) => {
|
||||
const s = window.__studio.getState()
|
||||
s.clearRun()
|
||||
window.__studio.setState({ run: { id: run, workflow_id: 'fx-stream3', status: 'succeeded', input: {}, output: {
|
||||
结论: '华东 1,204 单', _issuance: { tier: 'degraded', calibers: [], metrics_checked: 1, matched_numbers: 0,
|
||||
unmatched_numbers: [{ token: '1,204', context: '华东 1,204 单' }], missing_required: [], missing_expected: [], gaps: [] } },
|
||||
error: null, usage: {}, run_class: 'exploratory', version: null }, streaming: false, unsubscribe: () => {} })
|
||||
}, RUN)
|
||||
await feed([ev(1, 'run.started', null, { nodes: 3 }, 0), ev(2, 'node.started', 'out', { node_type: 'output', label: '成果' }, 0.1),
|
||||
ev(3, 'issuance', 'out', { tier: 'degraded', unmatched: 1, gaps: [], metrics_checked: 1, matched_numbers: 0 }, 0.2),
|
||||
ev(4, 'node.finished', 'out', { duration_ms: 2, preview: {} }, 0.3),
|
||||
ev(5, 'run.finished', null, { output: { 结论: '华东 1,204 单', _issuance: { tier: 'degraded', calibers: [], metrics_checked: 1,
|
||||
matched_numbers: 0, unmatched_numbers: [{ token: '1,204', context: '华东 1,204 单' }], missing_required: [], missing_expected: [], gaps: [] } },
|
||||
usage: {}, duration_ms: 400, timing: { wall_ms: 400, active_ms: 400, wait_ms: 0 } }, 0.4)])
|
||||
await page.waitForTimeout(400)
|
||||
await page.locator('.sheet-in').first().getByRole('button', { name: /助手/ }).first().click()
|
||||
await page.waitForTimeout(300)
|
||||
await page.evaluate((run) => window.dispatchEvent(new CustomEvent('agentlab:goto-issuance', { detail: { runId: run } })), RUN)
|
||||
await page.waitForTimeout(900)
|
||||
const banner = await page.evaluate(() => {
|
||||
const b = document.querySelector('.sheet-in [data-issuance-banner]')
|
||||
const s = b?.closest('[data-stream-scroll]')?.getBoundingClientRect()
|
||||
const r = b?.getBoundingClientRect()
|
||||
return { found: !!b, inView: !!r && !!s && r.top >= s.top - 1 && r.top < s.bottom, focused: !!b && document.activeElement === b }
|
||||
})
|
||||
check('goto-issuance：切到运行层、横幅在视野里', banner.found && banner.inView, JSON.stringify(banner))
|
||||
// 焦点带过去：读屏和键盘用户从画布那头被带过来，得落在横幅上，而不是留在画布的印章上
|
||||
check('goto-issuance：焦点落在出具横幅上', banner.focused)
|
||||
|
||||
// 回执：看过了没改（unchanged）不说「已更新画布」；修过几处数不出来时不写「几 处」
|
||||
await page.locator('.sheet-in').first().getByRole('button', { name: /助手/ }).first().click()
|
||||
await page.evaluate(() => {
|
||||
const s = window.__studio.getState()
|
||||
s.clearRun()
|
||||
const base = { explanation: '', error: '', phase: 'done' }
|
||||
window.__studio.setState({ copilotTurns: [
|
||||
{ ...base, id: 'u1', instruction: '检查一下有没有问题', outcome: 'unchanged',
|
||||
ops: [{ op: 'done', explanation: '' }, { op: 'check', status: 'passed', repaired: 0 }, { op: 'final', graph: { nodes: [{ id: 'in' }] } }] },
|
||||
{ ...base, id: 'u2', instruction: '把循环条件改对', outcome: 'applied', diff: { added: [], changed: ['review'], removed: [], total: 1 },
|
||||
ops: [{ op: 'update_node', id: 'review' }, { op: 'check', status: 'passed', repaired: 1 }, { op: 'final', graph: { nodes: [{ id: 'in' }] } }] },
|
||||
] })
|
||||
})
|
||||
await page.waitForTimeout(400)
|
||||
const turns = await page.locator('[data-turn]').allInnerTexts()
|
||||
check('unchanged：说「看过了，画布没有需要改的地方」', turns[0]?.includes('看过了，画布没有需要改的地方')
|
||||
&& !turns[0]?.includes('已更新画布'), turns[0]?.split('\n').slice(0, 3).join(' '))
|
||||
check('修过但数不出几处：不写「几 处」', turns[1]?.includes('自查发现的问题已自动修好') && !turns[1]?.includes('几 处'),
|
||||
turns[1]?.split('\n').slice(0, 3).join(' '))
|
||||
|
||||
// 改图时模型漏写了 tools：回执逐项列出去掉了什么；自查是通过的，「让 Copilot 再修」不管用，撤销是首选
|
||||
const dropped = { level: 'warning', node_id: 'query', code: 'tools_dropped', field: 'tools',
|
||||
message: '「数据查询」的工具从 db_query__shop 变成了空：这一轮的要求里没有提到去掉工具' }
|
||||
await page.evaluate((warn) => {
|
||||
window.__studio.setState({ copilotTurns: [{ explanation: '', error: '', phase: 'done', id: 'u3', instruction: '只查上月的数据',
|
||||
outcome: 'applied', diff: { added: [], changed: ['query'], removed: [], total: 1 },
|
||||
ops: [{ op: 'update_node', id: 'query', label: '数据查询', config: { system: '只查上月' } },
|
||||
{ op: 'check', status: 'passed', repaired: 0, warnings: [warn] },
|
||||
{ op: 'final', graph: { nodes: [{ id: 'in' }, { id: 'query' }] }, issues: [warn], tool_changes: [{ node_id: 'query',
|
||||
label: '数据查询', member: null, field: 'tools', before: ['db_query__shop'], after: [], added: [], removed: ['db_query__shop'] }] }] }] })
|
||||
}, dropped)
|
||||
await page.waitForTimeout(400)
|
||||
const loss = await page.locator('[data-tool-changes]').innerText().catch(() => '')
|
||||
check('回执逐项列出被去掉的工具', loss.includes('「数据查询」去掉了') && loss.includes('db_query__shop') && loss.includes('一个工具都没有'),
|
||||
loss.replace(/\n/g, ' '))
|
||||
const head3 = await page.locator('[data-turn="u3"] .sticky').innerText().catch(() => '')
|
||||
check('回执头部是提醒，写清几处工具被去掉', head3.includes('1 处工具被去掉了'), head3.replace(/\n/g, ' '))
|
||||
check('只丢了工具时不给「让 Copilot 再修」，撤销是主按钮',
|
||||
await page.getByRole('button', { name: /让 Copilot/ }).count() === 0
|
||||
&& await page.locator('button.btn-primary', { hasText: '撤销这次生成' }).count() === 1)
|
||||
if (SHOTS) {
|
||||
for (const theme of ['dark', 'light']) {
|
||||
await page.evaluate((t) => document.documentElement.setAttribute('data-theme', t), theme)
|
||||
await page.waitForTimeout(200)
|
||||
await page.locator('[data-assistant-panel]').screenshot({ path: `${SHOTS}/stream-panel-tools-lost-${theme}.png` })
|
||||
}
|
||||
await page.evaluate(() => document.documentElement.setAttribute('data-theme', 'dark'))
|
||||
}
|
||||
|
||||
// 模型把工具调用写成文字、判失败：报错和提醒行给「打开『数据查询』的设置」，点了就是那个节点的属性面板
|
||||
const MARKUP_ERROR = '模型输出了工具调用的原始标记，但没有真正调用工具，这一步一次都没查到数据。常见原因：节点没有绑定工具，或者模型、服务不支持工具调用。到画布里给这个节点绑定要用的工具；绑定了还这样，就换一个支持工具调用的模型'
|
||||
await page.evaluate((run) => {
|
||||
window.__studio.setState({ copilotTurns: [] })
|
||||
window.__studio.setState({ run: { id: run, workflow_id: 'fx-stream3', status: 'running', input: {}, output: {}, error: null,
|
||||
usage: {}, run_class: 'exploratory', version: null }, streaming: true, unsubscribe: () => {} })
|
||||
}, RUN)
|
||||
await feed([ev(1, 'run.started', null, { nodes: 4 }, 0), ev(2, 'node.started', 'query', { node_type: 'agent', label: '数据查询' }, 0.1),
|
||||
ev(3, 'llm.start', 'query', { model: 'demo-chat' }, 0.2),
|
||||
ev(4, 'llm.end', 'query', { model: 'demo-chat', duration_ms: 1800, input_tokens: 800, output_tokens: 90, cost_usd: 0.001 }, 2),
|
||||
ev(5, 'log', 'query', { level: 'warn', code: 'tool_markup_leak', message: '模型把工具调用写成了文字（<｜｜DSML｜｜invoke name="db_query__shop">…），没有真正调用工具，已提醒它重试一次' }, 2.1),
|
||||
ev(6, 'node.failed', 'query', { error: MARKUP_ERROR, duration_ms: 3900 }, 4),
|
||||
ev(7, 'run.failed', null, { error: MARKUP_ERROR, node_id: 'query', label: '数据查询', timing: { wall_ms: 4000, active_ms: 4000, wait_ms: 0 } }, 4)])
|
||||
await page.waitForTimeout(500)
|
||||
const runPanel = page.locator('.sheet-in').first()
|
||||
const alertText = await runPanel.locator('[data-turn-error]').innerText().catch(() => '')
|
||||
check('右栏报错说人话：模型没有真正调用工具', alertText.startsWith('模型没有真正调用工具') && !alertText.includes('DSML'),
|
||||
alertText.split('\n')[0])
|
||||
check('提醒行的下一步也给直达入口', await runPanel.locator('[data-step-code="tool_markup_leak"] [data-fix="canvas"]').count() === 1)
|
||||
if (SHOTS) {
|
||||
for (const theme of ['dark', 'light']) {
|
||||
await page.evaluate((t) => document.documentElement.setAttribute('data-theme', t), theme)
|
||||
await page.waitForTimeout(200)
|
||||
await page.locator('[data-assistant-panel]').screenshot({ path: `${SHOTS}/stream-panel-markup-${theme}.png` })
|
||||
}
|
||||
await page.evaluate(() => document.documentElement.setAttribute('data-theme', 'dark'))
|
||||
}
|
||||
const openSettings = runPanel.locator('[data-turn-error]').getByRole('button', { name: '打开「数据查询」的设置' })
|
||||
check('报错给「打开『数据查询』的设置」', await openSettings.count() === 1)
|
||||
if (await openSettings.count()) {
|
||||
await openSettings.click()
|
||||
await page.waitForTimeout(300)
|
||||
check('点「打开设置」就是那个节点的属性面板', await page.evaluate(() => window.__studio.getState().selectedId) === 'query')
|
||||
}
|
||||
|
||||
check('画布右栏没有运行时报错', errors.length === 0, errors.join(' | '))
|
||||
await ctx.close()
|
||||
}
|
||||
|
||||
await browser.close()
|
||||
console.log(failed ? `\n✗ ${failed} 项未通过` : '\n✓ 助手流渲染全部通过')
|
||||
process.exit(failed ? 1 : 0)
|
||||
|
||||
+371
-9
@@ -51,6 +51,34 @@ const GRAPH = {
|
||||
}
|
||||
const V1 = { nodes: GRAPH.nodes.slice(0, 3).concat([GRAPH.nodes[6]]), edges: [edge('start', 'lookup'), edge('lookup', 'gate'), edge('gate', 'done', 'fast')] }
|
||||
|
||||
// 查库的 agent 和带成员的协作团队：助手改节点时工具不能被悄悄改没（NI-1），提示词点名的
|
||||
// 工具没绑定要能定位到工具那一栏（NI-2）。全是通用示例名
|
||||
const QUERY_TOOLS = ['db_query__shop', 'db_schema__shop']
|
||||
const TEAM_GRAPH = {
|
||||
nodes: [
|
||||
node('start', 'input', 0, 200, '输入', { fields: [{ name: 'question', required: true }] }),
|
||||
node('fetch', 'agent', 300, 200, '数据查询', {
|
||||
system: '你是取数助手。', prompt: '先用 db_schema__shop 看表结构，再用 db_query__shop 查 orders 表里每个门店的订单数',
|
||||
tools: ['db_query__shop'], assign_to: 'rows', max_steps: 8,
|
||||
}),
|
||||
node('team', 'supervisor', 600, 200, '复核团队', {
|
||||
goal: '核对订单数', max_rounds: 4,
|
||||
agents: [
|
||||
{ name: 'fetcher', description: '负责查库', system: '只用 SQL 取数', tools: ['db_query__shop'] },
|
||||
{ name: 'writer', description: '写结论', system: '用 python_exec 算出占比再写结论', tools: [] },
|
||||
],
|
||||
}),
|
||||
node('done', 'output', 900, 200, '成果', { fields: [{ name: 'answer', value: '{{ vars.rows }}' }] }),
|
||||
],
|
||||
edges: [edge('start', 'fetch'), edge('fetch', 'team'), edge('team', 'done')],
|
||||
}
|
||||
/** 工具都绑好了的那一版：合并语义要守住的就是它 */
|
||||
const TEAM_BOUND = {
|
||||
...TEAM_GRAPH,
|
||||
nodes: TEAM_GRAPH.nodes.map((n) => (n.id === 'fetch'
|
||||
? { ...n, data: { ...n.data, config: { ...n.data.config, tools: [...QUERY_TOOLS] } } } : n)),
|
||||
}
|
||||
|
||||
const wf = (id, extra) => ({
|
||||
id, name: extra.name, description: extra.description ?? '检查脚本伪造的工作流', graph: extra.graph ?? GRAPH,
|
||||
tags: extra.tags ?? [], version: extra.version ?? 3, is_template: !!extra.is_template,
|
||||
@@ -62,6 +90,8 @@ const FAKES = {
|
||||
'st-main': wf('st-main', { name: '__studio_check__', version: 3, published_version: 2 }),
|
||||
'st-gov': wf('st-gov', { name: '__studio_check_gov__', status: 'governed', version: 3, published_version: 3, published_by: '张工' }),
|
||||
'st-tpl': wf('st-tpl', { name: '__studio_check_tpl__', is_template: true, tags: ['extracted'] }),
|
||||
'st-team': wf('st-team', { name: '__studio_check_team__', graph: TEAM_BOUND, version: 2, published_version: 2, status: 'published' }),
|
||||
'st-lint': wf('st-lint', { name: '__studio_check_lint__', graph: TEAM_GRAPH, version: 1 }),
|
||||
}
|
||||
const VERSIONS = [
|
||||
{ id: 'v3', version: 3, note: '', created_at: '2026-09-26T02:00:00Z' },
|
||||
@@ -182,7 +212,7 @@ async function open({ width = 1440, height = 900, path = '/studio/st-main', pref
|
||||
return { ctx, page, errors, state }
|
||||
}
|
||||
|
||||
const S = (page, fn) => page.evaluate(fn)
|
||||
const S = (page, fn, arg) => page.evaluate(fn, arg)
|
||||
const st = (page) => page.evaluate(() => {
|
||||
const s = window.__studio.getState()
|
||||
return { nodes: s.nodes.map((n) => n.id), edges: s.edges.length, selectedId: s.selectedId,
|
||||
@@ -675,9 +705,10 @@ console.log('=== 助手改图：一步撤销、回执、只回话、失败退回
|
||||
message: '模型写了一个不存在的节点类型「magic」，这一步已跳过' }] },
|
||||
]
|
||||
const past0 = (await st(page)).past
|
||||
await S(page, () => window.__studio.getState().runCopilot('给快速回答后面加一步润色', true))
|
||||
const began = await S(page, () => window.__studio.getState().runCopilot('给快速回答后面加一步润色', true))
|
||||
await page.waitForFunction(() => window.__studio.getState().copilotTurns.at(-1)?.phase !== 'running')
|
||||
const t1 = await S(page, () => window.__studio.getState().copilotTurns.at(-1))
|
||||
check('开始了就告诉输入框（runCopilot 返回 true）', began === true, String(began))
|
||||
check('收到 final：这一轮是「已应用」', t1.outcome === 'applied', t1.outcome)
|
||||
check('diff 记下新增和改动', t1.diff?.added.includes('polish') && t1.diff?.changed.includes('answer'), JSON.stringify(t1.diff))
|
||||
check('final.issues 进了这一轮（少了一步）', t1.issues?.some((i) => i.code === 'unknown_node_type'))
|
||||
@@ -767,8 +798,7 @@ console.log('=== 助手改图：一步撤销、回执、只回话、失败退回
|
||||
await page.waitForFunction(() => window.__studio.getState().copilotTurns.at(-1)?.phase !== 'running')
|
||||
state.stream = [{ op: 'model', model: 'fake-model' }, { op: 'reply', text: '好' }]
|
||||
const turnId = await S(page, () => window.__studio.getState().copilotTurns.at(-1).id)
|
||||
await S(page, (id) => window.__studio.getState().repairWithCopilot(id), turnId).catch(() => null)
|
||||
await page.evaluate((id) => window.__studio.getState().repairWithCopilot(id), turnId)
|
||||
await S(page, (id) => window.__studio.getState().repairWithCopilot(id), turnId)
|
||||
await page.waitForTimeout(400)
|
||||
const lastBody = state.bodies.at(-1)
|
||||
check('再修一次：指令里带着剩下的问题', lastBody?.instruction?.includes('还没选工具') && !!lastBody?.base_graph,
|
||||
@@ -798,7 +828,7 @@ console.log('=== 入口：?run=&focus=、选择器、发布弹窗 ===')
|
||||
await page.keyboard.type('__studio_check')
|
||||
await page.waitForTimeout(200)
|
||||
const rows = await count(page, '[role="listbox"][aria-label="工作流"] > li')
|
||||
check('选择器能搜索', rows === 3, `${rows} 条`)
|
||||
check('选择器能搜索', rows === Object.keys(FAKES).length, `${rows} 条`)
|
||||
check('当前工作流有底色和勾', await count(page, 'li[aria-selected="true"] .bg-accent-solid') === 1)
|
||||
check('模板行常驻「用它新建」', await count(page, 'li:has-text("__studio_check_tpl__") button:has-text("用它新建")') === 1)
|
||||
check('标签本地化（extracted → 从运行提取）', await count(page, 'li:has-text("__studio_check_tpl__") :text("从运行提取")') === 1)
|
||||
@@ -822,7 +852,8 @@ console.log('=== 入口：?run=&focus=、选择器、发布弹窗 ===')
|
||||
: route.fallback()))
|
||||
await S(pub.page, () => { const s = window.__studio.getState(); s.updateNode('lookup', { label: '背景检索2' }); s.undo() })
|
||||
const pb = await S(pub.page, () => { const s = window.__studio.getState(); return { future: s.future.length, fit: s.fitRequest, dirty: s.dirty } })
|
||||
await pub.page.locator('button:has-text("发布")').first().click()
|
||||
await pub.page.getByRole('button', { name: '发布', exact: true }).click()
|
||||
await pub.page.waitForSelector('[role="dialog"] button.btn-primary', { timeout: 5000 }).catch(() => {})
|
||||
await pub.page.waitForTimeout(300)
|
||||
await pub.page.locator('[role="dialog"] button.btn-primary').click()
|
||||
await pub.page.waitForFunction(() => window.__studio.getState().workflow?.status === 'published', null, { timeout: 5000 }).catch(() => {})
|
||||
@@ -843,17 +874,348 @@ console.log('=== 入口：?run=&focus=、选择器、发布弹窗 ===')
|
||||
await win.ctx.close()
|
||||
|
||||
const gov = await open({ path: '/studio/st-gov' })
|
||||
await gov.page.locator('button:has-text("发布")').first().click()
|
||||
// 按名字找：别的按钮（版本标签、提示）也可能带着「发布」两个字
|
||||
await gov.page.getByRole('button', { name: '发布', exact: true }).click()
|
||||
await gov.page.waitForSelector('[role="dialog"] [role="radio"]', { timeout: 5000 }).catch(() => {})
|
||||
await gov.page.waitForTimeout(300)
|
||||
const radio = await gov.page.locator('[role="radio"][aria-checked="true"]').innerText()
|
||||
// 只看发布弹窗里的：助手输入框的「在现有图上改 | 从头生成」也是一组 radio
|
||||
const radio = await gov.page.locator('[role="dialog"] [role="radio"][aria-checked="true"]').innerText()
|
||||
check('受管工作流的发布默认选「受管」', radio.includes('受管'), radio.slice(0, 20))
|
||||
check('选项用统一术语（已发布 — 可发起正式运行）', await gov.page.getByText('已发布 — 可发起正式运行').count() === 1)
|
||||
check('写明以谁的名义发布', await gov.page.getByText('检查脚本').count() >= 1)
|
||||
await gov.page.locator('[role="radio"]:has-text("已发布")').click()
|
||||
await gov.page.locator('[role="dialog"] [role="radio"]:has-text("已发布")').click()
|
||||
check('从受管降级要说出来', await gov.page.getByText(/会降级/).count() === 1)
|
||||
await gov.ctx.close()
|
||||
}
|
||||
|
||||
// ================================================================ 6b
|
||||
console.log('=== 助手改节点是合并，不是整体替换（NI-1）；出错留下「怎么办」 ===')
|
||||
{
|
||||
const { ctx, page, state, errors } = await open({ path: '/studio/st-team' })
|
||||
// 流在最后一步出错：画布退回这一轮之前，半成品进重做栈。读重做栈顶就是按操作流逐条
|
||||
// 落上去的样子——final 会拿后端合并好的整张图覆盖，只看 final 看不出前端合得对不对
|
||||
state.stream = [
|
||||
{ op: 'model', model: 'fake-model' },
|
||||
{ op: 'update_node', id: 'fetch', label: '数据查询',
|
||||
config: { system: '你是取数助手，按月份汇总。', prompt: '按月份统计 orders 的订单数' } },
|
||||
{ op: 'update_node', id: 'team', config: { goal: '核对按月的订单数', agents: [
|
||||
{ name: 'fetcher', system: '按月份取数' },
|
||||
{ name: 'writer', remove: true },
|
||||
{ name: 'checker', description: '复核口径', tools: ['db_schema__shop'], system: null },
|
||||
] } },
|
||||
{ op: 'update_node', id: 'fetch', config: { max_steps: null } },
|
||||
{ op: 'error', message: '助手这一轮没跑完：等待超时', hint: '换一个响应快的模型再试', detail: 'ReadTimeout: 120s' },
|
||||
]
|
||||
await S(page, () => window.__studio.getState().runCopilot('按月份拆一下', true))
|
||||
await page.waitForFunction(() => window.__studio.getState().copilotTurns.at(-1)?.phase !== 'running')
|
||||
const half = await S(page, () => {
|
||||
const s = window.__studio.getState()
|
||||
const nodes = s.future.at(-1)?.nodes ?? []
|
||||
const cfg = (id) => nodes.find((n) => n.id === id)?.data.config ?? null
|
||||
return { fetch: cfg('fetch'), team: cfg('team') }
|
||||
})
|
||||
const f = half.fetch ?? {}
|
||||
const members = half.team?.agents ?? []
|
||||
const fetcher = members.find((m) => m.name === 'fetcher') ?? {}
|
||||
const checker = members.find((m) => m.name === 'checker') ?? {}
|
||||
check('只改提示词：agent 的工具一个不少', JSON.stringify(f.tools) === JSON.stringify(QUERY_TOOLS), JSON.stringify(f.tools))
|
||||
check('写了的键改上去、没写的键保留', f.prompt === '按月份统计 orders 的订单数' && f.assign_to === 'rows',
|
||||
`${f.prompt} / ${f.assign_to}`)
|
||||
check('写 null 的键删掉', !('max_steps' in f), JSON.stringify(Object.keys(f)))
|
||||
check('成员按 name 合并：没写 tools 的保留原来的 tools 和别的字段',
|
||||
JSON.stringify(fetcher.tools) === '["db_query__shop"]' && fetcher.system === '按月份取数' && fetcher.description === '负责查库',
|
||||
JSON.stringify(fetcher))
|
||||
check('{name, remove:true} 删掉成员，新名字接在后面', members.map((m) => m.name).join(',') === 'fetcher,checker',
|
||||
members.map((m) => m.name).join(','))
|
||||
check('新成员里写 null 的字段不留', !('system' in checker) && JSON.stringify(checker.tools) === '["db_schema__shop"]',
|
||||
JSON.stringify(checker))
|
||||
check('团队没写的顶层键保留', half.team?.max_rounds === 4 && half.team?.goal === '核对按月的订单数')
|
||||
const cp = await S(page, () => window.__studio.getState().copilot)
|
||||
check('出错时保留「怎么办」和技术细节（Composer 的错误条要用）',
|
||||
cp.error.includes('等待超时') && cp.errorHint === '换一个响应快的模型再试' && cp.errorDetail === 'ReadTimeout: 120s',
|
||||
JSON.stringify({ hint: cp.errorHint, detail: cp.errorDetail }))
|
||||
const t0 = await S(page, () => window.__studio.getState().copilotTurns.at(-1))
|
||||
check('这一轮记下开始时刻（助手流头部的实时计时用）', typeof t0.startedAt === 'number'
|
||||
&& Math.abs(Date.now() - t0.startedAt) < 60_000, String(t0.startedAt))
|
||||
|
||||
// 改图回执：后端 final 带回的工具绑定变化。工具集合变小、指令又没让删的，warn 色、可一键撤销
|
||||
const dropped = {
|
||||
...TEAM_BOUND,
|
||||
nodes: TEAM_BOUND.nodes.map((n) => {
|
||||
if (n.id === 'fetch') return { ...n, data: { ...n.data, config: { ...n.data.config, prompt: '简短一点', tools: [] } } }
|
||||
if (n.id === 'team') {
|
||||
return { ...n, data: { ...n.data, config: { ...n.data.config,
|
||||
agents: n.data.config.agents.map((a) => (a.name === 'fetcher' ? { ...a, tools: [] } : a)) } } }
|
||||
}
|
||||
return n
|
||||
}),
|
||||
}
|
||||
const warn = (id, field, message) => ({ level: 'warning', node_id: id, edge_id: null, code: 'tools_dropped', field, message })
|
||||
const warnings = [
|
||||
warn('fetch', 'tools', '「数据查询」的工具从 db_query__shop、db_schema__shop 变成了空。这一轮的要求里没有提到去掉工具，确认一下是不是改漏了'),
|
||||
warn('team', 'agents[0].tools', '「复核团队」的成员「fetcher」的工具从 db_query__shop 变成了空。这一轮的要求里没有提到去掉工具，确认一下是不是改漏了'),
|
||||
]
|
||||
const changes = [
|
||||
{ node_id: 'fetch', label: '数据查询', member: null, field: 'tools', before: QUERY_TOOLS, after: [], added: [], removed: QUERY_TOOLS },
|
||||
{ node_id: 'team', label: '复核团队', member: 'fetcher', field: 'agents[0].tools', before: ['db_query__shop'], after: [], added: [], removed: ['db_query__shop'] },
|
||||
]
|
||||
state.stream = [
|
||||
{ op: 'model', model: 'fake-model' },
|
||||
{ op: 'update_node', id: 'fetch', config: { prompt: '简短一点', tools: [] } },
|
||||
{ op: 'update_node', id: 'team', config: { agents: [{ name: 'fetcher', tools: [] }] } },
|
||||
{ op: 'done', explanation: '改短了' },
|
||||
{ op: 'check', status: 'passed', repaired: 0, warnings },
|
||||
{ op: 'final', graph: dropped, explanation: '改短了', layout: { mode: 'keep', placed: [] },
|
||||
issues: [...warnings, { level: 'warning', node_id: 'done', edge_id: null, message: '成果字段引用的变量可能为空' }],
|
||||
tool_changes: changes },
|
||||
]
|
||||
await S(page, () => window.__studio.getState().runCopilot('提示词写短一点', true))
|
||||
await page.waitForFunction(() => window.__studio.getState().copilotTurns.at(-1)?.phase !== 'running')
|
||||
const t1 = await S(page, () => window.__studio.getState().copilotTurns.at(-1))
|
||||
check('回执记下后端给的工具绑定变化', t1.toolChanges?.length === 2 && t1.toolChanges[0].node_id === 'fetch',
|
||||
JSON.stringify(t1.toolChanges?.map((c) => c.node_id)))
|
||||
check('工具被删的提醒单独记，不混进普通校验提示', t1.toolWarnings?.length === 2
|
||||
&& t1.toolWarnings.every((w) => w.code === 'tools_dropped'), JSON.stringify(t1.toolWarnings?.map((w) => w.node_id)))
|
||||
check('自查那一步也带着这两条提醒', t1.check?.warnings?.length === 2, JSON.stringify(t1.check?.warnings))
|
||||
check('记下的 final 操作带着 tool_changes（右栏回执从它解码）',
|
||||
t1.ops.find((o) => o.op === 'final')?.tool_changes?.length === 2)
|
||||
const toastBox = page.locator('[role="status"] > div').filter({ hasText: '工具绑定变少了' })
|
||||
const toastText = (await toastBox.first().innerText().catch(() => '')).replace(/\s+/g, ' ')
|
||||
check('工具集合变小：回执是 warn 色，逐条列出前后', await toastBox.count() === 1
|
||||
&& toastText.includes('「数据查询」') && toastText.includes('db_query__shop、db_schema__shop → 空')
|
||||
&& toastText.includes('fetcher'), toastText.slice(0, 120))
|
||||
check('工具被删的回执用 warn 色边框', await toastBox.first().evaluate((el) =>
|
||||
el.getAttribute('style')?.includes('var(--warn)'), null, { timeout: 2000 }).catch(() => false))
|
||||
await toastBox.first().locator('button:has-text("撤销")').click({ timeout: 2000 }).catch(() => {})
|
||||
await page.waitForTimeout(250)
|
||||
const back = await S(page, () => window.__studio.getState().nodes.find((n) => n.id === 'fetch').data.config.tools)
|
||||
check('回执里一键撤销：工具回来了', JSON.stringify(back) === JSON.stringify(QUERY_TOOLS), JSON.stringify(back))
|
||||
|
||||
// 老后端的 final 不带 tool_changes：前端自己比对前后两张图
|
||||
state.stream = [
|
||||
{ op: 'model', model: 'fake-model' },
|
||||
{ op: 'update_node', id: 'fetch', config: { tools: ['db_query__shop', 'python_exec'] } },
|
||||
{ op: 'final', graph: { ...TEAM_BOUND, nodes: TEAM_BOUND.nodes.map((n) => (n.id === 'fetch'
|
||||
? { ...n, data: { ...n.data, config: { ...n.data.config, tools: ['db_query__shop', 'python_exec'] } } } : n)) },
|
||||
explanation: '', layout: { mode: 'keep', placed: [] }, issues: [] },
|
||||
]
|
||||
await S(page, () => window.__studio.getState().runCopilot('查表结构换成算一下', true))
|
||||
await page.waitForFunction(() => window.__studio.getState().copilotTurns.at(-1)?.phase !== 'running')
|
||||
const t2 = await S(page, () => window.__studio.getState().copilotTurns.at(-1))
|
||||
const c2 = t2.toolChanges?.[0]
|
||||
check('老后端不给 tool_changes：前端自己比对出来', t2.toolChanges?.length === 1 && c2.node_id === 'fetch'
|
||||
&& c2.added.join() === 'python_exec' && c2.removed.join() === 'db_schema__shop' && c2.member === null,
|
||||
JSON.stringify(t2.toolChanges))
|
||||
check('换了一个工具（数量没少）不算被删，没有提醒', !(t2.toolWarnings ?? []).length)
|
||||
check('没有运行时报错', errors.length === 0, errors.slice(0, 2).join(' | '))
|
||||
await ctx.close()
|
||||
|
||||
// 合并规则本身：和后端 merge_node_config 同一组用例（backend/tests/test_copilot_merge.py）
|
||||
const unit = await open({ path: '/studio/st-team' })
|
||||
const r = await unit.page.evaluate(async () => {
|
||||
const m = await import('/src/canvas/copilotMerge.ts')
|
||||
const old = { prompt: 'p', tools: ['db_query__shop', 'db_schema__shop'], assign_to: 'rows',
|
||||
agents: [{ name: '取数员', system: '只用 SQL 取数', tools: ['db_query__shop'] }, { system: '无名', tools: ['x'] }] }
|
||||
const snapshot = JSON.stringify(old)
|
||||
const a = m.mergeNodeConfig('agent', old, { assign_to: null, temperature: 0.2 })
|
||||
const b = m.mergeNodeConfig('agent', old, { tools: ['db_query__shop'] })
|
||||
const c = m.mergeNodeConfig('supervisor', old, { agents: [
|
||||
{ name: 'agent1', tools: null }, { name: '核对员', tools: ['db_schema__shop'], note: null }, { name: '取数员', remove: true }] })
|
||||
// agents 不是列表、或者不是 supervisor：整体替换这一个键（和后端一样只对 supervisor 的成员列表按 name 合并）
|
||||
const d = m.mergeNodeConfig('agent', old, { agents: [{ name: 'x' }] })
|
||||
return { a, b, c, d, untouched: JSON.stringify(old) === snapshot }
|
||||
})
|
||||
check('合并：null 删键，没写的留着', !('assign_to' in r.a) && r.a.temperature === 0.2 && r.a.tools.length === 2, JSON.stringify(r.a))
|
||||
check('合并：显式写 tools 就换成新的', r.b.tools.join() === 'db_query__shop')
|
||||
// 没名字的那个按 agent1 认出来、合并进去：条目里写的 name 也跟着落上（后端同样如此）
|
||||
check('合并：没名字的成员按 agentN 认；删成员、新名字接在后面', r.c.agents.map((x) => x.name).join() === 'agent1,核对员'
|
||||
&& !('tools' in r.c.agents[0]) && r.c.agents[0].system === '无名' && !('note' in r.c.agents[1]), JSON.stringify(r.c.agents))
|
||||
check('合并：不是 supervisor 时 agents 整体替换', JSON.stringify(r.d.agents) === '[{"name":"x"}]')
|
||||
check('合并：旧 config 不被原地改（比对工具变化还要用它）', r.untouched)
|
||||
await unit.ctx.close()
|
||||
}
|
||||
|
||||
// ================================================================ 6c
|
||||
console.log('=== 正式运行期间画布只读：调色板、粘贴、复制、排版、检查器、撤销都拦下 ===')
|
||||
{
|
||||
const { ctx, page, errors, state } = await open({ path: '/studio/st-team' })
|
||||
const layoutCalls = []
|
||||
page.on('request', (req) => { if (req.url().includes('/copilot/layout')) layoutCalls.push(req.url()) })
|
||||
const formal = (phase) => page.evaluate((p) => window.__studio.setState((s) => ({
|
||||
run: { id: 'st-formal-0001', workflow_id: 'st-team', status: p === 'succeeded' ? 'succeeded' : 'running',
|
||||
run_class: 'formal', version: 2, input: {}, output: {}, error: '', usage: {}, created_at: '2026-09-26T02:00:00Z' },
|
||||
runPhase: p, trace: { ...s.trace, phase: p, runClass: 'formal' },
|
||||
})), phase)
|
||||
// 先选中一个节点、复制下来，再开始正式运行
|
||||
await S(page, () => window.__studio.getState().select('fetch'))
|
||||
await blur(page)
|
||||
await S(page, () => window.__studio.getState().select('fetch'))
|
||||
await page.evaluate(() => document.activeElement instanceof HTMLElement && document.activeElement.blur())
|
||||
await page.keyboard.press(`${MOD}+c`)
|
||||
// 助手先改一轮（改成果节点的名字），正式运行开始后再点这一轮的「撤销这次生成」
|
||||
const renamed = { ...TEAM_BOUND, nodes: TEAM_BOUND.nodes.map((n) => (n.id === 'done'
|
||||
? { ...n, data: { ...n.data, label: '成果（按月）' } } : n)) }
|
||||
state.stream = [
|
||||
{ op: 'model', model: 'fake-model' },
|
||||
{ op: 'update_node', id: 'done', label: '成果（按月）' },
|
||||
{ op: 'final', graph: renamed, explanation: '', layout: { mode: 'keep', placed: [] }, issues: [] },
|
||||
]
|
||||
await S(page, () => window.__studio.getState().runCopilot('成果改个名字', true))
|
||||
await page.waitForFunction(() => window.__studio.getState().copilotTurns.at(-1)?.phase !== 'running')
|
||||
const turnId = await S(page, () => window.__studio.getState().copilotTurns.at(-1).id)
|
||||
await formal('running')
|
||||
await page.waitForTimeout(200)
|
||||
const s0 = await st(page)
|
||||
|
||||
await page.locator('button:has-text("模型调用"), button[aria-label="添加模型调用"]').first().click({ force: true, timeout: 2000 }).catch(() => {})
|
||||
await S(page, () => window.__studio.getState().addNode('llm'))
|
||||
check('调色板点击添加被拦下', (await st(page)).nodes.length === s0.nodes.length)
|
||||
check('调色板标成不可用并说明为什么', await count(page, '[aria-label^="节点库"] [aria-disabled="true"]') === 1
|
||||
&& (await page.locator('[aria-label^="节点库"]').innerText()).includes('只读'))
|
||||
|
||||
await page.evaluate(() => document.activeElement instanceof HTMLElement && document.activeElement.blur())
|
||||
await page.keyboard.press(`${MOD}+v`)
|
||||
await page.keyboard.press(`${MOD}+d`)
|
||||
await page.keyboard.press('Shift+KeyL')
|
||||
await page.keyboard.press(`${MOD}+z`)
|
||||
await page.waitForTimeout(400)
|
||||
const s1 = await st(page)
|
||||
check('⌘V 粘贴、⌘D 复制、⇧L 排版、⌘Z 撤销都不改图', s1.nodes.length === s0.nodes.length && s1.past === s0.past
|
||||
&& layoutCalls.length === 0, `${s1.nodes.length}/${s0.nodes.length} 节点 · 排版请求 ${layoutCalls.length}`)
|
||||
check('按了改图快捷键会说为什么不行', await page.getByText(/正式运行进行中/).count() >= 1)
|
||||
check('工具栏的撤销、自动排版置灰', await page.getByRole('button', { name: '自动排版' }).isDisabled()
|
||||
&& await page.getByRole('button', { name: '撤销', exact: true }).first().isDisabled())
|
||||
// 撤销被锁拦下、画布没动：这一轮不能被标成「已撤回」，回执也不能说撤掉了
|
||||
const undone = await S(page, (id) => window.__studio.getState().undoCopilotTurn(id), turnId)
|
||||
const kept = await S(page, (id) => {
|
||||
const s = window.__studio.getState()
|
||||
return { outcome: s.copilotTurns.find((t) => t.id === id)?.outcome,
|
||||
label: s.nodes.find((n) => n.id === 'done')?.data.label }
|
||||
}, turnId)
|
||||
check('「撤销这次生成」也拦下，这一轮不被标成已撤回', undone === false && kept.outcome === 'applied'
|
||||
&& kept.label === '成果（按月）', JSON.stringify({ undone, ...kept }))
|
||||
|
||||
await S(page, () => window.__studio.getState().select('fetch'))
|
||||
await page.waitForTimeout(300)
|
||||
check('检查器整块只读，并说明原因', await count(page, 'fieldset[disabled]:has([data-field="prompt"])') === 1
|
||||
&& await page.getByText('只能看，不能改').count() === 1)
|
||||
await S(page, () => window.__studio.getState().updateNode('fetch', { label: '改不动' }))
|
||||
check('store 这一层也拦着（直接调 updateNode 不生效）',
|
||||
(await S(page, () => window.__studio.getState().nodes.find((n) => n.id === 'fetch').data.label)) === '数据查询')
|
||||
const prompt0 = await S(page, () => window.__studio.getState().copilotTurns.length)
|
||||
// 返回 false：输入框据此留着用户刚写的那句，不能拦下了还把它清掉
|
||||
const started = await S(page, () => window.__studio.getState().runCopilot('加一步', true))
|
||||
check('正式运行期间不让助手改图（并告诉输入框没开始）', started === false
|
||||
&& (await S(page, () => window.__studio.getState().copilotTurns.length)) === prompt0
|
||||
&& !(await S(page, () => window.__studio.getState().copilot.active)), String(started))
|
||||
|
||||
// 跑完就能改了
|
||||
await formal('succeeded')
|
||||
await page.waitForTimeout(150)
|
||||
await S(page, () => window.__studio.getState().addNode('llm'))
|
||||
check('正式运行结束后恢复编辑', (await st(page)).nodes.length === s0.nodes.length + 1)
|
||||
check('没有运行时报错', errors.length === 0, errors.slice(0, 2).join(' | '))
|
||||
|
||||
// Shift+1 的名字要和它做的事一致：它回到打开时的取景（从入口看起），不是「看全所有节点」
|
||||
const fit = await page.evaluate(async () => (await import('/src/canvas/shortcuts.ts')).studioShortcut('fit').label)
|
||||
check('Shift+1 叫「重新取景」', fit.startsWith('重新取景') && !fit.includes('看全'), fit)
|
||||
await ctx.close()
|
||||
}
|
||||
|
||||
// ================================================================ 6d
|
||||
console.log('=== 提示词点名的工具没绑定：问题面板定位到工具那一栏，一键绑定（NI-2）；团队用完轮数（NI-5） ===')
|
||||
{
|
||||
const { ctx, page, errors } = await open({ path: '/studio/st-lint' })
|
||||
await page.waitForFunction(() => window.__studio.getState().analysis === 'ok', null, { timeout: 8000 })
|
||||
const lint = await S(page, () => window.__studio.getState().issues
|
||||
.filter((i) => /要求用「/.test(i.message)).map((i) => ({ node: i.node_id, field: i.field, msg: i.message.slice(0, 30) })))
|
||||
check('后端校验报出两处「提示词要求用 X 但没绑定」', lint.length === 2, JSON.stringify(lint))
|
||||
|
||||
await page.locator('button[aria-expanded]').filter({ hasText: /错|提示|可运行/ }).first().click()
|
||||
await page.waitForTimeout(250)
|
||||
const agentRow = page.locator('[data-problem]').filter({ hasText: '节点没有绑定它' })
|
||||
check('问题面板里这一条指向「可用工具」', (await agentRow.innerText().catch(() => '')).includes('可用工具'),
|
||||
(await agentRow.innerText().catch(() => '')).replace(/\s+/g, ' ').slice(-20))
|
||||
const memberRow = page.locator('[data-problem]').filter({ hasText: '没有给这个成员绑定它' })
|
||||
check('成员那一条指向「成员「writer」的工具」', (await memberRow.innerText().catch(() => '')).includes('成员「writer」的工具'),
|
||||
(await memberRow.innerText().catch(() => '')).replace(/\s+/g, ' ').slice(-20))
|
||||
|
||||
await agentRow.locator('button').first().click({ timeout: 2000 }).catch(() => {})
|
||||
await page.waitForTimeout(500)
|
||||
const tools = page.locator('[data-field="tools"]')
|
||||
check('点一下：检查器翻到工具那一栏，问题落在那儿', await S(page, () => window.__studio.getState().selectedId) === 'fetch'
|
||||
&& (await tools.innerText()).includes('db_schema__shop'))
|
||||
await tools.locator('button:has-text("绑定 db_schema__shop")').click({ timeout: 2000 }).catch(() => {})
|
||||
await page.waitForTimeout(200)
|
||||
const bound = await S(page, () => window.__studio.getState().nodes.find((n) => n.id === 'fetch').data.config.tools)
|
||||
check('一键绑定：工具加进这个节点', JSON.stringify(bound) === JSON.stringify(['db_query__shop', 'db_schema__shop']), JSON.stringify(bound))
|
||||
|
||||
await memberRow.locator('button').first().click({ timeout: 2000 }).catch(() => {})
|
||||
await page.waitForTimeout(500)
|
||||
const member = page.locator('[data-field="agents"] [data-item="1"]')
|
||||
check('成员那一条落在第 2 个成员的工具下面', (await member.locator('[data-sub="tools"]').innerText({ timeout: 2000 })
|
||||
.catch(() => '')).includes('没有给这个成员绑定它'))
|
||||
const inView = await member.locator('[data-sub="tools"]').evaluate((el) => {
|
||||
const r = el.getBoundingClientRect()
|
||||
return r.top >= 0 && r.bottom <= window.innerHeight
|
||||
}, null, { timeout: 2000 }).catch(() => false)
|
||||
check('镜头滚到那个成员的工具', inView)
|
||||
await member.locator('button:has-text("绑定 python_exec")').click({ timeout: 2000 }).catch(() => {})
|
||||
await page.waitForTimeout(200)
|
||||
const mt = await S(page, () => window.__studio.getState().nodes.find((n) => n.id === 'team').data.config.agents[1].tools)
|
||||
check('一键绑定到那个成员', JSON.stringify(mt) === '["python_exec"]', JSON.stringify(mt))
|
||||
await page.waitForFunction(() => !window.__studio.getState().issues.some((i) => /要求用「/.test(i.message)), null, { timeout: 6000 }).catch(() => {})
|
||||
check('绑好之后这两条问题消失', !(await S(page, () => window.__studio.getState().issues.some((i) => /要求用「/.test(i.message)))))
|
||||
|
||||
// 数据源的查询工具要能在选择器里挑到：不然「把它加进工具里」这句话做不到
|
||||
await member.locator('button:has-text("添加")').click({ timeout: 2000 }).catch(() => {})
|
||||
await page.waitForTimeout(300)
|
||||
const pick = await member.innerText({ timeout: 2000 }).catch(() => '')
|
||||
check('工具选择器列出数据源的 db_query__ / db_schema__ 工具', /db_query__/.test(pick) && /db_schema__/.test(pick))
|
||||
await member.locator('button:has-text("收起")').click({ timeout: 2000 }).catch(() => {})
|
||||
|
||||
// NI-5：协作团队用完轮数时怎么收场
|
||||
await S(page, () => window.__studio.getState().select('team'))
|
||||
await page.waitForTimeout(300)
|
||||
const ex = page.locator('[data-field="on_exhausted"]')
|
||||
const exText = await ex.innerText().catch(() => '')
|
||||
check('团队有「用完轮数时」，默认判为失败', exText.includes('用完轮数时')
|
||||
&& String(await ex.locator('select').evaluate((el) => el.options[el.selectedIndex].text, null, { timeout: 2000 })
|
||||
.catch(() => '')).includes('判为失败'), exText.slice(0, 40))
|
||||
check('选项说清两者差别', /降档/.test(exText) && /失败/.test(exText) && exText.length > 40)
|
||||
// 报错和右栏都叫人「调大『最多轮数』」：字段名得是这四个字，照着找得到
|
||||
check('轮数字段叫「最多轮数」，和报错里的说法一致',
|
||||
(await page.locator('[data-field="max_rounds"]').innerText().catch(() => '')).includes('最多轮数'))
|
||||
await ex.locator('select').selectOption({ label: '降档交付' }, { timeout: 2000 }).catch(() => {})
|
||||
check('选降档交付写进配置', (await S(page, () => window.__studio.getState().nodes.find((n) => n.id === 'team').data.config.on_exhausted)) === 'degrade')
|
||||
check('没有运行时报错', errors.length === 0, errors.slice(0, 2).join(' | '))
|
||||
await ctx.close()
|
||||
}
|
||||
|
||||
// ================================================================ 7b
|
||||
console.log('=== 工作流目录没取回来：不劝人新建（EmptyState source） ===')
|
||||
{
|
||||
// 目录还在读、或者读失败时，「还没有工作流 · 新建」是一句假话：库里可能有几十张
|
||||
for (const mode of ['loading', 'error']) {
|
||||
const ctx = await browser.newContext({ viewport: { width: 1280, height: 800 } })
|
||||
const page = await ctx.newPage()
|
||||
await page.route(/\/api\/workflows(\?.*)?$/, (route) => {
|
||||
if (route.request().method() !== 'GET') return route.fallback()
|
||||
return mode === 'loading' ? new Promise(() => {})
|
||||
: route.fulfill({ status: 500, contentType: 'application/json', body: JSON.stringify({ detail: '检查脚本伪造的 500' }) })
|
||||
})
|
||||
await page.goto(`${WEB}/studio`, { waitUntil: 'domcontentloaded' })
|
||||
// 目录一直不回来时，启动屏 5 秒后给「先进去看看」：进去之后才轮到编排页说话
|
||||
if (mode === 'loading') await page.getByRole('button', { name: '先进去看看' }).click({ timeout: 12000 }).catch(() => {})
|
||||
const hit = await page.waitForSelector(`[data-empty-unknown="${mode}"]`, { timeout: 8000 }).then(() => true).catch(() => false)
|
||||
check(mode === 'loading' ? '目录还在读：说「正在读取」' : '目录没取回来：说没取回来、给重新读取', hit
|
||||
&& (mode === 'loading' ? /正在读取/ : /没取回来/).test(await page.locator('[data-empty-unknown]').innerText()))
|
||||
check(mode === 'loading' ? '读的时候不劝人新建' : '读失败时不劝人新建',
|
||||
await page.getByRole('button', { name: /新建工作流/ }).count() === 0)
|
||||
await ctx.close()
|
||||
}
|
||||
}
|
||||
|
||||
// ================================================================ 8
|
||||
console.log('=== 删掉眼前这张 ===')
|
||||
{
|
||||
|
||||
@@ -541,6 +541,76 @@ console.log('\n=== 右栏收尾（decodeRun） ===')
|
||||
fold([start('甲', 1), start('乙', 2), end('甲', 3, 100), end('乙', 4, 300)]).savedMs === 100)
|
||||
}
|
||||
|
||||
console.log('\n=== 回放时的用量和工具数（逐刻度） ===')
|
||||
{
|
||||
// 回放到半路，卡片和 HUD 要写「那一刻」的 tokens 和工具数，而不是拿终值冒充
|
||||
const ev = (seq, type, node_id, data = {}) => ({ seq, type, node_id, ts: 5000 + seq, data })
|
||||
const events = [
|
||||
ev(1, 'run.started', null, { nodes: 2 }),
|
||||
ev(2, 'node.started', 'agent'),
|
||||
ev(3, 'llm.end', 'agent', { input_tokens: 100, output_tokens: 20, cost_usd: 0.001 }),
|
||||
ev(4, 'tool.start', 'agent', { tool: 'db_query__shop' }),
|
||||
ev(5, 'tool.end', 'agent', { tool: 'db_query__shop' }),
|
||||
ev(6, 'tool.start', 'agent', { tool: 'db_query__shop' }),
|
||||
ev(7, 'tool.end', 'agent', { tool: 'db_query__shop' }),
|
||||
ev(8, 'llm.end', 'agent', { input_tokens: 300, output_tokens: 80, cost_usd: 0.004 }),
|
||||
ev(9, 'node.finished', 'agent', { duration_ms: 7000 }),
|
||||
ev(10, 'node.started', 'out'), ev(11, 'node.finished', 'out'),
|
||||
ev(12, 'run.finished', null, { usage: { input_tokens: 400, output_tokens: 100, cost_usd: 0.005 } }),
|
||||
]
|
||||
const gr = g(['agent', 'out'], [['agent', 'out']])
|
||||
const t = run(events, gr)
|
||||
const at = (seq) => (5000 + seq) * 1000
|
||||
const p4 = project(t, at(4) + 1)
|
||||
check('第一次调工具那一刻：1 次工具、1 个在跑、tokens 只有第一轮',
|
||||
p4.nodes.agent.tools === 1 && p4.nodes.agent.toolsRunning === 1 && p4.nodes.agent.tokensIn === 100
|
||||
&& p4.nodes.agent.tokensOut === 20,
|
||||
JSON.stringify(p4.nodes.agent))
|
||||
const p6 = project(t, at(6) + 0.5)
|
||||
check('第二次调工具时：2 次，第一次已经回来', p6.nodes.agent.tools === 2 && p6.nodes.agent.toolsRunning === 1,
|
||||
JSON.stringify(p6.nodes.agent))
|
||||
const p9 = project(t, at(9) + 1)
|
||||
check('节点跑完：用量是这个节点的最终值、没有在跑的工具', p9.nodes.agent.tokensIn === 400
|
||||
&& p9.nodes.agent.tokensOut === 100 && p9.nodes.agent.tools === 2 && p9.nodes.agent.toolsRunning === 0,
|
||||
JSON.stringify(p9.nodes.agent))
|
||||
check('回放到开始之前：0，不是终值', project(t, at(1)).nodes.agent.tools === 0
|
||||
&& project(t, at(1)).nodes.agent.tokensIn === 0)
|
||||
check('整次运行的用量也按刻度走', p4.tokensIn === 100 && p4.tokensOut === 20 && p9.tokensIn === 400,
|
||||
`${p4.tokensIn}/${p4.tokensOut} → ${p9.tokensIn}`)
|
||||
const live = project(t, trace.lastEventAt(t))
|
||||
check('此刻（实时）读节点的当前值', live.nodes.agent.tokensIn === 400 && live.nodes.agent.tools === 2
|
||||
&& live.tokensIn === 400)
|
||||
// 老后端：agent 的调用不发 llm.end，run.finished 的总数比刻度加起来多——回放时那一刻的
|
||||
// 整次用量不可信，给 undefined，不拿终值冒充
|
||||
const old = run([...events.slice(0, 11),
|
||||
ev(12, 'run.finished', null, { usage: { input_tokens: 9000, output_tokens: 900, cost_usd: 0.05 } })], gr)
|
||||
const op = project(old, at(4) + 1)
|
||||
check('刻度对不上后端总数：回放时整次用量写不出', op.tokensIn === undefined && op.tokensOut === undefined,
|
||||
`${op.tokensIn}/${op.tokensOut}`)
|
||||
check('……实时时照样是后端的总数', project(old, trace.lastEventAt(old)).tokensIn === 9000)
|
||||
check('没有 tool / llm 事件的节点刻度为空、读数是 0',
|
||||
t.nodes.out.marks === undefined && project(t, at(11)).nodes.out.tools === 0)
|
||||
}
|
||||
|
||||
console.log('\n=== 最后一条事件的时刻（lastEventAt） ===')
|
||||
{
|
||||
const ev = (seq, type, node_id, data = {}) => ({ seq, type, node_id, ts: 6000 + seq, data })
|
||||
const gr = g([{ id: 'in', type: 'input' }, 'boom', 'after'], [['in', 'boom'], ['boom', 'after']])
|
||||
const t = run([
|
||||
ev(1, 'run.started', null, { nodes: 3 }),
|
||||
ev(2, 'node.started', 'in'), ev(3, 'node.finished', 'in'),
|
||||
ev(4, 'node.started', 'boom'), ev(5, 'node.failed', 'boom', { error: 'x' }),
|
||||
ev(6, 'run.failed', null, { error: 'x', timing: { wall_ms: 5000, active_ms: 4000, wait_ms: 0 } }),
|
||||
], gr)
|
||||
check('lastEventAt 是最后一条事件的 ts（毫秒）', trace.lastEventAt(t) === 6006 * 1000, `${trace.lastEventAt(t)}`)
|
||||
const settled = project(t, trace.lastEventAt(t))
|
||||
check('投影到这一刻 = 此刻：推导出的阻断还在', settled.nodes.after?.state === 'blocked', settled.nodes.after?.state)
|
||||
check('……后端给的权威时长也在', settled.elapsedMs === 5000 && settled.activeMs === 4000,
|
||||
`${settled.elapsedMs}/${settled.activeMs}`)
|
||||
check('早一毫秒就是按段重建的回放（阻断不在）', project(t, trace.lastEventAt(t) - 1).nodes.after?.state !== 'blocked')
|
||||
check('空航迹的 lastEventAt 是 0', trace.lastEventAt(emptyTrace()) === 0)
|
||||
}
|
||||
|
||||
console.log('\n=== 性能 ===')
|
||||
{
|
||||
// 256 段的大循环（真实运行里见过）：一次折叠 + 投影都要快，拖游标才不卡
|
||||
|
||||
+200
-7
@@ -1,5 +1,7 @@
|
||||
// 基础组件（components/ui.tsx）、前端工具库（lib/*）和连接状态（store/catalog.ts）
|
||||
// 的回归检查。
|
||||
// 基础组件（components/ui.tsx）、前端工具库（lib/*）、连接状态（store/catalog.ts）
|
||||
// 和 api/client 的上传进度的回归检查。管理页共用件（页头、连通胶囊、单选组、删除
|
||||
// 后撤销）和失败说明（lib/explain）也在这里：它们从页面挪进了公共层，页面级检查
|
||||
// 只走得到其中一两种情况。
|
||||
//
|
||||
// 这几样是全站的地基，坏了不会在哪一页上报错，只会悄悄变样：弹窗的焦点漏到页面
|
||||
// 上、组字时按 Esc 把半屏草稿关掉、toast 压住工具栏按钮、后端一抖整站说「还没有
|
||||
@@ -105,6 +107,11 @@ for (const [name, ok, detail] of await page.evaluate(() => {
|
||||
eq('相对时间：59.6 分钟进位到小时档', ago(59.6 * 60_000), '1 小时前')
|
||||
ok('相对时间：23.6 小时不写「24 小时前」', !/24 小时前/.test(ago(23.6 * 3_600_000)), ago(23.6 * 3_600_000))
|
||||
eq('短 id', f.shortId('66a5a6ff0011'), '#66a5a6')
|
||||
eq('文件大小：字节', f.formatBytes(820), '820 B')
|
||||
eq('文件大小：KB', f.formatBytes(12_700), '12.4 KB')
|
||||
eq('文件大小：MB', f.formatBytes(3.1 * 1024 * 1024), '3.1 MB')
|
||||
eq('文件大小拿不到写 —', f.formatBytes(undefined), '—')
|
||||
eq('短名去掉括号里的补充', f.shortLabel('OpenAI 兼容（DeepSeek、通义…）'), 'OpenAI 兼容')
|
||||
|
||||
// 状态
|
||||
eq('succeeded = 已完成', st.statusLabel('succeeded'), '已完成')
|
||||
@@ -157,6 +164,29 @@ for (const [name, ok, detail] of await page.evaluate(() => {
|
||||
ok('isNetworkError 对 HTTP 错误为假', !e.isNetworkError(new ApiError(500, 'NetworkError when attempting to fetch resource')))
|
||||
ok('测连接结果里的 ECONNREFUSED 不算连不上后端',
|
||||
e.humanizeError({ ok: false, error: 'connect ECONNREFUSED 10.0.0.5:5432' }).title !== '连不上后端服务')
|
||||
// 已经是「标题：原因」的中文人话，不再套一个「操作超时」、把整句当原因（标题说两遍）
|
||||
const zh = e.humanizeError(new ApiError(504, '数据库查询超时：超过 30 秒没有返回，已中断。缩小查询范围后重试。'))
|
||||
ok('中文人话里的「超时」不套「操作超时」', zh.title !== '操作超时' && !(zh.reason ?? '').includes('操作超时')
|
||||
&& !(zh.reason && zh.reason.includes(zh.title)), JSON.stringify(zh))
|
||||
eq('英文 timeout 原文照样翻成「操作超时」', e.humanizeError('ReadTimeout: timed out').title, '操作超时')
|
||||
ok('……原因里不带异常类名', !/ReadTimeout/.test(e.humanizeError('ReadTimeout: timed out').reason ?? ''))
|
||||
|
||||
// 失败说明（lib/explain）：记录页、助手流、问数据页共用一份
|
||||
const x = window.__ui.lib.explain
|
||||
const leak = x.explainRunError('模型输出了工具调用的原始标记，但没有真正调用工具。常见原因：节点没有绑定工具，或者模型、服务不支持工具调用。')
|
||||
ok('工具调用标记泄漏：说人话、不给接着跑、指到画布', leak.title === '模型没有真正调用工具' && !leak.continuable && leak.fix === 'canvas', JSON.stringify(leak))
|
||||
const tired = x.explainRunError('NodeError: 协作团队用完 4 轮仍未完成：还缺汇总员的结论')
|
||||
ok('团队轮数用完：标题带轮数，原因是调度者最后的理由', tired.title === '协作团队用完 4 轮仍未完成'
|
||||
&& tired.reason === '还缺汇总员的结论' && !tired.continuable, JSON.stringify(tired))
|
||||
const made = x.explainRunError('校验修复无效：修复时出现了原文没有的值：total_count=0')
|
||||
ok('校验修复编造数值：点出是哪个值、不给接着跑', made.reason.includes('total_count=0') && !made.continuable, JSON.stringify(made))
|
||||
eq('提示词点名的工具没绑定', x.explainRunError('提示词要求用 db_query__shop，但节点没有绑定它').title,
|
||||
'提示词要求用「db_query__shop」，但节点没有绑定它')
|
||||
eq('老运行的鉴权失败', x.explainRunError('NodeError: AuthenticationError: Error code: 401 - invalid_api_key').title, '模型鉴权没通过')
|
||||
const miss = x.explainRunError('缺少必填输入：region')
|
||||
ok('缺必填输入：给出是哪一项、重新发起', miss.missingInput === 'region' && miss.fix === 'rerun' && !miss.continuable)
|
||||
eq('没留原因', x.explainRunError(null).title, '运行失败，但没有留下原因')
|
||||
ok('兜底：句中的异常类名也去掉', !/ValueError/.test(x.explainRunError('结构不对：ValueError: bad').title))
|
||||
return out
|
||||
})) check(name, ok, detail)
|
||||
|
||||
@@ -305,12 +335,13 @@ await page.waitForTimeout(100)
|
||||
check('出错的在 assertive 区', (await page.locator('[role="alert"][aria-live="assertive"]').innerText()).includes('保存失败'))
|
||||
check('普通的在 polite 区', (await page.locator('[role="status"][aria-live="polite"]').first().innerText()).includes('已复制到剪贴板'))
|
||||
check('出错的能复制', (await page.getByRole('button', { name: '复制详情' }).count()) >= 1)
|
||||
// 位置：内容区（让出 56px 导航）居中，顶边在 48px 工具栏之下
|
||||
// 位置：内容区（让出 64px 的导航，App 里的 w-16）居中，顶边在 48px 工具栏之下
|
||||
const NAV_W = 64
|
||||
const stackBox = await page.locator('[role="alert"][aria-live="assertive"]').boundingBox()
|
||||
const vw = page.viewportSize().width
|
||||
check('toast 在工具栏下沿之下（≥ 56px）', stackBox.y >= 56, `top=${Math.round(stackBox.y)}`)
|
||||
check('toast 在内容区里居中', Math.abs(stackBox.x + stackBox.width / 2 - (56 + vw) / 2) < 4,
|
||||
`中线 ${Math.round(stackBox.x + stackBox.width / 2)}，内容区中线 ${(56 + vw) / 2}`)
|
||||
check('toast 在内容区里居中', Math.abs(stackBox.x + stackBox.width / 2 - (NAV_W + vw) / 2) < 2,
|
||||
`中线 ${Math.round(stackBox.x + stackBox.width / 2)}，内容区中线 ${(NAV_W + vw) / 2}`)
|
||||
const infoCard = page.getByText('已复制到剪贴板')
|
||||
await infoCard.hover()
|
||||
await page.waitForTimeout(4600)
|
||||
@@ -339,14 +370,24 @@ const emptyBlock = page.locator('[data-block="空态 EmptyState（模拟离线
|
||||
await page.click('#toggle-offline')
|
||||
await page.waitForTimeout(100)
|
||||
check('横幅出现', (await page.getByText('后端未连接').count()) === 1)
|
||||
check('空态改说拿不到数据', (await emptyBlock.getByText('暂时拿不到数据').count()) === 1)
|
||||
// 横幅早就挂着（连着时什么都不画），计时器是这一刻才走起来的：头一个数得按此刻算
|
||||
const firstSecs = Number((await page.getByText(/秒后自动重试/).innerText()).match(/\d+/)?.[0])
|
||||
check('刚断开时倒计时从此刻算（8 秒）', firstSecs === 8 || firstSecs === 7, `横幅 ${firstSecs} 秒`)
|
||||
check('空态改说拿不到数据', (await emptyBlock.getByText('暂时拿不到数据').count()) >= 1)
|
||||
check('离线空态不许诺「会自动刷新」', (await emptyBlock.getByText(/自动刷新/).count()) === 0)
|
||||
check('离线时收起「新建」', (await page.getByRole('button', { name: '新建工作流' }).count()) === 0)
|
||||
check('本地空态不受影响', (await page.getByText('没有匹配的工具').count()) === 1)
|
||||
// role=alert 整块重念：倒计时一秒一跳，放在里面读屏就一秒念一遍整条横幅
|
||||
const bannerAlert = page.locator('[data-offline-banner] [role="alert"]')
|
||||
check('横幅外层不是 role=alert', (await page.locator('[data-offline-banner]').getAttribute('role')) === null)
|
||||
check('横幅的播报区只有不变的那几句', (await bannerAlert.count()) === 1
|
||||
&& (await bannerAlert.innerText()).includes('后端未连接') && !/秒后自动重试/.test(await bannerAlert.innerText()),
|
||||
await bannerAlert.innerText().catch(() => ''))
|
||||
check('倒计时对读屏隐藏', (await page.getByText(/秒后自动重试/).getAttribute('aria-hidden')) === 'true')
|
||||
// 横幅把工具栏往下推了多少，toast 就跟着让多少
|
||||
await page.click('#toast-error')
|
||||
await page.waitForTimeout(80)
|
||||
const bannerH = (await page.locator('[role="alert"]', { hasText: '后端未连接' }).boundingBox()).height
|
||||
const bannerH = (await page.locator('[data-offline-banner]').boundingBox()).height
|
||||
const offTop = (await page.locator('[role="alert"][aria-live="assertive"]').boundingBox()).y
|
||||
check('离线时 toast 再让出横幅的高度', Math.abs(offTop - (56 + bannerH)) < 2, `top=${Math.round(offTop)}，横幅 ${Math.round(bannerH)}px`)
|
||||
await page.evaluate(() => window.__ui.toast.dismiss())
|
||||
@@ -364,6 +405,158 @@ const wantAria = await page.evaluate(() => window.__ui.lib.keys.ariaShortcut('Mo
|
||||
const gotAria = await page.getByRole('button', { name: '刷新列表' }).getAttribute('aria-keyshortcuts')
|
||||
check('IconButton 的 aria-keyshortcuts 按平台写', gotAria === wantAria && !/Ctrl|Mod/.test(gotAria), gotAria)
|
||||
|
||||
console.log('\n=== 管理页共用件：页头、连通胶囊、单选组、删除后撤销 ===')
|
||||
check('页头 48px 高', Math.round((await page.locator('#page-header-demo header').boundingBox()).height) === 48)
|
||||
const pills = await page.locator('#health-demo [data-health]').evaluateAll((els) => els.map((el) => el.textContent))
|
||||
// 「3 分钟前」是从预览页挂载时算的，跑到这里可能已经过了一分钟
|
||||
check('连通胶囊五种说法', pills[0] === '未测试' && /^正在测 · /.test(pills[1]) && /^已连通 · 42 ms · [34] 分钟前测$/.test(pills[2])
|
||||
&& pills[3] === '连不上 · 1 小时前测' && /已连通 · 380 ms.*配置改过了/.test(pills[4]), pills.join(' | '))
|
||||
check('不知道测的时刻：不留一个悬空的「 · 」', pills[5] === '连不上', JSON.stringify(pills[5]))
|
||||
const spoken = await page.locator('#health-demo [role="status"]').evaluateAll((els) => els.map((el) => el.textContent))
|
||||
check('读屏那一句不带跳动的计时和相对时间', spoken[1] === '正在测连接' && /^已连通 42 ms，\d\d:\d\d 测的$/.test(spoken[2]), spoken.join(' | '))
|
||||
|
||||
const radios = page.locator('#radio-demo [role="radio"]')
|
||||
check('单选组只占一个 Tab 位', JSON.stringify(await radios.evaluateAll((els) => els.map((el) => el.tabIndex))) === '[0,-1,-1]')
|
||||
const checkedRadio = () => page.locator('#radio-demo [role="radio"][aria-checked="true"]').innerText()
|
||||
const focusedRadio = () => page.evaluate(() => document.activeElement?.textContent)
|
||||
await radios.first().focus()
|
||||
await page.keyboard.press('ArrowRight')
|
||||
check('→ 选中下一项并聚焦', (await checkedRadio()) === 'sid' && (await focusedRadio()) === 'sid')
|
||||
await page.keyboard.press('End')
|
||||
check('End 到最后一项', (await checkedRadio()) === 'dsn')
|
||||
await page.keyboard.press('ArrowDown')
|
||||
check('↓ 首尾相接', (await checkedRadio()) === 'service_name')
|
||||
await page.keyboard.press('ArrowUp')
|
||||
await page.keyboard.press('Home')
|
||||
check('↑ 往回、Home 到第一项', (await checkedRadio()) === 'service_name' && (await focusedRadio()) === 'service_name')
|
||||
await page.keyboard.press('Alt+ArrowRight')
|
||||
check('带 Alt 的方向键不接（那是浏览器后退 / 前进）', (await checkedRadio()) === 'service_name')
|
||||
await page.getByRole('tab', { name: '模型接入' }).focus()
|
||||
await page.keyboard.press('End')
|
||||
check('标签页：End 到最后一个', (await page.getByRole('tab', { name: '偏好' }).getAttribute('aria-selected')) === 'true')
|
||||
await page.keyboard.press('Home')
|
||||
|
||||
let deletes = 0
|
||||
page.on('request', (r) => { if (r.method() === 'DELETE') deletes++ })
|
||||
const rowShown = (id) => page.locator(`#defer-demo [data-row="${id}"]`).count()
|
||||
const undoBtn = page.locator('[role="status"][aria-live="polite"]').getByRole('button', { name: '撤销' })
|
||||
await page.getByRole('button', { name: '删除 orders' }).click()
|
||||
await page.waitForTimeout(80)
|
||||
check('删除：行先拿掉，toast 给「撤销」', (await rowShown('orders')) === 0
|
||||
&& (await page.getByText('已删除「orders」').count()) === 1 && (await undoBtn.count()) === 1)
|
||||
await page.click('#defer-reload')
|
||||
await page.waitForTimeout(50)
|
||||
check('撤销窗口里重拉列表，删掉的行不回来', (await rowShown('orders')) === 0)
|
||||
await undoBtn.click()
|
||||
await page.waitForTimeout(80)
|
||||
check('撤销：行回来了，一个 DELETE 都没发', (await rowShown('orders')) === 1 && deletes === 0)
|
||||
await page.evaluate(() => window.__ui.toast.dismiss())
|
||||
|
||||
console.log('\n=== 空态：catalog 的表没取回来时不说「还没有」 ===')
|
||||
const sourceEmpty = page.locator('#empty-source-demo')
|
||||
await page.evaluate(() => window.__ui.useCatalog.getState().refresh())
|
||||
check('取回来了：照常说「还没有工作流」、给新建', (await sourceEmpty.getByText('还没有工作流').count()) === 1
|
||||
&& (await sourceEmpty.locator('#empty-source-action').count()) === 1)
|
||||
const realCatalog = await page.evaluate(() => {
|
||||
const s = window.__ui.useCatalog.getState()
|
||||
return { loadedAt: s.loadedAt, checks: s.checks }
|
||||
})
|
||||
await page.evaluate(() => window.__ui.useCatalog.setState({
|
||||
loadedAt: {}, checks: [{ key: 'workflows', label: '工作流', state: 'pending' }],
|
||||
}))
|
||||
await page.waitForTimeout(50)
|
||||
check('从没取回来过、还在路上：说「正在读取工作流」、收起新建', (await sourceEmpty.getByText('正在读取工作流').count()) === 1
|
||||
&& (await sourceEmpty.locator('#empty-source-action').count()) === 0
|
||||
&& (await sourceEmpty.locator('[data-empty-unknown="loading"]').getAttribute('role')) === 'status')
|
||||
await page.evaluate(() => window.__ui.useCatalog.setState({
|
||||
checks: [{ key: 'workflows', label: '工作流', state: 'error', error: '后端 15 秒没有响应，稍后重试' }],
|
||||
}))
|
||||
await page.waitForTimeout(50)
|
||||
check('超时或报错：说「工作流没取回来」和原因、给重新读取', (await sourceEmpty.getByText('工作流没取回来').count()) === 1
|
||||
&& (await sourceEmpty.getByText(/15 秒没有响应.*这里显示为空不代表没有数据/).count()) === 1
|
||||
&& (await sourceEmpty.getByRole('button', { name: '重新读取' }).count()) === 1
|
||||
&& (await sourceEmpty.locator('#empty-source-action').count()) === 0)
|
||||
await page.evaluate((c) => window.__ui.useCatalog.setState(c), realCatalog)
|
||||
|
||||
console.log('\n=== catalog：一张表卡住不拖累别的表 ===')
|
||||
// /tools 连得上但不回（MCP 服务不应答时就是这样）：以前 Promise.all 等它，工作流也一直空着
|
||||
const toolsBefore = await page.evaluate(() => window.__ui.useCatalog.getState().tools.length)
|
||||
// unroute 要拿同一个函数引用才摘得掉
|
||||
const isTools = (u) => new URL(u).pathname === '/api/tools'
|
||||
await page.route(isTools, () => { /* 永远不回 */ })
|
||||
const expectWf = await fetch(`${API}/workflows`).then((r) => r.json()).then((l) => l.length, () => null)
|
||||
await page.evaluate(() => window.__ui.useCatalog.setState({ workflows: [] }))
|
||||
const tHang = Date.now()
|
||||
let settled = false
|
||||
const hung = page.evaluate(() => window.__ui.useCatalog.getState().refresh()).then(() => { settled = true })
|
||||
await page.waitForFunction(() => window.__ui.useCatalog.getState().workflows.length > 0, null, { timeout: 8000 }).catch(() => {})
|
||||
let cat = await page.evaluate(() => {
|
||||
const s = window.__ui.useCatalog.getState()
|
||||
return { wf: s.workflows.length, wfAt: s.loadedAt.workflows, tools: s.checks.find((c) => c.key === 'tools')?.state }
|
||||
})
|
||||
if (expectWf) {
|
||||
check('工作流先回来先填，不等卡住的工具', cat.wf === expectWf && cat.wfAt >= tHang && cat.tools === 'pending',
|
||||
`${JSON.stringify(cat)}，后端有 ${expectWf} 个工作流`)
|
||||
}
|
||||
// 最多等 20 秒：没有超时的话 refresh 永远不落定，检查自己也不能跟着挂死
|
||||
await Promise.race([hung, new Promise((r) => setTimeout(r, 20_000))])
|
||||
const hangMs = Date.now() - tHang
|
||||
check('refresh 在超时之后落定', settled, `${hangMs}ms 还没落定`)
|
||||
cat = await page.evaluate(() => {
|
||||
const s = window.__ui.useCatalog.getState()
|
||||
const c = s.checks.find((x) => x.key === 'tools')
|
||||
return { state: c?.state, error: c?.error, backend: s.backend, tools: s.tools.length, loaded: s.loaded }
|
||||
})
|
||||
check('卡住的那张 15 秒超时，记成出错并说清', cat.state === 'error' && /15 秒/.test(cat.error ?? '') && hangMs > 14_500 && hangMs < 18_000,
|
||||
`${cat.state} ${cat.error} ${hangMs}ms`)
|
||||
check('一张表超时不算后端断开', cat.backend === 'ok', cat.backend)
|
||||
check('超时不清空手上已有的那份', cat.tools === toolsBefore, `${toolsBefore} → ${cat.tools}`)
|
||||
await page.unroute(isTools)
|
||||
|
||||
console.log('\n=== 上传进度（kb.upload 走 XHR） ===')
|
||||
const isUpload = (u) => new URL(u).pathname === '/api/kb/upload'
|
||||
const uploadVia = async (handler) => {
|
||||
await page.unroute(isUpload)
|
||||
await page.route(isUpload, handler)
|
||||
}
|
||||
const tryUpload = (size) => page.evaluate(async (size) => {
|
||||
const events = []
|
||||
try {
|
||||
const doc = await window.__ui.api.kb.upload(new File([new Uint8Array(size)], 'orders.csv', { type: 'text/csv' }), 'default',
|
||||
{ onProgress: (p) => events.push(p) })
|
||||
return { events, doc }
|
||||
} catch (e) {
|
||||
return { events, err: { isApi: e instanceof window.__ui.ApiError, status: e.status, kind: e.kind, message: e.message } }
|
||||
}
|
||||
}, size)
|
||||
// 上传进度只有真的走网络才有（page.route 直接 fulfill 时浏览器不发进度事件）。
|
||||
// 转给 dev server 上一个不存在的地址：字节真发出去，回 404，什么都不写
|
||||
await uploadVia((r) => r.continue({ url: `${WEB}/__upload_probe__` }))
|
||||
let up = await tryUpload(300_000)
|
||||
const lastUp = up.events.at(-1)
|
||||
check('字节发完报 sent，已发 = 总数，且不小于文件大小', !!lastUp?.sent && lastUp.loaded === lastUp.total && lastUp.total >= 300_000,
|
||||
JSON.stringify(up.events.slice(-2)))
|
||||
check('sent 只在最后', up.events.filter((p) => p.sent).length === 1 && !up.events.slice(0, -1).some((p) => p.sent))
|
||||
check('非 JSON 的错误响应：http 类的 ApiError', up.err?.isApi && up.err.kind === 'http' && up.err.status === 404, JSON.stringify(up.err))
|
||||
await uploadVia((r) => r.fulfill({
|
||||
status: 200, contentType: 'application/json',
|
||||
body: JSON.stringify({ id: 'doc_demo', collection: 'default', title: 'orders.csv', source: 'upload', mime: 'text/csv', chunk_count: 0, status: 'processing' }),
|
||||
}))
|
||||
up = await tryUpload(10)
|
||||
check('成功：返回值照常解析', up.doc?.id === 'doc_demo', JSON.stringify(up.err ?? up.doc))
|
||||
await uploadVia((r) => r.fulfill({
|
||||
status: 409, contentType: 'application/json', body: JSON.stringify({ detail: '同名文档已经在处理：等它处理完再传' }),
|
||||
}))
|
||||
up = await tryUpload(10)
|
||||
check('后端拒绝：和 fetch 那条路一样的 ApiError，message 是 detail', up.err?.isApi && up.err.status === 409
|
||||
&& up.err.kind === 'http' && up.err.message === '同名文档已经在处理：等它处理完再传', JSON.stringify(up.err))
|
||||
await uploadVia((r) => r.abort('connectionrefused'))
|
||||
up = await tryUpload(10)
|
||||
check('连不上：network 类的 ApiError', up.err?.isApi && up.err.kind === 'network' && up.err.status === 0, JSON.stringify(up.err))
|
||||
await page.unroute(isUpload)
|
||||
// 网络失败会顺手探一次活，等它落定再往下，免得下一段的探活计数多一次
|
||||
await page.evaluate(() => window.__ui.useCatalog.getState().checkBackend())
|
||||
|
||||
console.log('\n=== 连接状态：心跳、退避、恢复 ===')
|
||||
const expectRuns = await fetch(`${API}/runs?limit=3`).then((r) => r.json()).then((r) => r.length, () => null)
|
||||
await page.evaluate(() => window.__ui.useCatalog.getState().refresh())
|
||||
|
||||
Reference in New Issue
Block a user