agentworkshop 0.8.6 → 0.8.8

package/scripts/e2e-real-channel-mission.mjs

@@ -0,0 +1,543 @@
+#!/usr/bin/env node
+/**
+ * 真实 harness 全链路功能验收 —— Channel 任务中 Agent 真实使用五能力
+ *
+ * 真实模型:LM Studio qwen3.5-9b-64k(65536 ctx,lms ps 核实)经 omp harness(rpc 模式+运行级 overlay)。
+ * 场景:cast-film-physics 四协议产线,一条 aml_optimization 工艺优化频道(模板实例化),
+ * omp Agent 在任务中:数采真实抓数(R2) → HITL 挂起/批准/真实下发(R3) → 数控调优+PV 随动+账本归因(R4)。
+ * 另含:五能力工具面在册(R1)、AML 训练面+求解器门控(R5)、本体语义注入(R6)。
+ *
+ * 判据规范:docs/audit/2026-10-30-real-harness-acceptance-plan.md
+ * 运行:node scripts/e2e-real-channel-mission.mjs(自起隔离实例 3423 + 隔离模拟器 4038)
+ * 证据:docs/audit/evidence/2026-10-30-real-channel-mission/report.json
+ */
+import { spawn } from 'node:child_process'
+import { existsSync, mkdirSync, openSync, readFileSync, rmSync, writeFileSync } from 'node:fs'
+import { dirname, join, resolve } from 'node:path'
+import { pathToFileURL, fileURLToPath } from 'node:url'
+import { DatabaseSync } from 'node:sqlite'
+
+const REPO = resolve(dirname(fileURLToPath(import.meta.url)), '..')
+const HOME = process.env.AW_RHACC_HOME ?? 'D:/tmp/aw-rhacc-20261030'
+const PORT = Number(process.env.AW_RHACC_PORT ?? 3423)
+const BASE = `http://127.0.0.1:${PORT}`
+const SIM_PORT = Number(process.env.AW_RHACC_SIM_PORT ?? 4038)
+const SIM_BASE = `http://127.0.0.1:${SIM_PORT}`
+const SIM_SHADOW = process.env.AW_RHACC_SIM_SHADOW ?? 'D:/tmp/aw-rhacc-sim-shadow-4038'
+const EVIDENCE_DIR = join(REPO, 'docs', 'audit', 'evidence', '2026-10-30-real-channel-mission')
+const DEMO_MODEL = process.env.AW_DEMO_MODEL ?? 'qwen3.5-9b-64k'
+
+process.env.SIM_BASE = SIM_BASE
+process.env.SIM_PORT = String(SIM_PORT)
+process.env.SIM_SHADOW_DIR = SIM_SHADOW
+process.env.NO_PROXY = '127.0.0.1,localhost'
+process.env.no_proxy = '127.0.0.1,localhost'
+
+const results = []
+const sleep = ms => new Promise(r => setTimeout(r, ms))
+const ok = (id, name, cond, detail = '') => {
+  results.push({ id, name, pass: Boolean(cond), detail: String(detail).slice(0, 1800) })
+  console.log(`${cond ? '✔' : '✖'} [${id}] ${name}${cond ? '' : ` — ${String(detail).slice(0, 750)}`}`)
+  return Boolean(cond)
+}
+
+async function simApi(method, path, body) {
+  const res = await fetch(SIM_BASE + path, {
+    method,
+    headers: body !== undefined ? { 'content-type': 'application/json' } : {},
+    body: body === undefined ? undefined : JSON.stringify(body),
+    signal: AbortSignal.timeout(20_000),
+  })
+  const json = await res.json().catch(() => null)
+  return { status: res.status, data: json?.data ?? json, message: json?.message ?? '' }
+}
+const simNodes = async () => (await simApi('GET', '/api/nodes')).data ?? []
+const simSignal = (nodes, deviceId, signalId) => {
+  const sig = (nodes.find(n => n.id === deviceId)?.signals ?? []).find(s => s.id === signalId)
+  return sig?.value == null ? null : Number(sig.value)
+}
+
+function stopInstance(home) {
+  const pidFile = join(home, 'server.pid')
+  if (!existsSync(pidFile)) return
+  const pid = Number(readFileSync(pidFile, 'utf8').trim())
+  if (!Number.isInteger(pid) || pid <= 0) return
+  try {
+    spawn('taskkill', ['/PID', String(pid), '/T', '/F'], { windowsHide: true })
+  }
+  catch { /* 已退出 */ }
+}
+async function health() {
+  try {
+    const r = await fetch(`${BASE}/api/health`, { signal: AbortSignal.timeout(2500) })
+    if (!r.ok) return null
+    return (await r.json())?.data ?? null
+  }
+  catch { return null }
+}
+async function startInstance() {
+  stopInstance(HOME)
+  await sleep(1500)
+  if (await health()) throw new Error(`端口 ${PORT} 已被非本验收实例占用`)
+  if (!/aw-rhacc/.test(HOME)) throw new Error(`拒绝清理非验收目录:${HOME}`)
+  rmSync(HOME, { recursive: true, force: true })
+  mkdirSync(HOME, { recursive: true })
+  const log = openSync(join(HOME, 'server.log'), 'a')
+  const child = spawn(process.execPath, [join(REPO, '.output', 'server', 'index.mjs')], {
+    cwd: REPO,
+    env: {
+      ...process.env,
+      AW_MODE: 'home',
+      AW_HOME: HOME,
+      PORT: String(PORT),
+      NUXT_SESSION_PASSWORD: 'rhacc-20261030-session',
+      AW_TRIAL_INTERVAL_MS: '0',
+      AW_RATE_LIMIT_OFF: '1',
+      NO_PROXY: '127.0.0.1,localhost',
+    },
+    detached: true,
+    stdio: ['ignore', log, log],
+    windowsHide: true,
+  })
+  child.unref()
+  writeFileSync(join(HOME, 'server.pid'), String(child.pid))
+  for (let i = 0; i < 120; i++) {
+    await sleep(1000)
+    const h = await health()
+    if (h?.workshop === 'ready') return h
+  }
+  throw new Error('隔离实例未在 120s 内就绪')
+}
+
+let admin = ''
+async function apiOnce(method, path, body, token = admin, extraHeaders = {}, timeoutMs = 120_000) {
+  const res = await fetch(BASE + path, {
+    method,
+    headers: {
+      ...(body !== undefined ? { 'content-type': 'application/json' } : {}),
+      ...(token ? { authorization: `Bearer ${token}` } : {}),
+      ...extraHeaders,
+    },
+    body: body === undefined ? undefined : JSON.stringify(body),
+    signal: AbortSignal.timeout(timeoutMs),
+  })
+  let json = null
+  try {
+    json = await res.json()
+  }
+  catch { /* 非 JSON */ }
+  return { status: res.status, ...json }
+}
+async function api(method, path, body, token = admin, extraHeaders = {}, timeoutMs = 120_000) {
+  try {
+    return await apiOnce(method, path, body, token, extraHeaders, timeoutMs)
+  }
+  catch (err) {
+    if (err instanceof TypeError) {
+      await sleep(1500)
+      return apiOnce(method, path, body, token, extraHeaders, timeoutMs)
+    }
+    throw err
+  }
+}
+function must(resp, what) {
+  if (resp.status !== 200 || !resp?.data) throw new Error(`${what} 失败:${resp.status} ${JSON.stringify(resp).slice(0, 500)}`)
+  return resp.data
+}
+function agentTokenOf(agentId) {
+  const db = new DatabaseSync(join(HOME, 'data', 'workshop.sqlite'), { readOnly: true })
+  try {
+    return String(db.prepare('SELECT token FROM channel_agents WHERE id = ?').get(agentId)?.token ?? '')
+  }
+  finally {
+    db.close()
+  }
+}
+async function pendingOf(agentId) {
+  const r = await api('GET', `/api/workshop/agent-tools/approvals?agentId=${encodeURIComponent(agentId)}`)
+  return (r?.data?.approvals ?? []).filter(a => a.status === 'pending')
+}
+/** 以 agent 令牌直调工具(确定性路径);出现 HITL 卡自动批准(验收模拟人工) */
+async function invokeAs(agentId, token, tool, args, { timeoutMs = 180_000 } = {}) {
+  const cards = []
+  let resp = null
+  let err = null
+  const started = Date.now()
+  const p = api('POST', '/api/workshop/agent-tools/invoke', { agentId, tool, args }, undefined, { 'x-aw-agent-token': token }, timeoutMs)
+    .then((r) => {
+      resp = r
+    })
+    .catch((e) => {
+      err = e
+    })
+  while (!resp && !err && Date.now() - started < timeoutMs - 10_000) {
+    for (const card of await pendingOf(agentId)) {
+      if (cards.includes(card.id)) continue
+      cards.push(card.id)
+      await api('POST', `/api/workshop/agent-tools/approvals/${card.id}/decide`, { approved: true, comment: '验收:直调路径' })
+    }
+    await sleep(1000)
+  }
+  await p
+  const text = String(resp?.data?.result?.text ?? resp?.data?.text ?? '')
+  const isError = Boolean(resp?.data?.result?.isError ?? resp?.data?.isError) || Boolean(err)
+  return { resp, text, isError, cards }
+}
+
+const TERMINAL = new Set(['COMPLETED', 'FAILED', 'CANCELED', 'CANCELLED'])
+const taskOf = r => r?.data?.task ?? r?.data ?? null
+const deliverableOf = task => String((task?.artifacts ?? []).find(a => a.name === 'deliverable')?.parts?.map(p => p.text ?? '').join('\n') ?? '')
+const toolCallsOf = task => (task?.history ?? [])
+  .map(h => String(h?.parts?.map(p => p.text ?? '').join(' ') ?? ''))
+  .filter(t => t.includes('🔧'))
+  .map(t => t.replace(/\s+/g, ' ').slice(0, 300))
+/**
+ * 轮询任务到终态。onCard:新 HITL 卡回调(检查点),随后自动批准(验收模拟人工)。
+ * evidenceFn 命中后经 graceMs 宽限仍无终态且无挂起卡 → 收口 cancel(证据落定即止,不等慢模型)。
+ */
+async function waitTask(taskId, agentId, { timeoutMs = 8 * 60_000, evidenceFn = null, graceMs = 60_000, onCard = null } = {}) {
+  const started = Date.now()
+  let evidenceAt = 0
+  const seen = new Set()
+  while (Date.now() - started < timeoutMs) {
+    const task = taskOf(await api('GET', `/api/workshop/tasks/${taskId}`))
+    for (const card of await pendingOf(agentId)) {
+      if (seen.has(card.id)) continue
+      seen.add(card.id)
+      if (onCard) await onCard(card)
+      await api('POST', `/api/workshop/agent-tools/approvals/${card.id}/decide`, { approved: true, comment: '验收:任务路径' })
+    }
+    if (TERMINAL.has(String(task?.state ?? ''))) return { task, cancelled: false, cards: [...seen] }
+    if (evidenceFn && evidenceFn(task, [...seen]) && !evidenceAt) evidenceAt = Date.now()
+    if (evidenceAt && Date.now() - evidenceAt > graceMs) {
+      if (!(await pendingOf(agentId)).length) {
+        await api('POST', `/api/workshop/tasks/${taskId}/cancel`, { reason: '验收:证据收口' })
+        return { task: taskOf(await api('GET', `/api/workshop/tasks/${taskId}`)), cancelled: true, cards: [...seen] }
+      }
+      evidenceAt = Date.now() // 有卡挂着:重置宽限,继续等审批与任务前进
+    }
+    await sleep(2000)
+  }
+  return { task: taskOf(await api('GET', `/api/workshop/tasks/${taskId}`)), cancelled: false, cards: [...seen] }
+}
+const latestSample = async (nodeId, seconds = 30) => {
+  const to = Date.now()
+  const r = await api('GET', `/api/workshop/daq/${nodeId}/samples?from=${to - seconds * 1000}&to=${to}&bucketMs=1000&limit=200`)
+  const pts = (r?.data?.points ?? []).map(p => Number(p.value ?? p.avg)).filter(Number.isFinite)
+  return pts.length ? pts[pts.length - 1] : null
+}
+
+const evidence = { at: new Date().toISOString(), model: DEMO_MODEL, scenario: {}, assertions: results }
+let run3
+let run4
+let calls4
+let deliv4
+let deliv2
+let spBase4
+let injText
+let ctxText
+let capsText
+let solveText
+
+try {
+  // ═══ R0 · 隔离环境 ═══
+  const sim = await import(pathToFileURL(join(REPO, 'bench', 'lib', 'sim.mjs')).href)
+  const boot = await sim.ensureSimulator({ log: m => console.log(`    ${m}`) })
+  ok('R0.1', '隔离模拟器在线(4038)', boot.started || /already-up/.test(String(boot.reason)), JSON.stringify(boot))
+  await sim.applyPreset('cast-film-physics')
+  const nodes0 = await simNodes()
+  ok('R0.2', '四协议设备端点在线(cast-film-physics)', nodes0.filter(n => n.runtime?.running).length >= 5,
+    nodes0.map(n => `${n.id}:${n.protocol}`).join(', '))
+
+  const exportItems = async id => must(await simApi('GET', `/api/nodes/${id}/export`), `导出 ${id}`).items ?? []
+  const bySignal = async (id, signal) => {
+    const row = (await exportItems(id)).find(i => i.signal === signal)
+    if (!row) throw new Error(`导出缺少 ${id}/${signal}`)
+    return row.driverConfig
+  }
+  const cfgZone = await bySignal('dev-extruder-mbtcp', '加热区1SP')
+  const cfgTemp = await bySignal('dev-extruder-mbtcp', '熔体温度')
+
+  const h = await startInstance()
+  ok('R0.3', '生产构建隔离实例就绪(3423,全新 home)', h?.workshop === 'ready', `port=${PORT} version=${h?.version}`)
+  // 隔离新库:needsSetup=true 时首注册即 admin(setup 模式);已有库则直接 login
+  const setup = await api('GET', '/api/users/setup-status')
+  if (setup.data?.needsSetup) {
+    admin = (await api('POST', '/api/users/register', { name: 'rhacc-admin', email: 'admin@awshop.local', password: 'admin123' })).data?.token
+  }
+  else {
+    admin = (await api('POST', '/api/users/login', { email: 'admin@awshop.local', password: 'admin123' })).data?.token
+  }
+  ok('R0.4', 'admin 鉴权(隔离新库首注册即 admin)', Boolean(admin))
+
+  // ═══ 供给:产线 + DAQ + 一根 Modbus 数控节点 + 配方 + 开线 ═══
+  const line = must(await api('POST', '/api/workshop/dcw/lines', { name: '真实harness验收线', description: '五能力 Channel 任务验收' }), '建线').line?.id
+  const daq = must(await api('POST', '/api/workshop/daq', {
+    templateRef: 'daq-temp-tc', name: '熔体温度', driver: 'modbus-tcp', driverConfig: cfgTemp,
+    lineId: line, intervalMs: 1000, publishIntervalMs: 0, min: 0, max: 400, decimals: 2,
+  }), '建 DAQ').node?.id
+  const zone = must(await api('POST', '/api/workshop/dcw', {
+    templateRef: 'temp-sp', name: '加热区1设定', driver: 'modbus-tcp', driverConfig: cfgZone, lineId: line,
+    unit: 'C', min: 120, max: 260, decimals: 2, stepLimit: 5, writeLockSeconds: 0, readIntervalMs: 2000,
+  }), '建 zone').node?.id
+  const product = must(await api('POST', '/api/workshop/dcw/products', { name: '验收产品', lineId: line, paramLimits: {} }), '建产品').product
+  const recipe = must(await api('POST', '/api/workshop/dcw/recipes', {
+    name: '验收基线配方', productId: product.id, opIntervalMs: 0,
+    params: [{ nodeId: zone, value: 195, min: 120, max: 260, stepLimit: 5 }],
+  }), '建配方').recipe
+  await api('POST', '/api/workshop/daq/controller', { action: 'start' })
+  const started = await api('POST', `/api/workshop/dcw/lines/${line}/start`, { recipeId: recipe.id })
+  await api('PATCH', `/api/workshop/dcw/lines/${line}`, { controlMode: 'auto', confirm: true })
+  await sleep(12_000)
+  const devNow = await simNodes()
+  const daqNow = await latestSample(daq, 20)
+  ok('R0.5', '开线且基线落设备(modbus SP + DAQ 首采)', started.status === 200
+  && Number.isFinite(simSignal(devNow, 'dev-extruder-mbtcp', 'zone1-sp')) && daqNow != null,
+  `SP=${simSignal(devNow, 'dev-extruder-mbtcp', 'zone1-sp')} daqLast=${daqNow}`)
+
+  // ═══ R1 频道 + omp 真模型 Agent + 五能力工具面 ═══
+  const overlay = join(HOME, 'omp-overlay.yml')
+  writeFileSync(overlay, [
+    'modelRoles:',
+    ...['default', 'task', 'advisor', 'slow', 'designer'].map(r => `  ${r}: lm-studio/${DEMO_MODEL}`),
+    'tools:',
+    '  format: qwen3',
+    'compaction:',
+    '  enabled: true',
+    '',
+  ].join('\n'))
+  const tpl = await api('POST', '/api/workshop/channel-templates/chtpl-aml-optimization-default/instantiate', {
+    name: '真实harness-工艺优化', toolProfile: 'aml_optimization', optimizationMode: 'exploration',
+  })
+  const optCh = tpl.data ?? {}
+  const channelId = optCh?.channelId
+  ok('R1.1', 'aml_optimization 工艺优化频道(模板实例化)', tpl.status === 200 && Boolean(channelId), `ch=${channelId} lead=${optCh?.leadAgentId ?? '—'}`)
+  const omp = must(await api('POST', '/api/workshop/agents', {
+    name: 'omp-真实模型操作员', harness: 'omp',
+    config: { rpcMode: 'rpc', args: ['--no-tools', '--thinking=off', '--config', overlay] },
+  }), '建 omp Agent')
+  const ompTplId = omp.id ?? omp.data?.id // must() 已解包 data:omp 即模板对象
+  if (!ompTplId) throw new Error(`omp 模板创建回包缺 id:${JSON.stringify(omp).slice(0, 200)}`)
+  const joinResp = await api('POST', `/api/workshop/channels/${channelId}/agents`, { agentId: ompTplId, role: 'worker' })
+  let agentId = joinResp.data?.id ?? joinResp.data?.agentId ?? joinResp.data?.member?.id ?? joinResp.data?.agent?.id
+  if (!agentId) {
+    // 回包形状兜底:按 channel+名字直查实例表
+    const db = new DatabaseSync(join(HOME, 'data', 'workshop.sqlite'), { readOnly: true })
+    try {
+      agentId = String(db.prepare('SELECT id FROM channel_agents WHERE channel_id = ? AND name = ? ORDER BY created_at DESC LIMIT 1').get(channelId, 'omp-真实模型操作员')?.id ?? '')
+    }
+    finally { db.close() }
+  }
+  if (!agentId) {
+    const j = (v) => {
+      try {
+        return JSON.stringify(v) ?? String(v)
+      }
+      catch { return '(unserializable)' }
+    }
+    throw new Error(`omp 入队失败:join=${joinResp.status}/${j(joinResp).slice(0, 400)} ompKeys=${omp.data ? Object.keys(omp.data).join(',') : 'data-undefined'} omp=${j(omp).slice(0, 300)}`)
+  }
+  await api('PATCH', `/api/workshop/channels/${channelId}`, { llm: { provider: 'lm-studio', model: DEMO_MODEL, contextTokens: 65536 } })
+  const ompToken = agentTokenOf(agentId)
+  ok('R1.2', 'omp Agent(rpc+overlay,lm-studio 64k)入队且令牌可读', Boolean(agentId && ompToken), `agent=${agentId} token=${ompToken ? '***' + ompToken.slice(-4) : 'MISSING'}`)
+
+  const bindDcw = await api('POST', '/api/workshop/agent-tools/bindings', { agentId, nodeId: zone, kind: 'dcw', mode: 'manual' })
+  const bDaq = await api('POST', '/api/workshop/agent-tools/bindings', { agentId, nodeId: daq, kind: 'daq', mode: 'auto' })
+  if (bDaq?.data?.binding?.mode === 'manual') await api('PATCH', `/api/workshop/agent-tools/bindings/${bDaq.data.binding.id}`, { mode: 'auto', confirm: true })
+  const bindingMode = bindDcw?.data?.binding?.mode
+  ok('R1.3', '绑定面建立:dcw=manual(HITL 缺省)+daq=auto', bindDcw.status === 200 && bDaq.status === 200 && bindingMode === 'manual', `dcw mode=${bindingMode} daq=${bDaq.status}`)
+
+  // 9B 模型 + 64k 上下文:开放面(81 工具)的 schema 注入 ≈78k tokens,直接撑爆上下文
+  // (实测 OMP_LLM_ERROR n_keep:77925>=n_ctx:65536)。收窄到本验收所需的最小面。
+  await api('PUT', `/api/workshop/channels/${channelId}/tools`, {
+    mode: 'selected',
+    tools: ['my_industrial_nodes', 'line_context', 'daq_query', 'daq_frames', 'optimization_explore', 'dcw_judge', 'dcw_journal', 'optimizer_capabilities', 'complete_task', 'report_progress']
+      .map(name => ({ name, enabled: true })),
+  })
+  const toolList = (await api('GET', `/api/workshop/agent-tools/list?agentId=${encodeURIComponent(agentId)}`)).data?.tools ?? []
+  const toolNames = toolList.map(t => t.name)
+  const needFace = ['my_industrial_nodes', 'daq_query', 'daq_frames', 'optimization_explore', 'dcw_journal', 'dcw_judge', 'optimizer_capabilities']
+  ok('R1.4', '五能力工具面在册且已收窄(数采/数控治理/求解器)', needFace.every(n => toolNames.includes(n)),
+    `missing=${needFace.filter(n => !toolNames.includes(n)).join(',') || '无'}; face=${toolNames.length} 项`)
+
+  // ═══ R2 · 数采真实抓数(LLM 任务) ═══
+  const t2 = taskOf(await api('POST', `/api/workshop/channels/${channelId}/tasks`, {
+    title: 'R2-真实抓数',
+    description: '严格按两步执行,不要做别的:\n'
+      + `1) 调用 daq_query:node_id="${daq}",last_minutes=3,bucket_ms=1000,读出样本数与最新值;\n`
+      + '2) 立即调用 complete_task,汇报样本数与最新温度数值。\n'
+      + '只允许 daq_query 与 complete_task 两个工具;不要等待/睡眠。',
+    assigneeId: agentId,
+    parts: [{ text: `daq_query(node_id="${daq}", last_minutes=3, bucket_ms=1000) 后 complete_task 汇报数值` }],
+  }))
+  if (!t2?.id) throw new Error(`R2 任务创建失败:${JSON.stringify(t2).slice(0, 200)}`)
+  const run2 = await waitTask(t2.id, agentId, {
+    timeoutMs: 8 * 60_000,
+    evidenceFn: t => toolCallsOf(t).some(x => x.includes('daq_query')),
+    graceMs: 50_000,
+  })
+  const calls2 = toolCallsOf(run2.task)
+  const blob2 = deliverableOf(run2.task) + '\n' + (run2.task?.history ?? []).map(h2 => (h2?.parts ?? []).map(p2 => p2?.text ?? '').join(' ')).join('\n')
+  deliv2 = blob2
+  const truth2 = simSignal(await simNodes(), 'dev-extruder-mbtcp', 'melt-temp')
+  const nums2 = (blob2.match(/-?\d+(\.\d+)?/g) ?? []).map(Number).filter(v => v > 20 && v < 400)
+  const near2 = nums2.some(v => truth2 != null && Math.abs(v - truth2) <= 3)
+  ok('R2.1', 'LLM 任务中 daq_query 真实抓数(数值=设备真值±3℃)',
+    calls2.some(x => x.includes('daq_query')) && near2 && (run2.task?.state === 'COMPLETED' || run2.cancelled),
+    `state=${run2.task?.state} truth=${truth2} 交付数值=[${nums2.slice(0, 8).join(',')}]`)
+
+  // ═══ R3 · HITL 真实审批(批准前后设备双时点对照) ═══
+  const spBefore = simSignal(await simNodes(), 'dev-extruder-mbtcp', 'zone1-sp')
+  const step3 = 3
+  const t3 = taskOf(await api('POST', `/api/workshop/channels/${channelId}/tasks`, {
+    title: 'R3-HITL下发生效',
+    description: '严格按两步执行,第一步必须先做:\n'
+      + `1) 你的第一个动作就是调用 optimization_explore,参数:control_node_id="${zone}",target_node_id="${daq}",direction="up",step=${step3},settle_seconds=10,hypothesis="HITL 验收:上调 3C"。不要先做任何其他事;\n`
+      + '2) 该调用返回结果后(可能等待人工批准,耐心等它返回),调用 complete_task 汇报探索记录号与回读值。\n'
+      + '整个任务只允许这两个工具:optimization_explore、complete_task。禁止调用 search_memory、save_memory、daq_query 或任何其他工具;不要复述场景简报。',
+    assigneeId: agentId,
+    parts: [{ text: `第一个动作:optimization_explore(control_node_id="${zone}", target_node_id="${daq}", direction="up", step=${step3}, settle_seconds=10),返回后 complete_task` }],
+  }))
+  if (!t3?.id) throw new Error(`R3 任务创建失败:${JSON.stringify(t3).slice(0, 200)}`)
+  const checkpoint = { cardId: null, cardTool: '', spAtCard: null }
+  run3 = await waitTask(t3.id, agentId, {
+    timeoutMs: 10 * 60_000,
+    onCard: async (card) => {
+      checkpoint.cardId = card.id
+      checkpoint.cardTool = String(card.tool ?? card.toolName ?? '')
+      await sleep(3000) // 卡已挂起:给"若误直写"一个暴露窗,再取批准前时点
+      checkpoint.spAtCard = simSignal(await simNodes(), 'dev-extruder-mbtcp', 'zone1-sp')
+      console.log(`    · HITL 卡 ${card.id}(${checkpoint.cardTool})挂起,批准前 SP=${checkpoint.spAtCard}(基线 ${spBefore})`)
+    },
+  })
+  const unchangedBeforeApproval = checkpoint.spAtCard != null && Math.abs(checkpoint.spAtCard - spBefore) <= 0.05
+  ok('R3.1', 'HITL 挂起出现且批准前设备值不变(写被真实拦截)',
+    checkpoint.cardId && unchangedBeforeApproval,
+    `card=${checkpoint.cardId ?? '未出现'} tool=${checkpoint.cardTool} 批准前SP=${checkpoint.spAtCard} 基线=${spBefore}`)
+  if (checkpoint.cardId) {
+    let spAfter3 = null
+    const t03 = Date.now()
+    while (Date.now() - t03 < 60_000) {
+      await sleep(4000)
+      spAfter3 = simSignal(await simNodes(), 'dev-extruder-mbtcp', 'zone1-sp')
+      if (spAfter3 != null && Math.abs(spAfter3 - (spBefore + step3)) <= 0.5) break
+    }
+    ok('R3.2', '批准后真实下发:设备寄存器值=基线+3(±0.5)',
+      spAfter3 != null && Math.abs(spAfter3 - (spBefore + step3)) <= 0.5,
+      `批准后SP=${spAfter3} 期望=${spBefore + step3}`)
+    ok('R3.3', 'LLM 任务收口(COMPLETED/收口 cancel)', run3.task?.state === 'COMPLETED' || run3.cancelled,
+      `state=${run3.task?.state} calls=${toolCallsOf(run3.task).length}`)
+  }
+
+  // ═══ R4 · 数控真实下发 + 任务调优(auto 绑定,LLM 单步调优;9B 非确定性 → 至多两次派发) ═══
+  spBase4 = simSignal(await simNodes(), 'dev-extruder-mbtcp', 'zone1-sp')
+  await api('PATCH', `/api/workshop/agent-tools/bindings/${bindDcw.data.binding.id}`, { mode: 'auto', confirm: true })
+  const step4 = 2.5
+  const recsBefore4 = ((await api('GET', '/api/workshop/dcw/optimizations?limit=50')).data?.records ?? []).filter(r => r.nodeId === zone).length
+  const mkTask4 = () => api('POST', `/api/workshop/channels/${channelId}/tasks`, {
+    title: `R4-调优任务(尝试${Date.now() % 1000})`,
+    description: '你是工艺操作员,严格按四步执行,不要做别的:\n'
+      + `1) 调用 daq_query,只传三个参数:node_id="${daq}",last_minutes=2,bucket_ms=1000(不要传 recipe_id 或任何其他参数);\n`
+      + `2) 本任务的核心步骤:调用 optimization_explore,参数:control_node_id="${zone}",target_node_id="${daq}",direction="up",step=${step4},settle_seconds=10,hypothesis="调优:按计划上调 ${step4}C";\n`
+      + `3) 再调用一次 daq_query(node_id="${daq}",last_minutes=1,bucket_ms=1000) 观察响应;\n`
+      + '4) 调用 complete_task,汇报调优前后两个温度数值。\n'
+      + '只允许 daq_query、optimization_explore、complete_task 三个工具;optimization_explore 必须执行且只调用一次;禁止 search_memory/save_memory;不要等待/睡眠。',
+    assigneeId: agentId,
+    parts: [{ text: `调优:${zone} 上调 ${step4}C,daq="${daq}",完成后 complete_task` }],
+  })
+  let r4done = false
+  let spAfter4 = null
+  for (let attempt = 1; attempt <= 2 && !r4done; attempt++) {
+    const t4 = taskOf(await mkTask4())
+    if (!t4?.id) throw new Error(`R4 任务创建失败:${JSON.stringify(t4).slice(0, 200)}`)
+    run4 = await waitTask(t4.id, agentId, {
+      timeoutMs: 15 * 60_000,
+      evidenceFn: t => toolCallsOf(t).some(x => x.includes('optimization_explore')),
+      graceMs: 90_000,
+    })
+    calls4 = toolCallsOf(run4.task)
+    const t04 = Date.now()
+    while (Date.now() - t04 < 60_000) {
+      await sleep(4000)
+      spAfter4 = simSignal(await simNodes(), 'dev-extruder-mbtcp', 'zone1-sp')
+      if (spAfter4 != null && Math.abs(spAfter4 - (spBase4 + step4)) <= 0.75) break
+    }
+    r4done = spAfter4 != null && Math.abs(spAfter4 - (spBase4 + step4)) <= 0.75
+    if (!r4done && attempt === 1) {
+      spBase4 = simSignal(await simNodes(), 'dev-extruder-mbtcp', 'zone1-sp') ?? spBase4 // 二次尝试以当前 SP 为基线
+      console.log(`    · R4 第一次派发未落点(SP=${spAfter4}),重发任务(9B 非确定性容差)`)
+    }
+  }
+  ok('R4.1', '调优真实下发:设备 SP=基线+2.5(±0.75)',
+    spAfter4 != null && Math.abs(spAfter4 - (spBase4 + step4)) <= 0.75,
+    `SP=${spAfter4} 期望=${spBase4 + step4} state=${run4.task?.state} calls=${calls4.map(c => c.slice(0, 60)).join(' | ').slice(0, 260)}`)
+  await sleep(8000)
+  const pv4 = await latestSample(daq, 25)
+  ok('R4.2', 'PV 随动:熔体温度样本向新 SP 收敛(±4℃)', pv4 != null && spAfter4 != null && Math.abs(pv4 - spAfter4) <= 4,
+    `pv=${pv4} sp=${spAfter4}`)
+  const jr = await invokeAs(agentId, ompToken, 'dcw_journal', { node_id: zone, limit: 5 })
+  ok('R4.3', '参数账本 Agent 归因(dcw_journal 含 Agent 来源)', !jr.isError && /Agent|agent|优化|探索/.test(jr.text),
+    jr.text.replace(/\s+/g, ' ').slice(0, 300))
+  const optRecs4 = ((await api('GET', '/api/workshop/dcw/optimizations?limit=50')).data?.records ?? []).filter(r => r.nodeId === zone)
+  ok('R4.4', '优化记录链存在(本阶段治理写新开窗)', optRecs4.length > recsBefore4,
+    `R4 前本节点记录=${recsBefore4} · R4 后=${optRecs4.length}`)
+  ok('R4.5', 'LLM 调优任务收口(COMPLETED/收口 cancel)', run4.task?.state === 'COMPLETED' || run4.cancelled,
+    `state=${run4.task?.state}`)
+  deliv4 = deliverableOf(run4.task) + '\n' + (run4.task?.history ?? []).map(h4 => (h4?.parts ?? []).map(p4 => p4?.text ?? '').join(' ')).join('\n')
+  ok('R4.6', '交付文本含调优前后数值(≥2 个温度读数)', (deliv4.match(/\d+(\.\d+)?/g) ?? []).length >= 2,
+    deliv4.replace(/\s+/g, ' ').slice(0, 300))
+
+  // ═══ R5 · AML 训练面 + 求解器门控 ═══
+  const trainTpl = await api('POST', '/api/workshop/channel-templates/chtpl-aml-training-default/instantiate', { name: '真实harness-AML训练' })
+  const trainChId = trainTpl.data?.channelId
+  const trainLead = trainTpl.data?.leadAgentId ?? (trainTpl.data?.agents ?? []).find(a => a.role === 'lead')?.id
+  if (trainLead) await api('POST', '/api/workshop/agent-tools/bindings', { agentId: trainLead, nodeId: daq, kind: 'daq', mode: 'auto' })
+  const amf = trainLead ? await invokeAs(trainLead, agentTokenOf(trainLead), 'aml_model_find', {}) : { isError: true, text: '训练频道未就绪' }
+  ok('R5.1', 'AML 训练面:aml_model_find 可查询(模型清单可为空)', Boolean(trainChId) && !amf.isError,
+    amf.text.replace(/\s+/g, ' ').slice(0, 240))
+  const caps = await invokeAs(agentId, ompToken, 'optimizer_capabilities', {})
+  capsText = caps.text
+  ok('R5.2', '求解器:optimizer_capabilities 在优化频道可用', !caps.isError && caps.text.length > 20,
+    caps.text.replace(/\s+/g, ' ').slice(0, 240))
+  const solve = await invokeAs(agentId, ompToken, 'optimizer_solve', { objective: { target: 200 }, hypothesis: '验收:资格门负例' })
+  solveText = solve.text
+  ok('R5.3', '求解器:optimizer_solve 无合格 Profile 时被拒绝(负向断言,不放行执行)', solve.isError && solve.text.length > 0,
+    solve.text.replace(/\s+/g, ' ').slice(0, 240))
+
+  // ═══ R6 · 本体语义注入(节点工艺描述可达 Agent 面) ═══
+  const inj = await invokeAs(agentId, ompToken, 'my_industrial_nodes', {})
+  injText = inj.text
+  ok('R6.1', '本体语义:my_industrial_nodes 含工艺描述与量纲', !inj.isError && /加热区1设定/.test(inj.text) && /C/.test(inj.text),
+    inj.text.replace(/\s+/g, ' ').slice(0, 300))
+  const ctx = await invokeAs(agentId, ompToken, 'line_context', { line_id: line })
+  ctxText = ctx.text
+  ok('R6.2', 'line_context 产线语义可读(绑定谱系扩权)', !ctx.isError && String(ctx.text).length > 40,
+    String(ctx.text).replace(/\s+/g, ' ').slice(0, 300))
+
+  evidence.scenario = {
+    line, daq, zone, channelId, agentId, trainChId,
+    spTrace: { r3: { before: spBefore, atCard: checkpoint.spAtCard }, r4: { before: spBase4 } },
+  }
+  evidence.texts = {
+    r2: deliv2.slice(0, 900),
+    r3: toolCallsOf(run3.task ?? {}).join(' | ').slice(0, 900),
+    r4: (deliv4 || calls4.join(' | ')).slice(0, 900),
+    r5: { caps: capsText.slice(0, 400), solve: solveText.slice(0, 400) },
+    r6: { nodes: injText.slice(0, 600), line: ctxText.slice(0, 400) },
+  }
+}
+catch (err) {
+  ok('FATAL', '验收未捕获异常', false, err instanceof Error ? `${err.message}\n${err.stack ?? ''}` : String(err))
+}
+finally {
+  stopInstance(HOME)
+  mkdirSync(EVIDENCE_DIR, { recursive: true })
+  const passed = results.filter(r => r.pass).length
+  const total = results.length
+  evidence.summary = { total, passed, failed: total - passed, verdict: passed === total ? 'PASS' : 'FAIL' }
+  writeFileSync(join(EVIDENCE_DIR, 'report.json'), JSON.stringify(evidence, null, 2))
+  console.log(`\n结论:${passed === total ? 'PASS' : 'FAIL'}(${passed}/${total})`)
+  console.log(`证据:${join(EVIDENCE_DIR, 'report.json')}`)
+  console.log(`现场:平台 ${BASE}(AW_HOME=${HOME})/ 模拟器 ${SIM_BASE}`)
+  process.exit(passed === total ? 0 : 1)
+}