agentworkshop 0.7.21 → 0.7.23

package/server/data/line-runs.json

@@ -7,7 +7,7 @@
     "productId": "pd-6348a595",
     "productName": "0.8mm 光学膜",
     "startedAt": "2026-09-03T11:18:01.213Z",
-    "taggedSamples": 62326
+    "taggedSamples": 62397
   },
   {
     "lineId": "ln-9a60a0cd",
@@ -17,7 +17,7 @@
     "productId": "pd-e712875d",
     "productName": "闭环验证产品",
     "startedAt": "2026-09-03T16:03:27.304Z",
-    "taggedSamples": 1192
+    "taggedSamples": 1199
   },
   {
     "lineId": "ln-b33a3105",
@@ -27,7 +27,7 @@
     "productId": "pd-16eecb35",
     "productName": "闭环验证产品",
     "startedAt": "2026-09-03T16:39:22.962Z",
-    "taggedSamples": 1192
+    "taggedSamples": 1199
   },
   {
     "lineId": "ln-1aedc305",
@@ -37,7 +37,7 @@
     "productId": "pd-9c755781",
     "productName": "闭环验证产品",
     "startedAt": "2026-09-03T16:44:33.199Z",
-    "taggedSamples": 1192
+    "taggedSamples": 1199
   },
   {
     "lineId": "ln-758658dc",
@@ -47,7 +47,7 @@
     "productId": "pd-835aed90",
     "productName": "闭环验证产品",
     "startedAt": "2026-09-03T17:12:52.205Z",
-    "taggedSamples": 1192
+    "taggedSamples": 1199
   },
   {
     "lineId": "ln-1b318320",
@@ -57,7 +57,7 @@
     "productId": "pd-db432902",
     "productName": "探针产品",
     "startedAt": "2026-09-04T05:08:10.103Z",
-    "taggedSamples": 23141
+    "taggedSamples": 23170
   },
   {
     "lineId": "ln-6a34fbc7",
@@ -67,7 +67,7 @@
     "productId": "pd-10cfc2b3",
     "productName": "全协议产品-mtnynfph",
     "startedAt": "2026-09-05T05:47:26.453Z",
-    "taggedSamples": 82639
+    "taggedSamples": 82806
   },
   {
     "lineId": "ln-419463e0",
@@ -77,7 +77,7 @@
     "productId": "pd-b9b15b64",
     "productName": "全协议产品-mtnzg090",
     "startedAt": "2026-09-05T06:09:36.931Z",
-    "taggedSamples": 82586
+    "taggedSamples": 82751
   },
   {
     "lineId": "ln-3548f325",
@@ -87,7 +87,7 @@
     "productId": "pd-fd822802",
     "productName": "runp3-prod",
     "startedAt": "2026-09-05T07:23:51.972Z",
-    "taggedSamples": 9538
+    "taggedSamples": 9563
   },
   {
     "lineId": "ln-c3a7ba01",
@@ -107,7 +107,7 @@
     "productId": "pd-d547a376",
     "productName": "pp-prod",
     "startedAt": "2026-09-05T09:47:55.526Z",
-    "taggedSamples": 10851
+    "taggedSamples": 10880
   },
   {
     "lineId": "ln-1acc1a0c",
@@ -117,7 +117,7 @@
     "productId": "pd-3adf854f",
     "productName": "lx8-prod",
     "startedAt": "2026-09-06T02:22:23.209+08:00",
-    "taggedSamples": 10082
+    "taggedSamples": 10111
   },
   {
     "lineId": "ln-917490c1",
@@ -127,7 +127,7 @@
     "productId": "pd-75752beb",
     "productName": "lx8-prod",
     "startedAt": "2026-09-06T02:24:50.582+08:00",
-    "taggedSamples": 9941
+    "taggedSamples": 9970
   },
   {
     "lineId": "ln-24dedf82",
@@ -137,7 +137,7 @@
     "productId": "pd-f0436841",
     "productName": "模拟器产品-mtpsdote",
     "startedAt": "2026-09-06T20:27:23.270+08:00",
-    "taggedSamples": 22908
+    "taggedSamples": 23004
   },
   {
     "lineId": "ln-9d5f8b3d",
@@ -147,7 +147,7 @@
     "productId": "pd-8f6910a4",
     "productName": "模拟器产品-mtpt1b1w",
     "startedAt": "2026-09-06T20:45:45.440+08:00",
-    "taggedSamples": 23787
+    "taggedSamples": 23884
   },
   {
     "lineId": "ln-62776cae",
@@ -157,7 +157,7 @@
     "productId": "pd-e2b76cd7",
     "productName": "模拟器产品-mtptla26",
     "startedAt": "2026-09-06T21:01:14.334+08:00",
-    "taggedSamples": 20215
+    "taggedSamples": 20313
   },
   {
     "lineId": "ln-34c7c3c4",
@@ -167,7 +167,7 @@
     "productId": "pd-e883e141",
     "productName": "模拟器产品-mtpy8ms3",
     "startedAt": "2026-09-06T23:11:26.035+08:00",
-    "taggedSamples": 25255
+    "taggedSamples": 25382
   },
   {
     "lineId": "ln-b6fb7977",
@@ -197,7 +197,7 @@
     "productId": "pd-616de77a",
     "productName": "模拟产品-lv3ros",
     "startedAt": "2026-09-07T00:43:16.369+08:00",
-    "taggedSamples": 8064
+    "taggedSamples": 8136
   },
   {
     "lineId": "ln-0d66fdf4",
@@ -207,6 +207,6 @@
     "productId": "pd-7b25186a",
     "productName": "mh-prod-mhqc88",
     "startedAt": "2026-09-07T01:17:49.558+08:00",
-    "taggedSamples": 1371
+    "taggedSamples": 1385
   }
 ]

package/server/services/workshop/agents/industrial-context.ts

@@ -1,222 +1,231 @@
-/**
- * IndustrialContext —— Agent 工业语义上下文服务(**数据驱动,零硬编码**)。
- *
- * 语义来源优先级(逐层覆盖,全部为用户可编辑数据):
- *   节点 semantics(节点级备注)> 模板 semantics(模板编辑)> 结构化字段自动描述
- *   (量程/单位/预警带/采样周期由节点与模板元数据合成)。
- * 上下文 = 节点语义卡 × 产线工况 × 关联设备 —— 只描述「该 Agent 实际绑定的节点」。
- */
-
-import { getAgentNodeBindingRepo } from './node-bindings.repo'
-import { getDcwController } from '../dcw/dcw-controller'
-import { getDcwLineRepo } from '../dcw/dcw-line.repo'
-import { getActiveLineRun } from '../dcw/line-run'
-import { findDcwTemplate } from '../dcw/dcw-templates'
-import { getDaqNodeRepo } from '../daq/daq-node.repo'
-import { findDaqTemplate } from '../daq/daq-templates'
-import { getDeviceTwinRepo } from '../assets/device-twin.repo'
-import { getRecipeRollBackManager } from '../dcw/recipe-rollback-manager'
-
-/** 从属设备描述(名称/状态/实时遥测;数据驱动) */
-function describeTwin(bindId: string | null | undefined): string | null {
-  if (!bindId) return null
-  const t = getDeviceTwinRepo().findById(bindId)
-  if (!t) return null
-  return `${t.name}(id=${t.id},state=${t.state},telemetry=${JSON.stringify(t.telemetry)})`
-}
-
-/** 数控节点语义卡:模板 semantics(或节点覆盖)× 实时数据合成 */
-function dcwSemanticCard(nodeId: string, mode: string): string | null {
-  const node = getDcwController().byId(nodeId)
-  if (!node) return null
-  const tpl = findDcwTemplate(node.templateKey)
-  const sem = (node.semantics ?? tpl?.semantics ?? '').trim()
-  const line = node.lineId ? getDcwLineRepo().byId(node.lineId) : undefined
-  const run = node.lineId ? getActiveLineRun(node.lineId) : null
-  const recipe = run ? getDcwController().listRecipes().find(r => r.id === run.recipeId) : undefined
-  const param = recipe?.params.find(p => p.nodeId === nodeId)
-  const twinDesc = describeTwin(node.deviceBindingId)
-  const step = Math.max(10 ** -node.decimals, (node.max - node.min) * 0.02)
-
-  const lines = [
-    `#### ◆ ${node.name} [id=${nodeId}]`,
-    `- 物理量: ${tpl?.ch ?? node.templateKey},单位 ${node.unit},精度 ${node.decimals} 位小数`,
-  ]
-  if (sem) lines.push(`- 工艺语义: ${sem}`)
-  lines.push(`- 安全量程: [${node.min}, ${node.max}] ${node.unit}(硬联锁,越界即拒)`)
-  if (param && (param.min != null || param.max != null)) {
-    lines.push(`- 活动配方「${recipe!.name}」工艺窗口: [${param.min ?? '-∞'}, ${param.max ?? '+∞'}] ${node.unit}(软联锁,配方目标值 ${param.value}${node.unit ?? ''})`)
-  }
-  else if (run) {
-    lines.push(`- 活动配方「${run.recipeName}」未对本节点设窗口(全局量程生效)`)
-  }
-  lines.push(`- 当前设定: ${node.value != null ? `${node.value}${node.unit}` : '未下发'},状态 ${node.state},所属产线「${line?.name ?? '未分配'}」`)
-  if (twinDesc) lines.push(`- 从属设备: ${twinDesc};你的写入经 PLC 下发后反映到该设备的物理行为`)
-  lines.push(`- 操作守则: 单次调幅建议 ≤${Number(step.toFixed(node.decimals))}${node.unit}(量程 2%);下发后等待工艺响应再评估;目标值必须落在窗口内;驱动 ${node.driver};${mode === 'manual' ? '**手动确认模式**,每次下发会请求用户批准,请在下发前说明理由' : '自动模式,直接执行'}`)
-  return lines.join('\n')
-}
-
-/** 数采节点语义卡 */
-function daqSemanticCard(nodeId: string, mode: string): string | null {
-  const node = getDaqNodeRepo().byId(nodeId)
-  if (!node) return null
-  const tpl = findDaqTemplate(node.templateKey)
-  const sem = (node.semantics ?? tpl?.semantics ?? '').trim()
-  const line = node.lineId ? getDcwLineRepo().byId(node.lineId) : undefined
-  const run = node.lineId ? getActiveLineRun(node.lineId) : null
-  const recipe = run ? getDcwController().listRecipes().find(r => r.id === run.recipeId) : undefined
-  const win = recipe?.daqWindows?.find(w => w.nodeId === nodeId)
-  const twinDesc = describeTwin(node.deviceBindingId)
-  const fresh = node.lastAt ? Math.round((Date.now() - Date.parse(node.lastAt)) / 1000) : null
-
-  const lines = [
-    `#### ◇ ${node.name} [id=${nodeId}]`,
-    `- 物理量: ${tpl?.ch ?? node.templateKey},单位 ${node.unit},正常量程 [${node.min}, ${node.max}] ${node.unit}`,
-  ]
-  if (sem) lines.push(`- 采集语义: ${sem}`)
-  if (node.warnLow != null || node.warnHigh != null) {
-    lines.push(`- 预警带: [${node.warnLow ?? '-∞'}, ${node.warnHigh ?? '+∞'}] ${node.unit}(出带=warn,越量程=alarm)`)
-  }
-  if (win) lines.push(`- 活动配方「${recipe!.name}」监控窗口: [${win.min ?? '-∞'}, ${win.max ?? '+∞'}] ${node.unit}(实时值越限 → 节点标红 + 孪生告警)`)
-  lines.push(`- 实时值: ${node.value != null ? `${node.value}${node.unit}` : '暂无数据'}${fresh != null ? `( ${fresh}s 前更新)` : ''},状态 ${node.state},采样 ${node.effectiveInterval(1000)}ms,驱动 ${node.driver},所属产线「${line?.name ?? '未分配'}」`)
-  if (twinDesc) lines.push(`- 从属设备: ${twinDesc};量测的是该设备的真实过程量`)
-  lines.push(`- 数据获取: daq_query(node_id='${nodeId}');${mode === 'manual' ? '手动确认模式' : '查询自动执行'}`)
-  return lines.join('\n')
-}
-
-/** 工具:my_industrial_nodes 的语义卡视图 */
-export function nodeSemanticCards(agentId: string): { text: string, stale: number } {
-  const repo = getAgentNodeBindingRepo()
-  const bindings = repo.byAgent(agentId)
-  const cards: string[] = []
-  let stale = 0
-  for (const b of bindings) {
-    const card = b.kind === 'dcw' ? dcwSemanticCard(b.nodeId, b.mode) : daqSemanticCard(b.nodeId, b.mode)
-    if (card) cards.push(card)
-    else stale++
-  }
-  return { text: cards.join('\n\n'), stale }
-}
-
-/** 产线工况简报(注入每次回合 prompt;只描述 Agent 绑定节点所在的产线) */
-export function buildIndustrialContext(agentId: string): string {
-  const bindings = getAgentNodeBindingRepo().byAgent(agentId)
-  const dcwIds = bindings.filter(b => b.kind === 'dcw').map(b => b.nodeId)
-  const daqIds = bindings.filter(b => b.kind === 'daq').map(b => b.nodeId)
-  if (dcwIds.length === 0 && daqIds.length === 0) return ''
-
-  const lineIds = [...new Set([
-    ...dcwIds.map(id => getDcwController().byId(id)?.lineId ?? ''),
-    ...daqIds.map(id => getDaqNodeRepo().byId(id)?.lineId ?? ''),
-  ])].filter(Boolean)
-
-  const sections: string[] = []
-  for (const lineId of lineIds) {
-    const line = getDcwLineRepo().byId(lineId)
-    if (!line) continue
-    const run = getActiveLineRun(lineId)
-    const lines: string[] = [`### 产线「${line.name}」(id=${lineId},光晕色 ${line.color})`]
-    if (run) {
-      const elapsedMin = Math.round((Date.now() - Date.parse(run.startedAt)) / 60_000)
-      lines.push(`- 工况:**运行中**,批次 ${run.runId.slice(0, 8)},产品「${run.productName}」× 配方「${run.recipeName}」,已运行 ${elapsedMin} 分钟,已采集 ${run.taggedSamples} 打标样本`)
-      const recipe = getDcwController().listRecipes().find(r => r.id === run.recipeId)
-      if (recipe) {
-        const params = recipe.params
-          .map((p) => {
-            const n = getDcwController().byId(p.nodeId)
-            const win = (p.min != null || p.max != null) ? ` 窗口 ${p.min ?? '-∞'}~${p.max ?? '+∞'}${n?.unit ?? ''}` : ''
-            return `${n?.name ?? p.nodeId}=${p.value}${n?.unit ?? ''}${win}`
-          })
-          .join(';')
-        lines.push(`- 配方工艺参数: ${params}`)
-        const wins = (recipe.daqWindows ?? [])
-          .map((w) => {
-            const n = getDaqNodeRepo().byId(w.nodeId)
-            return `${n?.name ?? w.nodeId} ∈ [${w.min ?? '-∞'}, ${w.max ?? '+∞'}]${n?.unit ?? ''}`
-          })
-          .join(';')
-        if (wins) lines.push(`- 数采监控窗口(越限即报警): ${wins}`)
-      }
-    }
-    else {
-      lines.push('- 工况:**停线**(无活动批次;写入仅受节点全局量程约束,数采暂停)')
-    }
-    const twinNames = new Set<string>()
-    for (const id of [...dcwIds, ...daqIds]) {
-      const bindId = getDcwController().byId(id)?.deviceBindingId ?? getDaqNodeRepo().byId(id)?.deviceBindingId ?? null
-      const desc = describeTwin(bindId)
-      if (desc) twinNames.add(desc)
-    }
-    if (twinNames.size > 0) lines.push(`- 关联设备孪生: ${Array.from(twinNames).join(' | ')}`)
-    // 当前报警透出(只报本 Agent 绑定的数采节点):越限即报警是工况里最需要
-    // Agent 优先感知的事实 —— 不注入的话,Agent 只能靠 daq_query 事后发现
-    const alarms = daqIds
-      .map(id => getDaqNodeRepo().byId(id))
-      .filter(n => n && n.lineId === lineId && n.state === 'alarm')
-    if (alarms.length > 0) {
-      lines.push(`- **当前报警**: ${alarms.map(n => `${n.name} 实时 ${n.value ?? '?'}${n.unit} 处于报警态(越量程或越配方监控窗口)`)},应优先判读(设定-响应滞后 vs 真实异常)并处置`)
-    }
-    sections.push(lines.join('\n'))
-  }
-  if (sections.length === 0) return ''
-  return `## 产线工况简报(实时;每次回合自动注入)\n${sections.join('\n\n')}`
-}
-
-/**
- * 工业调控作业环(prompt 层纪律注入;有工业绑定才注入,零硬编码)。
- * 与节点语义卡(my_industrial_nodes 结果)互补:语义卡在工具调用后可见,
- * 作业环保证 Agent 在**任何回合开头**就知道调控方法论。
- * 七步闭环(v2.1):观察 → 理解 → 假设 → 设定 → 判定 → 回退/保持 → 复盘;
- * 每次下发自动开优化记录、判定入册、回退可审计 —— 闭环由系统记账,Agent 负责判断质量。
- */
-export function industrialLoopGuide(agentId: string): string {
-  const bindings = getAgentNodeBindingRepo().byAgent(agentId)
-  if (bindings.length === 0) return ''
-  const hasDcw = bindings.some(b => b.kind === 'dcw')
-  const hasManual = bindings.some(b => b.kind === 'dcw' && b.mode === 'manual')
-  const steps: string[] = [
-    '1. 观察:daq_query 获取绑定节点的真实时序数据(支持按产品/配方/时间窗过滤),结合返回的工况判读理解当前状态。',
-    '2. 理解:my_industrial_nodes 读取节点语义卡 —— 物理量含义/单位/安全量程/活动配方工艺窗口/单次调幅步进,以及调控闭环状态(进行中的优化记录/上次良好值/最近判定)。',
-  ]
-  if (hasDcw) {
-    steps.push(
-      '3. 假设:调参前先声明假设(写入 dcw_control 的 hypothesis 参数)—— 目标值 + 预期效果 + 观察计划;目标必须落在「安全量程 ∩ 活动配方工艺窗口」内,优先小步幅(≤量程 2%)、单向逼近。',
-      '4. 设定:dcw_control 下发(自动开优化记录);等待工艺响应(热惯性/传动惯量),勿连续大幅调整。',
-      '5. 判定:daq_query 复测窗口数据后,dcw_judge 落判定 —— keep(已验证经验)/ rollback(判应回退)/ uncertain(证据不足)。未判定前再下发,旧记录会被标记 superseded;判定必须引用具体数值与时间窗证据。',
-      '6. 回退/保持:判 rollback 后用 dcw_rollback(record_id) 执行回退(判定不自动改 PLC),回退后 daq_query 复测确认恢复;回退冷却期内禁止同向重写。收敛异常时分析根因或上报,而非盲目加码。',
-      '7. 复盘:dcw_journal 查看该节点参数变更史与判定结论;keep 的记录是已验证经验,rollback 的记录写进结论修正下次假设 —— 禁止重复已失败的调参方向。',
-    )
-  }
-  else {
-    steps.push('3. 判读:你是量测侧 —— 数据越过配方监控窗口会触发节点报警与孪生告警;判读时区分「同线数控设定-响应滞后」与「真实过程异常」,结论引用具体数值与时间窗。')
-  }
-  if (hasManual) {
-    steps.push('8. 权限边界:手动确认模式的节点,下发前必须说明理由并等待用户批准(用户备注会随结果返回,请响应其关切);自动模式节点直接执行,但同样受联锁约束;其判定回退会推请用户确认。')
-  }
-  steps.push(`${steps.length + 1}. 数值口径:工具返回的采集/设定值均为经标定钩子处理后的真实物理量纲;引用数值时带上单位与时间,便于人工复核。`)
-  const history = recentOptimizationNotes(agentId)
-  return `## 工业调控作业环(你的节点操作方法论,每回合生效)\n${steps.join('\n')}${history}`
-}
-
-/** 经验注入:该 Agent 绑定数控节点的最近优化记录结论(keep/rollback 统计),跨任务积累 */
-function recentOptimizationNotes(agentId: string): string {
-  try {
-    const rb = getRecipeRollBackManager()
-    const dcwIds = getAgentNodeBindingRepo().byAgent(agentId).filter(b => b.kind === 'dcw').map(b => b.nodeId)
-    if (dcwIds.length === 0) return ''
-    const lines: string[] = []
-    for (const nodeId of dcwIds.slice(0, 3)) {
-      const records = rb.records({ nodeId, limit: 20 })
-      if (records.length === 0) continue
-      const keep = records.filter(r => r.judge?.verdict === 'keep').length
-      const roll = records.filter(r => r.judge?.verdict === 'rollback' || r.status === 'rolled-back').length
-      const last = records.find(r => r.judge)
-      lines.push(`- ${records[0]!.nodeName}:近 ${records.length} 次调控(keep ${keep}/rollback ${roll})${last?.judge ? `,最近判定 ${last.judge.verdict}(${last.judge.reason.slice(0, 60)})` : ''}`)
-    }
-    if (lines.length === 0) return ''
-    return `\n\n## 优化经验台账(你绑定节点的近期调控结论;避免重复已失败方向)\n${lines.join('\n')}`
-  }
-  catch {
-    return ''
-  }
-}
+/**
+ * IndustrialContext —— Agent 工业语义上下文服务(**数据驱动,零硬编码**)。
+ *
+ * 语义来源优先级(逐层覆盖,全部为用户可编辑数据):
+ *   节点 semantics(节点级备注)> 模板 semantics(模板编辑)> 结构化字段自动描述
+ *   (量程/单位/预警带/采样周期由节点与模板元数据合成)。
+ * 上下文 = 节点语义卡 × 产线工况 × 关联设备 —— 只描述「该 Agent 实际绑定的节点」。
+ */
+
+import { getAgentNodeBindingRepo } from './node-bindings.repo'
+import { getDcwController } from '../dcw/dcw-controller'
+import { getDcwLineRepo } from '../dcw/dcw-line.repo'
+import { getActiveLineRun } from '../dcw/line-run'
+import { findDcwTemplate } from '../dcw/dcw-templates'
+import { getDaqNodeRepo } from '../daq/daq-node.repo'
+import { findDaqTemplate } from '../daq/daq-templates'
+import { getDeviceTwinRepo } from '../assets/device-twin.repo'
+import { getRecipeRollBackManager } from '../dcw/recipe-rollback-manager'
+
+/** 从属设备描述(名称/状态/实时遥测;数据驱动) */
+function describeTwin(bindId: string | null | undefined): string | null {
+  if (!bindId) return null
+  const t = getDeviceTwinRepo().findById(bindId)
+  if (!t) return null
+  return `${t.name}(id=${t.id},state=${t.state},telemetry=${JSON.stringify(t.telemetry)})`
+}
+
+/** 多对多绑定:节点全部绑定设备的描述(deviceIds 权威;legacy 别名兜底) */
+function describeTwins(node: { deviceIds?: string[], deviceBindingId?: string | null }): string[] {
+  const ids = node.deviceIds?.length ? node.deviceIds : (node.deviceBindingId ? [node.deviceBindingId] : [])
+  return ids.map(id => describeTwin(id)).filter((x): x is string => Boolean(x))
+}
+
+/** 数控节点语义卡:模板 semantics(或节点覆盖)× 实时数据合成 */
+function dcwSemanticCard(nodeId: string, mode: string): string | null {
+  const node = getDcwController().byId(nodeId)
+  if (!node) return null
+  const tpl = findDcwTemplate(node.templateKey)
+  const sem = (node.semantics ?? tpl?.semantics ?? '').trim()
+  const line = node.lineId ? getDcwLineRepo().byId(node.lineId) : undefined
+  const run = node.lineId ? getActiveLineRun(node.lineId) : null
+  const recipe = run ? getDcwController().listRecipes().find(r => r.id === run.recipeId) : undefined
+  const param = recipe?.params.find(p => p.nodeId === nodeId)
+  const twinDesc = describeTwins(node).join(' | ') || null
+  const step = Math.max(10 ** -node.decimals, (node.max - node.min) * 0.02)
+
+  const lines = [
+    `#### ◆ ${node.name} [id=${nodeId}]`,
+    `- 物理量: ${tpl?.ch ?? node.templateKey},单位 ${node.unit},精度 ${node.decimals} 位小数`,
+  ]
+  if (sem) lines.push(`- 工艺语义: ${sem}`)
+  lines.push(`- 安全量程: [${node.min}, ${node.max}] ${node.unit}(硬联锁,越界即拒)`)
+  if (param && (param.min != null || param.max != null)) {
+    lines.push(`- 活动配方「${recipe!.name}」工艺窗口: [${param.min ?? '-∞'}, ${param.max ?? '+∞'}] ${node.unit}(软联锁,配方目标值 ${param.value}${node.unit ?? ''})`)
+  }
+  else if (run) {
+    lines.push(`- 活动配方「${run.recipeName}」未对本节点设窗口(全局量程生效)`)
+  }
+  lines.push(`- 当前设定: ${node.value != null ? `${node.value}${node.unit}` : '未下发'},状态 ${node.state},所属产线「${line?.name ?? '未分配'}」`)
+  if (twinDesc) lines.push(`- 从属设备: ${twinDesc};你的写入经 PLC 下发后反映到该设备的物理行为`)
+  lines.push(`- 操作守则: 单次调幅建议 ≤${Number(step.toFixed(node.decimals))}${node.unit}(量程 2%);下发后等待工艺响应再评估;目标值必须落在窗口内;驱动 ${node.driver};${mode === 'manual' ? '**手动确认模式**,每次下发会请求用户批准,请在下发前说明理由' : '自动模式,直接执行'}`)
+  return lines.join('\n')
+}
+
+/** 数采节点语义卡 */
+function daqSemanticCard(nodeId: string, mode: string): string | null {
+  const node = getDaqNodeRepo().byId(nodeId)
+  if (!node) return null
+  const tpl = findDaqTemplate(node.templateKey)
+  const sem = (node.semantics ?? tpl?.semantics ?? '').trim()
+  const line = node.lineId ? getDcwLineRepo().byId(node.lineId) : undefined
+  const run = node.lineId ? getActiveLineRun(node.lineId) : null
+  const recipe = run ? getDcwController().listRecipes().find(r => r.id === run.recipeId) : undefined
+  const win = recipe?.daqWindows?.find(w => w.nodeId === nodeId)
+  const twinDesc = describeTwins(node).join(' | ') || null
+  const fresh = node.lastAt ? Math.round((Date.now() - Date.parse(node.lastAt)) / 1000) : null
+
+  const lines = [
+    `#### ◇ ${node.name} [id=${nodeId}]`,
+    `- 物理量: ${tpl?.ch ?? node.templateKey},单位 ${node.unit},正常量程 [${node.min}, ${node.max}] ${node.unit}`,
+  ]
+  if (sem) lines.push(`- 采集语义: ${sem}`)
+  if (node.warnLow != null || node.warnHigh != null) {
+    lines.push(`- 预警带: [${node.warnLow ?? '-∞'}, ${node.warnHigh ?? '+∞'}] ${node.unit}(出带=warn,越量程=alarm)`)
+  }
+  if (win) lines.push(`- 活动配方「${recipe!.name}」监控窗口: [${win.min ?? '-∞'}, ${win.max ?? '+∞'}] ${node.unit}(实时值越限 → 节点标红 + 孪生告警)`)
+  lines.push(`- 实时值: ${node.value != null ? `${node.value}${node.unit}` : '暂无数据'}${fresh != null ? `( ${fresh}s 前更新)` : ''},状态 ${node.state},采样 ${node.effectiveInterval(1000)}ms,驱动 ${node.driver},所属产线「${line?.name ?? '未分配'}」`)
+  if (twinDesc) lines.push(`- 从属设备: ${twinDesc};量测的是该设备的真实过程量`)
+  lines.push(`- 数据获取: daq_query(node_id='${nodeId}');${mode === 'manual' ? '手动确认模式' : '查询自动执行'}`)
+  return lines.join('\n')
+}
+
+/** 工具:my_industrial_nodes 的语义卡视图 */
+export function nodeSemanticCards(agentId: string): { text: string, stale: number } {
+  const repo = getAgentNodeBindingRepo()
+  const bindings = repo.byAgent(agentId)
+  const cards: string[] = []
+  let stale = 0
+  for (const b of bindings) {
+    const card = b.kind === 'dcw' ? dcwSemanticCard(b.nodeId, b.mode) : daqSemanticCard(b.nodeId, b.mode)
+    if (card) cards.push(card)
+    else stale++
+  }
+  return { text: cards.join('\n\n'), stale }
+}
+
+/** 产线工况简报(注入每次回合 prompt;只描述 Agent 绑定节点所在的产线) */
+export function buildIndustrialContext(agentId: string): string {
+  const bindings = getAgentNodeBindingRepo().byAgent(agentId)
+  const dcwIds = bindings.filter(b => b.kind === 'dcw').map(b => b.nodeId)
+  const daqIds = bindings.filter(b => b.kind === 'daq').map(b => b.nodeId)
+  if (dcwIds.length === 0 && daqIds.length === 0) return ''
+
+  const lineIds = [...new Set([
+    ...dcwIds.map(id => getDcwController().byId(id)?.lineId ?? ''),
+    ...daqIds.map(id => getDaqNodeRepo().byId(id)?.lineId ?? ''),
+  ])].filter(Boolean)
+
+  const sections: string[] = []
+  for (const lineId of lineIds) {
+    const line = getDcwLineRepo().byId(lineId)
+    if (!line) continue
+    const run = getActiveLineRun(lineId)
+    const lines: string[] = [`### 产线「${line.name}」(id=${lineId},光晕色 ${line.color})`]
+    if (run) {
+      const elapsedMin = Math.round((Date.now() - Date.parse(run.startedAt)) / 60_000)
+      lines.push(`- 工况:**运行中**,批次 ${run.runId.slice(0, 8)},产品「${run.productName}」× 配方「${run.recipeName}」,已运行 ${elapsedMin} 分钟,已采集 ${run.taggedSamples} 打标样本`)
+      const recipe = getDcwController().listRecipes().find(r => r.id === run.recipeId)
+      if (recipe) {
+        const params = recipe.params
+          .map((p) => {
+            const n = getDcwController().byId(p.nodeId)
+            const win = (p.min != null || p.max != null) ? ` 窗口 ${p.min ?? '-∞'}~${p.max ?? '+∞'}${n?.unit ?? ''}` : ''
+            return `${n?.name ?? p.nodeId}=${p.value}${n?.unit ?? ''}${win}`
+          })
+          .join(';')
+        lines.push(`- 配方工艺参数: ${params}`)
+        const wins = (recipe.daqWindows ?? [])
+          .map((w) => {
+            const n = getDaqNodeRepo().byId(w.nodeId)
+            return `${n?.name ?? w.nodeId} ∈ [${w.min ?? '-∞'}, ${w.max ?? '+∞'}]${n?.unit ?? ''}`
+          })
+          .join(';')
+        if (wins) lines.push(`- 数采监控窗口(越限即报警): ${wins}`)
+      }
+    }
+    else {
+      lines.push('- 工况:**停线**(无活动批次;写入仅受节点全局量程约束,数采暂停)')
+    }
+    const twinNames = new Set<string>()
+    for (const id of [...dcwIds, ...daqIds]) {
+      for (const desc of [
+        ...describeTwins(getDcwController().byId(id) ?? {}),
+        ...describeTwins(getDaqNodeRepo().byId(id) ?? {}),
+      ]) {
+        if (desc) twinNames.add(desc)
+      }
+    }
+    if (twinNames.size > 0) lines.push(`- 关联设备孪生: ${Array.from(twinNames).join(' | ')}`)
+    // 当前报警透出(只报本 Agent 绑定的数采节点):越限即报警是工况里最需要
+    // Agent 优先感知的事实 —— 不注入的话,Agent 只能靠 daq_query 事后发现
+    const alarms = daqIds
+      .map(id => getDaqNodeRepo().byId(id))
+      .filter(n => n && n.lineId === lineId && n.state === 'alarm')
+    if (alarms.length > 0) {
+      lines.push(`- **当前报警**: ${alarms.map(n => `${n.name} 实时 ${n.value ?? '?'}${n.unit} 处于报警态(越量程或越配方监控窗口)`)},应优先判读(设定-响应滞后 vs 真实异常)并处置`)
+    }
+    sections.push(lines.join('\n'))
+  }
+  if (sections.length === 0) return ''
+  return `## 产线工况简报(实时;每次回合自动注入)\n${sections.join('\n\n')}`
+}
+
+/**
+ * 工业调控作业环(prompt 层纪律注入;有工业绑定才注入,零硬编码)。
+ * 与节点语义卡(my_industrial_nodes 结果)互补:语义卡在工具调用后可见,
+ * 作业环保证 Agent 在**任何回合开头**就知道调控方法论。
+ * 七步闭环(v2.1):观察 → 理解 → 假设 → 设定 → 判定 → 回退/保持 → 复盘;
+ * 每次下发自动开优化记录、判定入册、回退可审计 —— 闭环由系统记账,Agent 负责判断质量。
+ */
+export function industrialLoopGuide(agentId: string): string {
+  const bindings = getAgentNodeBindingRepo().byAgent(agentId)
+  if (bindings.length === 0) return ''
+  const hasDcw = bindings.some(b => b.kind === 'dcw')
+  const hasManual = bindings.some(b => b.kind === 'dcw' && b.mode === 'manual')
+  const steps: string[] = [
+    '1. 观察:daq_query 获取绑定节点的真实时序数据(支持按产品/配方/时间窗过滤),结合返回的工况判读理解当前状态。',
+    '2. 理解:my_industrial_nodes 读取节点语义卡 —— 物理量含义/单位/安全量程/活动配方工艺窗口/单次调幅步进,以及调控闭环状态(进行中的优化记录/上次良好值/最近判定)。',
+  ]
+  if (hasDcw) {
+    steps.push(
+      '3. 假设:调参前先声明假设(写入 dcw_control 的 hypothesis 参数)—— 目标值 + 预期效果 + 观察计划;目标必须落在「安全量程 ∩ 活动配方工艺窗口」内,优先小步幅(≤量程 2%)、单向逼近。',
+      '4. 设定:dcw_control 下发(自动开优化记录);等待工艺响应(热惯性/传动惯量),勿连续大幅调整。',
+      '5. 判定:daq_query 复测窗口数据后,dcw_judge 落判定 —— keep(已验证经验)/ rollback(判应回退)/ uncertain(证据不足)。未判定前再下发,旧记录会被标记 superseded;判定必须引用具体数值与时间窗证据。',
+      '6. 回退/保持:判 rollback 后用 dcw_rollback(record_id) 执行回退(判定不自动改 PLC),回退后 daq_query 复测确认恢复;回退冷却期内禁止同向重写。收敛异常时分析根因或上报,而非盲目加码。',
+      '7. 复盘:dcw_journal 查看该节点参数变更史与判定结论;keep 的记录是已验证经验,rollback 的记录写进结论修正下次假设 —— 禁止重复已失败的调参方向。',
+    )
+  }
+  else {
+    steps.push('3. 判读:你是量测侧 —— 数据越过配方监控窗口会触发节点报警与孪生告警;判读时区分「同线数控设定-响应滞后」与「真实过程异常」,结论引用具体数值与时间窗。')
+  }
+  if (hasManual) {
+    steps.push('8. 权限边界:手动确认模式的节点,下发前必须说明理由并等待用户批准(用户备注会随结果返回,请响应其关切);自动模式节点直接执行,但同样受联锁约束;其判定回退会推请用户确认。')
+  }
+  steps.push(`${steps.length + 1}. 数值口径:工具返回的采集/设定值均为经标定钩子处理后的真实物理量纲;引用数值时带上单位与时间,便于人工复核。`)
+  const history = recentOptimizationNotes(agentId)
+  return `## 工业调控作业环(你的节点操作方法论,每回合生效)\n${steps.join('\n')}${history}`
+}
+
+/** 经验注入:该 Agent 绑定数控节点的最近优化记录结论(keep/rollback 统计),跨任务积累 */
+function recentOptimizationNotes(agentId: string): string {
+  try {
+    const rb = getRecipeRollBackManager()
+    const dcwIds = getAgentNodeBindingRepo().byAgent(agentId).filter(b => b.kind === 'dcw').map(b => b.nodeId)
+    if (dcwIds.length === 0) return ''
+    const lines: string[] = []
+    for (const nodeId of dcwIds.slice(0, 3)) {
+      const records = rb.records({ nodeId, limit: 20 })
+      if (records.length === 0) continue
+      const keep = records.filter(r => r.judge?.verdict === 'keep').length
+      const roll = records.filter(r => r.judge?.verdict === 'rollback' || r.status === 'rolled-back').length
+      const last = records.find(r => r.judge)
+      lines.push(`- ${records[0]!.nodeName}:近 ${records.length} 次调控(keep ${keep}/rollback ${roll})${last?.judge ? `,最近判定 ${last.judge.verdict}(${last.judge.reason.slice(0, 60)})` : ''}`)
+    }
+    if (lines.length === 0) return ''
+    return `\n\n## 优化经验台账(你绑定节点的近期调控结论;避免重复已失败方向)\n${lines.join('\n')}`
+  }
+  catch {
+    return ''
+  }
+}