agentworkshop 0.7.27 → 0.7.34

package/server/services/workshop/daq/drivers.ts

@@ -237,7 +237,7 @@ export async function closeModbusSafely(client: import('modbus-serial').ModbusRT
     Promise.resolve()
       .then(() => client.close())
       .catch(() => { /* 已断 */ }),
-    new Promise((r) => setTimeout(r, 500)),
+    new Promise(r => setTimeout(r, 500)),
   ])
 }
 
@@ -267,8 +267,14 @@ export function withModbusConn<R>(conn: ModbusConn, fn: () => Promise<R>): Promi
   ] as Promise<R>)
   void guarded.catch(() => { /* 超时分支先 reject,防 unhandled */ })
   conn.tail = guarded.then(
-    () => { clearTimeout(timer); conn.pending-- },
-    () => { clearTimeout(timer); conn.pending-- },
+    () => {
+      clearTimeout(timer)
+      conn.pending--
+    },
+    () => {
+      clearTimeout(timer)
+      conn.pending--
+    },
   ).then(() => {
     if (dead) conn.tail = Promise.resolve()
   })
@@ -311,6 +317,11 @@ export async function getModbusConn(cfg: Record<string, unknown>, transport: Mod
     throw err
   })
   client.setID(Number(cfg.unitId ?? 1))
+  // 吸收底层 socket 错误(write-after-end / ECONNRESET 等):超时作废重连路径关闭 socket 后,
+  // 队列中滞留的写入会触发 modbus-serial 内部 socket.write 异步错误 —— 若不挂 error 处理器,
+  // 会以 unhandledRejection 打崩进程(stability-guard fatal)。池层已有 evict 自愈,这里只需静默。
+  const sock = (client as unknown as { port?: { client?: { on?: (ev: string, cb: (e: Error) => void) => void } } }).port?.client
+  sock?.on?.('error', () => {})
   const conn: ModbusConn = { client, lastUsed: Date.now(), tail: Promise.resolve(), pending: 0, errors: 0 }
   modbusPool.set(key, conn)
   return conn
@@ -503,6 +514,13 @@ if (!opcuaSweepGlobal.__daqOpcUaSweep) {
   opcuaSweepGlobal.__daqOpcUaSweep.unref?.()
 }
 
+/** 已告警过的 OPC UA 端点(进程内,按 `类别|安全模式|endpoint` 键;同端点重连风暴只告警一次)。
+ *  与 logger 的同指纹限流互补:Set 管「同端点反复重连」,限流器管「大量不同端点」的同形文本。 */
+const gOpcUaSecWarned = globalThis as typeof globalThis & { __daqOpcUaSecWarned?: Set<string> }
+function opcUaSecWarned(): Set<string> {
+  return gOpcUaSecWarned.__daqOpcUaSecWarned ??= new Set()
+}
+
 export function opcuaKey(cfg: Record<string, unknown>): string {
   // S3:证书/安全模式必须纳入池 key——不同证书/安全策略绝不能复用同一连接
   return `${cfg.endpoint}|${cfg.username ?? ''}|${cfg.securityMode ?? 'None'}|${cfg.certificateFile ?? ''}|${cfg.privateKeyFile ?? ''}`
@@ -520,7 +538,13 @@ export async function getOpcUaConn(cfg: Record<string, unknown>): Promise<OpcUaC
     ? (cfg.securityMode as 'None')
     : 'None'
   if (securityMode === 'None' && process.env.NODE_ENV === 'production') {
-    log.warn(`[daq-opcua] WARN:节点 endpoint=${String(cfg.endpoint)} 使用 securityMode=None(匿名/明文),仅限测试环境;生产请配置 Sign/SignAndEncrypt + 证书`)
+    // P0:该告警原先每次建连都写一行(实测把生产日志灌到 18MB 且全是同一句)。
+    // 去重键 = endpoint:同端点反复重连只告警一次;warnThrottled 再兜一层同形文本洪水。
+    const noneKey = `none|${String(cfg.endpoint)}`
+    if (!opcUaSecWarned().has(noneKey)) {
+      opcUaSecWarned().add(noneKey)
+      log.warnThrottled(`[daq-opcua] WARN:节点 endpoint=${String(cfg.endpoint)} 使用 securityMode=None(匿名/明文),仅限测试环境;生产请配置 Sign/SignAndEncrypt + 证书`)
+    }
   }
   const clientOpts: Record<string, unknown> = {
     endpointMustExist: false,
@@ -540,7 +564,12 @@ export async function getOpcUaConn(cfg: Record<string, unknown>): Promise<OpcUaC
     }
     else {
       // 无证书时由 node-opcua 自动生成自签证书(仍加密,但身份不可信);生产建议显式证书
-      log.warn(`[daq-opcua] WARN:securityMode=${securityMode} 但未配置 certificateFile/privateKeyFile,将使用自动生成自签证书`)
+      // 同端点同模式只告警一次(重连不再重复);限流器兜同形文本洪水
+      const noCertKey = `nocert|${securityMode}|${String(cfg.endpoint)}`
+      if (!opcUaSecWarned().has(noCertKey)) {
+        opcUaSecWarned().add(noCertKey)
+        log.warnThrottled(`[daq-opcua] WARN:securityMode=${securityMode} 但未配置 certificateFile/privateKeyFile,将使用自动生成自签证书`)
+      }
     }
   }
   const client = opcua.OPCUAClient.create(clientOpts as Parameters<typeof opcua.OPCUAClient.create>[0])
@@ -871,11 +900,44 @@ export const httpDaqDriver: DaqDriver = {
   async available() {
     return true
   },
-  async sample({ driverConfig }) {
+  async sample({ driverConfig, signalKind }) {
     if (!driverConfig.url) throw new Error('缺少接口地址 url')
     const headers = parseHeaders(driverConfig.headersJSON)
     const res = await fetch(String(driverConfig.url), { headers, signal: AbortSignal.timeout(5000) })
     if (!res.ok) throw new Error(`HTTP ${res.status}`)
+    // v2 多形态分支(vector/image 由模板信号形态门控;标量端点零感知)
+    if (signalKind === 'vector') {
+      const j: unknown = JSON.parse(await res.text())
+      const path = driverConfig.jsonPath ? String(driverConfig.jsonPath) : 'points'
+      const target = path.split('.').reduce<unknown>((acc, k) => (acc == null ? undefined : (acc as Record<string, unknown>)[k]), j)
+      if (!Array.isArray(target)) throw new Error(`vector 帧报文缺少点列(jsonPath=${path})`)
+      const points = (target as unknown[]).map(Number)
+      if (points.length === 0 || points.length > 4096 || !points.every(Number.isFinite)) {
+        throw new Error(`vector 帧点列非法(1~4096 个有限数值,实际 ${points.length})`)
+      }
+      return { frame: { kind: 'vector', points } }
+    }
+    if (signalKind === 'image') {
+      const ct = res.headers.get('content-type') ?? ''
+      if (ct.includes('image/')) {
+        const blob = Buffer.from(await res.arrayBuffer())
+        if (blob.length === 0) throw new Error('image 帧响应体为空')
+        // PNG 尺寸兜底:IHDR 宽高在固定偏移(width@16, height@20;大端)——端点不必额外传
+        let width = Number(driverConfig.width ?? 0) || 0
+        let height = Number(driverConfig.height ?? 0) || 0
+        if ((!width || !height) && blob.length >= 24 && blob.readUInt32BE(12) === 0x49484452) {
+          width = width || blob.readUInt32BE(16)
+          height = height || blob.readUInt32BE(20)
+        }
+        return { frame: { kind: 'image', blob, mime: ct.split(';')[0]!, width, height } }
+      }
+      // JSON 报文:{ png: base64, width?, height? }
+      const j = JSON.parse(await res.text()) as Record<string, unknown>
+      const b64 = typeof j.png === 'string' ? j.png : typeof j.blob === 'string' ? j.blob : ''
+      const blob = Buffer.from(b64, 'base64')
+      if (blob.length === 0) throw new Error('image 帧报文缺少 png/blob(base64)字段')
+      return { frame: { kind: 'image', blob, mime: 'image/png', width: Number(j.width ?? 0) || 0, height: Number(j.height ?? 0) || 0 } }
+    }
     const text = await res.text()
     return extractNumeric(text, driverConfig.jsonPath ? String(driverConfig.jsonPath) : undefined)
   },
@@ -886,9 +948,28 @@ export const httpDaqDriver: DaqDriver = {
       const headers = parseHeaders(driverConfig.headersJSON)
       const res = await fetch(String(driverConfig.url), { headers, signal: AbortSignal.timeout(5000) })
       if (!res.ok) return { ok: false, message: `接口返回 HTTP ${res.status}` }
+      const ct = res.headers.get('content-type') ?? ''
+      // 帧形态感知测试:同一端点形态可能是 vector(JSON 点列)/ image(PNG 或 base64)/ 标量
+      if (ct.includes('image/')) {
+        const n = (await res.arrayBuffer()).byteLength
+        return { ok: true, message: `接口可达,image 帧 ${n}B(${ct.split(';')[0]})`, latencyMs: Date.now() - t0 }
+      }
       const text = await res.text()
-      const v = extractNumeric(text, driverConfig.jsonPath ? String(driverConfig.jsonPath) : undefined)
-      return { ok: true, message: `接口可达,取值 = ${v}`, sampleValue: v, latencyMs: Date.now() - t0 }
+      try {
+        const j = JSON.parse(text) as Record<string, unknown>
+        if (Array.isArray(j.points) && (j.points as unknown[]).length > 0) {
+          return { ok: true, message: `接口可达,vector 帧 ${(j.points as unknown[]).length} 点`, latencyMs: Date.now() - t0 }
+        }
+        if (typeof j.png === 'string' || typeof j.blob === 'string') {
+          return { ok: true, message: `接口可达,image 帧(base64, ${(j.png ?? j.blob as string).length} chars)`, latencyMs: Date.now() - t0 }
+        }
+        const v = extractNumeric(text, driverConfig.jsonPath ? String(driverConfig.jsonPath) : undefined)
+        return { ok: true, message: `接口可达,取值 = ${v}`, sampleValue: v, latencyMs: Date.now() - t0 }
+      }
+      catch {
+        const v = extractNumeric(text, driverConfig.jsonPath ? String(driverConfig.jsonPath) : undefined)
+        return { ok: true, message: `接口可达,取值 = ${v}`, sampleValue: v, latencyMs: Date.now() - t0 }
+      }
     }
     catch (err) {
       return { ok: false, message: `HTTP 请求失败: ${err instanceof Error ? err.message : String(err)}` }

package/server/services/workshop/db/aml-team-seeds.ts

@@ -0,0 +1,118 @@
+/**
+ * AML 自动建模平台内置种子(4 个 Agent 模板 + 1 个编组)。
+ * - harness 一律 'omp'(真实 LLM 执行器;config.systemPromptPrefix 注入角色场景提示)
+ * - id 固定,由 database.ts seedDefaultWorkshopData 以 INSERT OR IGNORE 幂等注入(重启安全)
+ * - 门禁 G1-G5 语义与默认阈值对齐 server/services/workshop/aml/gates.ts
+ *   (G1 0.10 / G2 0.25 / G3 20% / G4 500 行 × 3 批次 / G5 工件完整,均 live 可调)
+ * - 团队共同作业手册:.AgentWorkShop/prompts/aml-playbook.md;
+ *   各模板 systemPromptPrefix 只写角色差异,与手册互补不重复
+ */
+
+/** Agent 模板种子条目(与 database.ts DEFAULT_AGENT_TEMPLATES 元素同形) */
+export interface AmlAgentTemplateSeed {
+  id: string
+  name: string
+  harness: string
+  config: Record<string, unknown>
+}
+
+/** 编组种子条目(与 database.ts DEFAULT_TEAMS 元素同形) */
+export interface AmlTeamSeed {
+  id: string
+  name: string
+  description: string
+  members: Array<{ templateId: string, role: 'lead' | 'worker' }>
+}
+
+export const AML_AGENT_TEMPLATES: AmlAgentTemplateSeed[] = [
+  {
+    id: 'tpl-aml-lead',
+    name: 'AML 首席数据科学家',
+    harness: 'omp',
+    config: {
+      role: 'lead',
+      intro: 'AML 自动建模 lead;统筹使命拆解、派工与预算,唯一有权发起模型晋升(aml_model_promote)的人',
+      systemPromptPrefix: [
+        '你是「AML 影子建模团队」首席数据科学家(lead):统筹使命拆解、派工与预算,是门禁守门人,也是唯一有权发起模型晋升的人(aml_model_promote,内部走人工审批)。',
+        '## 作业流',
+        '1. 先 aml_node_catalog 盘点可建模资产(节点语义 × 近 24h 数据量),确认使命(产品/配方/horizon/预算)可满足;',
+        '2. 派数据工程师 aml_dataset_build 建数据集,核对其统计报告:行数≥500、批次≥3、清洗剔除率≤5%;',
+        '3. 派训练工程师迭代:首轮先线性基线,此后每轮 change_note 只声明单组件改动;',
+        '4. 派评测工程师审 aml_leaderboard 与门禁 G1-G5,重点防数据泄漏与过拟合;',
+        '5. 全过后由你调 aml_model_promote 发起晋升申请;未全过一律打回并写明缺哪项门禁。',
+        '## 诚实纪律',
+        '预算耗尽或连续两轮无改进:如实汇报瓶颈(数据不足/噪声/变量缺失),禁止虚报精度,禁止在人工批准前宣称模型已投产。',
+      ].join('\n'),
+    },
+  },
+  {
+    id: 'tpl-aml-data',
+    name: 'AML 数据工程师',
+    harness: 'omp',
+    config: {
+      role: 'worker',
+      intro: 'AML 数据工程师;时序清洗与统计判读,负责数据集构建与质量复核',
+      systemPromptPrefix: [
+        '你是 AML 团队数据工程师,精通工业时序清洗与统计判读。工具:aml_node_catalog / aml_dataset_build / aml_dataset_stats。',
+        '## 硬约束',
+        '隔离三元组:产品 × 热 Recipe × run 绝不混。建集前先 aml_node_catalog 核对本产品本配方下的授权节点;跨 Recipe 取数平台直接报错,不得尝试绕过。',
+        '## 窗口设计(时滞先行)',
+        '先 aml_dataset_stats 读控制→目标滞后估计:historySteps ≥ 时滞格数(时滞/beatMs 向上取整);horizon 按使命给定;beatMs 依数据密度选取并控制行数不超上限(必要时加粗 beatMs)。',
+        '## 清洗红线',
+        'Hampel 剔除率>5% 或缺失率>15% 时,先复核(换 beatMs / 收窄时间窗 / 上报 lead)再交付训练,不把脏数据甩给下游;单 run 缺失过高按平台策略整 run 丢弃并记录。',
+        '## 交付物',
+        'datasetId + 行数/批次数 + 清洗参数 + 时滞结论与 historySteps 建议,供训练与评测直接引用。',
+      ].join('\n'),
+    },
+  },
+  {
+    id: 'tpl-aml-trainer',
+    name: 'AML 训练工程师',
+    harness: 'omp',
+    config: {
+      role: 'worker',
+      intro: 'AML 训练工程师;写 train.py 提交训练作业,按排行榜反馈迭代模型结构',
+      systemPromptPrefix: [
+        '你是 AML 团队训练工程师:写 train.py → aml_job_submit 提交实验,aml_job_logs 排错,按 aml_leaderboard 反馈迭代。',
+        '## 代码契约',
+        'import amlkit;bundle = amlkit.load_bundle(job[\'datasetPath\']);x/u/y 为归一化 float32,形状 [N,H,nAll] / [N,F,nCtrl] / [N,F,nTgt];训练循环内用 amlkit.report_progress 上报进度;结束时 amlkit.export_torch_onnx(model, manifest) 导出契约工件——模型是一步预测:输入 [batch,H,nAll] → 输出 [batch,nTgt];梯度收敛后再 eval() + torch.no_grad() 评估,禁止在训练态导出。',
+        '## 迭代纪律(AIDE 树搜索)',
+        '1. 首轮必做线性基线(ridge/ARX 类)校准数据可学性;',
+        '2. 逐步升级:GBDT → MLP → LSTM/TCN;',
+        '3. 每轮只改一个组件(结构/特征/超参/采样),并在 change_note 写明(MLE-STAR);',
+        '4. 连续两轮指标变差→回溯最优父节点换方向(greedy-then-restore),不在坏分支上加码。',
+      ].join('\n'),
+    },
+  },
+  {
+    id: 'tpl-aml-eval',
+    name: 'AML 评测工程师',
+    harness: 'omp',
+    config: {
+      role: 'worker',
+      intro: 'AML 评测工程师;对抗性审查数据泄漏/过拟合/漂移,守门禁并沉淀案例',
+      systemPromptPrefix: [
+        '你是 AML 团队评测工程师,职责是对抗性审查——默认怀疑指标,不替训练工程师乐观。',
+        '## 审查清单',
+        '1. 数据泄漏:切分是否严格 byRun(同一 run 的窗口不得跨 train/test),归一化统计是否只用训练段;',
+        '2. 过拟合:G3 泛化差(|test−val|/val≤20%)逼近上限、train/test 曲线背离即打回;',
+        '3. 分布漂移:结合 aml_dataset_stats 逐 run 轮廓,判断 test 批次代表性是否不足;',
+        '4. 清洗异常:剔除率>5% 或缺失>15% 且未解释的实验直接存疑。',
+        '## 日常动作',
+        '用 aml_leaderboard 解读谱系树与门禁明细(G1 单步≤0.10 / G2 滚动≤0.25 / G4 行数≥500 且批次≥3 / G5 工件完整),输出「过 / 打回+原因」;通过后用 save_memory 沉淀「配方 × 有效结构 × 超参 × 陷阱」案例(DS-Agent 式)供全队复用。',
+      ].join('\n'),
+    },
+  },
+]
+
+export const AML_TEAM: AmlTeamSeed = {
+  id: 'team-aml-shadow',
+  name: 'AML 影子建模团队',
+  description: '内置自动建模编组:数据科学家(lead)+ 数据/训练/评测工程师;自主完成拉数→训练→评测→晋升申请',
+  members: [
+    { templateId: 'tpl-aml-lead', role: 'lead' },
+    { templateId: 'tpl-aml-data', role: 'worker' },
+    { templateId: 'tpl-aml-trainer', role: 'worker' },
+    { templateId: 'tpl-aml-eval', role: 'worker' },
+  ],
+}

package/server/services/workshop/db/channel-plugins.repo.ts

@@ -4,6 +4,8 @@
  * 语义:无显式行 = 未配置 → 全部启用插件对团队可见(向后兼容,存量 channel 不受影响);
  * 显式行(建队勾选/团队设置切换)写入后按行过滤 —— 关闭的插件其工具不注入该团队
  * Agent,dispatch 同源拒绝,防止不需要插件的 channel 上下文被污染。
+ * 主键列同时承载两种作用域 id:channel id(部署实例)与 team id(团队偏好,部署时传导),
+ * 故不设外键(migrateDropOwnerFks 同理:列仅存 id 引用;悬挂行无行为影响,读取即忽略)。
  * 自持连接(user.repository.ts 同模式):workshop.sqlite 单文件 WAL。
  */
 import { DatabaseSync } from 'node:sqlite'
@@ -18,16 +20,40 @@ function getDb(): DatabaseSync {
     db = new DatabaseSync(DB_PATH)
     db.exec('PRAGMA busy_timeout=4000')
     db.exec(`CREATE TABLE IF NOT EXISTS channel_plugins (
-      channel_id TEXT NOT NULL REFERENCES channels(id) ON DELETE CASCADE,
+      channel_id TEXT NOT NULL,
       plugin     TEXT NOT NULL,
       enabled    INTEGER NOT NULL DEFAULT 1,
       updated_at TEXT NOT NULL,
       PRIMARY KEY (channel_id, plugin)
     )`)
+    migrateDropChannelFk(db)
   }
   return db
 }
 
+/** 早期建表带 REFERENCES channels(id):team 作用域 id 写入会触发 FK 拒绝 → 重建剥壳 */
+function migrateDropChannelFk(db: DatabaseSync): void {
+  const fks = db.prepare('PRAGMA foreign_key_list(channel_plugins)').all() as Array<{ table: string }>
+  if (!fks.some(f => f.table === 'channels')) return
+  db.exec('PRAGMA foreign_keys = OFF;')
+  try {
+    db.exec(`CREATE TABLE channel_plugins_new (
+      channel_id TEXT NOT NULL,
+      plugin     TEXT NOT NULL,
+      enabled    INTEGER NOT NULL DEFAULT 1,
+      updated_at TEXT NOT NULL,
+      PRIMARY KEY (channel_id, plugin)
+    );
+    INSERT OR IGNORE INTO channel_plugins_new (channel_id, plugin, enabled, updated_at)
+      SELECT channel_id, plugin, enabled, updated_at FROM channel_plugins;
+    DROP TABLE channel_plugins;
+    ALTER TABLE channel_plugins_new RENAME TO channel_plugins;`)
+  }
+  finally {
+    db.exec('PRAGMA foreign_keys = ON;')
+  }
+}
+
 export interface ChannelPluginToggle {
   name: string
   enabled: boolean

package/server/services/workshop/db/database.ts

@@ -15,6 +15,7 @@
 import { createRequire } from 'node:module'
 import { DatabaseSync } from 'node:sqlite'
 import type * as sqliteVec from 'sqlite-vec'
+import { AML_AGENT_TEMPLATES, AML_TEAM } from './aml-team-seeds'
 
 const require = createRequire(import.meta.url)
 
@@ -279,7 +280,83 @@ CREATE TABLE IF NOT EXISTS approval_requests (
   decided_at    TEXT,
   comment       TEXT NOT NULL DEFAULT ''
 );
-CREATE INDEX IF NOT EXISTS idx_approval_requests_status ON approval_requests(status, requested_at DESC);`
+CREATE INDEX IF NOT EXISTS idx_approval_requests_status ON approval_requests(status, requested_at DESC);
+-- v15:AML 自动建模平台(数据集快照 / 训练作业 / 实验谱系 / 模型注册表)
+-- 级联:删数据集连带作业与实验行(模型挂实验下随之消失);GC 删除前代码层先查引用,
+-- CASCADE 仅作 FK 兜底(杜绝 ws.ts FK 毒化式永久错误)。
+CREATE TABLE IF NOT EXISTS aml_datasets (
+  id            TEXT PRIMARY KEY,
+  line_id       TEXT NOT NULL,
+  product_id    TEXT NOT NULL,
+  recipe_id     TEXT NOT NULL,
+  run_ids_json  TEXT NOT NULL DEFAULT '[]',
+  spec_json     TEXT NOT NULL,
+  sha256        TEXT NOT NULL,
+  row_count     INTEGER NOT NULL DEFAULT 0,
+  from_ms       INTEGER,
+  to_ms         INTEGER,
+  path          TEXT NOT NULL,
+  created_by    TEXT NOT NULL DEFAULT '',
+  created_by_kind TEXT NOT NULL DEFAULT 'user',
+  note          TEXT NOT NULL DEFAULT '',
+  created_at    TEXT NOT NULL
+);
+CREATE INDEX IF NOT EXISTS idx_aml_datasets_recipe ON aml_datasets(recipe_id, created_at DESC);
+CREATE TABLE IF NOT EXISTS aml_jobs (
+  id            TEXT PRIMARY KEY,
+  dataset_id    TEXT NOT NULL REFERENCES aml_datasets(id) ON DELETE CASCADE,
+  purpose       TEXT NOT NULL DEFAULT 'mpc_surrogate',
+  status        TEXT NOT NULL DEFAULT 'queued', -- queued|provisioning|training|evaluating|done|failed|cancelled|timeout|interrupted
+  stage         TEXT NOT NULL DEFAULT '',
+  progress      INTEGER NOT NULL DEFAULT 0,
+  budget_json   TEXT NOT NULL DEFAULT '{}',
+  metrics_json  TEXT,
+  gates_json    TEXT,
+  artifacts_path TEXT,
+  error         TEXT,
+  retry_count   INTEGER NOT NULL DEFAULT 0,
+  agent_id      TEXT NOT NULL DEFAULT '',
+  channel_id    TEXT NOT NULL DEFAULT '',
+  task_id       TEXT NOT NULL DEFAULT '',
+  created_at    TEXT NOT NULL,
+  started_at    TEXT,
+  ended_at      TEXT
+);
+CREATE INDEX IF NOT EXISTS idx_aml_jobs_status ON aml_jobs(status, created_at DESC);
+CREATE INDEX IF NOT EXISTS idx_aml_jobs_dataset ON aml_jobs(dataset_id, created_at DESC);
+CREATE TABLE IF NOT EXISTS aml_experiments (
+  id            TEXT PRIMARY KEY,
+  job_id        TEXT NOT NULL REFERENCES aml_jobs(id) ON DELETE CASCADE,
+  dataset_id    TEXT NOT NULL REFERENCES aml_datasets(id) ON DELETE CASCADE,
+  parent_experiment_id TEXT,
+  change_note   TEXT NOT NULL DEFAULT '',
+  config_json   TEXT NOT NULL DEFAULT '{}',
+  metrics_json  TEXT NOT NULL DEFAULT '{}',
+  gates_json    TEXT NOT NULL DEFAULT '{}',
+  status        TEXT NOT NULL DEFAULT 'running', -- running|gates_passed|gates_failed|failed
+  seed          INTEGER,
+  created_at    TEXT NOT NULL
+);
+CREATE INDEX IF NOT EXISTS idx_aml_experiments_dataset ON aml_experiments(dataset_id, created_at DESC);
+CREATE TABLE IF NOT EXISTS aml_models (
+  id            TEXT PRIMARY KEY,
+  experiment_id TEXT NOT NULL REFERENCES aml_experiments(id) ON DELETE CASCADE,
+  dataset_id    TEXT NOT NULL REFERENCES aml_datasets(id) ON DELETE CASCADE,
+  product_id    TEXT NOT NULL,
+  recipe_id     TEXT NOT NULL,
+  purpose       TEXT NOT NULL DEFAULT 'mpc_surrogate',
+  stage         TEXT NOT NULL DEFAULT 'candidate', -- candidate|shadow|production|retired
+  io_spec_json  TEXT NOT NULL DEFAULT '{}',
+  metrics_json  TEXT NOT NULL DEFAULT '{}',
+  path          TEXT NOT NULL DEFAULT '',
+  artifacts_pruned INTEGER NOT NULL DEFAULT 0,
+  created_by    TEXT NOT NULL DEFAULT '',
+  promoted_by   TEXT,
+  promoted_at   TEXT,
+  note          TEXT NOT NULL DEFAULT '',
+  created_at    TEXT NOT NULL
+);
+CREATE INDEX IF NOT EXISTS idx_aml_models_lookup ON aml_models(product_id, recipe_id, purpose, stage);`
 
 // ===== 默认种子数据(首轮初始化注入;owner NULL = 公共资源,所有登录用户只读共享) =====
 
@@ -430,6 +507,10 @@ function seedDefaultWorkshopData(db: DatabaseSync): void {
   for (const t of DEFAULT_AGENT_TEMPLATES) {
     insertAgent.run(t.id, t.name, t.harness, JSON.stringify(t.config), now, now)
   }
+  // v15:AML 自动建模团队种子(内置公共,同款 INSERT OR IGNORE 幂等)
+  for (const t of AML_AGENT_TEMPLATES) {
+    insertAgent.run(t.id, t.name, t.harness, JSON.stringify(t.config), now, now)
+  }
   const insertTeam = db.prepare(
     'INSERT OR IGNORE INTO teams (id, name, description, visibility, owner_user_id, created_at, updated_at) VALUES (?, ?, ?, \'public\', NULL, ?, ?)',
   )
@@ -442,6 +523,13 @@ function seedDefaultWorkshopData(db: DatabaseSync): void {
       insertMember.run(team.id, m.templateId, m.role, now)
     }
   }
+  // v15:AML 编组种子(team + team_members 同款幂等注入)
+  for (const team of [AML_TEAM]) {
+    insertTeam.run(team.id, team.name, team.description, now, now)
+    for (const m of team.members) {
+      insertMember.run(team.id, m.templateId, m.role, now)
+    }
+  }
   const insertChannelTpl = db.prepare(
     'INSERT OR IGNORE INTO channel_templates (id, name, description, scenario_prompt, workspace, lead_json, members_json, visibility, owner_user_id, created_at, updated_at) VALUES (?, ?, ?, ?, \'\', \'\', ?, \'public\', NULL, ?, ?)',
   )

package/server/services/workshop/db/memory.repo.ts

@@ -11,6 +11,12 @@ import type { MemoryKind, MemoryRow } from './database'
 /** 团队共享记忆域 sentinel(单 channel 内全员可读) */
 export const TEAM_AGENT_ID = '__team__'
 
+/** '2026-09' → '2026-10'(跨年安全);用于半开区间 [本月1日, 下月1日) 查询 */
+function nextMonth(month: string): string {
+  const [y, m] = month.split('-').map(Number)
+  return m === 12 ? `${(y ?? 0) + 1}-01` : `${y}-${String((m ?? 0) + 1).padStart(2, '0')}`
+}
+
 export interface MemoryUpsertInput {
   channelId: string
   agentId: string
@@ -194,6 +200,31 @@ export function createMemoryRepo(db: DatabaseSync) {
       return listByAgentRowidStmt.all(agentId, limit) as unknown as Array<MemoryRow & { rowid: number }>
     },
 
+    /**
+     * 按 (agent, kind, 年月) 精确取行 —— 供空闲反思使用。
+     * 替代原先「limit=1_000_000 拉全量再由 JS filter」的退化作法:
+     * 过滤条件下推到 SQL(走 agent_id+kind 前缀),返回量与实际增量同阶,
+     * 不再随记忆总量线性放大内存与 CPU。
+     * month 形如 '2026-09'(与 createdAt 的 ISO 前 7 位一致)。
+     */
+    listByAgentKindMonth(agentId: string, kind: MemoryKind, month: string): MemoryRow[] {
+      return db.prepare(
+        `SELECT ${COLS} FROM agent_memories m
+          WHERE m.agent_id = ? AND m.kind = ?
+            AND m.created_at >= ? AND m.created_at < ?
+          ORDER BY m.created_at ASC`,
+      ).all(agentId, kind, `${month}-01`, `${nextMonth(month)}-01`) as unknown as MemoryRow[]
+    },
+
+    /** 当月某 kind 的计数(反思触发阈值判断用;不取行体,零拷贝) */
+    countByAgentKindMonth(agentId: string, kind: MemoryKind, month: string): number {
+      const row = db.prepare(
+        `SELECT COUNT(*) AS c FROM agent_memories
+          WHERE agent_id = ? AND kind = ? AND created_at >= ? AND created_at < ?`,
+      ).get(agentId, kind, `${month}-01`, `${nextMonth(month)}-01`) as { c: number } | undefined
+      return Number(row?.c ?? 0)
+    },
+
     touch(id: string): void {
       touchStmt.run(new Date().toISOString(), id)
     },

package/server/services/workshop/db/user.repo.ts

@@ -41,6 +41,9 @@ export function createUserRepo(db: DatabaseSync) {
   const deleteMount = db.prepare(
     `DELETE FROM workspace_channels WHERE workspace_id = ? AND channel_id = ?`,
   )
+  const deleteMountsByChannel = db.prepare(
+    `DELETE FROM workspace_channels WHERE channel_id = ?`,
+  )
   const listMounts = db.prepare(
     `SELECT channel_id AS channelId FROM workspace_channels WHERE workspace_id = ? ORDER BY created_at ASC`,
   )
@@ -93,6 +96,12 @@ export function createUserRepo(db: DatabaseSync) {
     unmountChannel(workspaceId: string, channelId: string): void {
       deleteMount.run(workspaceId, channelId)
     },
+    /** channel 删除级联:清掉所有 workspace 对它的挂载引用(返回清理行数)。
+     *  早先没有这个级联,只能靠 listWorkspaces 在**读路径**里顺手 DELETE 兜底 ——
+     *  那让 GET /workspaces 变成有副作用的写操作。改为删除时主动收敛。 */
+    unmountChannelEverywhere(channelId: string): number {
+      return Number(deleteMountsByChannel.run(channelId).changes ?? 0)
+    },
     listMountedChannels(workspaceId: string): string[] {
       return (listMounts.all(workspaceId) as Array<{ channelId: string }>).map(r => r.channelId)
     },

package/server/services/workshop/dcw/dcw-controller.ts

@@ -372,9 +372,31 @@ class DcwController {
       node.transform = normalizeDataTransform(patch.transform)
     }
     if (patch.unit !== undefined) node.unit = patch.unit
-    if (patch.decimals !== undefined) node.decimals = patch.decimals
-    if (patch.min !== undefined) node.min = patch.min
-    if (patch.max !== undefined) node.max = patch.max
+    // decimals 必须 0..6 整数:它同时喂给 toFixed() —— 越界值会在**驱动已经写进 PLC 之后**
+    // 由 applyWriteResult/applyReadResult 抛 RangeError,造成「硬件已改、账本无记录、
+    // 调用方吃 500」(审计实测:decimals=101 → PLC 写入成功、无锚点、无写历史)。
+    // 在校验入口一次性拒掉,别让脏值走到写路径。
+    if (patch.decimals !== undefined) {
+      const d = Number(patch.decimals)
+      if (!Number.isInteger(d) || d < 0 || d > 6) {
+        throw new AppError(400, ErrorCodes.VALIDATION_ERROR, `decimals 必须为 0..6 的整数(当前: ${String(patch.decimals)})`)
+      }
+      node.decimals = d
+    }
+    if (patch.min !== undefined) {
+      const v = Number(patch.min)
+      if (!Number.isFinite(v)) throw new AppError(400, ErrorCodes.VALIDATION_ERROR, `min 必须为有限数字(当前: ${String(patch.min)})`)
+      node.min = v
+    }
+    if (patch.max !== undefined) {
+      const v = Number(patch.max)
+      if (!Number.isFinite(v)) throw new AppError(400, ErrorCodes.VALIDATION_ERROR, `max 必须为有限数字(当前: ${String(patch.max)})`)
+      node.max = v
+    }
+    // 量程交叉校验放在两者都应用之后:min>max 会让 writeTolerance 变负、越界判定恒真
+    if (node.max < node.min) {
+      throw new AppError(400, ErrorCodes.VALIDATION_ERROR, `量程无效:max(${node.max}) 不得小于 min(${node.min})`)
+    }
     if (patch.holdIntervalMs !== undefined) {
       node.holdIntervalMs = patch.holdIntervalMs == null ? null : Math.max(0, Math.min(3_600_000, Math.round(patch.holdIntervalMs)))
       rearm = true
@@ -480,15 +502,17 @@ class DcwController {
     return node
   }
 
-  /** 设备删除级联解绑(device-twins 删除路由调用) */
+  /** 设备删除级联解绑(device-twins 删除路由调用)。flushNow 移出循环(同 daq 侧) */
   unbindDevice(deviceId: string): void {
+    let touched = false
     for (const node of this.repo.all()) {
       if (!node.deviceIds.includes(deviceId)) continue
       node.deviceIds = node.deviceIds.filter(d => d !== deviceId)
       node.deviceBindingId = node.deviceIds[0] ?? null
-      this.repo.flushNow()
+      touched = true
       this.emitNodeChanged('updated', node)
     }
+    if (touched) this.repo.flushNow()
   }
 
   // ---------- 写命令(上位机核心操作)----------
@@ -972,8 +996,10 @@ class DcwController {
     await tsdbReady
     const bucketMs = endMs - startMs > 60_000 ? Math.max(1000, Math.round((endMs - startMs) / 200)) : undefined
     const daq = [] as Array<{ templateRef: string, nodeId: string, nodeName: string, ch: string, unit: string, latest: number | null, avg: number | null, min: number | null, max: number | null, cnt: number }>
-    // 并行查询(tsdb 往返一次/节点,互不依赖):N+1 串行会让 50 节点产线的批次视图放大 50 倍时延
-    const results = await Promise.all(getDaqNodeRepo().all().map(async (node) => {
+    // 并行查询(tsdb 往返一次/节点,互不依赖):N+1 串行会让 50 节点产线的批次视图放大 50 倍时延。
+    // 只查本批次产线的节点:批次归属单线,跨线节点的 tsdb 往返是纯浪费(节点数增长线性放大)
+    const lineNodes = getDaqNodeRepo().all().filter(n => !run.lineId || n.lineId === run.lineId)
+    const results = await Promise.all(lineNodes.map(async (node) => {
       try {
         const points = await getTsdb().query(node.id, { fromMs: startMs, toMs: endMs, bucketMs, limit: 500 })
         return { node, points }

package/server/services/workshop/dcw/dcw-line.repo.ts

@@ -6,12 +6,14 @@
  */
 
 import { randomUUID } from 'node:crypto'
-import path from 'node:path'
+import { join } from 'node:path'
 import { dcwLineColorFor, type LineInput, type LineView } from '../../../../shared/dcw-protocol'
+import { ensureDataDir } from '@/shared/config/home.mjs'
 import { AppError, ErrorCodes } from '../../../utils/errors'
-import { loadJsonFile, saveJsonFileAtomic } from '../json-store.mjs'
+import { loadJsonFile, saveJsonFileAtomic } from '../json-store.mjs'
 
-const DB_PATH = process.cwd().endsWith('server') ? 'data/dcw-lines.json' : path.join(process.cwd(), 'server', 'data', 'dcw-lines.json')
+// 配置根 .AgentWorkShop/data（ensureDataDir 自动迁移旧 cwd/server/data 位置）
+const DB_PATH = join(ensureDataDir(), 'dcw-lines.json')
 
 function load(): LineView[] {
   const parsed = loadJsonFile(DB_PATH, [])

package/server/services/workshop/dcw/dcw-node.repo.ts

@@ -5,15 +5,15 @@
  */
 
 import { createLogger } from '../logger'
-import path from 'node:path'
+import { join } from 'node:path'
 import { DcwNode } from './dcw-node'
+import { ensureDataDir } from '@/shared/config/home.mjs'
 import { loadJsonFile, saveJsonFileAtomic } from '../json-store.mjs'
 
 const log = createLogger('dcw.node-repo')
 
-const DB_PATH = process.cwd().endsWith('server')
-  ? 'data/dcws.json'
-  : path.join(process.cwd(), 'server', 'data', 'dcws.json')
+// 配置根 .AgentWorkShop/data（ensureDataDir 自动迁移旧 cwd/server/data 位置）
+const DB_PATH = join(ensureDataDir(), 'dcws.json')
 
 function load(): DcwNode[] {
   try {

package/server/services/workshop/dcw/dcw-product.repo.ts

@@ -3,18 +3,18 @@
  * 产品 = 数据隔离顶层维度:一个产品多个配方;批次/样本逐条携带 productId。
  */
 
-import { createLogger } from '../logger'
-import path from 'node:path'
+import { createLogger } from '../logger'
+import { join } from 'node:path'
 import { randomUUID } from 'node:crypto'
 import type { ProductInput, ProductView } from '../../../../shared/dcw-protocol'
+import { ensureDataDir } from '@/shared/config/home.mjs'
 import { AppError, ErrorCodes } from '../../../utils/errors'
-import { loadJsonFile, saveJsonFileAtomic } from '../json-store.mjs'
+import { loadJsonFile, saveJsonFileAtomic } from '../json-store.mjs'
 
-const log = createLogger('dcw.product-repo')
-
-const DB_PATH = process.cwd().endsWith('server')
-  ? 'data/dcw-products.json'
-  : path.join(process.cwd(), 'server', 'data', 'dcw-products.json')
+const log = createLogger('dcw.product-repo')
+
+// 配置根 .AgentWorkShop/data（ensureDataDir 自动迁移旧 cwd/server/data 位置）
+const DB_PATH = join(ensureDataDir(), 'dcw-products.json')
 
 function load(): ProductView[] {
   try {