agentworkshop 0.8.6 → 0.8.8

package/server/services/workshop/daq/storage/timescale.adapter.ts

@@ -29,7 +29,16 @@ export class TimescaleAdapter implements TsdbPort {
 
   async init(): Promise<void> {
     const mod = requirePg('pg') as typeof import('pg')
-    this.pool = new mod.Pool({ connectionString: this.url, max: 4 })
+    // 连接池/语句超时来自 daq.timescale.*(生产调优;异常查询不得长期占用连接)
+    const s = daqRuntimeSettings()
+    const timeoutMs = Math.max(1000, Math.min(300_000, Number(s.timescale.statementTimeoutMs)))
+    // pg 运行期支持 statement_timeout(连接参数),但 @types/pg 未声明 → 收窄后传入
+    const poolConfig = {
+      connectionString: this.url,
+      max: Math.max(1, Math.min(32, Number(s.timescale.poolMax))),
+      statement_timeout: timeoutMs,
+    } as unknown as import('pg').PoolConfig
+    this.pool = new mod.Pool(poolConfig)
     await this.pool.query(`
       CREATE TABLE IF NOT EXISTS daq_samples (
         node_id text        NOT NULL,
@@ -102,12 +111,13 @@ export class TimescaleAdapter implements TsdbPort {
   }
 
   private async sweepRetention(): Promise<void> {
-    if (!this.pool) return
+    // 2026-10-30 修复:门控此前错挂在样本保留期(retentionH)上;本函数只清帧,
+    // 应按帧保留期判定(0 = 永久保留帧,不清理)。
+    if (!this.pool || this.frameRetentionH <= 0) return
+    // 样本保留期由 TimescaleRollupStore 统一执行:它感知逐产线覆写(含"保留更久"的线),
+    // 而全局 drop_chunks 会按全局保留期整块丢弃,误删长保留产线的数据。
+    // 这里只保留帧元数据的全局清理(帧保留期不支持逐线覆写)。
     try {
-      await this.pool.query(
-        `SELECT drop_chunks('daq_samples', older_than => now() - ($1 || ' hours')::interval)`,
-        [this.retentionH],
-      )
       await this.pool.query(
         `SELECT drop_chunks('daq_frames', older_than => now() - ($1 || ' hours')::interval)`,
         [this.frameRetentionH],
@@ -146,12 +156,14 @@ export class TimescaleAdapter implements TsdbPort {
        ON CONFLICT (node_id, ts) DO NOTHING
        RETURNING node_id, ts`,
       [col.node, col.ts, col.value, col.state, col.line, col.product, col.recipe, col.run, col.ingested],
-    ) as { rows: Array<{ node_id: string, ts: string }> }
+    ) as { rows: Array<{ node_id: string, ts: string | Date }> }
     // 首写胜出:被忽略的键里,值/状态不同则计数(可观测,不静默更正)
-    const insertedKeys = new Set(inserted.rows.map(r => `${r.node_id}\u0000${r.ts}`))
+    // 2026-10-30 修复:pg 对 timestamptz RETURNING 回传 Date 对象(类型标注不改变运行时),
+    // 旧实现用 toISOString() 拼 key 恒不匹配 → 冲突快路径失效,每行多打一次 SELECT
+    // (写入热路径 N+1)。统一用 epoch ms 做 key,两侧行同源。
+    const insertedKeys = new Set(inserted.rows.map(r => `${r.node_id}\u0000${tsToMs(r.ts)}`))
     for (const r of rows) {
-      const key = `${r.nodeId}\u0000${new Date(r.tsMs).toISOString()}`
-      if (insertedKeys.has(key)) continue
+      if (insertedKeys.has(`${r.nodeId}\u0000${r.tsMs}`)) continue
       const prev = await this.pool.query('SELECT value, state FROM daq_samples WHERE node_id = $1 AND ts = $2', [r.nodeId, new Date(r.tsMs).toISOString()]) as { rows: Array<{ value: number, state: string }> }
       const row = prev.rows[0]
       if (row && (Math.abs(Number(row.value) - Number(r.value)) > 1e-9 || String(row.state) !== String(r.state))) this.correctionDropped++
@@ -252,6 +264,10 @@ export class TimescaleAdapter implements TsdbPort {
    * 训练/导出在 Timescale 上退化为 query() 截断)。
    * 语义与 sqlite.adapter 完全一致:稳定 ASC、游标严格大于 afterTs、每页限额、
    * done=false 时 nextCursor = 本页最后 at。
+   * 2026-10-29:支持批次标签等值过滤(line/product/recipe/run;SQL 全静态参数化,
+   * 参数为 NULL 即不过滤该列;提供值即等值过滤,天然排除未打标 NULL 行)——
+   * 此前训练分页不过滤标签,在 Timescale 上会把窗口内全部样本当目标批次,
+   * 产品隔离退化为时间窗推断(串味隐患)。
    */
   async queryTaggedPage(nodeId: string, q: import('./tsdb-port').TsdbPageQuery = {}): Promise<import('./tsdb-port').TsdbPage> {
     if (!this.pool) return { nodeId, points: [], nextCursor: null, done: true }
@@ -259,14 +275,31 @@ export class TimescaleAdapter implements TsdbPort {
     const from = Math.max(q.fromMs ?? 0, (q.afterTs ?? -1) + 1)
     const to = q.toMs ?? Date.now()
     const withState = q.withState !== false
-    const args = [nodeId, new Date(from).toISOString(), new Date(to).toISOString(), limit]
+    const tagOf = (v: string | undefined): string | null => (v != null && String(v).trim() !== '' ? String(v) : null)
+    const args: unknown[] = [
+      nodeId, new Date(from).toISOString(), new Date(to).toISOString(),
+      tagOf(q.lineId), tagOf(q.productId), tagOf(q.recipeId), tagOf(q.runId), limit,
+    ]
+    // 静态 SQL:占位符 $4..$7 为 NULL 时该列不过滤,提供值时等值过滤(NULL 行不匹配)
     const rows = withState
       ? (await this.pool.query(
-          'SELECT ts, value, state, ingested_at FROM daq_samples WHERE node_id = $1 AND ts >= $2 AND ts <= $3 ORDER BY ts ASC LIMIT $4',
+          `SELECT ts, value, state, ingested_at FROM daq_samples
+           WHERE node_id = $1 AND ts >= $2 AND ts <= $3
+             AND ($4::text IS NULL OR line_id = $4)
+             AND ($5::text IS NULL OR product_id = $5)
+             AND ($6::text IS NULL OR recipe_id = $6)
+             AND ($7::text IS NULL OR run_id = $7)
+           ORDER BY ts ASC LIMIT $8`,
           args,
         )).rows
       : (await this.pool.query(
-          'SELECT ts, value FROM daq_samples WHERE node_id = $1 AND ts >= $2 AND ts <= $3 ORDER BY ts ASC LIMIT $4',
+          `SELECT ts, value FROM daq_samples
+           WHERE node_id = $1 AND ts >= $2 AND ts <= $3
+             AND ($4::text IS NULL OR line_id = $4)
+             AND ($5::text IS NULL OR product_id = $5)
+             AND ($6::text IS NULL OR recipe_id = $6)
+             AND ($7::text IS NULL OR run_id = $7)
+           ORDER BY ts ASC LIMIT $8`,
           args,
         )).rows
     const points = (rows as Array<{ ts: string, value: number, state?: string | null, ingested_at?: string | number | null }>).map(r => ({

package/server/services/workshop/daq/storage/tsdb-port.ts

@@ -75,6 +75,15 @@ export interface TsdbPageQuery {
   limit?: number
   /** 置 false 可省 state 列(默认 true) */
   withState?: boolean
+  /**
+   * 批次标签过滤(2026-10-29):任一提供即按**等值**过滤(等值天然排除未打标 NULL 行)。
+   * 训练/导出等按批次隔离取数的调用方必须传三元组(line/product/recipe[+run]);
+   * 留空 = 纯时间窗原始分页(可用性水位等全量观测语义)。
+   */
+  lineId?: string
+  productId?: string
+  recipeId?: string
+  runId?: string
 }
 
 export interface TsdbPage {
@@ -90,9 +99,104 @@ export interface TsdbStats {
   correctionDropped: number
 }
 
+// ===== 按产线分表(生产布局)=====
+
+/** 布局形态:single = 升级前的单表;per-line = 一产线一表(物理隔离 + 独立保留期) */
+export type TsdbLayoutKind = 'single' | 'per-line'
+
+/** 单条产线的物理表与保留策略 */
+export interface TsdbLineTableInfo {
+  /** 空 = 未分配产线(哨兵行为 __unassigned__) */
+  lineId: string
+  samplesTable: string
+  framesTable?: string | null
+  rollupTable?: string | null
+  rawRetentionH: number
+  minuteRetentionH?: number | null
+  hourRetentionH?: number | null
+  /** 近似行数(pg_class.reltuples;SQLite 为 NULL) */
+  rows?: number | null
+  /** 物理占用字节(pg_total_relation_size) */
+  bytes?: number | null
+}
+
+/** 存储布局/保留期/rollup 水位(运维可见面;不含数据本身) */
+export interface TsdbLayoutInfo {
+  backend: string
+  layout: TsdbLayoutKind
+  rawRetentionH: number
+  frameRetentionH: number
+  rollup: {
+    enabled: boolean
+    minuteRetentionH: number
+    /** 0 = 永久 */
+    hourRetentionH: number
+    /** key = `${lineId}|${resMs}`;value = 已聚合到的上界 epoch ms */
+    watermarks: Record<string, number>
+  }
+  lines: TsdbLineTableInfo[]
+  legacy?: { table: string, rows: number | null, bytes: number | null, migrated: boolean } | null
+}
+
+/** 历史查询(跨原始保留期自动走 rollup;起止时间必填) */
+export interface TsdbHistoryQuery {
+  fromMs: number
+  toMs: number
+  /** 期望桶宽(≥100ms;1 分钟以上可命中 rollup) */
+  bucketMs: number
+  limit?: number
+  lineId?: string
+  productId?: string
+  recipeId?: string
+  runId?: string
+}
+
+export interface TsdbHistoryResult {
+  points: TsdbPoint[]
+  /** raw = 精确样本;minute/hour = 聚合历史;mixed = 边界两侧拼接 */
+  resolution: 'raw' | 'minute' | 'hour' | 'mixed'
+  /** 实际可查范围与说明(超出保留期必须显式说明,不得静默返回空) */
+  coverage: {
+    requestedFromMs: number
+    requestedToMs: number
+    /** 实际最早可查时刻(null = 无数据/未知) */
+    availableFromMs: number | null
+    /** 原始样本可用下界(该时刻之前只有聚合) */
+    rawFromMs: number
+    truncated: boolean
+    note: string
+  }
+}
+
+export interface TsdbRollupResult {
+  at: string
+  lines: number
+  minuteRows: number
+  hourRows: number
+  /** 因单次预算/窗口未闭合而留待下轮的水位说明 */
+  notes: string[]
+}
+
+export interface TsdbRetentionPatch {
+  rawRetentionH?: number | null
+  minuteRetentionH?: number | null
+  hourRetentionH?: number | null
+}
+
+/** 轻量保留期信息(Agent 工具输出用,不做任何统计扫描) */
+export interface TsdbRetentionInfo {
+  rawRetentionH: number
+  rollupEnabled: boolean
+  rollupResolutions: number[]
+}
+
 export interface TsdbPort {
   /** 后端标识(sqlite-emulated | timescale)——REST meta 直报前端 */
   readonly backend: string
+  /** 布局形态(single | per-line;缺省 single) */
+  readonly layout?: TsdbLayoutKind
+  /** 单批建议行数(生产者攒批上限;缺省由调用方决定) */
+  readonly writeBatchHint?: number
   init(): Promise<void>
   writeSamples(rows: DaqSampleRow[]): Promise<void>
   query(nodeId: string, opts: DaqQueryOpts): Promise<TsdbPoint[]>
@@ -108,6 +212,22 @@ export interface TsdbPort {
   // ===== 多形态帧(v2:向量/图像;像素 blob 在对象存储,这里只存元数据+派生指标)=====
   writeFrames(rows: DaqFrameRow[]): Promise<void>
   queryFrames(nodeId: string, opts: DaqFrameQueryOpts): Promise<DaqFrameRecord[]>
+  // ===== 按产线分表后的运维面(可选实现:后端不支持时调用方降级)=====
+  /** 布局/保留期/rollup 水位 */
+  layoutInfo?(): Promise<TsdbLayoutInfo>
+  /** 轻量保留期信息(不扫表) */
+  retentionInfo?(): TsdbRetentionInfo
+  /**
+   * 历史查询:窗口早于原始保留期时自动改走聚合表。
+   * 未实现的后端由调用方退化为 query()/queryTagged()。
+   */
+  queryHistory?(nodeId: string, q: TsdbHistoryQuery): Promise<TsdbHistoryResult>
+  /** 增量聚合(幂等:仅处理已闭合窗口;返回本轮处理量) */
+  runRollup?(): Promise<TsdbRollupResult>
+  /** 逐产线保留期覆写(null = 回到全局缺省) */
+  setLineRetention?(lineId: string, patch: TsdbRetentionPatch): Promise<void>
+  /** 立即执行一轮保留期清理(运维/测试;定时器本身仍按 30min 跑) */
+  sweepNow?(): Promise<void>
   /** 释放底层连接池/句柄(rebuild 换池时由工厂调用;可选) */
   close?(): Promise<void> | void
 }

package/server/services/workshop/db/chat-message.repo.ts

@@ -229,5 +229,33 @@ export function createChatMessageRepo(db: DatabaseSync) {
          WHERE channel_id = ? AND target_agent_id = ? AND status IN ('pending', 'delivered')`,
       ).run(new Date().toISOString(), channelId, targetAgentId).changes ?? 0)
     },
+
+    /** 崩溃恢复扫描(整改计划 2026-10-30 阶段2):pending 且 mailbox 未落库、到期可重试的投递 */
+    listRetryableDeliveries(nowIso: string, limit = 50): Array<{ id: string, chatMessageId: string, channelId: string, targetAgentId: string, attempts: number }> {
+      return db.prepare(
+        `SELECT id, chat_message_id AS chatMessageId, channel_id AS channelId, target_agent_id AS targetAgentId, attempts
+         FROM chat_deliveries
+         WHERE status = 'pending' AND mailbox_message_id IS NULL
+           AND (next_retry_at IS NULL OR next_retry_at <= ?)
+         ORDER BY created_at ASC LIMIT ?`,
+      ).all(nowIso, limit) as unknown as Array<{ id: string, chatMessageId: string, channelId: string, targetAgentId: string, attempts: number }>
+    },
+
+    /** 记一次重试(claim + 退避排程);返回累计 attempts */
+    bumpDeliveryRetry(id: string, nextRetryAt: string): number {
+      const now = new Date().toISOString()
+      db.prepare(
+        `UPDATE chat_deliveries SET attempts = attempts + 1, next_retry_at = ?, claimed_at = ?, updated_at = ? WHERE id = ?`,
+      ).run(nextRetryAt, now, now, id)
+      const row = db.prepare(`SELECT attempts FROM chat_deliveries WHERE id = ?`).get(id) as { attempts?: number } | undefined
+      return Number(row?.attempts ?? 0)
+    },
+
+    /** 重试超限转死信(failed;仅 pending 可转) */
+    markDeliveryDead(id: string, error: string): boolean {
+      return db.prepare(
+        `UPDATE chat_deliveries SET status = 'failed', error = ?, updated_at = ? WHERE id = ? AND status = 'pending'`,
+      ).run(error, new Date().toISOString(), id).changes > 0
+    },
   }
 }

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

@@ -42,6 +42,12 @@ export function initWorkshopDb(db: DatabaseSync): void {
   migrateAddColumn(db, 'channels', 'llm_json', 'TEXT NOT NULL DEFAULT \'\'')
   // 派发路由理由(koda RouteDecision 借鉴):lead 派发留痕"为什么派给他",供审计与前端呈现
   migrateAddColumn(db, 'tasks', 'route_reason', 'TEXT NOT NULL DEFAULT \'\'')
+  // 2026-10-30 子任务显式替代关系:FAILED/CANCELED 分支被 replacement 覆盖后不再阻塞父 Goal 收口
+  migrateAddColumn(db, 'tasks', 'superseded_by_task_id', 'TEXT')
+  // 2026-10-30 chat 投递崩溃恢复:pending 且 mailbox 未落库的投递由补偿器退避重投(阶段2)
+  migrateAddColumn(db, 'chat_deliveries', 'attempts', 'INTEGER NOT NULL DEFAULT 0')
+  migrateAddColumn(db, 'chat_deliveries', 'next_retry_at', 'TEXT')
+  migrateAddColumn(db, 'chat_deliveries', 'claimed_at', 'TEXT')
   // v10:模板可见性列(既有库补列;默认 private,仅属主可见)
   migrateAddColumn(db, 'agents', 'visibility', 'TEXT NOT NULL DEFAULT \'private\'')
   migrateAddColumn(db, 'teams', 'visibility', 'TEXT NOT NULL DEFAULT \'private\'')

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

@@ -322,6 +322,8 @@ export interface TaskRow {
   sourceChatDeliveryId: string | null
   closeReason: string | null
   deadlineAt: string | null
+  /** 显式替代关系:本任务(FAILED/CANCELED)被 replacement 覆盖 */
+  supersededByTaskId?: string | null
   /** 执行交接代次(§5.1;每次重分配 +1) */
   assignmentGeneration: number
   /** 当前执行租约(§5.1/§5.2;旧 worker 迟到事件据此丢弃) */

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

@@ -83,6 +83,8 @@ CREATE TABLE IF NOT EXISTS tasks (
   source_chat_delivery_id TEXT,
   close_reason   TEXT,
   deadline_at    TEXT,
+  -- 显式替代关系:FAILED/CANCELED 分支被 replacement 覆盖后仍保留失败事实,但不再阻塞父任务收口
+  superseded_by_task_id TEXT,
   -- 执行交接栅栏(§2.1/§5.1):每次新分配/重分配 generation+1 并换新 lease;
   -- worker 的 message/artifact/progress/complete/failed 事件携带 generation+lease,
   -- 与当前值不符的迟到事件被丢弃,避免旧 worker 覆盖重分配后的新执行结果。
@@ -460,10 +462,16 @@ CREATE TABLE IF NOT EXISTS chat_deliveries (
   mailbox_message_id  TEXT,
   status              TEXT NOT NULL DEFAULT 'pending', -- pending|delivered|consumed|failed|cancelled
   error               TEXT NOT NULL DEFAULT '',
+  -- 崩溃恢复(整改计划 2026-10-30 阶段2):pending 且 mailbox 未落库的投递由补偿器重投;
+  -- attempts 退避,超限转 failed 死信;mailbox 已落库的 pending(第二窗口)按确定性消息 id 去重,不重复投递
+  attempts            INTEGER NOT NULL DEFAULT 0,
+  next_retry_at       TEXT,
+  claimed_at          TEXT,
   created_at          TEXT NOT NULL,
   updated_at          TEXT NOT NULL,
   UNIQUE(chat_message_id, target_agent_id)
 );
+CREATE INDEX IF NOT EXISTS idx_chat_deliveries_retry ON chat_deliveries(status, next_retry_at);
 CREATE INDEX IF NOT EXISTS idx_chat_deliveries_message ON chat_deliveries(chat_message_id);
 CREATE INDEX IF NOT EXISTS idx_chat_deliveries_agent ON chat_deliveries(target_agent_id, status);
 -- Agent 回复关联链热路径:agentReplyToChat 按 mailboxMessageId 反查群聊来源

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

@@ -48,7 +48,7 @@ export const DEFAULT_CHANNEL_TEMPLATES: Array<{
 
 ## 团队分工(每位成员按此定位协作)
 - 生产主管(lead):理解用户优化目标 → 拆解为「数据分析」与「工艺调整」子任务并派发 → 复核成员结论(必须带数据证据)→ 汇总汇报;不直接操作节点,协调跨成员信息(如把数据分析师的越限发现转给工艺工程师处置)。
-- 工艺工程师(worker):持有数控节点授权 —— 用 daq_query 取证 → 判定设定-响应关系 → 在「安全量程 ∩ 配方窗口」内小步幅下发 dcw_control(单变量调整);手动确认模式节点先说明理由等用户批准;调整后等待热惯性响应再复测。
+- 工艺工程师(worker):持有数控节点授权 —— 用 daq_query 取证 → 判定设定-响应关系 → 参数调整一律走配方链路:在「安全量程 ∩ 配方窗口」内定小步幅候选(单变量调整),recipe_trial 整批试验,达标后 recipe_update 固化、recipe_apply 整批下发,劣化用 recipe_rollback(dispatch=true) 统一回退;手动确认模式先说明理由等用户批准;调整后等待热惯性响应再复测。
 - 数据分析师(worker):持有数采节点授权 —— 用 daq_query 获取时序数据(支持按产品/配方/时间窗过滤),输出趋势/均值/极值/越限统计与工况判读(区分设定-响应滞后与真实异常);发现越限或异常趋势第一时间用 send_message_to_agent 通报 lead 与工艺工程师。
 协作规则:引用数值必带单位与时间窗;结论不确定时先补数据再下判断;所有阶段性结论沉淀为任务交付物。
 
@@ -56,7 +56,7 @@ export const DEFAULT_CHANNEL_TEMPLATES: Array<{
 数据先行(先看数再动手)→ 窗口内小步幅(单次 ≤ 量程 2%)→ 单变量调整 → 复测闭环(调整后等待工艺响应再评估,不连续大幅调整)→ 异常先判读再行动 → 结论必附数据证据。任何成员不得越权操作未绑定节点。`,
     members: [
       { templateId: 'tpl-default-lead', role: 'lead' },
-      { inline: { name: '工艺工程师', harness: 'omp', config: { rpcMode: 'rpc', systemPromptPrefix: '你是光学薄膜涂布产线的工艺工程师,精通烘箱多段温控与涂层质量的关联。操作数控节点的标准流程:my_industrial_nodes 理解节点 → daq_query 取证 → dcw_control 在配方窗口内小步幅下发 → 等待热惯性后复测。手动确认模式下发前必须说明理由。结论一律引用带单位与时间窗的数据。' } }, role: 'worker' },
+      { inline: { name: '工艺工程师', harness: 'omp', config: { rpcMode: 'rpc', systemPromptPrefix: '你是光学薄膜涂布产线的工艺工程师,精通烘箱多段温控与涂层质量的关联。操作数控节点的标准流程:my_industrial_nodes 理解节点 → daq_query 取证 → 在配方窗口内定小步幅候选(单变量)→ recipe_trial 整批试验 → 等待热惯性后复测,达标 recipe_update 固化并 recipe_apply 整批下发,劣化 recipe_rollback(dispatch=true) 统一回退。手动确认模式下发前必须说明理由。结论一律引用带单位与时间窗的数据。' } }, role: 'worker' },
       { inline: { name: '数据分析师', harness: 'omp', config: { rpcMode: 'rpc', systemPromptPrefix: '你是光学薄膜涂布产线的数据分析师,擅长时序数据判读与工况诊断。用 daq_query 获取授权数采节点数据,输出趋势/统计/越限分析,判读时结合同产线数控设定考虑设定-响应滞后;发现越限或异常趋势立即通报 lead 与工艺工程师。结论一律引用具体数值。' } }, role: 'worker' },
     ],
   },
@@ -78,7 +78,7 @@ export const DEFAULT_CHANNEL_TEMPLATES: Array<{
 数据先行 → 窗口内小步幅(单次 ≤ 量程 2%)→ 单变量调整(温度与转速禁止同时调)→ 复测闭环 → 压力异常优先判读(滤网堵塞倾向 vs 温度耦合 vs 真实波动)→ 结论必附数据证据。任何成员不得越权操作未绑定节点。`,
     members: [
       { templateId: 'tpl-default-lead', role: 'lead' },
-      { inline: { name: '工艺工程师', harness: 'omp', config: { rpcMode: 'rpc', systemPromptPrefix: '你是薄膜挤出流延产线的工艺工程师,精通熔体压力/温度耦合控制与流延质量。操作数控节点:my_industrial_nodes → daq_query 取证 → dcw_control 窗口内小步幅(温度与转速禁同调)→ 等响应后复测。结论引用带单位与时间窗的数据。' } }, role: 'worker' },
+      { inline: { name: '工艺工程师', harness: 'omp', config: { rpcMode: 'rpc', systemPromptPrefix: '你是薄膜挤出流延产线的工艺工程师,精通熔体压力/温度耦合控制与流延质量。操作数控节点走配方链路:my_industrial_nodes → daq_query 取证 → recipe_trial 窗口内小步幅整批试验(温度与转速禁同调,单变量)→ 等响应后复测,达标 recipe_update 固化并 recipe_apply 整批下发,劣化 recipe_rollback(dispatch=true) 统一回退。结论引用带单位与时间窗的数据。' } }, role: 'worker' },
       { inline: { name: '数据分析师', harness: 'omp', config: { rpcMode: 'rpc', systemPromptPrefix: '你是薄膜挤出流延产线的数据分析师,擅长熔体压力/温度时序判读。用 daq_query 输出趋势/统计/越限分析,识别基线漂移与周期波动;结合同线数控设定考虑耦合与滞后;异常立即通报 lead 与工艺工程师。结论引用具体数值。' } }, role: 'worker' },
     ],
   },
@@ -129,15 +129,15 @@ export const DEFAULT_CHANNEL_TEMPLATES: Array<{
 - 生产主管(lead):统一调配 —— 把 goal 拆解为「数据分析」「知识调优」「参数调整」子任务并派发(worker 缺节点权限时,用 team_grant_nodes 把 lead 已绑定的节点授予对应 worker,或在 dispatch_task 里带 grant_node_ids 随任务授予;只授予自己绑定的节点);复核成员结论(必须带数据证据);复测劣化时决策回退(指示工艺工程师 dcw_rollback / 配方回退);对拿不准的现象先查知识库(若已启用)。
 - 数据分析师(worker):**持续读数** —— 按固定节拍(每批次或每 5 分钟)daq_query 拉取授权数采节点时窗数据,滑动窗口对比输出趋势/均值/越限统计与工况判读;异常第一时间通报 lead、工艺工程师与知识调优工程师。
 - 知识调优工程师(worker):**知识↔数据 → 调优方案** —— 用 kb_agent(mode=sync) 检索知识库中本场景的机理知识与历史经验(检索词带产线/产品/关键物理量),结合数据分析师的判读结论,产出带知识依据的调优方案(参数/方向/步幅/预期响应,引用知识条目与数据);知识库未启用或检索为空时退化为纯数据驱动建议并说明。收口后用 kb_agent(mode=async) 沉淀「知识↔调整↔响应」经验入库。
-- 工艺工程师(worker):持有数控节点授权 —— daq_query 取证 → 按调优方案在「安全量程 ∩ 配方窗口 ∩ 单步限幅」内用 dcw_control/param_control 小步幅下发(单变量);manual 模式节点先说明理由等用户批准;调整后等工艺惯性响应再复测;复测劣化立即 dcw_rollback 回退本次调整并通报 lead。
+- 工艺工程师(worker):持有数控节点授权 —— daq_query 取证 → 按调优方案转成配方候选,在「安全量程 ∩ 配方窗口 ∩ 单步限幅」内 recipe_trial 整批试验(单变量小步幅);manual 模式节点先说明理由等用户批准;调整后等工艺惯性响应再复测;复测劣化立即 dcw_rollback 回退本次调整并通报 lead。
 
 ## 作业纪律
-数据先行 → 知识增强(检索在动手前)→ **产线上下文先行**(频道已绑定产线时,开工先用 line_context 读取绑定产线全景:运行状态/活动批次/当前配方与参数窗口 —— 频道绑定=只读授权,日志/配方史全员可查,写操作仍需节点授权)→ **参数变更一律走配方链路**:多参数候选 → recipe_trial 整批试验(一次带上所有要改的参数,**不写配方版本**;同线节拍 ≥5 分钟)→ 等工艺惯性后复测判读 → 有进步/达标 **recipe_update 固化**(同一配方 id,版本+1,更新而非新建)→ 需要正式落地用 recipe_apply 整批下发;无进步/劣化 **recipe_rollback(dispatch=true) 统一回退**(定义回退+PLC 整批恢复)→ 复测闭环 → 结论必附带单位与时间窗的数据证据;优化闭环内禁用 dcw_control/param_control 逐参数直调(防单参数震荡;AML 探索模式的小步激励除外);知识闭环 = 检索 → 增强 → 试验 → 判读 → 固化/回退 → 沉淀;不得操作未授权节点;治理审批与硬约束被拒时按指引降步重试,不绕行。`,
+数据先行 → 知识增强(检索在动手前)→ **产线上下文先行**(频道已绑定产线时,开工先用 line_context 读取绑定产线全景:运行状态/活动批次/当前配方与参数窗口 —— 频道绑定=只读授权,日志/配方史全员可查,写操作仍需节点授权)→ **参数变更一律走配方链路**:多参数候选 → recipe_trial 整批试验(一次带上所有要改的参数,**不写配方版本**;同线节拍 ≥5 分钟)→ 等工艺惯性后复测判读 → 有进步/达标 **recipe_update 固化**(同一配方 id,版本+1,更新而非新建)→ 需要正式落地用 recipe_apply 整批下发;无进步/劣化 **recipe_rollback(dispatch=true) 统一回退**(定义回退+PLC 整批恢复)→ 复测闭环 → 结论必附带单位与时间窗的数据证据;优化闭环内禁用逐参数直调(直写工具已禁用,防单参数震荡;AML 探索模式经 optimization_explore 的小步激励除外);知识闭环 = 检索 → 增强 → 试验 → 判读 → 固化/回退 → 沉淀;不得操作未授权节点;治理审批与硬约束被拒时按指引降步重试,不绕行。`,
     members: [
       { inline: { name: '生产主管', harness: 'omp', config: { rpcMode: 'rpc', role: 'lead', systemPromptPrefix: '你是通用闭环优化团队的生产主管(lead),统一调配三人(数据分析师/知识调优工程师/工艺工程师):把用户 goal 拆解为「数据分析」「知识调优」「参数调整」子任务并用 dispatch_task 派发(可带 grant_node_ids 随任务把你自己已绑定的节点授权给 worker,或用 team_grant_nodes 执行中授予;只能授予自己绑定的节点);复核成员结论(必须带数据证据);优化动作一律走配方链路(trial 试验 → 判读 → update 固化 / 统一回退),复测指标劣化时立即指示工艺工程师 recipe_rollback(dispatch=true) 统一回退后再调整方向;频道绑定的产线只优化其当前活动配方,不新建配方 id;若知识库已启用,开工前先用 kb_agent(mode=sync) 检索场景相关数据分析与物理机理,收口后用 kb_agent(mode=async) 沉淀本次经验(可交知识调优工程师执行)。' } }, role: 'lead' },
       { inline: { name: '数据分析师', harness: 'omp', config: { rpcMode: 'rpc', systemPromptPrefix: '你是通用闭环优化团队的数据分析师,负责持续读数与工况判读:按固定节拍(每批次或每 5 分钟,以任务要求为准)用 daq_query 拉取授权数采节点数据(按批次/时间窗),做滑动窗口对比(当前窗 vs 基线窗),输出趋势/均值/极值/越限统计;结合数控设定考虑设定-响应滞后;发现越限或异常趋势立即用 send_message_to_agent 通报 lead、工艺工程师与知识调优工程师。结论一律引用具体数值与时间窗。' } }, role: 'worker' },
       { inline: { name: '知识调优工程师', harness: 'omp', config: { rpcMode: 'rpc', systemPromptPrefix: '你是通用闭环优化团队的知识调优工程师,把「知识库经验」与「数据分析结论」转化为可执行调优方案:先用 kb_agent(mode=sync) 检索知识库中本场景的机理知识与历史经验(检索词带产线/产品/关键物理量);结合数据分析师的最新判读,产出带知识依据的调优方案(调哪个参数/方向/步幅/预期响应/依据哪条知识,引用知识条目与数据证据);知识库未启用或检索为空时退化为纯数据驱动建议并明确说明;方案交工艺工程师治理下发(或经 lead 授权后自行 param_control 小步验证);复测有效则用 kb_agent(mode=async) 把「知识↔调整↔响应」经验沉淀入库(异步提交后无需等待);复测劣化则立即建议回退并同步 lead。' } }, role: 'worker' },
-      { inline: { name: '工艺工程师', harness: 'omp', config: { rpcMode: 'rpc', systemPromptPrefix: '你是通用闭环优化团队的工艺工程师,按知识调优工程师的方案走配方链路执行:my_industrial_nodes 确认授权 → daq_query 取证 → 把方案转成 recipe_trial(多参数候选整批试验,一次带上所有要改的参数,带 hypothesis 声明依据;不写配方版本)→ 等待工艺惯性后 daq_query 复测 → 有进步/达标用 recipe_update 把候选值固化进配方(同 id 版本+1),需要正式落地用 recipe_apply 整批下发;无进步/劣化用 recipe_rollback(dispatch=true)统一回退(定义回退+PLC 整批恢复)。优化闭环内禁用 dcw_control/param_control 逐参数直调(防震荡)。你的节点授权可能由 lead 在派发时或执行中授予(权限变化以 my_industrial_nodes 实时查询为准)。manual 模式节点下发前说明理由等批准。结论一律引用带单位与时间窗的数据。' } }, role: 'worker' },
+      { inline: { name: '工艺工程师', harness: 'omp', config: { rpcMode: 'rpc', systemPromptPrefix: '你是通用闭环优化团队的工艺工程师,按知识调优工程师的方案走配方链路执行:my_industrial_nodes 确认授权 → daq_query 取证 → 把方案转成 recipe_trial(多参数候选整批试验,一次带上所有要改的参数,带 hypothesis 声明依据;不写配方版本)→ 等待工艺惯性后 daq_query 复测 → 有进步/达标用 recipe_update 把候选值固化进配方(同 id 版本+1),需要正式落地用 recipe_apply 整批下发;无进步/劣化用 recipe_rollback(dispatch=true)统一回退(定义回退+PLC 整批恢复)。优化闭环内禁用逐参数直调(直写工具已禁用,防震荡)。你的节点授权可能由 lead 在派发时或执行中授予(权限变化以 my_industrial_nodes 实时查询为准)。manual 模式节点下发前说明理由等批准。结论一律引用带单位与时间窗的数据。' } }, role: 'worker' },
     ],
   },
   {

package/server/services/workshop/db/schema.sql

@@ -81,6 +81,8 @@ CREATE TABLE IF NOT EXISTS tasks (
   source_chat_delivery_id TEXT,
   close_reason   TEXT,
   deadline_at    TEXT,
+  -- 显式替代关系:FAILED/CANCELED 分支被 replacement 覆盖后仍保留失败事实,但不再阻塞父任务收口
+  superseded_by_task_id TEXT,
   assignment_generation      INTEGER NOT NULL DEFAULT 0,  -- 执行交接代次(重分配 +1)
   execution_lease_id         TEXT,                         -- 当前执行租约 id(旧 worker 迟到事件据此丢弃)
   execution_lease_agent_id   TEXT,                         -- 租约持有者
@@ -243,10 +245,15 @@ CREATE TABLE IF NOT EXISTS chat_deliveries (
   mailbox_message_id  TEXT,
   status              TEXT NOT NULL DEFAULT 'pending', -- pending|delivered|consumed|failed|cancelled
   error               TEXT NOT NULL DEFAULT '',
+  -- 崩溃恢复(整改计划 2026-10-30 阶段2):pending 且 mailbox 未落库由补偿器重投;超限转 failed 死信
+  attempts            INTEGER NOT NULL DEFAULT 0,
+  next_retry_at       TEXT,
+  claimed_at          TEXT,
   created_at          TEXT NOT NULL,
   updated_at          TEXT NOT NULL,
   UNIQUE(chat_message_id, target_agent_id)
 );
+CREATE INDEX IF NOT EXISTS idx_chat_deliveries_retry ON chat_deliveries(status, next_retry_at);
 CREATE INDEX IF NOT EXISTS idx_chat_deliveries_message ON chat_deliveries(chat_message_id);
 CREATE INDEX IF NOT EXISTS idx_chat_deliveries_agent ON chat_deliveries(target_agent_id, status);
 -- Agent 回复关联链热路径:按 mailboxMessageId 反查群聊来源

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

@@ -7,7 +7,7 @@ import type { DatabaseSync } from 'node:sqlite'
 import type { TaskRow } from './database'
 
 const COLS
-  = 'id, channel_id AS channelId, parent_id AS parentId, root_queue_seq AS rootQueueSeq, assignee_id AS assigneeId, creator_id AS creatorId, title, description, state, progress, retry_count AS retryCount, artifacts_json AS artifactsJson, history_json AS historyJson, route_reason AS routeReason, source_chat_message_id AS sourceChatMessageId, source_chat_delivery_id AS sourceChatDeliveryId, close_reason AS closeReason, deadline_at AS deadlineAt, assignment_generation AS assignmentGeneration, execution_lease_id AS executionLeaseId, execution_lease_agent_id AS executionLeaseAgentId, execution_lease_started_at AS executionLeaseStartedAt, execution_lease_revoked_at AS executionLeaseRevokedAt, created_at AS createdAt, updated_at AS updatedAt'
+  = 'id, channel_id AS channelId, parent_id AS parentId, root_queue_seq AS rootQueueSeq, assignee_id AS assigneeId, creator_id AS creatorId, title, description, state, progress, retry_count AS retryCount, artifacts_json AS artifactsJson, history_json AS historyJson, route_reason AS routeReason, source_chat_message_id AS sourceChatMessageId, source_chat_delivery_id AS sourceChatDeliveryId, close_reason AS closeReason, deadline_at AS deadlineAt, superseded_by_task_id AS supersededByTaskId, assignment_generation AS assignmentGeneration, execution_lease_id AS executionLeaseId, execution_lease_agent_id AS executionLeaseAgentId, execution_lease_started_at AS executionLeaseStartedAt, execution_lease_revoked_at AS executionLeaseRevokedAt, created_at AS createdAt, updated_at AS updatedAt'
 
 const NON_TERMINAL_STATES = `'SUBMITTED', 'ASSIGNED', 'WORKING', 'WAITING'`
 
@@ -29,6 +29,7 @@ export interface TaskCreateInput {
   sourceChatDeliveryId?: string | null
   closeReason?: string | null
   deadlineAt?: string | null
+  supersededByTaskId?: string | null
   assignmentGeneration?: number
   executionLeaseId?: string | null
   executionLeaseAgentId?: string | null
@@ -51,6 +52,7 @@ export interface TaskPatch {
   sourceChatDeliveryId?: string | null
   closeReason?: string | null
   deadlineAt?: string | null
+  supersededByTaskId?: string | null
   assignmentGeneration?: number
   executionLeaseId?: string | null
   executionLeaseAgentId?: string | null
@@ -76,6 +78,7 @@ export interface TaskMetaRow {
   sourceChatDeliveryId: string | null
   closeReason: string | null
   deadlineAt: string | null
+  supersededByTaskId: string | null
   assignmentGeneration: number
   executionLeaseId: string | null
   executionLeaseAgentId: string | null
@@ -131,12 +134,12 @@ export function createTaskRepo(db: DatabaseSync) {
     `SELECT ${COLS} FROM tasks WHERE state IN (${NON_TERMINAL_STATES}) ORDER BY createdAt ASC, rowid ASC`,
   )
   const META_COLS
-    = 'id, channel_id AS channelId, parent_id AS parentId, root_queue_seq AS rootQueueSeq, assignee_id AS assigneeId, creator_id AS creatorId, title, description, state, progress, retry_count AS retryCount, route_reason AS routeReason, source_chat_message_id AS sourceChatMessageId, source_chat_delivery_id AS sourceChatDeliveryId, close_reason AS closeReason, deadline_at AS deadlineAt, assignment_generation AS assignmentGeneration, execution_lease_id AS executionLeaseId, execution_lease_agent_id AS executionLeaseAgentId, execution_lease_started_at AS executionLeaseStartedAt, execution_lease_revoked_at AS executionLeaseRevokedAt, created_at AS createdAt, updated_at AS updatedAt'
+    = 'id, channel_id AS channelId, parent_id AS parentId, root_queue_seq AS rootQueueSeq, assignee_id AS assigneeId, creator_id AS creatorId, title, description, state, progress, retry_count AS retryCount, route_reason AS routeReason, source_chat_message_id AS sourceChatMessageId, source_chat_delivery_id AS sourceChatDeliveryId, close_reason AS closeReason, deadline_at AS deadlineAt, superseded_by_task_id AS supersededByTaskId, assignment_generation AS assignmentGeneration, execution_lease_id AS executionLeaseId, execution_lease_agent_id AS executionLeaseAgentId, execution_lease_started_at AS executionLeaseStartedAt, execution_lease_revoked_at AS executionLeaseRevokedAt, created_at AS createdAt, updated_at AS updatedAt'
   const selectByChannelMeta = db.prepare(`SELECT ${META_COLS} FROM tasks WHERE channel_id = ? ORDER BY COALESCE(root_queue_seq, 2147483647) ASC, created_at ASC, rowid ASC`)
   const selectByChannelAssigneeMeta = db.prepare(`SELECT ${META_COLS} FROM tasks WHERE channel_id = ? AND assignee_id = ? ORDER BY created_at ASC, rowid ASC`)
   const selectChildrenMeta = db.prepare(`SELECT ${META_COLS} FROM tasks WHERE channel_id = ? AND parent_id = ? ORDER BY created_at ASC, rowid ASC`)
   const updateStmt = db.prepare(
-    `UPDATE tasks SET parent_id = ?, assignee_id = ?, creator_id = ?, title = ?, description = ?, state = ?, progress = ?, retry_count = ?, artifacts_json = ?, history_json = ?, source_chat_message_id = ?, source_chat_delivery_id = ?, close_reason = ?, deadline_at = ?, assignment_generation = ?, execution_lease_id = ?, execution_lease_agent_id = ?, execution_lease_started_at = ?, execution_lease_revoked_at = ?, updated_at = ? WHERE id = ?`,
+    `UPDATE tasks SET parent_id = ?, assignee_id = ?, creator_id = ?, title = ?, description = ?, state = ?, progress = ?, retry_count = ?, artifacts_json = ?, history_json = ?, source_chat_message_id = ?, source_chat_delivery_id = ?, close_reason = ?, deadline_at = ?, superseded_by_task_id = ?, assignment_generation = ?, execution_lease_id = ?, execution_lease_agent_id = ?, execution_lease_started_at = ?, execution_lease_revoked_at = ?, updated_at = ? WHERE id = ?`,
   )
 
   const buildRow = (input: TaskCreateInput, rootQueueSeq: number | null): TaskRow => {
@@ -272,6 +275,7 @@ export function createTaskRepo(db: DatabaseSync) {
         sourceChatDeliveryId: patch.sourceChatDeliveryId !== undefined ? patch.sourceChatDeliveryId : current.sourceChatDeliveryId,
         closeReason: patch.closeReason !== undefined ? patch.closeReason : current.closeReason,
         deadlineAt: patch.deadlineAt !== undefined ? patch.deadlineAt : current.deadlineAt,
+        supersededByTaskId: patch.supersededByTaskId !== undefined ? patch.supersededByTaskId : current.supersededByTaskId,
         assignmentGeneration: patch.assignmentGeneration ?? current.assignmentGeneration,
         executionLeaseId: patch.executionLeaseId !== undefined ? patch.executionLeaseId : current.executionLeaseId,
         executionLeaseAgentId: patch.executionLeaseAgentId !== undefined ? patch.executionLeaseAgentId : current.executionLeaseAgentId,
@@ -282,6 +286,7 @@ export function createTaskRepo(db: DatabaseSync) {
       updateStmt.run(
         next.parentId, next.assigneeId, next.creatorId, next.title, next.description, next.state, next.progress, next.retryCount,
         next.artifactsJson, next.historyJson, next.sourceChatMessageId, next.sourceChatDeliveryId, next.closeReason, next.deadlineAt,
+        next.supersededByTaskId ?? null,
         next.assignmentGeneration, next.executionLeaseId, next.executionLeaseAgentId, next.executionLeaseStartedAt, next.executionLeaseRevokedAt,
         next.updatedAt, id,
       )

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

@@ -12,6 +12,7 @@ import { findDcwTemplate } from '../dcw-templates'
 import { getDcwRecipeRepo } from '../dcw-recipe.repo'
 import { randomUUID } from 'node:crypto'
 import { resolveDcwDriver } from '../drivers'
+import { beginGovernanceDriverCall, endGovernanceDriverCall } from '../../agents/governance-harness-audit'
 
 /** 驱动未显式上报 ack 时的旧口径归一(ok+回读→设备证实;ok 无回读→链路受理;失败→未证实) */
 function normalizeAck(outcome: { ok: boolean, readback: number | null, ack?: WriteAckLevel }): WriteAckLevel {
@@ -32,6 +33,11 @@ export abstract class DcwControllerActuation extends DcwControllerState {
     const t = node.transform
     const plcValue = inverseTransform(eng, t)
     const plcTolerance = Math.max(tolerance, 1e-9) / (t?.kind === 'linear' && Math.abs(t.scale ?? 1) > 0 ? Math.abs(t.scale!) : 1)
+    const driverAudit = beginGovernanceDriverCall({
+      node_id: node.id, line_id: node.lineId ?? null, driver: node.driver,
+      requested_value: eng, plc_value: plcValue, tolerance: plcTolerance,
+      recipe_run_id: recipeRunId, driver_config: node.driverConfig,
+    })
     let outcome: { ok: boolean, message: string, raw: number | null, readback: number | null, ack?: WriteAckLevel }
     try {
       outcome = await resolveDcwDriver(node.driver).write({
@@ -87,6 +93,7 @@ export abstract class DcwControllerActuation extends DcwControllerState {
       at,
     })
     this.emitNodeChanged('updated', node)
+    endGovernanceDriverCall(driverAudit, outcome)
     return outcome
   }
 

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

@@ -13,6 +13,7 @@ import { getDcwLineRepo } from '../dcw-line.repo'
 import { getDcwParamRepo } from '../param-map.repo'
 import { requireDcwDriverKind } from '../drivers'
 import { assertDriverRawRange } from '../drivers/shared'
+import { validateMqttWriteConfig } from '../mqtt-payload'
 import { randomUUID } from 'node:crypto'
 
 function normalizeStepLimitInput(value: unknown): number | null | undefined {
@@ -60,6 +61,9 @@ export abstract class DcwControllerCrud extends DcwControllerControl {
     // S15:Modbus 原始量程必须在配置期可被 dataType 表示(越界运行期会拒绝写,但提前拦更好)
     try {
       assertDriverRawRange(node.driver, node.driverConfig)
+      // MQTT 下发载荷模板配置期校验:jsonTemplate 必须可渲染且渲染后是合法 JSON
+      // (坏模板若留到写入期才炸,一次下发失败会占用写锁/留下半套审计)
+      if (node.driver === 'mqtt') validateMqttWriteConfig(node.driverConfig)
     }
     catch (err) {
       throw new AppError(400, ErrorCodes.VALIDATION_ERROR, err instanceof Error ? err.message : String(err))
@@ -96,6 +100,7 @@ export abstract class DcwControllerCrud extends DcwControllerControl {
     // S15:配置期校验(失败回滚 driver/driverConfig,避免内存里留下半改状态)
     try {
       assertDriverRawRange(node.driver, node.driverConfig)
+      if (node.driver === 'mqtt') validateMqttWriteConfig(node.driverConfig)
     }
     catch (err) {
       node.driver = prevDriver