agentworkshop 0.7.34 → 0.7.38

package/server/api/workshop/aml/jobs/[id].delete.ts

@@ -0,0 +1,44 @@
+/**
+ * DELETE /api/workshop/aml/jobs/:id —— 删除作业(元数据行 + 实验谱系 + 实体目录)。
+ * 保护:运行中的作业拒绝删除(须先 cancel);已产出模型的作业拒绝删除(模型工件溯源依赖它)。
+ * 权限:admin/editor。
+ */
+import { getRouterParam } from 'h3'
+import { defineApiHandler } from '@/server/utils/response'
+import { requireRole } from '@/server/api/workshop/caller'
+import { AppError } from '@/server/utils/errors'
+import { getAmlRuntime } from '@/server/services/workshop/aml/runtime'
+import { removeEntityDir } from '@/server/services/workshop/aml/entity'
+import { recordOps } from '@/server/services/workshop/ops/ops'
+
+const ACTIVE = new Set(['queued', 'provisioning', 'training', 'evaluating'])
+
+export default defineApiHandler(async (event) => {
+  const user = requireRole(event, ['admin', 'editor'])
+  const id = getRouterParam(event, 'id') ?? ''
+  const rt = getAmlRuntime()
+  const job = rt.repo.job.get(id)
+  if (!job) throw new AppError(404, 'AML_JOB_MISSING', `作业 ${id} 不存在`)
+  if (ACTIVE.has(job.status)) {
+    throw new AppError(409, 'AML_JOB_ACTIVE', `作业 ${id} 仍在 ${job.status} 状态,请先取消(cancel)再删除。`)
+  }
+  const ds = rt.repo.dataset.get(job.datasetId)
+  if (ds) {
+    const expIds = rt.repo.experiment.listByDataset(job.datasetId, 1000).filter(e => e.jobId === id).map(e => e.id)
+    const models = rt.repo.model.list({ limit: 1000 }).filter(m => expIds.includes(m.experimentId))
+    if (models.length > 0) {
+      throw new AppError(409, 'AML_JOB_HAS_MODEL',
+        `作业 ${id} 已产出模型 ${models.length} 个(${models.map(m => m.id).join(', ')}),拒绝删除:模型的可复现性溯源依赖该作业记录。请先删除对应模型。`)
+    }
+  }
+  const ent = removeEntityDir(rt, 'job', id)
+  const exps = rt.repo.job.removeExperimentsOf(id)
+  rt.repo.job.remove(id)
+  recordOps({
+    actor: user.id, actorName: user.name, actorKind: 'user',
+    action: 'aml.job.delete', kind: 'system',
+    summary: `删除作业 ${id}(状态 ${job.status},级联实验 ${exps} 条)${ent.removed ? ',含实体目录' : ''}`,
+    targetKind: 'aml_job', targetId: id,
+  })
+  return { deleted: true, entityRemoved: ent.removed, experimentsRemoved: exps, dir: ent.dir }
+})

package/server/api/workshop/aml/models/[id].delete.ts

@@ -0,0 +1,35 @@
+/**
+ * DELETE /api/workshop/aml/models/:id —— 删除模型(元数据行 + 实体目录)。
+ * 保护:production 阶段的模型拒绝直接删除(在线预测依赖它),须先退役为 retired。
+ * 权限:admin/editor + 产线 operate。
+ */
+import { getRouterParam } from 'h3'
+import { defineApiHandler } from '@/server/utils/response'
+import { requireRole } from '@/server/api/workshop/caller'
+import { requireLineMode } from '@/server/services/workshop/permissions'
+import { AppError } from '@/server/utils/errors'
+import { getAmlRuntime } from '@/server/services/workshop/aml/runtime'
+import { removeEntityDir } from '@/server/services/workshop/aml/entity'
+import { recordOps } from '@/server/services/workshop/ops/ops'
+
+export default defineApiHandler(async (event) => {
+  const user = requireRole(event, ['admin', 'editor'])
+  const id = getRouterParam(event, 'id') ?? ''
+  const rt = getAmlRuntime()
+  const model = rt.repo.model.get(id)
+  if (!model) throw new AppError(404, 'AML_MODEL_MISSING', `模型 ${id} 不存在`)
+  requireLineMode(user, rt.repo.dataset.get(model.datasetId)?.lineId, 'operate')
+  if (model.stage === 'production') {
+    throw new AppError(409, 'AML_MODEL_IN_PRODUCTION',
+      `模型 ${id} 处于 production 阶段,在线预测依赖它,拒绝删除。请先晋升同组新模型或将其退役(retired)后再删。`)
+  }
+  const ent = removeEntityDir(rt, 'model', id)
+  rt.repo.model.remove(id)
+  recordOps({
+    actor: user.id, actorName: user.name, actorKind: 'user',
+    action: 'aml.model.delete', kind: 'system',
+    summary: `删除模型 ${id}(阶段 ${model.stage})${ent.removed ? ',含实体目录' : ''}`,
+    targetKind: 'aml_model', targetId: id,
+  })
+  return { deleted: true, entityRemoved: ent.removed, dir: ent.dir }
+})

package/server/api/workshop/aml/models/[id].patch.ts

@@ -0,0 +1,30 @@
+/**
+ * PATCH /api/workshop/aml/models/:id —— 更新模型元数据(备注)。
+ * body: { note: string }
+ * 阶段变更不走这里(必须走 /promote,那里有 ONNX 深检与同组唯一性保护)。
+ */
+import { getRouterParam, readBody } from 'h3'
+import { defineApiHandler } from '@/server/utils/response'
+import { requireRole } from '@/server/api/workshop/caller'
+import { requireLineMode } from '@/server/services/workshop/permissions'
+import { AppError } from '@/server/utils/errors'
+import { getAmlRuntime } from '@/server/services/workshop/aml/runtime'
+import { recordOps } from '@/server/services/workshop/ops/ops'
+
+export default defineApiHandler(async (event) => {
+  const user = requireRole(event, ['admin', 'editor'])
+  const id = getRouterParam(event, 'id') ?? ''
+  const body = await readBody<{ note?: string }>(event) ?? {}
+  const rt = getAmlRuntime()
+  const model = rt.repo.model.get(id)
+  if (!model) throw new AppError(404, 'AML_MODEL_MISSING', `模型 ${id} 不存在`)
+  requireLineMode(user, rt.repo.dataset.get(model.datasetId)?.lineId, 'operate')
+  if (typeof body.note === 'string') rt.repo.model.updateNote(id, body.note.slice(0, 2000))
+  recordOps({
+    actor: user.id, actorName: user.name, actorKind: 'user',
+    action: 'aml.model.update', kind: 'system',
+    summary: `更新模型 ${id} 元数据(备注)`,
+    targetKind: 'aml_model', targetId: id,
+  })
+  return { model: rt.repo.model.get(id) }
+})

package/server/plugins/dev-stability-guard.ts

@@ -1,63 +1,40 @@
-/**
- * Server 稳定性护栏:浏览器断开(WS/keep-alive socket 硬断)产生 ECONNRESET/EPIPE
- * 类异步异常 —— Node 默认策略直接击穿进程,拖垮所有并发调试/运行会话。
- *
- * 严重性过滤(关键设计):只抑制 socket 级断连噪声;其余异常保留致命语义
- * (记日志后按非零码退出)—— 安全插件的 fail-fast(如生产密钥校验失败拒绝启动)
- * 依赖异常中止进程,绝不能被护栏吞掉。
- */
-const SOCKET_NOISE = /ECONNRESET|EPIPE|ECONNABORTED|ETIMEDOUT|ERR_STREAM_PREMATURE_CLOSE|ERR_STREAM_WRITE_AFTER_END/
-// SQLite 忙/锁(hardening ST-1):多写竞争的瞬态错误,记录并继续;
-// 真正的库损坏/磁盘错误不经此路径,仍走致命语义。
-const SQLITE_BUSY = /SQLITE_BUSY|SQLITE_LOCKED|database is locked/i
-// Agent 引擎边界错误(harness 容错):opencode/omp/codex/dsh 等外部引擎 RPC 失败
-// (HTTP 4xx/5xx、连接拒绝)只影响对应 agent 会话,应降级重试而非退进程 ——
-// 服务端同时承载 DAQ 数采/DCW 写控,引擎故障带走监控平台属于可用性事故。
-const ENGINE_BOUNDARY = /\b(opencode|omp|codex|dsh) API\b/i
-
-function isSocketNoise(reason: unknown): boolean {
-  if (reason == null) return false
-  const err = reason as NodeJS.ErrnoException
-  // code 命中即噪声(经典 errno 形态)
-  if (typeof err.code === 'string' && SOCKET_NOISE.test(err.code)) return true
-  // undici 等包装错误:code 可能是 UND_ERR_*/undefined,但 message/stack 仍带
-  // "read ECONNRESET" 文本 —— 全文兜底再测一次(四引擎并发下 keep-alive socket
-  // 复位是高频瞬态,不可判死进程)
-  return SOCKET_NOISE.test(describe(reason))
-}
-
-function describe(reason: unknown): string {
-  return reason instanceof Error
-    ? `${reason.message}\n${reason.stack?.slice(0, 600) ?? ''}`
-    : String(reason)
-}
-
-export default defineNitroPlugin(() => {
-  const isTransientDbBusy = (reason: unknown): boolean => SQLITE_BUSY.test(describe(reason))
-  const isEngineBoundary = (reason: unknown): boolean => ENGINE_BOUNDARY.test(describe(reason).split('\n')[0] ?? '')
-  let engineFaults = 0
-  const onFatalCandidate = (kind: string, reason: unknown): void => {
-    if (isSocketNoise(reason)) {
-      console.warn(`[stability-guard] socket noise suppressed: ${describe(reason).slice(0, 120)}`)
-      return
-    }
-    if (isTransientDbBusy(reason)) {
-      console.warn(`[stability-guard] sqlite busy(瞬态,不退出): ${describe(reason).slice(0, 120)}`)
-      return
-    }
-    if (isEngineBoundary(reason)) {
-      engineFaults += 1
-      console.error(`[stability-guard] 引擎边界错误(累计 ${engineFaults},不退出): ${describe(reason).slice(0, 400)}`)
-      return
-    }
-    console.error(`[stability-guard] fatal ${kind}, exiting:\n${describe(reason)}`)
-    process.exit(1)
-  }
-  process.on('unhandledRejection', (reason) => {
-    onFatalCandidate('unhandledRejection', reason)
-  })
-  process.on('uncaughtException', (err) => {
-    onFatalCandidate('uncaughtException', err)
-  })
-  console.warn('[stability-guard] armed (socket 断连/sqlite 瞬态忙/引擎边界错误 → 记录继续;其余真实错误 → exit 1)')
-})
+/**
+ * Server 稳定性护栏:浏览器断开(WS/keep-alive socket 硬断)产生 ECONNRESET/EPIPE
+ * 类异步异常 —— Node 默认策略直接击穿进程,拖垮所有并发调试/运行会话。
+ *
+ * 严重性判定集中在 `server/utils/stability-severity.ts`(可被纯 node 回归测试直接断言):
+ * socket 断连 / SQLite 瞬态忙 / 外部设备不可达 / Agent 引擎边界 → 记录并继续;
+ * 其余判为 fatal,记日志后按非零码退出 —— 安全插件的 fail-fast(如生产密钥校验
+ * 失败拒绝启动)依赖异常中止进程,绝不能被护栏吞掉。
+ *
+ * "外部设备不可达"是**运维常态**(设备下电、模拟器未启、容器重启):
+ * 曾实测到一台 Modbus TCP 设备未启动(`connect ECONNREFUSED 127.0.0.1:1502`)
+ * 使 unhandledRejection 走到 fatal 分支,整个生产实例 exit 1 ——
+ * 数采/数控/数字孪生全平台被一台设备带走,属可用性事故。
+ */
+import { classifyFatalCandidate, describeReason } from '../utils/stability-severity'
+
+export default defineNitroPlugin(() => {
+  // 'db-busy' 必须引号 → 按 @stylistic/quote-props 的 consistent 口径,同对象内全部加引号
+  const counts: Record<string, number> = { 'socket': 0, 'db-busy': 0, 'upstream': 0, 'engine': 0 }
+  const onFatalCandidate = (kind: string, reason: unknown): void => {
+    const severity = classifyFatalCandidate(reason)
+    if (severity === 'fatal') {
+      console.error(`[stability-guard] fatal ${kind}, exiting:\n${describeReason(reason)}`)
+      process.exit(1)
+    }
+    counts[severity] += 1
+    const line = `[stability-guard] ${severity}(累计 ${counts[severity]},不退出): ${describeReason(reason).slice(0, 200).replace(/\n/g, ' ')}`
+    // 外部设备不可达与引擎边界属"要有人知道"的信号 → error;其余瞬态噪声 → warn
+    if (severity === 'upstream' || severity === 'engine') console.error(line)
+    else console.warn(line)
+  }
+
+  process.on('unhandledRejection', (reason) => {
+    onFatalCandidate('unhandledRejection', reason)
+  })
+  process.on('uncaughtException', (err) => {
+    onFatalCandidate('uncaughtException', err)
+  })
+  console.warn('[stability-guard] armed (socket 断连/sqlite 瞬态忙/外部设备不可达/引擎边界错误 → 记录继续;其余真实错误 → exit 1)')
+})

package/server/services/workshop/agents/agent-interface.ts

@@ -18,6 +18,9 @@ export interface AgentInfo {
   token?: string
   /** 实例启停(1 启用 / 0 禁用;模板视图无此字段) */
   enabled?: number
+  /** 来源 Agent 模板 id(成员实例专有;手工创建的成员为 null)。
+   *  Agent↔工业节点绑定必须以实例 id 为主体,调用方靠它把模板映射到已部署实例。 */
+  templateId?: string | null
 }
 
 /** 运行请求:平台向 Agent 投递的一次消息输入 */

package/server/services/workshop/agents/codex-agent.ts

@@ -81,13 +81,10 @@ export interface CodexAgentConfig {
 export class CodexAgentImpl extends BaseAgentImpl implements AgentInterface {
   private readonly config: CodexAgentConfig
   private agentInfo: AgentInfo | null = null
-  private readonly bridgeCtx = {
-    identity: { agentId: '', channelId: '', role: 'worker' as 'lead' | 'worker', name: 'agent' },
-    state: this.toolState,
-    getWorkspace: () => this.workspace,
-  }
-
-  private agentRole: 'lead' | 'worker' = 'worker'
+  // 不要重新声明 bridgeCtx / agentRole(由 BaseAgentImpl 持有)。
+  // 子类字段初始化器在 super() 之后执行,会用空 identity 遮蔽基类已装配好的 bridgeCtx,
+  // 使 host tool 桥的 agentId 恒为 '' → 按 agentId 的鉴权全部误判(节点绑定/工业工具)。
+  // 同理 agentRole 会被重置为 'worker',lead 身份丢失。
 
   private client: StdioJsonRpcClient | null = null
   private clientStarting: Promise<void> | null = null

package/server/services/workshop/agents/dsh-agent.ts

@@ -79,13 +79,10 @@ export interface DshAgentConfig {
 export class DshAgentImpl extends BaseAgentImpl implements AgentInterface {
   private readonly config: DshAgentConfig
   private agentInfo: AgentInfo | null = null
-  private readonly bridgeCtx = {
-    identity: { agentId: '', channelId: '', role: 'worker' as 'lead' | 'worker', name: 'agent' },
-    state: this.toolState,
-    getWorkspace: () => this.workspace,
-  }
-
-  private agentRole: 'lead' | 'worker' = 'worker'
+  // 不要重新声明 bridgeCtx / agentRole(由 BaseAgentImpl 持有)。
+  // 子类字段初始化器在 super() 之后执行,会用空 identity 遮蔽基类已装配好的 bridgeCtx,
+  // 使 host tool 桥的 agentId 恒为 '' → 按 agentId 的鉴权全部误判(节点绑定/工业工具)。
+  // 同理 agentRole 会被重置为 'worker',lead 身份丢失。
   private client: StdioJsonRpcClient | null = null
   private clientStarting: Promise<void> | null = null
   private sessionId: string | null = null

package/server/services/workshop/agents/opencode-agent.ts

@@ -110,13 +110,11 @@ interface PendingHitl {
 export class OpenCodeAgentImpl extends BaseAgentImpl implements AgentInterface {
   private readonly config: OpenCodeAgentConfig
   private agentInfo: AgentInfo | null = null
-  private readonly bridgeCtx = {
-    identity: { agentId: '', channelId: '', role: 'worker' as 'lead' | 'worker', name: 'agent' },
-    state: this.toolState,
-    getWorkspace: () => this.workspace,
-  }
-
-  private agentRole: 'lead' | 'worker' = 'worker'
+  // 注意:不要在本类重新声明 bridgeCtx / agentRole —— 它们由 BaseAgentImpl 持有。
+  // 子类字段初始化器在 super() **之后**执行,会用一个空的 identity 覆盖基类已装配好的
+  // bridgeCtx(class field 遮蔽),导致 host tool 桥拿到的 agentId 恒为 '' ——
+  // 所有按 agentId 的鉴权(节点绑定/工业工具)都会误判为"未绑定"。
+  // 同理 agentRole 被重置为 'worker',lead 身份传不到桥里。
 
   // 服务进程与 API 面
   private child: ReturnType<typeof spawnLineProcess> | null = null

package/server/services/workshop/aml/aml.repo.ts

@@ -209,6 +209,10 @@ export function createAmlRepo(db: DatabaseSync) {
       remove(id: string): void {
         db.prepare('DELETE FROM aml_datasets WHERE id = ?').run(id)
       },
+      /** 元数据局部更新(备注;CRUD 的 U) */
+      updateNote(id: string, note: string): void {
+        db.prepare('UPDATE aml_datasets SET note = ? WHERE id = ?').run(note, id)
+      },
     },
 
     // ----- jobs -----
@@ -268,6 +272,14 @@ export function createAmlRepo(db: DatabaseSync) {
       requeue(id: string): void {
         jobRetryBump.run(id)
       },
+      remove(id: string): void {
+        db.prepare('DELETE FROM aml_jobs WHERE id = ?').run(id)
+      },
+      /** 删除作业时级联清实验谱系(否则实验行成为悬挂元数据) */
+      removeExperimentsOf(jobId: string): number {
+        const r = db.prepare('DELETE FROM aml_experiments WHERE job_id = ?').run(jobId)
+        return Number(r.changes)
+      },
     },
 
     // ----- experiments -----
@@ -361,6 +373,12 @@ export function createAmlRepo(db: DatabaseSync) {
       markPruned(id: string): void {
         modelPrune.run(id)
       },
+      remove(id: string): void {
+        db.prepare('DELETE FROM aml_models WHERE id = ?').run(id)
+      },
+      updateNote(id: string, note: string): void {
+        db.prepare('UPDATE aml_models SET note = ? WHERE id = ?').run(note, id)
+      },
     },
   }
 }

package/server/services/workshop/aml/dataset-builder.ts

@@ -403,6 +403,10 @@ export async function buildDataset(spec: AmlDatasetSpec, by: { id: string, kind:
     beatMs: beat,
   }
   writeFileSync(join(dir, 'manifest.json'), JSON.stringify(manifest, null, 2))
+  // spec.json:取数规格随实体落盘。元数据行里本来就有 spec_json,但那是**索引**;
+  // 把规格同时写进实体目录,./aml 才是自描述、可整体拷贝的资产 ——
+  // 拷到另一台机器(或 sqlite 丢了)时,仅凭 datasets/<id>/ 仍能还原"这份数据是怎么取的"。
+  writeFileSync(join(dir, 'spec.json'), JSON.stringify(spec, null, 2))
 
   // 7. 统计报告(基于 train 网格序列)
   const nodeSummaries: SeriesSummary[] = []