agentworkshop 0.7.49 → 0.7.50

package/server/services/workshop/aml/job-orchestrator/stages.ts

@@ -1,96 +1,108 @@
-/**
- * 阶段与状态读取 / stub 运行 / 结题
- * (由 server/services/workshop/aml/job-orchestrator.ts 按职责拆出;内容逐行原文搬运)
- */
-import type { AmlJobRow } from '../aml.repo'
-import type { PlatformMetrics } from '../gates'
-import type { RunningJob } from './shared'
-import { evaluateGates } from '../gates'
-import { finishRun, safeParse } from './failure'
-import { getAmlRuntime } from '../runtime'
-import { join } from 'node:path'
-import { loadManifest } from '../dataset-builder'
-import { log } from './shared'
-import { readFileSync, writeFileSync } from 'node:fs'
-import { recordOps } from '../../ops/ops'
-import { registerModelFromJob } from '../model-registry'
-import { wsJob, wsStage } from './broadcast'
-
-export function currentStatus(jobId: string): AmlJobRow['status'] {
-  const row = getAmlRuntime().repo.job.get(jobId)
-  return row?.status ?? 'training'
-}
-export function currentStage(jobId: string): string {
-  return getAmlRuntime().repo.job.get(jobId)?.stage ?? ''
-}
-
-/** 存根运行器(AML_STUB=1):不依赖 Python,模拟训练并产出可达标 metrics + STUB 标记 */
-export async function runStub(run: RunningJob, artifacts: string): Promise<void> {
-  const rt = getAmlRuntime()
-  rt.repo.job.setState(run.row.id, 'training', 'train', 30)
-  wsStage(run.row, 'train')
-  await new Promise(r => setTimeout(r, 500))
-  const params = safeParse(rt.repo.job.get(run.row.id)?.budget ?? {}).params ?? {}
-  const p = params as { stubNrmse?: number, stubRolloutNrmse?: number }
-  const oneStep = Number(p.stubNrmse ?? 0.07)
-  const rollout = Number(p.stubRolloutNrmse ?? 0.18)
-  rt.repo.job.setState(run.row.id, 'evaluating', 'evaluate', 80)
-  wsStage(run.row, 'evaluate')
-  await new Promise(r => setTimeout(r, 200))
-  const metrics: PlatformMetrics = {
-    oneStepVal: { windows: 100, nrmse: oneStep * 0.9 },
-    oneStepTest: { windows: 100, nrmse: oneStep },
-    rolloutTest: { windows: 50, nrmse: rollout, horizon: 12 },
-  }
-  writeFileSync(join(artifacts, 'metrics.json'), JSON.stringify({ ...params, ...metrics }, null, 2))
-  writeFileSync(join(artifacts, 'STUB'), new Date().toISOString())
-  concludeJob(run, artifacts)
-}
-
-export function concludeJob(run: RunningJob, artifactsDir: string): void {
-  const rt = getAmlRuntime()
-  const dataset = rt.repo.dataset.get(run.row.datasetId)
-  if (!dataset) {
-    rt.repo.job.fail(run.row.id, 'failed', '数据集已被删除', new Date().toISOString())
-    finishRun(run)
-    return
-  }
-  let metrics: PlatformMetrics | undefined
-  try {
-    // 存储形态:平台指标平铺顶层(消费方契约)+ 保留 agent 自报段;
-    // 兼容读取:平台评估器嵌套于 platform.* 时展平
-    const parsed = JSON.parse(readFileSync(join(artifactsDir, 'metrics.json'), 'utf8')) as { platform?: PlatformMetrics, agent?: unknown } & PlatformMetrics
-    metrics = parsed.platform
-      ? { ...parsed.platform, agent: parsed.agent }
-      : parsed
-  }
-  catch { /* 缺 metrics → 门禁全挂 */ }
-  const manifest = loadManifest(dataset)
-  const gates = evaluateGates(run.row.purpose, dataset, manifest, metrics, artifactsDir)
-  const gatesJson = JSON.stringify(gates)
-  const metricsJson = JSON.stringify(metrics ?? {})
-  rt.repo.job.finish(run.row.id, gates.passed ? 'done' : 'failed', metricsJson, gatesJson, artifactsDir, new Date().toISOString())
-  if (!gates.passed) {
-    rt.repo.job.fail(run.row.id, 'failed', `评测门未通过:${gates.checks.filter(c => !c.pass).map(c => `${c.id} ${c.detail}`).join('; ')}`, new Date().toISOString())
-    // 门禁未过=实验记录 gates_failed(非作业错误语义,但作业 fail 承载原因)
-  }
-  // 实验与模型登记(实验 id = job id 首跑对应;重试复用)
-  try {
-    registerModelFromJob(run.row.id, gates, metricsJson)
-  }
-  catch (err) {
-    log.warn(`[aml-job] ${run.row.id} 模型登记失败:${err instanceof Error ? err.message : String(err)}`)
-  }
-  const datasetFresh = rt.repo.dataset.get(run.row.datasetId)
-  const byKind = safeParse(run.row.budget).byKind
-  const actorKind: 'user' | 'agent' | 'system' = byKind === 'user' || byKind === 'agent' ? byKind : run.row.agentId ? 'agent' : 'system'
-  recordOps({
-    actor: run.row.agentId || 'aml', actorName: run.row.agentId || 'AML', actorKind,
-    action: 'aml.job.finish', kind: 'system',
-    summary: `作业 ${run.row.id} 结束:门禁${gates.passed ? '全部通过' : '未通过'}`,
-    targetKind: 'aml_job', targetId: run.row.id,
-    lineId: datasetFresh?.lineId, productId: datasetFresh?.productId, recipeId: datasetFresh?.recipeId,
-  })
-  wsJob(rt.repo.job.get(run.row.id) ?? run.row, datasetFresh)
-  finishRun(run)
-}
+/**
+ * 阶段与状态读取 / stub 运行 / 结题
+ * (由 server/services/workshop/aml/job-orchestrator.ts 按职责拆出;内容逐行原文搬运)
+ */
+import type { AmlJobRow } from '../aml.repo'
+import type { PlatformMetrics } from '../gates'
+import type { RunningJob } from './shared'
+import { evaluateGates } from '../gates'
+import { finishRun, safeParse } from './failure'
+import { getAmlRuntime } from '../runtime'
+import { join } from 'node:path'
+import { loadManifest } from '../dataset-builder'
+import { log } from './shared'
+import { readFileSync, writeFileSync } from 'node:fs'
+import { recordOps } from '../../ops/ops'
+import { registerModelFromJob } from '../model-registry'
+import { wsJob, wsStage } from './broadcast'
+
+export function currentStatus(jobId: string): AmlJobRow['status'] {
+  const row = getAmlRuntime().repo.job.get(jobId)
+  return row?.status ?? 'training'
+}
+export function currentStage(jobId: string): string {
+  return getAmlRuntime().repo.job.get(jobId)?.stage ?? ''
+}
+
+/** 存根运行器(AML_STUB=1):不依赖 Python,模拟训练并产出可达标 metrics + STUB 标记 */
+export async function runStub(run: RunningJob, artifacts: string): Promise<void> {
+  const rt = getAmlRuntime()
+  rt.repo.job.setState(run.row.id, 'training', 'train', 30)
+  wsStage(run.row, 'train')
+  await new Promise(r => setTimeout(r, 500))
+  const params = safeParse(rt.repo.job.get(run.row.id)?.budget ?? {}).params ?? {}
+  const p = params as { stubNrmse?: number, stubRolloutNrmse?: number }
+  const oneStep = Number(p.stubNrmse ?? 0.07)
+  const rollout = Number(p.stubRolloutNrmse ?? 0.18)
+  rt.repo.job.setState(run.row.id, 'evaluating', 'evaluate', 80)
+  wsStage(run.row, 'evaluate')
+  await new Promise(r => setTimeout(r, 200))
+  const metrics: PlatformMetrics = {
+    oneStepVal: { windows: 100, nrmse: oneStep * 0.9 },
+    oneStepTest: { windows: 100, nrmse: oneStep },
+    rolloutTest: { windows: 50, nrmse: rollout, horizon: 12 },
+  }
+  writeFileSync(join(artifacts, 'metrics.json'), JSON.stringify({ ...params, ...metrics }, null, 2))
+  writeFileSync(join(artifacts, 'STUB'), new Date().toISOString())
+  concludeJob(run, artifacts)
+}
+
+export function concludeJob(run: RunningJob, artifactsDir: string): void {
+  const rt = getAmlRuntime()
+  const dataset = rt.repo.dataset.get(run.row.datasetId)
+  if (!dataset) {
+    rt.repo.job.fail(run.row.id, 'failed', '数据集已被删除', new Date().toISOString())
+    finishRun(run)
+    return
+  }
+  let metrics: PlatformMetrics | undefined
+  try {
+    // 存储形态:平台指标平铺顶层(消费方契约)+ 保留其余自报块(hybrid/physics/uncertainty
+    // 等 Twin Gate 消费)与嵌套 platform 原样 —— 两种读取形状同源,不丢块。
+    const parsed = JSON.parse(readFileSync(join(artifactsDir, 'metrics.json'), 'utf8')) as { platform?: PlatformMetrics, agent?: unknown } & PlatformMetrics
+    metrics = { ...parsed, ...(parsed.platform ?? {}) } as PlatformMetrics
+  }
+  catch { /* 缺 metrics → 门禁全挂 */ }
+  const manifest = loadManifest(dataset)
+  const gates = evaluateGates(run.row.purpose, dataset, manifest, metrics, artifactsDir)
+  const gatesJson = JSON.stringify(gates)
+  const metricsJson = JSON.stringify(metrics ?? {})
+  rt.repo.job.finish(run.row.id, gates.passed ? 'done' : 'failed', metricsJson, gatesJson, artifactsDir, new Date().toISOString())
+  if (!gates.passed) {
+    rt.repo.job.fail(run.row.id, 'failed', `评测门未通过:${gates.checks.filter(c => !c.pass).map(c => `${c.id} ${c.detail}`).join('; ')}`, new Date().toISOString())
+    // 门禁未过=实验记录 gates_failed(非作业错误语义,但作业 fail 承载原因)
+  }
+  // 实验与模型登记(实验 id = job id 首跑对应;重试复用)
+  let registeredModelId: string | undefined
+  try {
+    registeredModelId = registerModelFromJob(run.row.id, gates, metricsJson).modelId
+  }
+  catch (err) {
+    log.warn(`[aml-job] ${run.row.id} 模型登记失败:${err instanceof Error ? err.message : String(err)}`)
+  }
+  // 建模任务收口:该作业若由 aml_training_plans 派生,回写 plan 终态与模型 id。
+  void import('../twin/training-plans')
+    .then(m => m.settleTrainingPlanByJob(run.row.id, gates.passed ? 'done' : 'failed', registeredModelId, gates.checks.filter(c => !c.pass).map(c => `${c.id} ${c.detail}`).join('; ')))
+    .catch(() => { /* plan 模块不可用不阻断作业收口 */ })
+  // 持续校准 worker 收口:该作业若由 twin_update_runs 派生,更新其终态(成功/失败)。
+  // 动态 import 规避 stages ↔ calibration-worker ↔ job-orchestrator 的循环装载。
+  void import('../twin/calibration-worker')
+    .then(m => m.markTwinUpdateRunFinished(
+      run.row.id,
+      gates.passed,
+      gates.checks.filter(c => !c.pass).map(c => `${c.id} ${c.detail}`).join('; '),
+    ))
+    .catch(() => { /* worker 不可用不阻断作业收口 */ })
+  const datasetFresh = rt.repo.dataset.get(run.row.datasetId)
+  const byKind = safeParse(run.row.budget).byKind
+  const actorKind: 'user' | 'agent' | 'system' = byKind === 'user' || byKind === 'agent' ? byKind : run.row.agentId ? 'agent' : 'system'
+  recordOps({
+    actor: run.row.agentId || 'aml', actorName: run.row.agentId || 'AML', actorKind,
+    action: 'aml.job.finish', kind: 'system',
+    summary: `作业 ${run.row.id} 结束:门禁${gates.passed ? '全部通过' : '未通过'}`,
+    targetKind: 'aml_job', targetId: run.row.id,
+    lineId: datasetFresh?.lineId, productId: datasetFresh?.productId, recipeId: datasetFresh?.recipeId,
+  })
+  wsJob(rt.repo.job.get(run.row.id) ?? run.row, datasetFresh)
+  finishRun(run)
+}

package/server/services/workshop/aml/job-orchestrator/start-job.ts

@@ -6,15 +6,18 @@ import type { AmlJobRow } from '../aml.repo'
 import type { RunningJob } from './shared'
 import { concludeJob, runStub } from './stages'
 import { ensureVenv, platformPythonDir } from '../python-runtime'
-import { existsSync, mkdirSync, writeFileSync } from 'node:fs'
+import { existsSync, mkdirSync, readFileSync, writeFileSync } from 'node:fs'
 import { experimentFail, failPermanentOrRetry, finishRun, lastErrLine, safeParse } from './failure'
 import { getAmlRuntime } from '../runtime'
 import { hashDatasetDir } from '../dataset-builder'
+import { createLogger } from '../../logger'
 import { join } from 'node:path'
 import { runPython } from './process'
 import { state } from './shared'
 import { wsStage } from './broadcast'
 
+const log = createLogger('aml.job')
+
 export async function startJob(row: AmlJobRow): Promise<void> {
   const rt = getAmlRuntime()
   const st = state()
@@ -42,6 +45,18 @@ export async function startJob(row: AmlJobRow): Promise<void> {
       writeFileSync(join(workspace, 'train.py'), inline)
       trainFile = 'train.py'
     }
+    else if ((safeParse(row.budget).jobKind ?? 'supervised') === 'hybrid_residual') {
+      // hybrid_residual 模板回退:平台参考训练器(物理参数校准→有界残差集成→UQ 校准)。
+      // Agent 无需自带训练代码即可把冻结 PhysicsSpec + 数据集训练成 hybrid 模型。
+      const template = join(platformPythonDir(), 'train-hybrid-example.py')
+      if (!existsSync(template)) {
+        rt.repo.job.fail(row.id, 'failed', `hybrid_residual 模板缺失:${template}(平台安装不完整,永久错误)`, new Date().toISOString())
+        return
+      }
+      writeFileSync(join(workspace, 'train.py'), readFileSync(template, 'utf8'))
+      trainFile = 'train.py'
+      log.warn(`[aml-job] ${row.id} hybrid_residual 未携带 code,使用平台参考训练器`)
+    }
     else {
       rt.repo.job.fail(row.id, 'failed', 'workspace/train.py 不存在(Agent 用 aml_job_submit 携带 code 提交,或 REST 提交时带 code)', new Date().toISOString())
       return
@@ -60,17 +75,26 @@ export async function startJob(row: AmlJobRow): Promise<void> {
     sceneVersion: budget.sceneVersion ?? null,
     objectiveId: budget.objectiveId ?? null,
     jobKind: budget.jobKind ?? 'supervised',
+    providerId: budget.providerId ?? null,
+    providerVersion: budget.providerVersion ?? null,
+    providerHash: budget.providerHash ?? null,
+    providerGeneration: budget.providerGeneration ?? null,
   }, null, 2))
-  // Hybrid Twin lineage is immutable job input. Persist these sidecars in the
-  // isolated workspace so amlkit.load_* can consume them without trusting REST
-  // callers or reaching into the live database during training.
+  // Hybrid Twin lineage is immutable job input. Persist these sidecars where
+  // amlkit.load_* actually reads them (AML_JOB_DIR root) so training/evaluation
+  // never trusts REST callers or reaches into the live database; the workspace
+  // copy is a provenance archive of the exact bytes the job was submitted with.
   for (const [name, key] of [
     ['physics_manifest.json', 'physicsManifest'],
+    ['physics_spec.json', 'physicsSpec'],
     ['twin_snapshot.json', 'twinSnapshot'],
     ['objective_profile.json', 'objectiveProfile'],
   ] as const) {
     const value = budget[key]
-    if (value && typeof value === 'object') writeFileSync(join(workspace, name), JSON.stringify(value, null, 2))
+    if (value && typeof value === 'object') {
+      writeFileSync(join(jobDir, name), JSON.stringify(value, null, 2))
+      writeFileSync(join(workspace, name), JSON.stringify(value, null, 2))
+    }
   }
   const run: RunningJob = {
     row,

package/server/services/workshop/aml/job-orchestrator/submit.ts

@@ -10,6 +10,8 @@ import { getAmlRuntime } from '../runtime'
 import { recordOps } from '../../ops/ops'
 import { state } from './shared'
 import { wsJob } from './broadcast'
+import { createTwinRepo } from '../twin/repo'
+import { amlTwinFeatureFlags } from '../twin/feature-flags'
 
 export interface SubmitJobInput {
   datasetId: string
@@ -28,13 +30,22 @@ export interface SubmitJobInput {
   objectiveId?: string
   jobKind?: 'supervised' | 'physics_calibration' | 'hybrid_residual' | 'uncertainty_calibration'
   physicsManifest?: Record<string, unknown>
+  physicsSpec?: Record<string, unknown>
   twinSnapshot?: Record<string, unknown>
   objectiveProfile?: Record<string, unknown>
+  providerId?: string
+  providerVersion?: string
+  providerHash?: string
+  providerGeneration?: number
   /** 发起者;byKind 决定审计归属(actorKind) */
   agent?: { id: string, channelId?: string, taskId?: string }
   byKind?: 'user' | 'agent'
   /** 实验谱系:由重试等内部路径复用既有 experiment 行 */
   experimentId?: string
+  /** 模型人读标识(注册时作 label 前缀;缺省由平台按 谱系·配方·目标 派生) */
+  modelName?: string
+  /** 建模意图描述(注册时落模型 description;缺省派生自谱系) */
+  modelDescription?: string
 }
 
 export function submitJob(input: SubmitJobInput): AmlJobRow {
@@ -43,6 +54,23 @@ export function submitJob(input: SubmitJobInput): AmlJobRow {
   const dataset = rt.repo.dataset.get(input.datasetId)
   if (!dataset) throw new AppError(404, 'AML_DATASET_MISSING', `数据集 ${input.datasetId} 不存在`)
 
+  // Hybrid Twin 作业必须绑定已冻结的 SceneContract。场景发现/编译可以由 Agent 自动完成，
+  // 但训练是不可逆的谱系写入，平台不能允许未确认 draft 进入校准或残差训练。
+  const jobKind = input.jobKind ?? 'supervised'
+  if (jobKind !== 'supervised' && !amlTwinFeatureFlags().trainingEnabled) throw new AppError(503, 'AML_TWIN_TRAINING_DISABLED', 'AML Twin 训练开关已关闭，拒绝提交物理校准/残差/UQ 作业')
+  if (jobKind !== 'supervised') {
+    if (!input.sceneId || !input.sceneVersion) throw new AppError(422, 'TWIN_SCENE_REQUIRED', 'Hybrid Twin 作业必须提供 sceneId 和 sceneVersion')
+    const scene = createTwinRepo(rt.db).getScene(input.sceneId, input.sceneVersion)
+    if (!scene) throw new AppError(404, 'TWIN_SCENE_MISSING', `SceneContract ${input.sceneId}@${input.sceneVersion} 不存在`)
+    if (scene.status !== 'frozen') throw new AppError(409, 'TWIN_SCENE_NOT_FROZEN', `SceneContract ${input.sceneId}@${input.sceneVersion} 尚未冻结，禁止进入 ${jobKind}`)
+    if (scene.lineId !== dataset.lineId || (scene.productId && scene.productId !== dataset.productId) || (scene.recipeId && scene.recipeId !== dataset.recipeId)) {
+      throw new AppError(409, 'TWIN_SCENE_DATASET_MISMATCH', 'SceneContract 与 AML 数据集的 line/product/recipe 谱系不一致')
+    }
+    if (jobKind === 'hybrid_residual' && (!input.providerId || !input.providerVersion || !input.providerHash)) {
+      throw new AppError(422, 'TWIN_PROVIDER_LINEAGE_REQUIRED', 'hybrid_residual 必须提供 providerId/providerVersion/providerHash')
+    }
+  }
+
   // 预算硬上限(ADR-3):单数据集实验数耗尽即拒绝,防 Agent 无限迭代
   const maxExperiments = input.budget?.maxExperiments ?? 12
   const used = rt.repo.experiment.countByDataset(input.datasetId)
@@ -81,10 +109,17 @@ export function submitJob(input: SubmitJobInput): AmlJobRow {
       sceneId: input.sceneId ?? null,
       sceneVersion: input.sceneVersion ?? null,
       objectiveId: input.objectiveId ?? null,
-      jobKind: input.jobKind ?? 'supervised',
+      jobKind,
       physicsManifest: input.physicsManifest ?? null,
+      physicsSpec: input.physicsSpec ?? null,
       twinSnapshot: input.twinSnapshot ?? null,
       objectiveProfile: input.objectiveProfile ?? null,
+      providerId: input.providerId ?? null,
+      providerVersion: input.providerVersion ?? null,
+      providerHash: input.providerHash ?? null,
+      providerGeneration: input.providerGeneration ?? null,
+      modelName: input.modelName?.trim().slice(0, 80) || null,
+      modelDescription: input.modelDescription?.trim().slice(0, 500) || null,
     },
     metricsJson: null,
     gatesJson: null,