agentworkshop 0.7.47 → 0.7.48

package/scripts/_aw0748-failclosed-probe.mjs

@@ -0,0 +1,90 @@
+/**
+ * 0.7.48 打包系统验收 · 硬约束失败关闭探针(无需 LLM,秒级)
+ * 场景:同一份真实快照 + 同一组候选,只改 constraints[].id
+ *   ① id = weight(可映射)→ 试验正常给出约束结论
+ *   ② id = part_weight(不可映射)→ 必须**判不通过**并给出"无法映射"明细(v0.7.47 会假通过)
+ * 用法:AW_BASE=http://127.0.0.1:3001 [AW_HOME=<home>] node scripts/_aw0748-failclosed-probe.mjs
+ */
+import { existsSync, readFileSync, readdirSync, statSync } from 'node:fs'
+
+const BASE = process.env.AW_BASE ?? 'http://127.0.0.1:3001'
+const HOME = process.env.AW_HOME ?? `${process.cwd()}/.aw0746-home`
+const ADMIN = { email: process.env.AW_ADMIN_EMAIL ?? 'admin@awshop.local', password: process.env.AW_ADMIN_PASSWORD ?? 'Awshop@123!' }
+
+let pass = 0
+const fails = []
+const check = (name, ok, detail = '') => {
+  console.log(`  ${ok ? 'PASS' : 'FAIL'}  ${name}${detail ? ` — ${detail}` : ''}`)
+  if (ok) pass += 1
+  else fails.push(name)
+}
+const j = async (method, path, body, token) => {
+  const r = await fetch(BASE + path, {
+    method,
+    headers: { 'content-type': 'application/json', ...(token ? { authorization: `Bearer ${token}` } : {}) },
+    body: body === undefined ? undefined : JSON.stringify(body),
+    signal: AbortSignal.timeout(120_000),
+  })
+  return { status: r.status, ...(await r.json().catch(() => ({}))) }
+}
+const firstJson = (text) => {
+  const a = text.indexOf('{')
+  const b = text.lastIndexOf('}')
+  return a >= 0 && b > a ? JSON.parse(text.slice(a, b + 1)) : null
+}
+
+const token = (await j('POST', '/api/users/login', ADMIN)).data?.token
+console.log(`\n═══ 0.7.48 硬约束失败关闭探针 @ ${BASE} ═══`)
+
+// 复用最近一次真实快照(阶段 7 / 阶段 5 产物)
+const snapRoot = `${HOME}/aml/twins/snapshots`
+const dirs = existsSync(snapRoot)
+  ? readdirSync(snapRoot).map(d => ({ d, m: statSync(`${snapRoot}/${d}`).mtimeMs })).sort((a, b) => b.m - a.m)
+  : []
+check('存在真实 TwinSnapshot 工件(取最新)', dirs.length > 0, dirs[0]?.d ?? '(无)')
+if (!dirs.length) {
+  console.log('\n★ 无可复用快照,先跑阶段 7/5 生成')
+  process.exit(2)
+}
+const snapshot = JSON.parse(readFileSync(`${snapRoot}/${dirs[0].d}/snapshots.json`, 'utf8'))
+check('快照新鲜且水位>0(可直接跑试验)', snapshot.dataQuality?.fresh === true && Number(snapshot.daqWatermark) > 0, `fresh=${snapshot.dataQuality?.fresh} watermark=${snapshot.daqWatermark}`)
+
+const chans = (await j('GET', '/api/workshop/channels', undefined, token)).data ?? []
+const twin = chans.find(c => /闭环优化-hybridtwin/.test(c.name))
+const members = (await j('GET', `/api/workshop/channels/${twin.id}/agents`, undefined, token)).data ?? []
+const lead = members.find(m => m.role === 'lead')
+const invoke = async (tool, args) => (await j('POST', '/api/workshop/agent-tools/invoke', { agentId: lead.id, tool, args }, token)).data?.result ?? {}
+
+const baseScene = firstJson(String((await invoke('twin_scene_read', {})).text ?? ''))
+const controls = { hold_pressure: 65, hold_time: 8, melt_temperature_setpoint: 247 }
+const objective = { schemaVersion: 1, createdAt: new Date().toISOString(), createdBy: lead.id, objectiveId: 'weight-quality', targets: { weight: 32.5 }, weights: { weight: 1 }, controlCosts: {}, horizonSteps: 1, trustRegion: {} }
+const trialOf = async (scene) => {
+  const res = await invoke('twin_trial_run', { scene_json: scene, snapshot_json: snapshot, baseline_controls: controls, candidate_controls: [{ ...controls, hold_pressure: 67 }], objective })
+  const text = String(res.text ?? '')
+  const id = (text.match(/trial_id:\s*(\S+)/) ?? [])[1]
+  const artifact = id ? JSON.parse(readFileSync(`${HOME}/aml/twins/trials/${id}/trials.json`, 'utf8')) : null
+  return { text, trial: artifact?.trial ?? null }
+}
+
+// ① 可映射 id:baseline 预测克重 32.5 g 落在 31~34 内 → 应通过
+const mapped = { ...baseScene, lineId: snapshot.lineId, productId: snapshot.productId, recipeId: snapshot.recipeId, constraints: [{ id: 'weight', kind: 'hard_range', min: 31, max: 34 }] }
+const okTrial = await trialOf(mapped)
+const okWeight = okTrial.trial?.constraintResults?.[0]
+check('可映射约束 id(weight)正常评估', Boolean(okTrial.trial) && okWeight?.id === 'weight' && okWeight.passed === true, `weight=${okTrial.trial?.predictedTrajectory?.[0]?.weight?.toFixed?.(3)} detail=${okWeight?.detail}`)
+
+// ② 不可映射 id:必须失败关闭(旧版本这里会返回"全轨迹通过")
+const unmapped = { ...mapped, constraints: [{ id: 'part_weight', kind: 'hard_range', min: 31, max: 34 }] }
+const badTrial = await trialOf(unmapped)
+const badCon = badTrial.trial?.constraintResults?.[0]
+const allPassed = (badTrial.trial?.constraintResults ?? []).every(c => c.passed)
+check('不可映射约束 id 判为不通过(失败关闭)', Boolean(badTrial.trial) && badCon?.passed === false && /无法映射/.test(String(badCon?.detail)), `passed=${badCon?.passed} detail=${String(badCon?.detail).slice(0, 80)}`)
+check('候选不再被当作安全(constraints_passed=false)', allPassed === false, `constraints_passed=${allPassed}`)
+check('工具文本同步反映未通过', /constraints_passed:\s*false/.test(badTrial.text), (badTrial.text.match(/constraints_passed:.*/) ?? [''])[0])
+
+// ③ 空轨迹(候选为空)同样失败关闭
+const empty = await invoke('twin_trial_run', { scene_json: mapped, snapshot_json: snapshot, baseline_controls: controls, candidate_controls: [], objective })
+const emptyTrialId = (String(empty.text ?? '').match(/trial_id:\s*(\S+)/) ?? [])[1]
+const emptyTrial = emptyTrialId ? JSON.parse(readFileSync(`${HOME}/aml/twins/trials/${emptyTrialId}/trials.json`, 'utf8')).trial : null
+check('空轨迹判为不通过(不再默认通过)', emptyTrial ? emptyTrial.constraintResults.every(c => c.passed === false) : false, `details=${(emptyTrial?.constraintResults ?? []).map(c => c.detail).join(' | ').slice(0, 80)}`)
+
+console.log(`\n★ 0.7.48 探针:${pass} 通过 / ${fails.length} 失败${fails.length ? ` (${fails.join('; ')})` : ''}`)

package/server/services/workshop/aml/twin/physics-runtime.ts

@@ -94,14 +94,29 @@ export class InjectionGreyboxProvider implements PhysicsModelProvider {
   }
 
   evaluateConstraints(scene: SceneContract, trajectory: PhysicsTrajectory): Array<{ id: string, passed: boolean, detail: string, firstViolationStep?: number }> {
+    // 失败关闭(fail-closed):约束 id 必须能映射到轨迹观测量/守卫量。
+    // 旧行为对映射不到的 id 直接跳过并返回「全轨迹通过」(value == null 分支),
+    // 于是 scene_json 里写成 part_weight / 质量窗口 这类 id 时,硬约束门禁形同虚设
+    // —— 实测被真实 Agent 团队两次抓出:预测 weight 30.3 g 明显低于窗口下界仍报通过。
+    const known = [...new Set([
+      ...Object.keys(trajectory.steps.at(-1)?.observations ?? {}),
+      ...Object.keys(trajectory.steps.at(-1)?.guards ?? {}),
+    ])]
+    if (trajectory.steps.length === 0) {
+      return scene.constraints.map(constraint => ({ id: constraint.id, passed: false, detail: '空轨迹:未执行任何校验' }))
+    }
     return scene.constraints.map((constraint) => {
+      let unmapped = true
       for (let i = 0; i < trajectory.steps.length; i++) {
         const obs = trajectory.steps[i]?.observations ?? {}
         const value = obs[constraint.id] ?? trajectory.steps[i]?.guards[constraint.id]
-        if (constraint.kind === 'hard_range' && value != null && ((constraint.min != null && value < constraint.min) || (constraint.max != null && value > constraint.max))) {
+        if (value == null) continue
+        unmapped = false
+        if (constraint.kind === 'hard_range' && ((constraint.min != null && value < constraint.min) || (constraint.max != null && value > constraint.max))) {
           return { id: constraint.id, passed: false, detail: `${constraint.id}=${value} 越过 ${constraint.min ?? '-∞'}~${constraint.max ?? '∞'}`, firstViolationStep: i }
         }
       }
+      if (unmapped) return { id: constraint.id, passed: false, detail: `约束 ${constraint.id} 无法映射到观测量/守卫量(可用:${known.join(', ') || '无'}),按失败关闭处理` }
       return { id: constraint.id, passed: true, detail: '全轨迹通过' }
     })
   }

package/.output/public/_nuxt/builds/meta/1b18aaf2-ccbc-45e0-b9c3-107655b226f9.json

@@ -1 +0,0 @@
-{"id":"1b18aaf2-ccbc-45e0-b9c3-107655b226f9","timestamp":1790321904277,"prerendered":[]}

package/scripts/_probe-density.cjs

@@ -1,15 +0,0 @@
-/** 临时:探测 daq_samples 采样密度(评估 AML beatMs 可下调空间) */
-const { Client } = require('pg')
-;(async () => {
-  const c = new Client({ host: '127.0.0.1', port: 5432, user: 'postgres', password: 'awshop', database: 'awshop' })
-  await c.connect()
-  const q = (sql) => c.query(sql).then(r => r.rows)
-  const rows = await q(`select node_id, count(*)::int n, min(ts) a, max(ts) b,
-      round(extract(epoch from (max(ts)-min(ts)))/greatest(count(*)-1,1),2) avg_gap_s
-    from daq_samples where ts > now() - interval '45 minutes'
-    group by node_id order by n desc limit 12`)
-  console.table(rows)
-  const t = await q(`select count(*)::int total, count(distinct node_id)::int nodes from daq_samples where ts > now() - interval '45 minutes'`)
-  console.log('total', t)
-  await c.end()
-})().catch(e => { console.error(e.message); process.exit(1) })

package/scripts/_probe-gate-model.mjs

@@ -1,34 +0,0 @@
-/** 临时:看 twin_gate_evaluate 的 modelUpdate 尾部 + 最近作业来源 */
-import { DatabaseSync } from 'node:sqlite'
-const BASE = 'http://127.0.0.1:3001'
-const ADMIN = { email: 'admin@awshop.local', password: 'Awshop@123!' }
-
-const j = async (method, path, body, token) => {
-  const r = await fetch(BASE + path, { method, headers: { 'content-type': 'application/json', ...(token ? { authorization: `Bearer ${token}` } : {}) }, body: body === undefined ? undefined : JSON.stringify(body) })
-  return { status: r.status, ...(await r.json().catch(() => ({}))) }
-}
-const token = (await j('POST', '/api/users/login', ADMIN)).data?.token
-const chans = (await j('GET', '/api/workshop/channels', undefined, token)).data ?? []
-const twin = chans.find(c => /闭环优化-hybridtwin/.test(c.name))
-const members = (await j('GET', `/api/workshop/channels/${twin.id}/agents`, undefined, token)).data ?? []
-const lead = members.find(m => m.role === 'lead')
-const models = ((await j('GET', '/api/workshop/aml/models', undefined, token)).data?.models ?? [])
-const prod = models.find(m => m.stage === 'production')
-console.log('production model =', prod?.id, 'ds =', prod?.datasetId)
-
-const res = await j('POST', '/api/workshop/agent-tools/invoke', {
-  agentId: lead.id,
-  tool: 'twin_gate_evaluate',
-  args: { rows: 1724, runs: 10, one_step_nrmse: 0.0763, rollout_nrmse: 0.0964, val_test_gap: 0.1072, calibration_rows: 1724, coverage: 1, candidate_trials: [], physics_failure_rate: 0, model_id: prod?.id ?? '' },
-}, token)
-const text = String(res.data?.result?.text ?? '')
-const tail = text.slice(-700)
-console.log('--- gate tail ---')
-console.log(tail)
-console.log('--- parsed keys ---', Object.keys(JSON.parse(text)).join(','))
-
-const db = new DatabaseSync('D:/codes/ABO/AgentWorkShop/.aw0746-home/data/workshop.sqlite', { readOnly: true })
-const cols = db.prepare('pragma table_info(aml_jobs)').all().map(c => c.name)
-console.log('\naml_jobs columns:', cols.join(','))
-const pick = ['id', 'status', 'kind', 'created_by', 'change_note', 'created_at'].filter(c => cols.includes(c))
-for (const r of db.prepare(`select ${pick.join(',')} from aml_jobs order by created_at desc limit 6`).all()) console.log('  ', JSON.stringify(r).slice(0, 220))

package/scripts/_probe-model-eligibility.cjs

@@ -1,13 +0,0 @@
-/** 临时:检查 AML 模型 metrics_json 里的 twinEligibility 是否回写 */
-const { DatabaseSync } = require('node:sqlite')
-const db = new DatabaseSync('D:/codes/ABO/AgentWorkShop/.aw0746-home/data/workshop.sqlite', { readOnly: true })
-const rows = db.prepare('select id, stage, dataset_id, experiment_id, metrics_json from aml_models order by created_at desc limit 4').all()
-for (const r of rows) {
-  let m = {}
-  try { m = JSON.parse(r.metrics_json || '{}') } catch { m = { parse: 'fail' } }
-  console.log(`${r.id} stage=${r.stage} ds=${r.dataset_id} exp=${r.experiment_id}`)
-  console.log(`   twinEligibility=${JSON.stringify(m.twinEligibility ?? null)} oneStep=${m.oneStepTest?.nrmse}`)
-}
-const jobs = db.prepare('select id, status, created_by_kind, created_by, substr(change_note,1,60) note, created_at from aml_jobs order by created_at desc limit 6').all()
-console.log('\njobs:')
-for (const j of jobs) console.log(`  ${j.id} ${j.status} by=${j.created_by_kind}/${j.created_by} ${j.created_at} ${j.note}`)