agentworkshop 0.7.34 → 0.7.37

package/scripts/acceptance-suite.py

@@ -1,207 +0,0 @@
-# -*- coding: utf-8 -*-
-"""AgentWorkShop 全功能综合验收(真实 omp + 真实 LLM,5 场景):
-
-  A 简单任务      —— lead 分解 → worker 执行 → 父任务收口(回归)
-  B A2A 通信      —— w2 阻塞 poll_messages 期间 w1 require-reply 提问(信箱优先路径)
-  C 多任务队列    —— 4 任务直发 2 worker:FIFO + 单 WORKING 不变量
-  D busy-peer     —— w1 执行长任务期间 w2 发 require-reply 消息 → turn 结束 peer 回合处理
-  E 任务取消      —— 提交后立即取消 → CANCELED + 无交付浪费 + 信箱排空
-
-终局:token 经济性指标(supervise 轮次/LLM 回合数/防重复命中/空轮询计数)。
-"""
-import json, os, sys, time, urllib.request
-sys.stdout.reconfigure(encoding='utf-8', errors='replace')
-BASE = os.environ.get('AW_BASE', 'http://127.0.0.1:3000')
-PROVIDER, MODEL = 'zhipu-coding-plan', 'glm-5-turbo'
-PASS = FAIL = 0
-
-def check(name, ok, detail=''):
-    global PASS, FAIL
-    print(f'  [{"PASS" if ok else "FAIL"}] {name}{f" — {detail}" if detail else ""}', flush=True)
-    PASS, FAIL = PASS + (1 if ok else 0), FAIL + (0 if ok else 1)
-
-def log(m): print(f'[{time.strftime("%H:%M:%S")}] {m}', flush=True)
-
-def api(method, path, body=None, token=None, timeout=30):
-    h = {'Content-Type': 'application/json'}
-    if token: h['Authorization'] = 'Bearer ' + token
-    req = urllib.request.Request(BASE + path, data=json.dumps(body).encode() if body is not None else None, method=method, headers=h)
-    with urllib.request.urlopen(req, timeout=timeout) as r:
-        d = json.loads(r.read().decode())
-    if isinstance(d, dict) and d.get('code') not in (None, 0, 'ok'):
-        raise RuntimeError(f'{path} -> {str(d)[:200]}')
-    return d.get('data', d)
-
-def tasks_of(t, cid): return api('GET', f'/api/workshop/channels/{cid}/tasks', token=t)
-def events_of(t, cid, n=1000): return api('GET', f'/api/workshop/channels/{cid}/events?limit={n}', token=t)['items']
-def msgs_of(t, cid, n=300): return sorted(api('GET', f'/api/workshop/channels/{cid}/messages?limit={n}', token=t), key=lambda m: m['createdAt'])
-def ptext(m): return ' '.join(p.get('text', '') for p in (m.get('parts') or []) if isinstance(p, dict))
-def wait_task(t, cid, tid, timeout, want=None):
-    want = want or ('COMPLETED', 'FAILED', 'CANCELED')
-    dl = time.monotonic() + timeout
-    while time.monotonic() < dl:
-        st = next((x['state'] for x in tasks_of(t, cid) if x['id'] == tid), None)
-        if st in want: return st
-        time.sleep(4)
-    return st
-
-def team(token, name, workers=2):
-    ch = api('POST', '/api/workshop/channels', body={'name': name, 'leadAgent': {'name': f'{name}-lead', 'harness': 'omp', 'config': {'provider': PROVIDER, 'model': MODEL}}}, token=token)
-    cid = ch['channelId']
-    ws = [api('POST', f'/api/workshop/channels/{cid}/agents', body={'name': f'{name}-w{i+1}', 'harness': 'omp', 'role': 'worker', 'config': {'provider': PROVIDER, 'model': MODEL}}, token=token)['id'] for i in range(workers)]
-    log(f'团队就绪 {name}: channel={cid[:8]} workers={[w[:8] for w in ws]}')
-    return cid, ws
-
-def submit(t, cid, title, desc, assignee=None):
-    b = {'title': title, 'description': desc}
-    if assignee: b['assigneeId'] = assignee
-    return api('POST', f'/api/workshop/channels/{cid}/tasks', body=b, token=t)['id']
-
-# ═══ 场景 A:简单任务 ═══
-def scenario_a(token):
-    log('── A 简单任务:lead 分解 → worker 执行 → 收口 ──')
-    cid, ws = team(token, 'sv-a', workers=1)
-    t0 = time.monotonic()
-    pid = submit(token, cid, 'A:巴黎问题', '写出句子「巴黎是法国的首都」,调用 complete_task 提交交付物。不要多余操作。')
-    st = wait_task(token, cid, pid, 600)
-    check('A 父任务 COMPLETED', st == 'COMPLETED', f'state={st} 耗时={int(time.monotonic()-t0)}s')
-    ch = [x for x in tasks_of(token, cid) if x.get('parentId') == pid]
-    done = next((c for c in ch if c['state'] == 'COMPLETED'), None)
-    check('A 子任务完成带交付', done is not None and len(done.get('artifacts', [])) > 0, f'children={len(ch)}')
-    return cid
-
-# ═══ 场景 B:A2A(poll 阻塞期 require-reply) ═══
-def scenario_b(token):
-    log('── B A2A:w2 阻塞 poll 期间 w1 require-reply 提问 ──')
-    cid, (w1, w2) = team(token, 'sv-b', workers=2)
-    t2 = submit(token, cid, 'B-等待者', f'立即调用 poll_messages(limit=5, wait_seconds=120) 阻塞等待。收到消息后:1) send_message_to_agent 回发送者,message="ANSWER-B-OK"。2) complete_task(summary=已回复)。', assignee=w2)
-    log('等 25s 让 w2 进入 poll 阻塞…'); time.sleep(25)
-    t1 = submit(token, cid, 'B-提问者', f'立即用 send_message_to_agent 向 <{w2}> 发消息(message="请回复:TOKEN-B9", require_reply=true),然后 complete_task(summary=已发送)。', assignee=w1)
-    s1, s2 = wait_task(token, cid, t1, 480), wait_task(token, cid, t2, 480, want=('COMPLETED',))
-    check('B 双任务 COMPLETED', s1 == 'COMPLETED' and s2 == 'COMPLETED', f'{s1}/{s2}')
-    ms = msgs_of(token, cid)
-    ask = [m for m in ms if m.get('fromAgentId') == w1 and m.get('toAgentId') == w2]
-    rep = [m for m in ms if m.get('fromAgentId') == w2 and m.get('toAgentId') == w1 and m.get('metadata', {}).get('x-aw-in-reply-to')]
-    check('B 回信自动关联 in_reply_to', len(rep) >= 1, f'ask={len(ask)} reply={len(rep)}')
-    errs = [e for e in events_of(token, cid) if e['type'] == 'error']
-    check('B 无 error 事件(无工具调用被打断)', len(errs) == 0, f'errors={len(errs)}')
-    return cid
-
-# ═══ 场景 C:多任务队列 ═══
-def scenario_c(token):
-    log('── C 多任务队列:4 任务直发 2 worker,FIFO + 单 WORKING ──')
-    cid, (w1, w2) = team(token, 'sv-c', workers=2)
-    assigns = [w1, w2, w1, w2]
-    ids = [submit(token, cid, f'C-队列-{i+1}(回 C{i+1}-OK)', f'只回复一句:C{i+1}-OK,然后 complete_task(summary=该句)。', assignee=assigns[i]) for i in range(4)]
-    dl = time.monotonic() + 720
-    states = {}
-    while time.monotonic() < dl:
-        states = {tid: next((x['state'] for x in tasks_of(token, cid) if x['id'] == tid), None) for tid in ids}
-        if all(v == 'COMPLETED' for v in states.values()): break
-        time.sleep(5)
-    check('C 4 任务全部 COMPLETED', all(v == 'COMPLETED' for v in states.values()), f'{list(states.values())}')
-    # 单 WORKING 重放
-    evs = events_of(token, cid)
-    tset = set(ids); sim = {}; mx = 0
-    tmap = {x['id']: x for x in tasks_of(token, cid)}
-    for e in evs:
-        if e['type'] == 'task.status' and e.get('taskId') in tset:
-            st = (e.get('payload') or {}).get('state'); ag = e.get('agentId') or ''
-            if tmap[e['taskId']]['assigneeId'] != ag: continue
-            if st == 'WORKING': sim[ag] = sim.get(ag, 0) + 1; mx = max(mx, sim[ag])
-            elif st in ('COMPLETED', 'FAILED', 'CANCELED', 'WAITING'): sim[ag] = max(0, sim.get(ag, 0) - 1)
-    check('C 单 WORKING 不变量(并发≤1)', mx <= 1, f'max={mx}')
-    # FIFO:按 events 中每任务 COMPLETED 的出现序,校验每 worker 的完成序 = 提交序
-    order = {}
-    for e in evs:
-        if e['type'] == 'task.status' and (e.get('payload') or {}).get('state') == 'COMPLETED' and e.get('taskId') in tset:
-            order.setdefault(e['taskId'], len(order))
-    seq_of = {tid: i for i, tid in enumerate(ids)}
-    per = {}
-    for tid in ids:
-        a = tmap[tid]['assigneeId']; per.setdefault(a, []).append(tid)
-    fifo = all([seq_of[x] for x in sorted(lst, key=lambda x: order.get(x, 99))] == sorted(seq_of[x] for x in lst) for lst in per.values())
-    check('C 每 worker FIFO', fifo, '; '.join(f'{k[:8]}:{len(v)}' for k, v in per.items()))
-    return cid
-
-# ═══ 场景 D:busy 期间 peer 消息(turn-end 处理) ═══
-def scenario_d(token):
-    log('── D busy-peer:w1 长任务执行中收到 w2 require-reply → turn 结束处理 ──')
-    cid, (w1, w2) = team(token, 'sv-d', workers=2)
-    t1 = submit(token, cid, 'D-长任务', '写一篇 120 字左右的短文(主题:协作),中间调用一次 report_progress(50%),最后 complete_task(summary=短文全文)。', assignee=w1)
-    log('等 20s 让 w1 进入执行…'); time.sleep(20)
-    t2 = submit(token, cid, 'D-发信者', f'立即用 send_message_to_agent 向 <{w1}> 发消息(message="请回复:D-ACK-55", require_reply=true),然后 complete_task(summary=已发送)。', assignee=w2)
-    s1, s2 = wait_task(token, cid, t1, 600), wait_task(token, cid, t2, 600)
-    check('D 双任务 COMPLETED', s1 == 'COMPLETED' and s2 == 'COMPLETED', f'{s1}/{s2}')
-    dl = time.monotonic() + 300
-    rep = []
-    while time.monotonic() < dl:
-        rep = [m for m in msgs_of(token, cid) if m.get('fromAgentId') == w1 and m.get('toAgentId') == w2 and m.get('metadata', {}).get('x-aw-in-reply-to')]
-        if rep: break
-        time.sleep(5)
-    check('D busy w1 turn 结束后回信(带 in_reply_to)', len(rep) >= 1, f'reply={len(rep)}')
-    errs = [e for e in events_of(token, cid) if e['type'] == 'error']
-    check('D 无 error 事件', len(errs) == 0, f'errors={len(errs)}')
-    return cid
-
-# ═══ 场景 E:任务取消 ═══
-def scenario_e(token):
-    log('── E 取消:提交后立即取消,不浪费执行 ──')
-    cid, ws = team(token, 'sv-e', workers=1)
-    tid = submit(token, cid, 'E-应被取消(回 E-NEVER)', '只回复一句:E-NEVER,然后 complete_task。', assignee=ws[0])
-    api('POST', f'/api/workshop/tasks/{tid}/cancel', body={}, token=token)
-    st = wait_task(token, cid, tid, 180, want=('COMPLETED', 'FAILED', 'CANCELED'))
-    task = next(x for x in tasks_of(token, cid) if x['id'] == tid)
-    check('E 任务已 CANCELED', st == 'CANCELED', f'state={st}')
-    if st == 'CANCELED':
-        check('E 取消后无交付(未烧完整回合)', len(task.get('artifacts', [])) <= 1, f'artifacts={len(task.get("artifacts", []))}')
-    time.sleep(8)
-    pending = [m for m in msgs_of(token, cid) if m.get('state') == 'pending']
-    check('E 信箱无 pending 残留(作废投递被清)', len(pending) == 0, f'pending={len(pending)}')
-    return cid
-
-# ═══ token 经济性审计 ═══
-def audit(token, cids):
-    log('── token 经济性审计(全部场景频道) ──')
-    total_supervise = 0; total_turns = 0; total_tasks = 0; total_tool = 0; dup_hits = 0
-    for cid in cids:
-        evs = events_of(token, cid)
-        lead = None
-        ch = api('GET', f'/api/workshop/channels/{cid}', token=token)
-        lead = ch.get('leadAgentId')
-        # supervise 轮次 ≈ lead 的 status.message 帧数(每轮 supervise 至少一帧文本)
-        supervise = sum(1 for e in evs if e['type'] == 'agent.status.message' and e.get('agentId') == lead)
-        # LLM 回合 ≈ agent_end 映射的 artifact 'output' 帧数 + error 帧(近似)
-        turns = sum(1 for e in evs if e['type'] == 'a2a.artifact' and (e.get('payload') or {}).get('artifact', {}).get('name') == 'output')
-        tools = sum(1 for e in evs if e['type'] == 'agent.status.message' and str((e.get('payload') or {}).get('text', '')).startswith('🔧'))
-        errs = [e for e in evs if e['type'] == 'error']
-        total_supervise += supervise; total_turns += turns; total_tool += tools
-        total_tasks += len([t for t in tasks_of(token, cid)])
-        log(f'  {cid[:8]}: tasks={len(tasks_of(token, cid))} lead状态帧={supervise} LLM回合≈{turns} 工具调用={tools} error={len(errs)}')
-    log(f'合计: 任务={total_tasks} lead监督帧={total_supervise} LLM回合≈{total_turns} 工具调用={total_tool}')
-    ratio = total_turns / max(1, total_tasks)
-    check('token 经济性:每任务平均 LLM 回合 ≤ 4', ratio <= 4, f'回合/任务={ratio:.1f}')
-    check('token 经济性:lead 监督帧/任务 ≤ 6', total_supervise / max(1, total_tasks) <= 6, f'{total_supervise/max(1,total_tasks):.1f}')
-
-def main():
-    which = sys.argv[1] if len(sys.argv) > 1 else 'all'
-    user = api('POST', '/api/users/register', body={'name': f'sv-{int(time.time())%1000000}', 'email': f'sv{int(time.time()*1000)%10**10}@t.local', 'password': 'Test1234!'})
-    token = user['token']
-    parts = set(which.split(','))
-    cids = []
-    # 清理上一轮中断遗留的测试频道(sv- 前缀)
-    for c in api('GET', '/api/workshop/channels', token=token):
-        if str(c.get('name', '')).startswith('sv-'):
-            api('DELETE', f"/api/workshop/channels/{c['id']}", token=token)
-    if 'a' in parts: cids.append(scenario_a(token))
-    if 'b' in parts: cids.append(scenario_b(token))
-    if 'c' in parts: cids.append(scenario_c(token))
-    if 'd' in parts: cids.append(scenario_d(token))
-    if 'e' in parts: cids.append(scenario_e(token))
-    audit(token, cids)
-    for cid in cids:
-        api('DELETE', f'/api/workshop/channels/{cid}', token=token)
-    log(f'★ 结果: {PASS} passed, {FAIL} failed(测试频道已清理)')
-    sys.exit(1 if FAIL else 0)
-
-main()

package/scripts/aml-agent-bridge.ts

@@ -1,230 +0,0 @@
-/**
- * AML × AgentTeam 集成实测(真实 agent 身份 + 工具桥 + HITL):
- *   建频道 → 建队(内置 AML 模板)→ 部署 → 绑定 daq 节点 →
- *   以 worker 令牌走 aml_node_catalog / aml_dataset_build / aml_job_submit / aml_leaderboard →
- *   以 lead 令牌发起 aml_model_promote(HITL 人工批准)→ aml_model_reference 影子预测。
- *
- * 前置:活服 3001;AW_E2E_TOKEN(admin);工业闭环脚本已产线/节点/数据。
- * 运行:AW_E2E_TOKEN=… npx tsx --tsconfig .nuxt/tsconfig.server.json scripts/aml-agent-bridge.ts
- */
-/* eslint-disable @typescript-eslint/no-explicit-any -- 集成实测:响应体按动态 JSON 探查 */
-import { DatabaseSync } from 'node:sqlite'
-import { readFileSync } from 'node:fs'
-import { join } from 'node:path'
-
-const BASE = 'http://127.0.0.1:3001'
-const TOKEN = process.env.AW_E2E_TOKEN ?? ''
-if (!TOKEN) {
-  console.error('需要 AW_E2E_TOKEN')
-  process.exit(1)
-}
-
-let passed = 0
-let failures = 0
-function check(name: string, ok: boolean, detail = ''): void {
-  console.log(`  ${ok ? 'PASS' : 'FAIL'}  ${name}${detail ? ` — ${detail}` : ''}`)
-  if (ok) passed++
-  else failures++
-}
-const sleep = async (ms: number): Promise<void> => {
-  await new Promise(r => setTimeout(r, ms))
-}
-
-async function api(method: string, path: string, opts: { body?: unknown, token?: string, agentToken?: string } = {}): Promise<any> {
-  const headers: Record<string, string> = { 'content-type': 'application/json' }
-  if (opts.agentToken) headers['x-aw-agent-token'] = opts.agentToken
-  else headers.authorization = `Bearer ${opts.token ?? TOKEN}`
-  const res = await fetch(`${BASE}${path}`, {
-    method,
-    headers,
-    body: opts.body === undefined ? undefined : JSON.stringify(opts.body),
-  })
-  return res.json().catch(() => null)
-}
-
-const suffix = Date.now().toString(36).slice(-5)
-
-/** 只读查 channel_agents 实例令牌(部署克隆时签发) */
-function agentTokenOf(channelId: string, role: string): { id: string, name: string, token: string } | null {
-  const db = new DatabaseSync('.AgentWorkShop/data/workshop.sqlite', { readOnly: true })
-  try {
-    const row = db.prepare(
-      `SELECT id, name, token FROM channel_agents WHERE channel_id = ? AND role = ? AND enabled = 1 ORDER BY created_at DESC LIMIT 1`,
-    ).get(channelId, role) as { id: string, name: string, token: string } | undefined
-    return row ?? null
-  }
-  finally { db.close() }
-}
-
-async function main(): Promise<void> {
-  console.log('━━━ 1. 找到工业闭环产线与节点 ━━━')
-  const lines = await api('GET', '/api/workshop/dcw/lines')
-  const line = (lines?.data?.lines ?? []).filter((l: any) => l.name?.startsWith('涂布烘干 AML 实测线')).pop()
-  check('定位实测产线', !!line, line?.id)
-  const daqList = await api('GET', '/api/workshop/daq')
-  const daqNodes = (daqList?.data?.nodes ?? []).filter((n: any) => n.lineId === line.id)
-  const spDaq = daqNodes.find((n: any) => n.name?.includes('SP 读'))
-  const pvDaq = daqNodes.find((n: any) => n.name?.includes('PV 读'))
-  check('定位 SP/PV 数采节点', !!spDaq && !!pvDaq, `${spDaq?.id} / ${pvDaq?.id}`)
-  const recipes = await api('GET', '/api/workshop/dcw/recipes')
-  const recipe = (recipes?.data?.recipes ?? []).filter((r: any) => r.name?.startsWith('烘干温度配方')).pop()
-  check('定位配方', !!recipe?.id, recipe?.id)
-
-  console.log('━━━ 2. 频道 + 内置 AML 团队部署 ━━━')
-  const ch = await api('POST', '/api/workshop/channels', { body: { name: `AML 实测频道 ${suffix}`, scenario: `对配方 ${recipe?.id} 建立影子模型` } })
-  const channelId = ch?.data?.channelId ?? ch?.data?.channel?.id ?? ch?.data?.id
-  check('创建频道', !!channelId, channelId)
-  const team = await api('POST', '/api/workshop/teams', { body: { name: `AML 实测队 ${suffix}`, description: '集成实测' } })
-  const teamId = team?.data?.id ?? team?.data?.team?.id
-  check('创建团队', !!teamId, teamId)
-  for (const m of [{ agentId: 'tpl-aml-data', role: 'worker' }, { agentId: 'tpl-aml-lead', role: 'lead' }]) {
-    const add = await api('POST', `/api/workshop/teams/${teamId}/members`, { body: m })
-    check(`入队成员 ${m.agentId}(${m.role})`, add?.code === 0 || add?.ok === true, add?.message ?? '')
-  }
-  const deploy = await api('POST', `/api/workshop/teams/${teamId}/deploy`, { body: { channelId } })
-  check('团队部署到频道', deploy?.code === 0 || deploy?.ok === true, JSON.stringify(deploy?.data ?? deploy?.message)?.slice(0, 120))
-
-  const worker = agentTokenOf(channelId, 'worker')
-  const lead = agentTokenOf(channelId, 'lead')
-  check('worker 实例令牌签发', !!worker, worker?.name)
-  check('lead 实例令牌签发', !!lead, lead?.name)
-
-  console.log('━━━ 3. 节点绑定(Agent 权限边界)与试点门禁标定 ━━━')
-  const gateRelax = await api('PATCH', '/api/system/settings', { body: { overrides: { 'aml.gates.minRows': 150, 'aml.gates.valTestGap': 0.5 } } })
-  const gatesCalibrated = gateRelax?.code === 0 || gateRelax?.ok === true
-  const restoreGates = async (): Promise<void> => {
-    if (!gatesCalibrated) return
-    await api('PATCH', '/api/system/settings', { body: { overrides: { 'aml.gates.minRows': 500, 'aml.gates.valTestGap': 0.2 } } })
-  }
-  check('试点门禁标定(G1/G2 保持严格)', gatesCalibrated, `code=${gateRelax?.code}`)
-  for (const n of [spDaq, pvDaq]) {
-    const b = await api('POST', '/api/workshop/agent-tools/bindings', { body: { agentId: worker.id, nodeId: n.id, kind: 'daq', mode: 'auto' } })
-    check(`绑定 daq 节点 → worker(${n.name?.slice(0, 12)})`, b?.code === 0, b?.message ?? '')
-  }
-  for (const n of [spDaq, pvDaq]) {
-    await api('POST', '/api/workshop/agent-tools/bindings', { body: { agentId: lead.id, nodeId: n.id, kind: 'daq', mode: 'auto' } })
-  }
-
-  const invoke = async (agentId: string, token: string, tool: string, args: Record<string, unknown>) => {
-    const j = await api('POST', '/api/workshop/agent-tools/invoke', { agentToken: token, body: { agentId, tool, args } })
-    return j?.data ?? j // 信封拆包:{code,message,data:{result}} → {result}
-  }
-
-  console.log('━━━ 4. worker 视角:工具链闭环 ━━━')
-  const catalog = await invoke(worker.id, worker.token, 'aml_node_catalog', {})
-  check('aml_node_catalog(含绑定节点语义卡)', catalog?.result?.text?.includes(spDaq.id) || catalog?.result?.text?.length > 50,
-    `${String(catalog?.result?.text ?? '').slice(0, 60)}…`)
-  const dsBuild = await invoke(worker.id, worker.token, 'aml_dataset_build', {
-    line_id: line.id, product_id: recipe?.productId ?? '', recipe_id: recipe?.id ?? '',
-    nodes: [{ node_id: spDaq.id, role: 'control' }, { node_id: pvDaq.id, role: 'target' }],
-    beat_ms: 1000, history_steps: 6, horizon_steps: 3,
-    split_seed: 42, val_ratio: 0.34, test_ratio: 0.33, purpose: 'mpc_surrogate',
-    note: 'Agent 自主构建(PLC 模拟器数据)',
-  })
-  const dsId = String(dsBuild?.result?.text ?? '').match(/ds-[a-z0-9-]+/)?.[0]
-  check('aml_dataset_build(Agent 自主构建数据集)', !!dsId, dsId ?? String(dsBuild?.result?.text ?? dsBuild?.message ?? '').slice(0, 100))
-
-  console.log('━━━ 5. worker 提交训练(AIDE 式两轮迭代) ━━━')
-  const trainCode = readFileSync(join(process.cwd(), 'server', 'services', 'workshop', 'aml', 'python', 'train-example.py'), 'utf8')
-  const submit = async (datasetId: string, changeNote: string, params: Record<string, unknown>) => {
-    const r = await invoke(worker.id, worker.token, 'aml_job_submit', {
-      dataset_id: datasetId, code: trainCode, change_note: changeNote, params, seed: 42,
-    })
-    return String(r?.result?.text ?? '').match(/job-[a-z0-9-]+/)?.[0]
-  }
-  const waitDone = async (jobId: string): Promise<string | null> => {
-    const t0 = Date.now()
-    while (Date.now() - t0 < 8 * 60_000) {
-      await sleep(4000)
-      const st = await invoke(worker.id, worker.token, 'aml_job_status', { job_id: jobId })
-      const statusLine = String(st?.result?.text ?? '')
-      const m = statusLine.match(/状态[：:]\s*(\w+)/) ?? statusLine.match(/(queued|provisioning|training|evaluating|done|failed|timeout)/i)
-      if (m && ['done', 'failed', 'timeout'].includes(m[1].toLowerCase())) return m[1].toLowerCase()
-    }
-    return null
-  }
-
-  // 第一轮:小窗基线(H=6/F=3,h=64)——预期校准数据可学性
-  const jobId1 = await submit(dsId, 'Agent 首轮:MLP h=64 小窗基线', { hidden: 64, epochs: 40 })
-  check('第一轮作业提交', !!jobId1, jobId1)
-  const final1 = jobId1 ? await waitDone(jobId1) : null
-  console.log(`    第一轮终态=${final1}(基线校准;未达标即进入第二轮靶向精修)`)
-
-  // 第二轮:更大历史窗 + 更大容量(依据第一轮门禁反馈的靶向精修)
-  const dsBuild2 = await invoke(worker.id, worker.token, 'aml_dataset_build', {
-    line_id: line.id, product_id: recipe?.productId ?? '', recipe_id: recipe?.id ?? '',
-    nodes: [{ node_id: spDaq.id, role: 'control' }, { node_id: pvDaq.id, role: 'target' }],
-    beat_ms: 1000, history_steps: 10, horizon_steps: 4,
-    split_seed: 42, val_ratio: 0.2, test_ratio: 0.2, purpose: 'mpc_surrogate',
-    note: 'Agent 第二轮:加大历史窗(H=10)覆盖 τ≈8s 惯性',
-  })
-  const dsId2 = String(dsBuild2?.result?.text ?? '').match(/ds-[a-z0-9-]+/)?.[0]
-  check('第二轮数据集(靶向:加大历史窗)', !!dsId2, dsId2)
-  const jobId2 = dsId2 ? await submit(dsId2, 'Agent 第二轮:加大历史窗 H=10 + 容量 h=192(靶向精修)', { hidden: 192, lr: 0.0008, epochs: 80 }) : null
-  check('第二轮作业提交', !!jobId2, jobId2)
-  const final2 = jobId2 ? await waitDone(jobId2) : null
-  check('第二轮训练达标(done)', final2 === 'done', `final=${final2}`)
-
-  console.log('━━━ 6. 排行榜与越权守卫 ━━━')
-  const lb = await invoke(worker.id, worker.token, 'aml_leaderboard', { dataset_id: dsId2 ?? dsId })
-  check('aml_leaderboard(谱系+门禁)', String(lb?.result?.text ?? '').length > 100, String(lb?.result?.text ?? '').slice(0, 80))
-  const models = await api('GET', '/api/workshop/aml/models')
-  const cand = (models?.data?.models ?? []).filter((m: any) => m.stage === 'candidate' && m.recipeId === recipe?.id)
-    .sort((a: any, b: any) => String(b.createdAt).localeCompare(String(a.createdAt)))[0]
-  check('候选模型在册(登记自动完成)', !!cand, cand?.id)
-  if (!cand) {
-    await restoreGates()
-    process.exit(failures > 0 ? 1 : 0)
-    return
-  }
-  const workerPromote = await invoke(worker.id, worker.token, 'aml_model_promote', { model_id: cand?.id ?? 'x', to_stage: 'production' })
-  const workerBlocked = workerPromote?.result?.isError === true || /lead 专属|无权/.test(String(workerPromote?.result?.text ?? workerPromote?.message ?? ''))
-  check('worker 调晋升被拒(lead 专属硬守卫)', workerBlocked, String(workerPromote?.result?.text ?? workerPromote?.message ?? '').slice(0, 80))
-
-  console.log('━━━ 7. lead 发起晋升 → HITL 人工批准 ━━━')
-  const promotePromise = invoke(lead.id, lead.token, 'aml_model_promote', { model_id: cand.id, to_stage: 'production' })
-  let approved = false
-  for (let i = 0; i < 20 && !approved; i++) {
-    await sleep(1000)
-    const pending = await api('GET', '/api/workshop/hitl/pending')
-    const items = pending?.data?.items ?? pending?.data ?? []
-    const hit = (Array.isArray(items) ? items : []).find((x: any) => x.kind === 'dcw-approval' && String(x.title ?? '').includes('AML'))
-    if (hit?.id) {
-      const resp = await api('POST', '/api/workshop/hitl/respond', { body: { kind: 'dcw-approval', id: hit.id, confirmed: true, comment: '集成实测批准' } })
-      approved = resp?.data?.ok === true || resp?.ok === true || resp?.code === 0
-      check('HITL 待办出现并批准', approved, hit.id)
-    }
-  }
-  const promoteResult = await Promise.race([promotePromise, sleep(30_000).then(() => null)])
-  check('晋升执行完成(HITL 批准后)', String(promoteResult?.result?.text ?? '').includes('production'),
-    String(promoteResult?.result?.text ?? '').slice(0, 120))
-  const prodList = await api('GET', `/api/workshop/aml/models?recipeId=${recipe.id}&stage=production`)
-  check('production 模型唯一在册', (prodList?.data?.models?.length ?? 0) === 1, `n=${prodList?.data?.models?.length}`)
-
-  console.log('━━━ 8. 影子预测(Agent 调参参考) ━━━')
-  const now = Date.now()
-  const fetchSeries = async (nodeId: string) => {
-    const r = await api('GET', `/api/workshop/daq/${nodeId}/samples?from=${now - 30_000}&to=${now}&bucketMs=1000&limit=100`)
-    return (r?.data?.points ?? []).map((p: any) => p.avg ?? p.value)
-  }
-  const spSeries = await fetchSeries(spDaq.id)
-  const pvSeries = await fetchSeries(pvDaq.id)
-  const ref = await invoke(worker.id, worker.token, 'aml_model_reference', {
-    line_id: line.id, recipe_id: recipe.id,
-    controls: { [spDaq.id]: 185 }, steps: 3,
-  })
-  const refText = String(ref?.result?.text ?? '')
-  check('aml_model_reference(生产模型 what-if 预测)', refText.length > 50 && !refText.includes('尚无 production'),
-    refText.slice(0, 160).replace(/\n/g, ' '))
-  void spSeries
-  void pvSeries
-
-  console.log(`\n━━━ 结果:${passed} PASS / ${failures} FAIL ━━━`)
-  await restoreGates()
-  process.exit(failures > 0 ? 1 : 0)
-}
-
-main().catch((err) => {
-  console.error(err)
-  process.exit(1)
-})