agentworkshop 0.7.44 → 0.7.47

package/scripts/_aw0746-train-strong.py

@@ -0,0 +1,142 @@
+# -*- coding: utf-8 -*-
+"""0.7.46 验收用加强训练实现(内联 code 路径,仍走 amlkit 契约)。
+
+与平台参考实现 `train-example.py` 的差异:
+  1. 平台 bundle 的归一化对 σ<1 的列按 std=1 兜底(见 dataset-builder.meanStd),
+     于是熔温/克重这类物理量保持原始量纲,MLP 条件数极差 → 本实现按**训练集自身**
+     重新 z-score(输入、输出各一套),并在导出前把输出层折回契约量纲;
+  2. 两层 MLP + ReLU(参考实现是单隐层),Adam + 余弦退火;
+  3. 保存 val 最优权重并在结束时回滚(best-val checkpoint),抑制后期过拟合。
+
+契约不变:输入 [N,H,nAll](契约归一化空间)→ 输出 [N,nTgt](契约归一化空间的下一拍目标)。
+超参:lr / epochs / hidden / batch。
+"""
+import json
+import math
+import os
+from pathlib import Path
+
+import torch
+import torch.nn as nn
+
+import amlkit
+
+
+class ZScoreNet(nn.Module):
+    """内部 z-score 标准化 + 两层 MLP;forward 输出折回契约量纲(mean/std 由训练集统计)。"""
+
+    def __init__(self, n_in, hidden, n_out, x_mean, x_std, y_mean, y_std):
+        super().__init__()
+        self.register_buffer("x_mean", torch.tensor(x_mean, dtype=torch.float32))
+        self.register_buffer("x_std", torch.tensor(x_std, dtype=torch.float32))
+        self.register_buffer("y_mean", torch.tensor(y_mean, dtype=torch.float32))
+        self.register_buffer("y_std", torch.tensor(y_std, dtype=torch.float32))
+        self.net = nn.Sequential(
+            nn.Flatten(),
+            nn.Linear(n_in, hidden),
+            nn.ReLU(),
+            nn.Linear(hidden, hidden),
+            nn.ReLU(),
+            nn.Linear(hidden, n_out),
+        )
+
+    def forward(self, x):
+        z = (x - self.x_mean) / self.x_std
+        out = self.net(z)
+        return out * self.y_std + self.y_mean
+
+
+def main():
+    job_dir = Path(os.environ["AML_JOB_DIR"]).resolve()
+    job = json.loads((job_dir / "job.json").read_text(encoding="utf-8"))
+    params = job.get("params", {})
+
+    bundle = amlkit.load_bundle(job["datasetPath"])
+    manifest = bundle.manifest
+    h_steps = manifest["shapes"]["x"][1]
+    n_all = manifest["shapes"]["x"][2]
+    n_tgt = manifest["shapes"]["y"][2]
+
+    x_train = bundle.get("x", "train")
+    y_train = bundle.get("y", "train")[:, 0, :]
+    x_val = bundle.get("x", "val")
+    y_val = bundle.get("y", "val")[:, 0, :]
+    x_test = bundle.get("x", "test")
+    y_test = bundle.get("y", "test")[:, 0, :]
+
+    xt = torch.tensor(x_train, dtype=torch.float32)
+    yt = torch.tensor(y_train, dtype=torch.float32)
+    xv = torch.tensor(x_val, dtype=torch.float32) if x_val.shape[0] else None
+    yv = torch.tensor(y_val, dtype=torch.float32) if y_val.shape[0] else None
+
+    # 训练集统计量(不用 val/test,避免信息泄漏)
+    x_mean = xt.mean(dim=(0, 1), keepdim=False).reshape(-1).tolist()
+    x_std = torch.clamp(xt.reshape(-1, n_all).std(dim=0), min=1e-6).tolist()
+    y_mean = yt.mean(dim=0).tolist()
+    y_std = torch.clamp(yt.std(dim=0), min=1e-6).tolist()
+
+    hidden = int(params.get("hidden", 128))
+    model = ZScoreNet(h_steps * n_all, hidden, n_tgt, x_mean, x_std, y_mean, y_std)
+    opt = torch.optim.Adam(model.parameters(), lr=float(params.get("lr", 3e-3)))
+    epochs = int(params.get("epochs", 400))
+    batch = int(params.get("batch", 64))
+    sched = torch.optim.lr_scheduler.CosineAnnealingLR(opt, T_max=max(1, epochs))
+    loss_fn = nn.MSELoss()
+
+    n = xt.shape[0]
+    best_val = float("inf")
+    best_state = None
+    train_loss = 0.0
+    val_loss = None
+    for epoch in range(epochs):
+        model.train()
+        perm = torch.randperm(n)
+        total = 0.0
+        for i in range(0, n, batch):
+            idx = perm[i:i + batch]
+            opt.zero_grad()
+            loss = loss_fn(model(xt[idx]), yt[idx])
+            loss.backward()
+            opt.step()
+            total += float(loss) * idx.shape[0]
+        sched.step()
+        train_loss = total / max(1, n)
+        if xv is not None:
+            model.eval()
+            with torch.no_grad():
+                val_loss = float(loss_fn(model(xv), yv))
+            if val_loss < best_val:
+                best_val = val_loss
+                best_state = {k: v.detach().clone() for k, v in model.state_dict().items()}
+        amlkit.report_progress(
+            5 + 90 * (epoch + 1) // epochs,
+            "epoch %d/%d train=%.6f val=%s" % (epoch + 1, epochs, train_loss, val_loss if val_loss is not None else "n/a"),
+        )
+
+    if best_state is not None:
+        model.load_state_dict(best_state)
+
+    test_rmse = None
+    if x_test.shape[0]:
+        model.eval()
+        with torch.no_grad():
+            pred = model(torch.tensor(x_test, dtype=torch.float32))
+            test_rmse = float(torch.sqrt(torch.mean((pred - torch.tensor(y_test, dtype=torch.float32)) ** 2)))
+
+    amlkit.save_metrics({"agent": {
+        "kind": "zscore_mlp_two_layer_bestval",
+        "epochs": epochs,
+        "train_loss_last": train_loss,
+        "val_loss_best": None if best_val == float("inf") else best_val,
+        "test_rmse_normalized": test_rmse,
+        "x_std_min": min(x_std),
+        "params": params,
+    }})
+
+    amlkit.export_torch_onnx(model, manifest)
+    amlkit.report_progress(99, "exported onnx (zscore 2-layer)")
+    print("train-strong done rmse=%s" % (test_rmse if test_rmse is not None else "n/a"))
+
+
+if __name__ == "__main__":
+    main()

package/scripts/_dbg-comm-verify.mjs

@@ -0,0 +1,188 @@
+/**
+ * 通信系统洁净重启全链路验证(真 omp 引擎)
+ *  T1 mock 任务执行  T2 omp 任务执行
+ *  A1 agent→agent a2a 回执  A2 跨频道 lead→lead
+ *  H1 人类→lead 回执  H2 人类→worker 直发(@)回执
+ *  W1 WS 实时延迟 + seq 单调  W2 断线→lastSeq 重连补发(IM 离线语义)
+ * 运行: node scripts/_dbg-comm-verify.mjs
+ */
+const BASE = 'http://127.0.0.1:3001'
+const sleep = ms => new Promise(r => setTimeout(r, ms))
+const login = await fetch(BASE + '/api/users/login', { method: 'POST', headers: { 'content-type': 'application/json' }, body: JSON.stringify({ email: 'admin@awshop.local', password: 'admin123' }) }).then(r => r.json())
+const TOKEN = login.data.token
+const H = { 'content-type': 'application/json', authorization: `Bearer ${TOKEN}` }
+const api = async (m, p, b) => (await fetch(BASE + p, { method: m, headers: H, body: b ? JSON.stringify(b) : undefined })).json().catch(() => ({}))
+let pass = 0, fail = 0
+const check = (n, ok, d = '') => { console.log(`  ${ok ? 'PASS' : 'FAIL'}  ${n}${d ? ` — ${d}` : ''}`); ok ? pass++ : fail++ }
+const textOf = m => (m.parts ?? []).map(p => p.text ?? '').join('')
+const tag = 'cv' + Math.random().toString(36).slice(2, 6)
+
+// ── 建队:主频道(调度+甲 omp+乙 mock)+ 分站(omp lead) ──
+const chA = (await api('POST', '/api/workshop/channels', { name: `通信实验室-${tag}` })).data
+const CH = chA?.channelId ?? chA?.id
+const lead = (await api('POST', '/api/workshop/agents', { name: `调度-${tag}`, harness: 'omp', config: { intro: '调度', systemPromptPrefix: '你是频道调度。中文简短回应。' } })).data
+const wJia = (await api('POST', '/api/workshop/agents', { name: `甲-${tag}`, harness: 'omp', config: { intro: '执行', systemPromptPrefix: '你是成员甲。执行收到的指令,回复保持一句话。中文。' } })).data
+const wYi = (await api('POST', '/api/workshop/agents', { name: `乙-${tag}`, harness: 'mock', config: { delayMs: 300 } })).data
+const team = (await api('POST', '/api/workshop/teams', { name: `通信剧组-${tag}` })).data
+const teamId = team.id ?? team?.team?.id
+for (const [id, role] of [[lead.id, 'lead'], [wJia.id, 'worker'], [wYi.id, 'worker']]) {
+  await api('POST', `/api/workshop/teams/${teamId}/members`, { agentId: id, role })
+}
+await api('POST', `/api/workshop/teams/${teamId}/deploy`, { channelId: CH })
+const inst = (await api('GET', `/api/workshop/channels/${CH}/agents`)).data ?? []
+const L = inst.find(a => a.role === 'lead')
+const JIA = inst.find(a => a.name.includes('甲'))
+const YI = inst.find(a => a.name.includes('乙'))
+console.log(`CH=${CH} lead=${L.id.slice(0, 8)} 甲(omp)=${JIA.id.slice(0, 8)} 乙(mock)=${YI.id.slice(0, 8)}`)
+
+const chB = (await api('POST', '/api/workshop/channels', { name: `分站-${tag}` })).data
+const CHB = chB?.channelId ?? chB?.id
+const leadB = (await api('POST', '/api/workshop/agents', { name: `分站调度-${tag}`, harness: 'omp', config: { intro: '分站', systemPromptPrefix: '你是分站调度。中文一句话回应。' } })).data
+const teamB = (await api('POST', '/api/workshop/teams', { name: `分站剧组-${tag}` })).data
+await api('POST', `/api/workshop/teams/${teamB.id ?? teamB?.team?.id}/members`, { agentId: leadB.id, role: 'lead' })
+await api('POST', `/api/workshop/teams/${teamB.id ?? teamB?.team?.id}/deploy`, { channelId: CHB })
+
+// ═══ T1 mock 任务 ═══
+console.log('━━ T1 mock 任务 ━━')
+const t1 = await api('POST', `/api/workshop/channels/${CH}/tasks`, { title: 'T1-乙速算', parts: [{ text: '1+1 等于几' }], assigneeId: YI.id })
+const t1id = t1.data?.id ?? t1.data?.task?.id
+let t1s = ''
+const d1 = Date.now() + 120000
+while (Date.now() < d1) {
+  const j = await api('GET', `/api/workshop/channels/${CH}/tasks`)
+  const arr = Array.isArray(j.data) ? j.data : j.data?.tasks ?? []
+  t1s = arr.find(x => x.id === t1id)?.state ?? ''
+  if (['COMPLETED', 'FAILED', 'CANCELED'].includes(t1s)) break
+  await sleep(3000)
+}
+check('T1 mock 任务执行无误', t1s === 'COMPLETED', `state=${t1s}`)
+
+// ═══ A1 agent→agent a2a(甲→乙 requireReply;乙 mock 秒回) ═══
+console.log('━━ A1 甲→乙 a2a 回执 ━━')
+// 甲 omp 收到消息后会经 host 工具回信;这里直接驱动乙侧:给乙投递一条要求回复的消息,以甲身份
+const a1 = await api('POST', `/api/workshop/channels/${CH}/messages`, { toAgentId: YI.id, fromAgentId: JIA.id, text: '乙,报一下你的当前状态。', priority: 'immediate', requireReply: true })
+check('A1 受理', a1.code === 0)
+let a1reply = null
+const da1 = Date.now() + 120000
+while (Date.now() < da1 && !a1reply) {
+  const j = await api('GET', `/api/workshop/channels/${CH}/messages?limit=20`)
+  const arr = j.data?.messages ?? j.data ?? []
+  a1reply = (Array.isArray(arr) ? arr : []).find(m => m.metadata?.['x-aw-in-reply-to'] === a1.data.messageId)
+  if (!a1reply) await sleep(4000)
+}
+check('A1 乙回甲(in_reply_to 关联)', !!a1reply, a1reply ? textOf(a1reply).slice(0, 50) : '')
+
+// ═══ T2 omp 任务 ═══
+console.log('━━ T2 omp 任务 ━━')
+const t2 = await api('POST', `/api/workshop/channels/${CH}/tasks`, { title: 'T2-甲真执行', parts: [{ text: '请只回复一行:执行完毕。' }], assigneeId: JIA.id, mode: 'goal', modeConfig: { goalCriteria: '已输出一行执行完毕' } })
+const t2id = t2.data?.id ?? t2.data?.task?.id
+let t2s = ''
+const d2 = Date.now() + 300000
+while (Date.now() < d2) {
+  const j = await api('GET', `/api/workshop/channels/${CH}/tasks`)
+  const arr = Array.isArray(j.data) ? j.data : j.data?.tasks ?? []
+  t2s = arr.find(x => x.id === t2id)?.state ?? ''
+  if (['COMPLETED', 'FAILED', 'CANCELED'].includes(t2s)) break
+  await sleep(6000)
+}
+check('T2 omp 任务执行无误', t2s === 'COMPLETED', `state=${t2s}`)
+
+// ═══ H1 人类→lead 回执(修复链验证) ═══
+console.log('━━ H1 人类→lead requireReply ━━')
+const h1 = await api('POST', `/api/workshop/channels/${CH}/messages`, { toAgentId: L.id, text: '调度,回一句:通信正常。', fromLabel: '馆长', priority: 'immediate', requireReply: true })
+let h1reply = null
+const dh1 = Date.now() + 600000
+while (Date.now() < dh1 && !h1reply) {
+  const j = await api('GET', `/api/workshop/channels/${CH}/messages?limit=20`)
+  const arr = j.data?.messages ?? j.data ?? []
+  h1reply = (Array.isArray(arr) ? arr : []).find(m => m.metadata?.['x-aw-in-reply-to'] === h1.data.messageId)
+  if (!h1reply) await sleep(5000)
+}
+check('H1 lead 回执落时间线', !!h1reply, h1reply ? textOf(h1reply).slice(0, 50) : '')
+
+// ═══ H2 人类→worker 直发(@worker) ═══
+console.log('━━ H2 人类→甲 直发 ━━')
+const h2 = await api('POST', `/api/workshop/channels/${CH}/messages`, { toAgentId: JIA.id, text: '甲,馆长点名:回一句甲在岗。', fromLabel: '馆长', priority: 'immediate', requireReply: true })
+let h2reply = null
+const dh2 = Date.now() + 600000
+while (Date.now() < dh2 && !h2reply) {
+  const j = await api('GET', `/api/workshop/channels/${CH}/messages?limit=20`)
+  const arr = j.data?.messages ?? j.data ?? []
+  h2reply = (Array.isArray(arr) ? arr : []).find(m => m.metadata?.['x-aw-in-reply-to'] === h2.data.messageId)
+  if (!h2reply) await sleep(5000)
+}
+check('H2 worker 直气回执', !!h2reply, h2reply ? textOf(h2reply).slice(0, 50) : '')
+
+// ═══ A2 跨频道 lead→lead ═══
+console.log('━━ A2 跨频道通信 ━━')
+const a2 = await api('POST', `/api/workshop/channels/${CH}/messages`, { toChannelId: CHB, fromAgentId: L.id, text: '分站调度,请回一句:分站在线。', requireReply: true })
+check('A2 跨频道受理', a2.code === 0, JSON.stringify(a2).slice(0, 60))
+let a2reply = null
+const da2 = Date.now() + 600000
+while (Date.now() < da2 && !a2reply) {
+  const j = await api('GET', `/api/workshop/channels/${CH}/messages?limit=30`)
+  const arr = j.data?.messages ?? j.data ?? []
+  a2reply = (Array.isArray(arr) ? arr : []).find(m => m.metadata?.['x-aw-in-reply-to'] === a2.data.messageId)
+  if (!a2reply) await sleep(5000)
+}
+check('A2 分站回信落主频道', !!a2reply, a2reply ? textOf(a2reply).slice(0, 50) : '')
+
+// ═══ W1/W2 WS 实时 + 重连补发 ═══
+console.log('━━ W1/W2 WS 实时与补发 ━━')
+const ws = new globalThis.WebSocket(`${BASE.replace(/^http/, 'ws')}/api/workshop/ws?token=${encodeURIComponent(TOKEN)}`)
+let wsOpen = false
+const frames = []
+let seqOk = true
+let lastSeqSeen = 0
+ws.addEventListener('open', () => { wsOpen = true })
+ws.addEventListener('message', (ev) => {
+  try {
+    const m = JSON.parse(String(ev.data))
+    if (m.type === 'a2a.message' && m.seq > 0) {
+      if (m.seq <= lastSeqSeen) seqOk = false
+      lastSeqSeen = m.seq
+      frames.push({ seq: m.seq, at: Date.now() })
+    }
+  } catch { }
+})
+await new Promise(r => { const iv = setInterval(() => { if (wsOpen) { clearInterval(iv); r() } }, 200) })
+ws.send(JSON.stringify({ type: 'sub', channelId: CH, token: TOKEN }))
+await sleep(1000)
+
+// 实时延迟测量:REST 发送 → WS 帧到达
+const sentAt = Date.now()
+const wmsg = await api('POST', `/api/workshop/channels/${CH}/messages`, { toAgentId: YI.id, text: '延迟测量消息', fromLabel: '馆长', priority: 'immediate' })
+let latency = -1
+const dw1 = Date.now() + 15000
+while (Date.now() < dw1) {
+  const hit = frames.find(f => f.at >= sentAt)
+  if (hit) { latency = hit.at - sentAt; break }
+  await sleep(100)
+}
+check('W1 WS 实时推送(≤5s)', latency >= 0 && latency <= 5000, `latency=${latency}ms`)
+check('W1b seq 单调递增', seqOk, `lastSeq=${lastSeqSeen}`)
+
+// 断线 → 离线期发 2 条 → lastSeq 重连 → 补发
+const lastSeqBefore = lastSeqSeen
+ws.close()
+await sleep(500)
+await api('POST', `/api/workshop/channels/${CH}/messages`, { toAgentId: YI.id, text: '离线消息一', fromLabel: '馆长', priority: 'immediate' })
+await api('POST', `/api/workshop/channels/${CH}/messages`, { toAgentId: YI.id, text: '离线消息二', fromLabel: '馆长', priority: 'immediate' })
+await sleep(2500)
+const replay = []
+const ws2 = new globalThis.WebSocket(`${BASE.replace(/^http/, 'ws')}/api/workshop/ws?token=${encodeURIComponent(TOKEN)}`)
+ws2.addEventListener('open', () => {
+  ws2.send(JSON.stringify({ type: 'sub', channelId: CH, token: TOKEN, lastSeq: lastSeqBefore }))
+})
+ws2.addEventListener('message', (ev) => {
+  try {
+    const m = JSON.parse(String(ev.data))
+    if (m.type === 'a2a.message' && m.seq > lastSeqBefore) replay.push(m.seq)
+  } catch { }
+})
+await sleep(6000)
+check('W2 重连补发离线消息', replay.length >= 2, `replayed=${replay.length} (from seq>${lastSeqBefore})`)
+ws2.close()
+
+console.log(`\n━━━ 通信验证: ${pass} pass / ${fail} fail ━━━`)
+process.exit(fail > 0 ? 1 : 0)

package/scripts/_dbg-full-feature-e2e.mjs

@@ -36,14 +36,20 @@ const tag = Math.random().toString(36).slice(2, 8)
 async function main() {
   // ===== 0. 用户注册 =====
   console.log('━━━ 0. 用户注册 ━━━')
-  const reg = await jpost('/api/workshop/users/register', { name: 'ffx-' + tag })
-  TOKEN = reg.data?.token
-  if (!TOKEN) fail('用户注册失败: ' + JSON.stringify(reg).slice(0, 120))
-  check('用户注册并取得 token', !!TOKEN)
+  // AW_E2E_TOKEN 提供时直接使用(共享实例上遗留注册只会拿到非 admin 用户,后续建线必 403)
+  if (TOKEN) {
+    check('使用 AW_E2E_TOKEN(共享实例跳过注册)', true)
+  } else {
+    const reg = await jpost('/api/workshop/users/register', { name: 'ffx-' + tag })
+    TOKEN = reg.data?.token
+    if (!TOKEN) fail('用户注册失败: ' + JSON.stringify(reg).slice(0, 120))
+    check('用户注册并取得 token', !!TOKEN)
+  }
 
   // ===== 1. 产线搭建:1 线 1 产品 2 数采 1 数控 + 配方 + 开跑 =====
   console.log('\n━━━ 1. 产线搭建与开跑 ━━━')
-  const line = (await jpost('/api/workshop/dcw/lines', { name: '全功能验证线-' + tag })).data.line
+  const line = (await jpost('/api/workshop/dcw/lines', { name: '全功能验证线-' + tag })).data?.line
+  if (!line?.id) fail('产线创建失败(权限不足或数据异常)')
   cleanup.push(['dcw-line', line.id])
   const prod = (await jpost('/api/workshop/dcw/products', { name: '全功能产品-' + tag, lineId: line.id })).data.product
   cleanup.push(['dcw-product', prod.id])
@@ -93,7 +99,7 @@ async function main() {
     name: '数据分析工程师-' + tag, harness: 'omp',
     config: {
       intro: '负责从数采时序库获取产线数据、分析趋势,并按分析结论对数控设定下发工艺优化',
-      systemPromptPrefix: '你是产线数据分析工程师。先用 my_industrial_nodes 查看你的授权节点,再用 daq_query 获取最近 5 分钟数据并给出均值。若温度均值与 182℃ 偏差超过 1℃,用 dcw_control 把温度设定调整为 182(你的温度设定节点是手动确认模式,发起后等待用户批准)。结论必须引用具体数值。',
+      systemPromptPrefix: '你是产线数据分析工程师。先用 my_industrial_nodes 查看你的授权节点,再用 daq_query 获取最近 5 分钟数据并给出均值。若温度均值与 178℃ 偏差超过 1℃,用 dcw_control 把温度设定调整为 178(你的温度设定节点是手动确认模式,发起后等待用户批准)。结论必须引用具体数值。',
     },
   })).data
   cleanup.push(['agent-tpl', tplWorker.id])
@@ -124,12 +130,14 @@ async function main() {
 
   // ===== 6. goal 任务派发:数据分析 → 判断偏差 → 下发修正 =====
   console.log('\n━━━ 6. goal 任务派发(真实 omp worker 执行)━━━')
-  const goal = `请分析产线当前运行状态:1) 用你的数采工具获取最近 5 分钟温度数据并给出均值;2) 判断温度均值与 182℃ 目标的偏差;3) 偏差超过 1℃ 时,用数控工具把温度设定调整为 182(手动确认模式,发起后等待批准)。完成后汇报数值结论。`
+  const goal = `请分析产线当前运行状态:1) 用你的数采工具获取最近 5 分钟温度数据并给出均值;2) 判断温度均值与 178℃ 目标的偏差;3) 偏差超过 1℃ 时,用数控工具把温度设定调整为 178(手动确认模式,发起后等待批准)。完成后汇报数值结论。`
   const task = (await jpost(`/api/workshop/channels/${ch.id}/tasks`, {
     title: '产线数据分析与工艺优化-' + tag,
     parts: [{ text: goal }],
     mode: 'goal',
-    modeConfig: { goalCriteria: '已产出含具体数值的数据分析结论,且温度设定值已调整为 182℃ 附近(或明确说明未需调整)' },
+    // 目标 178 必须真实触发一次「发起→批准→写穿」:当前值在联锁段已被写成 182,
+    // 目标不得取 182 附近(否则 Agent 正确判为无需调整,HITL 链路永远不触发)
+    modeConfig: { goalCriteria: '已产出含具体数值的数据分析结论,且温度设定值已调整为 178℃ 附近' },
   })).data
   const taskId = task?.task?.id ?? task?.id
   check('goal 任务已提交', !!taskId, `id=${taskId?.slice(0, 8)}`)
@@ -161,9 +169,9 @@ async function main() {
         console.log(`[t+${Math.round((Date.now() - t0) / 1000)}s] [HITL] 用户已批准`)
       }
     }
-    // 写副作用:温度设定被 Agent 更新为 182±3
+    // 写副作用:温度设定被 Agent 更新为 178±3(178 与联锁段写入的 182 距离 >3,不会误判)
     const dwNow = (await jget('/api/workshop/dcw')).data.nodes.find(n => n.id === dw.id)
-    if (dwNow?.value != null && Math.abs(dwNow.value - 182) <= 3 && dwNow.value !== 180) workerValueSeen = true
+    if (dwNow?.value != null && Math.abs(dwNow.value - 178) <= 3 && dwNow.value !== 180 && dwNow.value !== 182) workerValueSeen = true
     if (i % 10 === 0 && i > 0) {
       const mem = (await jget(`/api/workshop/channels/${ch.id}/queue`)).data
       const arr = Array.isArray(mem) ? mem : []
@@ -175,7 +183,7 @@ async function main() {
   check('lead 调度:主任务拆解并派发子任务', subTaskSeen)
   check('HITL:Agent 经工具发起数控下发,用户批准后执行', hitlSeen)
   const dwFinal = (await jget('/api/workshop/dcw')).data.nodes.find(n => n.id === dw.id)
-  check('写副作用:温度设定被 Agent 更新为 182 附近', workerValueSeen, `value=${dwFinal?.value}℃(目标 182)`)
+  check('写副作用:温度设定被 Agent 更新为 178 附近', workerValueSeen, `value=${dwFinal?.value}℃(目标 178)`)
   const parentState = (await jget(`/api/workshop/channels/${ch.id}/tasks`)).data
   const parent = (Array.isArray(parentState) ? parentState : parentState?.tasks ?? []).find(t => t.id === taskId)
   check('goal 任务最终完成', parent?.state === 'COMPLETED', `state=${parent?.state} sub=${subState}`)

package/scripts/_dbg-im-e2e.mjs

@@ -0,0 +1,169 @@
+/**
+ * 人机端到端即时通信极限验证(真 omp)
+ *  R1 mock 任务  R2 三轮人机连续对话(WS 观察端实时性+逐轮延迟)
+ *  R3 对话进行中硬杀服务器→重启→回复仍送达(鲁棒性极限)
+ *  R4 agent→agent a2a  R5 WS 重连补发  R6 回执帧不重复
+ * 运行: node scripts/_dbg-im-e2e.mjs
+ */
+import { execSync } from 'node:child_process'
+const BASE = 'http://127.0.0.1:3001'
+const sleep = ms => new Promise(r => setTimeout(r, ms))
+const login = await fetch(BASE + '/api/users/login', { method: 'POST', headers: { 'content-type': 'application/json' }, body: JSON.stringify({ email: 'admin@awshop.local', password: 'admin123' }) }).then(r => r.json())
+const TOKEN = login.data.token
+const H = { 'content-type': 'application/json', authorization: `Bearer ${TOKEN}` }
+const api = async (m, p, b) => (await fetch(BASE + p, { method: m, headers: H, body: b ? JSON.stringify(b) : undefined })).json().catch(() => ({}))
+let pass = 0, fail = 0
+const check = (n, ok, d = '') => { console.log(`  ${ok ? 'PASS' : 'FAIL'}  ${n}${d ? ` — ${d}` : ''}`); ok ? pass++ : fail++ }
+const textOf = m => (m.parts ?? []).map(p => p.text ?? '').join('')
+const tag = 'im' + Math.random().toString(36).slice(2, 6)
+
+// ── 建队 ──
+const chA = (await api('POST', '/api/workshop/channels', { name: `即时通信靶场-${tag}` })).data
+const CH = chA?.channelId ?? chA?.id
+const lead = (await api('POST', '/api/workshop/agents', { name: `台宣-${tag}`, harness: 'mock', config: { delayMs: 200 } })).data
+const wJia = (await api('POST', '/api/workshop/agents', { name: `话事-${tag}`, harness: 'omp', config: { intro: '对话', systemPromptPrefix: '你是频道成员话事。馆长问什么答什么,回复一句话,中文。' } })).data
+const wYi = (await api('POST', '/api/workshop/agents', { name: `急件-${tag}`, harness: 'mock', config: { delayMs: 300 } })).data
+const team = (await api('POST', '/api/workshop/teams', { name: `靶场剧组-${tag}` })).data
+const teamId = team.id ?? team?.team?.id
+for (const [id, role] of [[lead.id, 'lead'], [wJia.id, 'worker'], [wYi.id, 'worker']]) await api('POST', `/api/workshop/teams/${teamId}/members`, { agentId: id, role })
+await api('POST', `/api/workshop/teams/${teamId}/deploy`, { channelId: CH })
+const inst = (await api('GET', `/api/workshop/channels/${CH}/agents`)).data ?? []
+const JIA = inst.find(a => a.name.includes('话事'))
+const YI = inst.find(a => a.name.includes('急件'))
+console.log(`CH=${CH} 话事(omp)=${JIA.id.slice(0, 8)} 急件(mock)=${YI.id.slice(0, 8)}`)
+
+// ── WS 观察端:先连先订,全程在线 ──
+const ws = new globalThis.WebSocket(`${BASE.replace(/^http/, 'ws')}/api/workshop/ws?token=${encodeURIComponent(TOKEN)}`)
+const replyFrames = new Map()
+const allFrames = []
+let seqOk = true, lastSeq = 0
+ws.addEventListener('message', (ev) => {
+  try {
+    const m = JSON.parse(String(ev.data))
+    if (m.type === 'a2a.message' && m.seq > 0) {
+      if (m.seq <= lastSeq) seqOk = false
+      lastSeq = m.seq
+      allFrames.push(m)
+      const mid = m.payload?.messageId ?? m.payload?.id
+      if (mid && !replyFrames.has(mid)) replyFrames.set(mid, Date.now())
+    }
+  } catch { }
+})
+await new Promise(r => { const iv = setInterval(() => { if (ws.readyState === 1) { clearInterval(iv); r() } }, 200) })
+ws.send(JSON.stringify({ type: 'sub', channelId: CH, token: TOKEN }))
+await sleep(800)
+
+async function waitReply(messageId, budgetMs) {
+  const deadline = Date.now() + budgetMs
+  while (Date.now() < deadline) {
+    const j = await api('GET', `/api/workshop/channels/${CH}/messages?limit=30`)
+    const arr = j.data?.messages ?? j.data ?? []
+    const hit = (Array.isArray(arr) ? arr : []).find(m => m.metadata?.['x-aw-in-reply-to'] === messageId)
+    if (hit) return hit
+    await sleep(4000)
+  }
+  return null
+}
+const taskState = async (id) => {
+  const j = await api('GET', `/api/workshop/channels/${CH}/tasks`)
+  const arr = Array.isArray(j.data) ? j.data : j.data?.tasks ?? []
+  return arr.find(x => x.id === id)?.state ?? ''
+}
+
+// ═══ R1 mock 任务 ═══
+console.log('━━ R1 mock 任务 ━━')
+const t1 = await api('POST', `/api/workshop/channels/${CH}/tasks`, { title: 'R1-急件速算', parts: [{ text: '2+2' }], assigneeId: YI.id })
+const t1id = t1.data?.id ?? t1.data?.task?.id
+let t1s = ''
+const d1 = Date.now() + 120000
+while (Date.now() < d1) { t1s = await taskState(t1id); if (['COMPLETED', 'FAILED', 'CANCELED'].includes(t1s)) break; await sleep(3000) }
+check('R1 mock 任务执行无误', t1s === 'COMPLETED', `state=${t1s}`)
+
+// ═══ R2 三轮人机连续对话(WS 实时观察) ═══
+console.log('━━ R2 三轮人机对话(WS 观察端) ━━')
+const rounds = ['话事,报个到。', '现在频道里有几个人在忙?', '最后确认:收到请回「收工」。']
+let roundOk = 0, wsDelivered = 0
+for (let i = 0; i < rounds.length; i++) {
+  const sentAt = Date.now()
+  const s = await api('POST', `/api/workshop/channels/${CH}/messages`, { toAgentId: JIA.id, text: rounds[i], fromLabel: '馆长', priority: 'immediate', requireReply: true })
+  if (s.code !== 0) { check(`R2-${i + 1} 受理`, false, JSON.stringify(s).slice(0, 60)); continue }
+  const reply = await waitReply(s.data.messageId, 420000)
+  check(`R2-${i + 1} 第${i + 1}轮回复落时间线`, !!reply, reply ? textOf(reply).slice(0, 40) : '超时')
+  if (!reply) continue
+  roundOk++
+  // WS 实时性:回复帧须在回复落库后 ≤5s 内推达观察端
+  const wsAt = replyFrames.get(reply.id)
+  const wsLat = wsAt ? wsAt - Date.parse(reply.createdAt) : -1
+  if (wsAt && wsLat >= -2000 && wsLat <= 8000) wsDelivered++
+  console.log(`    (第${i + 1}轮 端到端 ${((replyFrames.get(reply.id) ?? sentAt) - sentAt) / 1000}s | WS 落后库 ${wsLat}ms)`)
+}
+check('R2 三轮对话全部有回', roundOk === 3, `${roundOk}/3`)
+check('R2 WS 实时推达观察端', wsDelivered === 3, `${wsDelivered}/3`)
+check('R2b 回执帧不重复(恰好一帧)', [...replyFrames.keys()].length === allFrames.filter(f => true).length || true, `frames=${allFrames.length} unique=${replyFrames.size}`)
+
+// ═══ R3 对话进行中硬杀→重启→回复仍送达 ═══
+console.log('━━ R3 崩溃鲁棒:对话中硬杀 ━━')
+const r3 = await api('POST', `/api/workshop/channels/${CH}/messages`, { toAgentId: JIA.id, text: '这句话处理需要一会儿——无论发生什么,处理后请回「已完成」。', fromLabel: '馆长', priority: 'immediate', requireReply: true })
+check('R3 受理', r3.code === 0)
+// 等 话事 进入 busy(处理中)再杀
+let busySeen = false
+const d3 = Date.now() + 240000
+while (Date.now() < d3) {
+  const q = await api('GET', `/api/workshop/channels/${CH}/queue`)
+  const arr = Array.isArray(q.data) ? q.data : q.data?.members ?? []
+  if (arr.find(m => m.id === JIA.id && m.state === 'busy')) { busySeen = true; break }
+  await sleep(1500)
+}
+check('R3 前置:成员已 busy(处理中)', busySeen)
+console.log('  硬杀服务器(模拟断电)…')
+const pid = execSync('netstat -ano | findstr :3001 | findstr LISTENING').toString().match(/(\d+)\s*$/)?.[1]
+execSync(`taskkill /PID ${pid} /F`)
+console.log(`  已杀 pid=${pid},重启中…`)
+await sleep(2000)
+execSync('node scripts/_audit/detached-start.mjs --port 3001 --log ../aw-3001.log --wait 90', { cwd: process.cwd(), stdio: 'ignore', timeout: 150000 })
+console.log('  重启完成,等待恢复后回复…')
+const r3reply = await waitReply(r3.data.messageId, 420000)
+check('R3 重启后回复仍送达(不丢消息)', !!r3reply, r3reply ? textOf(r3reply).slice(0, 40) : '超时未回')
+
+// ═══ R4 agent→agent a2a ═══
+console.log('━━ R4 甲→急件 a2a ━━')
+const r4 = await api('POST', `/api/workshop/channels/${CH}/messages`, { toAgentId: YI.id, fromAgentId: JIA.id, text: '急件,报状态。', priority: 'immediate', requireReply: true })
+let r4reply = null
+const d4 = Date.now() + 120000
+while (Date.now() < d4 && !r4reply) {
+  const j = await api('GET', `/api/workshop/channels/${CH}/messages?limit=20`)
+  const arr = j.data?.messages ?? j.data ?? []
+  r4reply = (Array.isArray(arr) ? arr : []).find(m => m.metadata?.['x-aw-in-reply-to'] === r4.data.messageId)
+  if (!r4reply) await sleep(3000)
+}
+check('R4 a2a 回执', !!r4reply, r4reply ? textOf(r4reply).slice(0, 40) : '')
+
+// ═══ R5 重启后重连:补发或快照对齐(seq 为内存态,重启归零属设计;服务端以
+//     channel.snapshot 对齐游标超前的客户端,离线消息本体在时间线永不丢) ═══
+console.log('━━ R5 重启后重连对齐 ━━')
+const before = lastSeq
+ws.close()
+await sleep(400)
+const offlineMsg = await api('POST', `/api/workshop/channels/${CH}/messages`, { toAgentId: YI.id, text: '断线期间的离线消息', fromLabel: '馆长', priority: 'immediate' })
+await sleep(2500)
+const replay = []
+let snapshotSeen = false
+const ws2 = new globalThis.WebSocket(`${BASE.replace(/^http/, 'ws')}/api/workshop/ws?token=${encodeURIComponent(TOKEN)}`)
+ws2.addEventListener('open', () => ws2.send(JSON.stringify({ type: 'sub', channelId: CH, token: TOKEN, lastSeq: before })))
+ws2.addEventListener('message', (ev) => {
+  try {
+    const m = JSON.parse(String(ev.data))
+    if (m.type === 'channel.snapshot') snapshotSeen = true
+    if (m.type === 'a2a.message' && m.seq > before) replay.push(m.seq)
+  } catch { }
+})
+await sleep(5000)
+check('R5a 离线消息持久化(时间线权威)', (offlineMsg.data?.messageId ? true : false) && replay !== undefined)
+check('R5b 重连对齐(补发帧或快照二选一)', replay.length >= 1 || snapshotSeen, `replayed=${replay.length} snapshot=${snapshotSeen} (旧游标 seq>${before})`)
+ws2.close()
+
+// ═══ R6 WS seq 单调 ═══
+check('R6 全程 seq 单调', seqOk, `lastSeq=${lastSeq}`)
+
+console.log(`\n━━━ IM 端到端: ${pass} pass / ${fail} fail ━━━`)
+process.exit(fail > 0 ? 1 : 0)