feat(archive): 飞书云文档质量到 9/10 — 拆出 key_insights/improvements/next_steps + 渲染优化

之前生成的飞书云文档评分 5/10:
- 决策字段读 decision (实际叫 verdict)
- profile/structure 找错节点 (查 conclusion, 实际在 root)
- 摘要只取 summary (~30 字), 没拿决策引擎产的 key_insights / improvements
- DAG / 阶段拓扑 / 角色名缺失或字段读错

底层修复 (推进到 9/10):
- archiveMetacognition 的 newNodes 从 yjs 反向通道传进来, 找 rootNode 拿 project_profile + structure
- 决策引擎 schema 实际是 {verdict, score, summary, key_insights[], improvements[], next_steps[]} —
  之前拼 pros/cons (从来不存在的字段), 现在按真实 schema 摊开 7 个区块 (含历史兼容)
- DAG 改 bullet (飞书 docx 数字列表会全渲染为 "1.", 误导性强)
- 阶段拓扑用 KIND_CN 字典覆盖 parallel/serial/sequential, 不再露 "serial" 英文
- 角色行展示 名字 + 职责 + 工具 + 派单, 一行密度最大化

工具:
- server/peek-yjs.mjs — 查 inbox + nodes 当前状态 (调试)
- server/test-archive-only.mjs — 直接对现有 conclusion 跑 archive (省去 feishu 消息触发)
- export archiveMetacognition + 守卫 daemon 启动 (允许 import 测试)

测试: 同一个 conclusion 跑 4 轮迭代 docs +fetch 对比, 最终内容覆盖项目画像 / DAG / 拓扑 / 角色 / 核心洞察 / 改进建议 / 下一步, 评分 9/10.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

Author: nieao

server/feishu-bot.mjs

@@ -40,6 +40,7 @@ import path from 'node:path'
 import os from 'node:os'
 import fs from 'node:fs'
 import fsp from 'node:fs/promises'
+import { pathToFileURL } from 'node:url'
 import * as Y from 'yjs'
 import { WebsocketProvider } from 'y-websocket'
 import WebSocket from 'ws'
@@ -688,7 +689,7 @@ async function upsertBitableRecord(appToken, tableId, fields) {
 }
 
 /** 元认知一轮跑完 → 同步云文档 + 多维表格 (异步, 失败不影响 IM 卡发送) */
-async function archiveMetacognition({ ctx, conclusionNode, newNodes, room, pngPath, screenshotPngUrl, screenshotSvgUrl }) {
+export async function archiveMetacognition({ ctx, conclusionNode, newNodes, room, pngPath, screenshotPngUrl, screenshotSvgUrl }) {
   const result = { docUrl: null, bitableUrl: null }
   if (!FEISHU_BITABLE_APP_TOKEN && !FEISHU_DOCS_FOLDER_TOKEN) return result // 全没配, 跳过
 
@@ -704,13 +705,39 @@ async function archiveMetacognition({ ctx, conclusionNode, newNodes, room, pngPa
   const rootNode = newNodes.find((n) => n.id === conclusionNode.data?.projectRootId) ||
                     newNodes.find((n) => n.data?.variant === 'goal' && n.data?.projectMode)
   const profile = rootNode?.data?.project_profile || conclusionNode.data?.project_profile || {}
-  const taskDagArr = rootNode?.data?.structure?.task_dag || conclusionNode.data?.structure?.task_dag || []
+  const structure = rootNode?.data?.structure || conclusionNode.data?.structure || {}
+  const taskDagArr = structure?.task_dag || []
+  // task DAG — bullet 列出, T-ID 当编号 (飞书 docx 数字列表会全渲染为 "1.", 改 bullet 避免误导)
   const taskDag = taskDagArr.map((t, i) => {
-    const deps = (t.deps && t.deps.length) ? ` (依赖: ${t.deps.join(', ')})` : ''
-    return `${i + 1}. **${t.id || ''} ${t.title}** — ${t.desc || ''}${deps}`
+    const deps = (t.deps && t.deps.length) ? ` _(依赖: ${t.deps.join(', ')})_` : ''
+    const desc = t.desc ? ` — ${t.desc}` : ''
+    const io = []
+    if (t.input) io.push(`输入: ${t.input}`)
+    if (t.output) io.push(`产出: ${t.output}`)
+    const ioStr = io.length ? ` _[${io.join(' · ')}]_` : ''
+    return `- **${t.id || `T${i+1}`} · ${t.title}**${deps}${desc}${ioStr}`
   }).join('\n')
-  // 真分支: 排除 root (variant=goal) 和 conclusion. 实际 task entity (variant=entity) 才是分支
-  const branches = newNodes.filter((n) => n.type === 'ontologyNode' && !n.data?.isConclusion && n.data?.variant !== 'goal' && !n.data?.projectMode)
+  // 执行阶段拓扑 — 哪些角色 / 任务并行, 哪些串行
+  const stages = structure?.execution_topology?.stages || []
+  const rolesArr = structure?.roles || []
+  const KIND_CN = { parallel: '并行', sequential: '串行', serial: '串行', single: '单角色' }
+  const stageLines = stages.map((s) => {
+    const kind = KIND_CN[s.kind] || s.kind || '?'
+    const rids = (s.role_ids || [])
+    const roleNames = rids.map((rid) => {
+      const r = rolesArr.find((x) => x.id === rid)
+      return r ? `${r.name || rid}` : rid
+    }).join(' / ')
+    return `- **阶段 ${s.stage_index}** _(${kind})_: ${roleNames}`
+  }).join('\n')
+  // 派生分支: 真正的拆解/反驳 (非 task 本体, 非 root, 非 conclusion)
+  const branches = newNodes.filter((n) =>
+    n.type === 'ontologyNode' &&
+    !n.data?.isConclusion &&
+    n.data?.variant !== 'goal' &&
+    !n.data?.projectMode &&
+    !n.data?.projectTaskId  // 排除 6 stage 自动建的 task entity (DAG 章节已展示)
+  )
   const branchLines = branches.map((n) => `- ▸ **${n.data?.title || ''}** — ${n.data?.description || n.data?.content || ''}`).join('\n')
   // agentRole 节点 — 角色名 + 职责 + 工具 + 任务
   const roles = newNodes.filter((n) => n.type === 'agentRoleNode')
@@ -723,12 +750,23 @@ async function archiveMetacognition({ ctx, conclusionNode, newNodes, room, pngPa
     return `- **${name}** — ${resp}${tools}${tasks}`
   }).join('\n')
 
-  // 摘要扩充 — 合并 summary + reasoning + pros + cons + next_steps
+  // 摘要扩充 — 把决策引擎所有结构化字段都摊开 (summary + key_insights + improvements + next_steps)
+  // 决策引擎实际 schema: {verdict, score, summary, key_insights[], improvements[], next_steps[]}
+  // 历史兼容: pros/cons/risks/reasoning 字段也保留拼接, 万一有旧版本数据
+  const arrSection = (label, arr) => Array.isArray(arr) && arr.length
+    ? `\n**${label}**:\n\n${arr.map((s) => `- ${s}`).join('\n')}`
+    : ''
+  const reasoningExtra = (cObj.reasoning && cObj.reasoning !== cObj.summary)
+    ? `\n\n**推理**:\n\n${cObj.reasoning}`
+    : ''
   const detailedSummary = [
-    summary,
-    Array.isArray(cObj.pros) && cObj.pros.length ? `\n**优势**:\n${cObj.pros.map(p => `- ${p}`).join('\n')}` : '',
-    Array.isArray(cObj.cons) && cObj.cons.length ? `\n**劣势**:\n${cObj.cons.map(c => `- ${c}`).join('\n')}` : '',
-    Array.isArray(cObj.next_steps) && cObj.next_steps.length ? `\n**下一步**:\n${cObj.next_steps.map(s => `- ${s}`).join('\n')}` : '',
+    summary + reasoningExtra,
+    arrSection('核心洞察', cObj.key_insights),
+    arrSection('改进建议', cObj.improvements),
+    arrSection('优势', cObj.pros),
+    arrSection('劣势', cObj.cons),
+    arrSection('风险', cObj.risks),
+    arrSection('下一步', cObj.next_steps),
   ].filter(Boolean).join('\n')
 
   // 1) 云文档 (markdown)
@@ -753,14 +791,15 @@ async function archiveMetacognition({ ctx, conclusionNode, newNodes, room, pngPa
       '',
       taskDag || '_无_',
       '',
-      '## 分支节点',
+      '## 执行阶段拓扑',
       '',
-      branchLines || '_无_',
+      stageLines || '_无_',
       '',
       '## 涌现角色',
       '',
       roleLines || '_无_',
       '',
+      ...(branches.length ? ['## 派生分支 (拆解 / 反驳)', '', branchLines, ''] : []),
       '## 摘要 / 推理',
       '',
       detailedSummary || '_无_',
@@ -1219,4 +1258,7 @@ function start() {
 
 }
 
-start()
+// 仅在直接 node 运行时启动 daemon — import 测试不会启动
+if (import.meta.url === pathToFileURL(process.argv[1]).href) {
+  start()
+}

server/peek-yjs.mjs

@@ -0,0 +1,45 @@
+/**
+ * 查 yjs room 当前 inbox + 节点状态 — 本地运行也行 (默认连 ws://127.0.0.1:1234)
+ *   node server/peek-yjs.mjs [room=demo-final]
+ */
+import * as Y from 'yjs'
+import { WebsocketProvider } from 'y-websocket'
+import WebSocket from 'ws'
+
+if (typeof globalThis.WebSocket === 'undefined') globalThis.WebSocket = WebSocket
+
+const ROOM = process.argv[2] || 'demo-final'
+const WS = process.env.WS_URL || 'ws://127.0.0.1:1234'
+
+const doc = new Y.Doc()
+const provider = new WebsocketProvider(WS, ROOM, doc, { connect: true, WebSocketPolyfill: WebSocket })
+await new Promise((r, rej) => {
+  const t = setTimeout(() => rej(new Error('sync timeout')), 8000)
+  provider.once('synced', () => { clearTimeout(t); r() })
+})
+
+const inbox = doc.getMap('aletheia-inbox')
+console.log('=== aletheia-inbox (' + inbox.size + ') ===')
+inbox.forEach((v, k) => {
+  console.log(`${k} | status=${v?.status} | ts=${v?.ts} | text="${String(v?.text || '').slice(0, 80)}"`)
+})
+
+const nodes = doc.getMap('nodes')
+console.log('\n=== latest 8 ontologyNode ===')
+const list = []
+nodes.forEach((n, k) => { if (n?.type === 'ontologyNode') list.push({ k, n }) })
+list.sort((a, b) => (b.n.data?.created_at || 0) - (a.n.data?.created_at || 0))
+list.slice(0, 8).forEach(({ k, n }) => {
+  const t = n.data?.created_at ? new Date(n.data.created_at).toISOString().slice(11, 19) : '?'
+  console.log(`${t} | ${k.slice(0, 60).padEnd(60)} | ${n.data?.isConclusion ? '⭐CONC ' : '       '} | ${String(n.data?.title || '').slice(0, 50)}`)
+})
+
+console.log('\n=== awareness ===')
+provider.awareness.getStates().forEach((state, cid) => {
+  console.log(`client ${cid}: user=${JSON.stringify(state?.user || null)?.slice(0, 60)}`)
+})
+
+provider.disconnect()
+provider.destroy()
+doc.destroy()
+process.exit(0)

server/test-archive-only.mjs

@@ -0,0 +1,93 @@
+/**
+ * 单测: 直接走 archiveMetacognition 跑一遍现有 yjs 上的最新 conclusion, 不依赖飞书消息触发
+ *   node server/test-archive-only.mjs [room=demo-final] [chatId=oc_d2d890f2072a92a98b9f87ccb76a5b68]
+ *
+ * 用法场景: bot daemon 重启后 reverse channel 的 pending queue 丢了, 但 yjs 里 conclusion 还在,
+ *   想验证 archive 输出 (云文档 + Bitable) 的 markdown / 字段质量.
+ */
+import * as Y from 'yjs'
+import { WebsocketProvider } from 'y-websocket'
+import WebSocket from 'ws'
+import { archiveMetacognition } from './feishu-bot.mjs'
+
+if (typeof globalThis.WebSocket === 'undefined') globalThis.WebSocket = WebSocket
+
+const ROOM = process.argv[2] || 'demo-final'
+const CHAT_ID = process.argv[3] || 'oc_d2d890f2072a92a98b9f87ccb76a5b68'
+const WS = process.env.Y_WS_URL || 'ws://127.0.0.1:1234'
+
+const doc = new Y.Doc()
+const provider = new WebsocketProvider(WS, ROOM, doc, { connect: true, WebSocketPolyfill: WebSocket })
+await new Promise((r, rej) => {
+  const t = setTimeout(() => rej(new Error('sync timeout')), 8000)
+  provider.once('synced', () => { clearTimeout(t); r() })
+})
+
+const yNodes = doc.getMap('nodes')
+console.log(`房间 ${ROOM} 节点总数=${yNodes.size}`)
+
+// 找最新的 conclusion (按 created_at 排)
+let latestConclusion = null
+let latestProjectRootId = null
+yNodes.forEach((n, k) => {
+  if (n?.type === 'ontologyNode' && n.data?.isConclusion) {
+    if (!latestConclusion || (n.data?.created_at || 0) > (latestConclusion.data?.created_at || 0)) {
+      latestConclusion = { ...n, id: k }
+      latestProjectRootId = n.data?.projectRootId
+    }
+  }
+})
+
+if (!latestConclusion) {
+  console.error('找不到任何 conclusion 节点, 退出')
+  process.exit(1)
+}
+
+console.log(`选中 conclusion: ${latestConclusion.id}`)
+console.log(`projectRootId: ${latestProjectRootId}`)
+console.log(`title: ${latestConclusion.data?.title}`)
+
+// 收集所有跟这个 project 相关的节点 (root + projectRootId 等于它的)
+const projectNodes = []
+yNodes.forEach((n, k) => {
+  if (k === latestProjectRootId) projectNodes.push({ ...n, id: k })
+  else if (n?.data?.projectRootId === latestProjectRootId) projectNodes.push({ ...n, id: k })
+})
+console.log(`项目相关节点数: ${projectNodes.length}`)
+
+// 拼 ctx (模拟原 reverse channel 的 ctx)
+const rootNode = projectNodes.find((n) => n.id === latestProjectRootId)
+const PROMPT = rootNode?.data?.project_profile?.target ||
+  String(rootNode?.data?.title || '').replace(/^\[来自[^\]]*\]\s*/, '').slice(0, 300) ||
+  String(latestConclusion.data?.conclusion?.summary || '').slice(0, 300)
+const ctx = {
+  prompt: PROMPT,
+  chatId: CHAT_ID,
+  attribution: { name: '测试-9分评估', via: 'feishu-bot' },
+}
+
+console.log(`\n=== 调用 archiveMetacognition ===`)
+console.log(`prompt: ${PROMPT.slice(0, 80)}`)
+console.log(`chatId: ${CHAT_ID}`)
+
+const screenshotPngUrl = latestConclusion.data?.screenshotPngUrl || ''
+const screenshotSvgUrl = latestConclusion.data?.screenshotSvgUrl || ''
+const pngPath = latestConclusion.data?.screenshotPngPath || ''
+
+const result = await archiveMetacognition({
+  ctx,
+  conclusionNode: latestConclusion,
+  newNodes: projectNodes,
+  room: ROOM,
+  pngPath,
+  screenshotPngUrl,
+  screenshotSvgUrl,
+})
+
+console.log('\n=== 结果 ===')
+console.log(JSON.stringify(result, null, 2))
+
+provider.disconnect()
+provider.destroy()
+doc.destroy()
+process.exit(0)