岗位会商

frontend-web/docs/position-roundtable-delivery-workflow-plan.md

@@ -4,6 +4,29 @@
 >
 > 目标：在现有岗位会商基础上补齐“产物交付、查看、驳回、方案总结、行动规划”的完整闭环，同时继续保持原多智能体会商页面不受影响。
 
+## 实施核对（2026-07-21）
+
+本计划已按“岗位会商独立路由、无总控、各席位直接问答”的边界完成第一轮闭环。实际流程调整为：
+
+```text
+任务信息 → 情报分析岗识别并确认意图 → 启动校验并冻结业务链
+→ 分阶段席位直接问答/交付 Markdown → 全部有效交付
+→ 方案总结智能体（roundtable-summary）→ 行动规划智能体（position-action-planner）
+→ 可随时返回交付产物驳回 → 后续阶段及内置产物级联失效 → 重做后重新收口
+```
+
+- [x] 情报意图、业务链选择、席位岗位归属和席位智能体配置均在启动前校验；启动后保存业务链快照。
+- [x] 普通席位直接调用各自分配的智能体，并注入任务、意图及已完成上游席位的问答/产物；不经过总控智能体。
+- [x] 普通席位必须在 `outputs` 中交付 Markdown 才会完成；情报分析岗以确认的结构化意图完成为准。
+- [x] 节点、交付卡片、流程图和大图弹窗均展示 `locked / ready / running / done / rejected / stale / error` 状态，并可回看节点会话和产物。
+- [x] 驳回要求标题和理由；当前节点转为 `rejected`，后续阶段按依赖转为 `stale/locked`，总结和行动规划同时失效，历史内容保留可查看。
+- [x] 被驳回或失效的普通席位必须写入一份**新的或已改写** Markdown 才能再次完成；单纯继续问答且没有新交付不会误使下游产物失效。
+- [x] 方案总结复用 `roundtable-summary` 的真实流式线程；行动规划新增 `position-action-planner`，强制输出 Markdown 报告和 `action-plan-subtasks.json`。
+- [x] 两个内置智能体的线程、产物快照和来源节点版本栅栏均持久化；流式期间若上游版本变化，旧运行不能覆盖最新状态。
+- [x] 右侧“交付产物”可查看内置节点结果、继续同一线程问答/修改交付，并在打开 Markdown 沙箱时自动收起右栏。
+
+设计收敛：当前版本以节点的 `artifact_manifest`、`latest_answer`、版本、驳回记录及 Session 内置快照作为唯一事实来源，已能覆盖统一产物卡片和级联逻辑，因此不新增重复的 `position_roundtable_deliverables` 表。若后续需要“单个文件独立驳回、跨会话检索或完整版本对比”，再将清单归一化为独立产物表，避免现在的双写一致性风险。
+
 ## 0. 核心结论
 
 - 方案总结智能体建议复用现有会商里的 `roundtable-summary` 能力，但不要直接搬整套 Step 3 页面流程。
@@ -75,32 +98,32 @@
 
 ### 3.1 Session 扩展
 
-`position_roundtable_sessions` 建议增加：
+`position_roundtable_sessions` 已使用下列持久化字段：
 
 | 字段 | 含义 |
 | --- | --- |
-| `workflow_status` | `intent_pending / active / summarizing / planning / completed / archived` |
+| `status` | `intent_pending / active / completed / archived`；总结/规划运行态由前端瞬时状态与对应快照的 `status` 表示，避免一次运行中断后留下伪运行态 |
 | `summary_thread_id` | 方案总结智能体线程 |
 | `summary_snapshot` | 最新有效总结报告快照 |
 | `action_plan_thread_id` | 行动规划智能体线程 |
 | `action_plan_snapshot` | 最新有效行动规划快照 |
-| `invalidated_at` | 最近一次级联失效时间，可选 |
+| `invalidated_at` | 记录在 `summary_snapshot` / `action_plan_snapshot` 内，包含 `invalidated_by`，无需重复写 Session 字段 |
 
 ### 3.2 Node 扩展
 
-`position_roundtable_nodes` 建议增加：
+`position_roundtable_nodes` 已使用下列字段（展示状态由前端按状态映射）：
 
 | 字段 | 含义 |
 | --- | --- |
-| `display_status` | 面向前端的聚合状态，可后端计算也可前端映射 |
+| `status` | `locked / ready / running / done / stale / rejected / error`，前端直接映射为展示状态 |
 | `rejection_count` | 当前节点累计驳回次数 |
 | `last_rejection` | 最新驳回标题、理由、操作者、时间 |
-| `valid_revision` | 当前有效交付版本 |
+| `revision` | 当前有效交付版本 |
 | `invalidated_by` | 导致当前节点失效的上游节点 key |
 
-### 3.3 Deliverable 新表
+### 3.3 Deliverable 新表（后续可选）
 
-建议新增 `position_roundtable_deliverables`，不要只依赖 `artifact_manifest`。
+当需要“单个文件独立驳回、跨会话检索或完整版本对比”时，再新增 `position_roundtable_deliverables`，不要只依赖 `artifact_manifest`。
 
 | 字段 | 含义 |
 | --- | --- |
@@ -121,7 +144,7 @@
 | `created_at` | 创建时间 |
 | `updated_at` | 更新时间 |
 
-这样右侧产物卡片、驳回记录、总结输入材料都能从同一张表聚合。
+当前第一轮不新增这张表：右侧产物卡片、驳回记录、总结输入材料从节点 `artifact_manifest`、节点版本/驳回记录和 Session 快照聚合，避免节点表与产物表双写。后续引入该表时，应把它提升为唯一事实来源，而不是并行维护两套状态。
 
 ## 4. 驳回与级联失效规则
 
@@ -173,25 +196,27 @@
 
 | 方法 | 路径 | 用途 |
 | --- | --- | --- |
-| `GET` | `/api/position-roundtable/sessions/{id}/deliverables` | 获取右侧产物列表 |
+| `GET` | `/api/position-roundtable/sessions/{id}/results` | 获取右侧统一产物来源节点（前端聚合文件与内置快照） |
-| `GET` | `/api/position-roundtable/deliverables/{deliverable_id}` | 获取产物详情 |
+| `GET` | `/api/position-roundtable/sessions/{id}/nodes/{node_key}` | 获取节点详情；文件正文由既有线程产物预览接口读取 |
-| `POST` | `/api/position-roundtable/deliverables/{deliverable_id}/reject` | 驳回产物并触发级联失效 |
+| `POST` | `/api/position-roundtable/sessions/{id}/nodes/{node_key}/reject` | 驳回当前节点交付并触发级联失效 |
-| `POST` | `/api/position-roundtable/nodes/{node_id}/complete-turn` | 节点回答结束后同步回答与产物 |
+| `POST` | `/api/position-roundtable/sessions/{id}/nodes/{node_key}/complete-turn` | 节点回答结束后同步回答与产物 |
 
 ### 5.3 方案总结与行动规划
 
 | 方法 | 路径 | 用途 |
 | --- | --- | --- |
-| `POST` | `/api/position-roundtable/sessions/{id}/summary/start` | 启动或重启方案总结 |
+| `POST` | `/api/position-roundtable/sessions/{id}/summary/prepare` | 服务端准备版本锁定的总结材料和线程信息 |
-| `POST` | `/api/position-roundtable/sessions/{id}/summary/ask` | 对方案总结继续提问或修改 |
+| `POST` | `/api/position-roundtable/sessions/{id}/summary/ask` | 服务端准备总结的同线程续问材料 |
-| `POST` | `/api/position-roundtable/sessions/{id}/action-plan/start` | 基于总结和产物生成行动规划 |
+| `POST` | `/api/position-roundtable/sessions/{id}/summary/complete` | 校验 Markdown 交付与来源版本后持久化总结快照 |
-| `POST` | `/api/position-roundtable/sessions/{id}/action-plan/ask` | 对行动规划继续提问或修改 |
+| `POST` | `/api/position-roundtable/sessions/{id}/action-plan/prepare` | 服务端准备行动规划材料和线程信息 |
+| `POST` | `/api/position-roundtable/sessions/{id}/action-plan/ask` | 服务端准备行动规划同线程续问材料 |
+| `POST` | `/api/position-roundtable/sessions/{id}/action-plan/complete` | 校验 Markdown + JSON 交付与来源版本后持久化行动规划快照 |
 
 可以复用现有 `streamMultiAgent` 流式协议，但建议在岗位会商下包一层 API，由后端负责拼装岗位会商材料，前端不直接拼长 prompt。
 
 ## 6. 前端实施步骤
 
-### 阶段 1：状态与启动校验
+### 阶段 1：状态与启动校验（已完成）
 
 - 扩展 `PositionNodeStatus` 和状态显示映射。
 - “确认意图并启动业务链”前补齐校验提示。
@@ -201,7 +226,7 @@
 
 验收标准：未选择业务链、岗位未配齐、意图未完成时都能准确阻止启动；启动后流程图能显示各节点状态。
 
-### 阶段 2：流程图弹窗
+### 阶段 2：流程图弹窗（已完成）
 
 - 左侧保留小型从上到下流程图。
 - 流程图右上角增加“查看”按钮。
@@ -210,7 +235,7 @@
 
 验收标准：小图不挤占左侧空间，大图能完整展示业务链和两个内置收口节点。
 
-### 阶段 3：产物卡片与预览
+### 阶段 3：产物卡片与预览（已完成）
 
 - 右侧新增或改造“分析结果”Tab 为“交付产物”列表。
 - 卡片展示智能体名称、产物名称、岗位名称、时间、状态。
@@ -220,7 +245,7 @@
 
 验收标准：节点回答和文件产物能进入统一产物列表；点击可查看详情。
 
-### 阶段 4：驳回弹窗与级联失效
+### 阶段 4：驳回弹窗与级联失效（已完成）
 
 - 新增 `RejectDeliverableDialog`。
 - 驳回表单包含标题和理由，均必填。
@@ -230,7 +255,7 @@
 
 验收标准：驳回第一阶段任意节点后，第二阶段及之后产物都显示未完成或已失效；旧内容可查看但不能作为总结输入。
 
-### 阶段 5：方案总结智能体
+### 阶段 5：方案总结智能体（已完成）
 
 - 封装 `usePositionSummary`，内部复用 `roundtable-summary` 的线程初始化、流式输出和 Markdown 产物预览能力。
 - 输入材料改为岗位会商材料：
@@ -245,7 +270,7 @@
 
 验收标准：总结报告使用现有方案总结能力生成 Markdown；刷新页面后可恢复；上游驳回后总结标记失效。
 
-### 阶段 6：行动规划智能体
+### 阶段 6：行动规划智能体（已完成）
 
 - 新增内置智能体 ID，建议先命名为 `position-action-planner`。
 - 新增行动规划 prompt：
@@ -261,7 +286,7 @@
 
 验收标准：行动规划能生成可读报告和结构化子任务；刷新可恢复；上游驳回后自动失效。
 
-### 阶段 7：返回修改与重新生成
+### 阶段 7：返回修改与重新生成（已完成）
 
 - 用户在方案总结或行动规划视图中仍可切回“交付产物”。
 - 驳回后自动显示需要重做的节点。
@@ -270,7 +295,7 @@
 
 验收标准：不会因为进入总结/规划阶段而锁死前序产物；前序重做后能形成新版本闭环。
 
-### 阶段 8：测试与回归
+### 阶段 8：测试与回归（已完成）
 
 - 后端测试：
   - activate 校验。

frontend-web/src/components/workspace/workspace-nav-chat-list.tsx

@@ -14,7 +14,7 @@ import {
 } from "@/components/ui/sidebar";
 import { useI18n } from "@/core/i18n/hooks";
 // import { BotIcon, BrainIcon, ListChecks, MessageSquarePlus, MessagesSquare, SparklesIcon, UsersIcon } from "lucide-react";
-import { BookOpen, BotIcon, BrainIcon, ListChecks, MessageSquarePlus, MessagesSquare, PenLine, Plus, SparklesIcon, Users } from "lucide-react";
+import { BookOpen, BotIcon, BrainIcon, BriefcaseBusiness, ListChecks, MessageSquarePlus, MessagesSquare, PenLine, Plus, SparklesIcon, Users } from "lucide-react";
 
 import { KnowledgeRagToggle } from "./knowledge-rag-toggle";
 import { SidebarCollapseButton } from "./workspace-header";
@@ -184,6 +184,18 @@ export function WorkspaceNavChatList({
                 </Link>
               </SidebarMenuButton>
             </SidebarMenuItem>
+            <SidebarMenuItem>
+              <SidebarMenuButton
+                isActive={pathname.startsWith("/page/workspace/roundtable/position")}
+                asChild
+                className={NAV_ITEM_CLASS}
+              >
+                <Link href="/page/workspace/roundtable/position">
+                  <BriefcaseBusiness />
+                  <span>岗位会商</span>
+                </Link>
+              </SidebarMenuButton>
+            </SidebarMenuItem>
             <SidebarMenuItem>
               <SidebarMenuButton
                 isActive={pathname === "/page/workspace/scheduled/tasks"}

frontend-web/src/pages/WorkspaceRoutes.tsx

@@ -54,6 +54,7 @@ import AIWritingPage from '@/open-canvas/pages/AIWritingPage';
 import StudioNotebooksPage from './StudioNotebooksPage';
 import StudioNotebookDetailPage from './StudioNotebookDetailPage';
 import RoundtablePlanningPage from '@/roundtable-planning/pages/RoundtablePlanningPage';
+import PositionRoundtablePage from '@/position-roundtable/pages/PositionRoundtablePage';
 import BusinessChainEditorPage from '@/roundtable-planning/pages/BusinessChainEditorPage';
 import BusinessChainListPage from '@/roundtable-planning/pages/BusinessChainListPage';
 import DispatchChainPreviewPage from '@/roundtable-planning/pages/DispatchChainPreviewPage';
@@ -433,6 +434,16 @@ export default function WorkspaceRoutes({
             </WorkspaceLayout>
           }
         />
+        <Route
+          path="roundtable/position"
+          element={
+            <WorkspaceLayout {...layoutProps}>
+              <ChatRuntime>
+                <PositionRoundtablePage compactChrome={headerHidden} />
+              </ChatRuntime>
+            </WorkspaceLayout>
+          }
+        />
         <Route
           path="roundtable/chains"
           element={

frontend-web/src/position-roundtable/api/sessions.ts

@@ -45,7 +45,9 @@ export interface PositionRoundtableSession {
   chainSnapshot: Record<string, unknown> | null;
   intentThreadId: string | null;
   intentSnapshot: Record<string, unknown> | null;
+  summaryThreadId: string | null;
   summarySnapshot: Record<string, unknown> | null;
+  actionPlanThreadId: string | null;
   actionPlanSnapshot: Record<string, unknown> | null;
   status: "intent_pending" | "active" | "completed" | "archived";
   createdAt?: string;
@@ -82,7 +84,9 @@ interface ServerSession {
   chain_snapshot?: Record<string, unknown> | null;
   intent_thread_id: string | null;
   intent_snapshot?: Record<string, unknown> | null;
+  summary_thread_id?: string | null;
   summary_snapshot?: Record<string, unknown> | null;
+  action_plan_thread_id?: string | null;
   action_plan_snapshot?: Record<string, unknown> | null;
   status: PositionRoundtableSession["status"];
   created_at?: string;
@@ -132,7 +136,9 @@ function toSession(session: ServerSession): PositionRoundtableSession {
     chainSnapshot: session.chain_snapshot ?? null,
     intentThreadId: session.intent_thread_id,
     intentSnapshot: session.intent_snapshot ?? null,
+    summaryThreadId: session.summary_thread_id ?? null,
     summarySnapshot: session.summary_snapshot ?? null,
+    actionPlanThreadId: session.action_plan_thread_id ?? null,
     actionPlanSnapshot: session.action_plan_snapshot ?? null,
     status: session.status,
     createdAt: session.created_at,
@@ -301,22 +307,104 @@ export async function rejectPositionNode(
   return toSession(result);
 }
 
-export async function generatePositionSummary(sessionId: string): Promise<PositionRoundtableSession> {
+export interface PositionBuiltinRunPreparation {
-  const result = await readJson<ServerSession>(
+  agentId: string;
-    await apiFetch(`${BASE()}/sessions/${encodeURIComponent(sessionId)}/summary/start`, {
+  displayName: string;
+  threadId: string | null;
+  prompt: string;
+  sourceRevisions: Record<string, number>;
+}
+
+interface ServerBuiltinRunPreparation {
+  agent_id: string;
+  display_name: string;
+  thread_id?: string | null;
+  prompt: string;
+  source_revisions?: Record<string, number>;
+}
+
+function toBuiltinRunPreparation(value: ServerBuiltinRunPreparation): PositionBuiltinRunPreparation {
+  return {
+    agentId: value.agent_id,
+    displayName: value.display_name,
+    threadId: value.thread_id ?? null,
+    prompt: value.prompt,
+    sourceRevisions: value.source_revisions ?? {},
+  };
+}
+
+type BuiltinKind = "summary" | "action_plan";
+
+async function prepareBuiltinRun(
+  sessionId: string,
+  kind: BuiltinKind,
+  message = "",
+): Promise<PositionBuiltinRunPreparation> {
+  const path = kind === "summary" ? "summary/prepare" : "action-plan/prepare";
+  const result = await readJson<ServerBuiltinRunPreparation>(
+    await apiFetch(`${BASE()}/sessions/${encodeURIComponent(sessionId)}/${path}`, {
       method: "POST",
+      headers: { "Content-Type": "application/json" },
+      body: JSON.stringify({ message }),
     }),
-    "generate position summary",
+    `prepare position ${kind}`,
+  );
+  return toBuiltinRunPreparation(result);
+}
+
+async function completeBuiltinRun(
+  sessionId: string,
+  kind: BuiltinKind,
+  input: {
+    threadId: string;
+    content: string;
+    artifactManifest: Array<Record<string, unknown>>;
+    sourceRevisions: Record<string, number>;
+    model?: string;
+  },
+): Promise<PositionRoundtableSession> {
+  const path = kind === "summary" ? "summary/complete" : "action-plan/complete";
+  const result = await readJson<ServerSession>(
+    await apiFetch(`${BASE()}/sessions/${encodeURIComponent(sessionId)}/${path}`, {
+      method: "POST",
+      headers: { "Content-Type": "application/json" },
+      body: JSON.stringify({
+        thread_id: input.threadId,
+        content: input.content,
+        artifact_manifest: input.artifactManifest,
+        source_revisions: input.sourceRevisions,
+        model: input.model || null,
+      }),
+    }),
+    `complete position ${kind}`,
   );
   return toSession(result);
 }
 
-export async function generatePositionActionPlan(sessionId: string): Promise<PositionRoundtableSession> {
+export function preparePositionSummary(
-  const result = await readJson<ServerSession>(
+  sessionId: string,
-    await apiFetch(`${BASE()}/sessions/${encodeURIComponent(sessionId)}/action-plan/start`, {
+  message = "",
-      method: "POST",
+): Promise<PositionBuiltinRunPreparation> {
-    }),
+  return prepareBuiltinRun(sessionId, "summary", message);
-    "generate position action plan",
+}
-  );
+
-  return toSession(result);
+export function completePositionSummary(
+  sessionId: string,
+  input: Parameters<typeof completeBuiltinRun>[2],
+): Promise<PositionRoundtableSession> {
+  return completeBuiltinRun(sessionId, "summary", input);
+}
+
+export function preparePositionActionPlan(
+  sessionId: string,
+  message = "",
+): Promise<PositionBuiltinRunPreparation> {
+  return prepareBuiltinRun(sessionId, "action_plan", message);
+}
+
+export function completePositionActionPlan(
+  sessionId: string,
+  input: Parameters<typeof completeBuiltinRun>[2],
+): Promise<PositionRoundtableSession> {
+  return completeBuiltinRun(sessionId, "action_plan", input);
 }

frontend-web/src/position-roundtable/components/PositionNodeChat.tsx

@@ -1,11 +1,12 @@
 import { useCallback, useEffect, useMemo, useRef, useState } from "react";
-import { Bot, FileText, LockKeyhole, Loader2 } from "lucide-react";
+import { Bot, FileText, LockKeyhole } from "lucide-react";
 import { toast } from "sonner";
-
 import type { PromptInputMessage } from "@/components/ai-elements/prompt-input";
 import { InputBox } from "@/components/workspace/input-box";
 import { ThreadContext } from "@/components/workspace/messages/context";
 import { MessageList } from "@/components/workspace/messages";
+import { useArtifactsOptional } from "@/components/workspace/artifacts";
+import { useAgent } from "@/core/agents";
 import { useLocalSettings } from "@/core/settings";
 import { useThreadStream } from "@/core/threads/hooks";
 import { textOfMessage } from "@/core/threads/utils";
@@ -17,15 +18,26 @@ import {
   type PositionRoundtableNode,
 } from "@/position-roundtable/api/sessions";
 import { listSessionArtifacts } from "@/roundtable-planning/api/multi-agent";
-import { ArtifactPreviewModal } from "@/roundtable-planning/components/ArtifactPreviewModal";
+import { desensitize } from "@/lib/desensitize";
 
 interface PositionNodeChatProps {
   sessionId: string;
+  /** Parent task id: all node threads are indexed under this task. */
+  taskId: string;
   chainId: string | null;
   node: PositionRoundtableNode;
   currentPositionId: string;
   sessionStatus: "intent_pending" | "active" | "completed" | "archived";
+  agentName?: string | null;
   onNodeUpdate: (node: PositionRoundtableNode) => void;
+  /** The workspace collapses the configuration panel on the first real turn. */
+  onConversationStart?: () => void;
+  onPreviewArtifact?: (artifact: {
+    threadId: string;
+    path: string;
+    name: string;
+    mimeType: string | null;
+  }) => void;
 }
 
 interface NodeArtifact {
@@ -52,6 +64,26 @@ function readNodeArtifacts(manifest: Array<Record<string, unknown>>): NodeArtifa
   });
 }
 
+function hasMarkdownArtifact(manifest: Array<Record<string, unknown>>): boolean {
+  return manifest.some((artifact) => {
+    const path = typeof artifact.path === "string" ? artifact.path.toLowerCase() : "";
+    const name = typeof artifact.name === "string" ? artifact.name.toLowerCase() : "";
+    const mime =
+      typeof artifact.mime_type === "string"
+        ? artifact.mime_type.toLowerCase()
+        : typeof artifact.mimeType === "string"
+          ? artifact.mimeType.toLowerCase()
+          : "";
+    return (
+      mime.includes("markdown") ||
+      path.endsWith(".md") ||
+      path.endsWith(".markdown") ||
+      name.endsWith(".md") ||
+      name.endsWith(".markdown")
+    );
+  });
+}
+
 /**
  * Direct single-agent chat for one frozen business-chain node.
  *
@@ -61,18 +93,24 @@ function readNodeArtifacts(manifest: Array<Record<string, unknown>>): NodeArtifa
  */
 export function PositionNodeChat({
   sessionId,
+  taskId,
   chainId,
   node,
   currentPositionId,
   sessionStatus,
+  agentName,
   onNodeUpdate,
+  onConversationStart,
+  onPreviewArtifact,
 }: PositionNodeChatProps) {
   const [settings, setSettings] = useLocalSettings();
+  const { agent } = useAgent(node.agentId);
+  const agentDisplayName = desensitize(agentName || agent?.name || node.agentId);
+  const artifactsCtx = useArtifactsOptional();
   const [boundThreadId, setBoundThreadId] = useState<string | null>(node.threadId);
   const boundThreadIdRef = useRef<string | null>(node.threadId);
   const draftThreadIdRef = useRef(node.threadId ?? uuid());
   const upstreamRevisionRef = useRef<Record<string, unknown> | null>(null);
-  const [previewArtifact, setPreviewArtifact] = useState<NodeArtifact | null>(null);
   const nodeArtifacts = useMemo(() => readNodeArtifacts(node.artifactManifest), [node.artifactManifest]);
   const canWrite =
     sessionStatus !== "archived" &&
@@ -81,6 +119,19 @@ export function PositionNodeChat({
     node.status !== "error" &&
     !node.invalidatedBy;
 
+  useEffect(() => {
+    if (artifactsCtx?.open && artifactsCtx.selectedArtifact) {
+      onPreviewArtifact?.({
+        threadId: boundThreadIdRef.current ?? boundThreadId ?? draftThreadIdRef.current,
+        path: artifactsCtx.selectedArtifact,
+        name: artifactsCtx.selectedArtifact.split("/").filter(Boolean).at(-1) || "Markdown 报告",
+        mimeType: artifactsCtx.selectedArtifact.toLowerCase().endsWith(".md")
+          ? "text/markdown"
+          : null,
+      });
+    }
+  }, [artifactsCtx?.open, artifactsCtx?.selectedArtifact, boundThreadId, onPreviewArtifact]);
+
   useEffect(() => {
     setBoundThreadId(node.threadId);
     boundThreadIdRef.current = node.threadId;
@@ -99,14 +150,21 @@ export function PositionNodeChat({
     [node.nodeKey, onNodeUpdate, sessionId],
   );
 
-  const [thread, sendMessage] = useThreadStream({
+  const [thread, sendMessage, , actualThreadId] = useThreadStream({
     threadId: boundThreadId,
     assistantId: node.agentId,
     context: { ...settings.context, agent_id: node.agentId },
     threadMetadata: {
       thread_type: "position_roundtable",
+      // ``taskId`` is the shared task-workspace convention.  The two
+      // position-specific identifiers retain the exact child node for audits
+      // and restore, even when one position owns more than one chain seat.
+      taskId,
+      position_roundtable_task_id: taskId,
+      position_task_id: `${taskId}:${node.positionId ?? "unassigned"}:${node.nodeKey}`,
       position_session_id: sessionId,
       position_node_id: node.id,
+      position_node_key: node.nodeKey,
       position_id: node.positionId,
       chain_id: chainId,
       agent_id: node.agentId,
@@ -128,6 +186,7 @@ export function PositionNodeChat({
             path: file.path,
             mime_type: file.mimeType,
             modified_at: file.modifiedAt,
+            size_bytes: file.sizeBytes,
           }));
         } catch {
           // Artifact listing is best-effort: a valid text answer must still be
@@ -140,6 +199,9 @@ export function PositionNodeChat({
             upstreamRevision: upstreamRevisionRef.current,
           });
           onNodeUpdate(updated);
+          if (updated.status !== "done" && !hasMarkdownArtifact(artifactManifest)) {
+            toast.info("已保存本轮回答；该席位需生成并交付一份 Markdown 报告文档后才会标记完成。");
+          }
         } catch (error) {
           toast.error(`节点结果保存失败：${error instanceof Error ? error.message : String(error)}`);
         }
@@ -147,6 +209,14 @@ export function PositionNodeChat({
     },
   });
 
+  // SDK thread creation is asynchronous.  ``onStart`` normally binds it, but
+  // watching the actual stream id as well closes the last race where a page
+  // refresh could happen between creation and the async node-index write.
+  useEffect(() => {
+    if (!actualThreadId || actualThreadId === boundThreadIdRef.current) return;
+    bindCreatedThread(actualThreadId);
+  }, [actualThreadId, bindCreatedThread]);
+
   const handleSubmit = useCallback(
     async (message: PromptInputMessage) => {
       if (!canWrite) {
@@ -162,6 +232,7 @@ export function PositionNodeChat({
         return;
       }
       try {
+        onConversationStart?.();
         const taskExternalContext = await getPositionNodeContext(sessionId, node.nodeKey);
         const upstreamRevision = taskExternalContext.upstream_revision;
         upstreamRevisionRef.current =
@@ -176,7 +247,16 @@ export function PositionNodeChat({
         toast.error(`发送失败：${error instanceof Error ? error.message : String(error)}`);
       }
     },
-    [canWrite, node.agentId, node.nodeKey, node.status, sendMessage, sessionId, sessionStatus],
+    [
+      canWrite,
+      node.agentId,
+      node.nodeKey,
+      node.status,
+      onConversationStart,
+      sendMessage,
+      sessionId,
+      sessionStatus,
+    ],
   );
 
   const status = thread.error ? "error" : thread.isLoading ? "streaming" : "ready";
@@ -201,13 +281,12 @@ export function PositionNodeChat({
           </span>
           <div className="min-w-0 flex-1">
             <h2 className="truncate text-sm font-extrabold text-slate-900 dark:text-[#e8edf5]">
-              {node.agentId}
+              {agentDisplayName}
             </h2>
             <p className="truncate text-[10px] text-muted-foreground dark:text-[#8aa0c4]">
               {readOnlyHint || `节点状态：${node.status}`}
             </p>
           </div>
-          {thread.isLoading && <Loader2 className="size-4 animate-spin text-primary" />}
         </header>
 
         {nodeArtifacts.length > 0 && (
@@ -225,7 +304,16 @@ export function PositionNodeChat({
                   key={artifact.path}
                   type="button"
                   title={`预览 ${artifact.name}`}
-                  onClick={() => setPreviewArtifact(artifact)}
+                  onClick={() => {
+                    if (boundThreadId && onPreviewArtifact) {
+                      onPreviewArtifact({
+                        threadId: boundThreadId,
+                        path: artifact.path,
+                        name: artifact.name,
+                        mimeType: artifact.mimeType,
+                      });
+                    }
+                  }}
                   className="inline-flex max-w-full items-center gap-1 rounded-md border border-slate-200 bg-white px-2 py-1 text-left text-[11px] text-slate-600 transition hover:border-primary/45 hover:text-primary dark:border-[#2b8ffe]/35 dark:bg-[#0a315d]/80 dark:text-[#c5cfe0]"
                 >
                   <FileText className="size-3 shrink-0" />
@@ -242,6 +330,7 @@ export function PositionNodeChat({
             threadId={boundThreadId}
             thread={thread}
             paddingBottom={24}
+            scrollButtonClassName="bottom-6"
             onEditResubmit={(text) => void handleSubmit({ text, files: [] })}
           />
         ) : (
@@ -275,15 +364,6 @@ export function PositionNodeChat({
             onStop={() => void thread.stop()}
           />
         </div>
-        {previewArtifact && boundThreadId && (
-          <ArtifactPreviewModal
-            threadId={boundThreadId}
-            path={previewArtifact.path}
-            name={previewArtifact.name}
-            mimeType={previewArtifact.mimeType}
-            onClose={() => setPreviewArtifact(null)}
-          />
-        )}
       </div>
     </ThreadContext.Provider>
   );

frontend-web/src/roundtable-planning/api/intent.ts

@@ -97,11 +97,16 @@ function parseIntentErrorDetail(value: unknown): IntentErrorDetail | null {
  * `detail.trace_tail` to help diagnose intranet deployment issues without
  * shell access to the server.
  */
-export async function initIntent(model?: string): Promise<IntentInitResponse> {
+export async function initIntent(
+  model?: string,
+  taskContext?: Record<string, unknown> | null,
+): Promise<IntentInitResponse> {
   const res = await apiFetch(`${getBackendBaseURL()}/api/intent/init`, {
     method: "POST",
     headers: { "Content-Type": "application/json" },
-    body: JSON.stringify({ model }),
+    // 岗位会商在创建意图线程时就写入父任务标识。这样即使用户尚未确认
+    // 业务链，第一阶段的对话也已经可以按 taskId 恢复。
+    body: JSON.stringify({ model, task_context: taskContext ?? undefined }),
   });
   if (!res.ok) {
     const raw = await res.text().catch(() => "");
@@ -159,6 +164,8 @@ export interface StreamIntentRequest {
   threadId: string;
   message: string;
   model?: string;
+  /** Optional hidden business/task context injected into the model run. */
+  taskContext?: Record<string, unknown> | null;
   /** Step 1 用户上传的参考文件（可选）。 */
   files?: IntentFileRef[];
   signal?: AbortSignal;
@@ -333,6 +340,7 @@ export async function streamIntent(req: StreamIntentRequest): Promise<StreamInte
       thread_id: req.threadId,
       message: req.message,
       model: req.model,
+      task_context: req.taskContext ?? undefined,
       // 仅在有文件时带上，空数组后端也安全忽略。
       files: req.files && req.files.length > 0 ? req.files : undefined,
     }),

frontend-web/src/roundtable-planning/api/multi-agent.ts

@@ -1054,6 +1054,36 @@ export async function initSummaryThread(
   return { agentName: json.agent_name, threadId: json.thread_id };
 }
 
+/**
+ * Create a per-session thread for the position-roundtable action-planning
+ * built-in agent.  The agent id is fixed; callers persist/reuse the returned
+ * thread id for subsequent plan refinements.
+ */
+export async function initPositionActionPlanThread(
+  options: { signal?: AbortSignal } = {},
+): Promise<ReportThreadResponse> {
+  const res = await apiFetch(`${getBackendBaseURL()}/api/multi-agent/position-action-plan/init`, {
+    method: "POST",
+    headers: { "Content-Type": "application/json" },
+    body: "{}",
+    signal: options.signal,
+  });
+  if (!res.ok) {
+    const text = await res.text().catch(() => "");
+    reportRoundtableError({
+      stage: "position_action_plan",
+      event: "init_thread_http_error",
+      message: `行动规划线程创建失败 (HTTP ${res.status}): ${text.slice(0, 500)}`,
+      httpStatus: res.status,
+      agentId: "position-action-planner",
+      agentName: "行动规划智能体",
+    });
+    throw new Error(`init position action-plan thread failed (${res.status}): ${text}`);
+  }
+  const json = (await res.json()) as { agent_name: string; thread_id: string };
+  return { agentName: json.agent_name, threadId: json.thread_id };
+}
+
 /**
  * Create a thread for the Step 3「入库」structure agent (`roundtable-structure`,
  * a built-in singleton — the same agent that extracts the flowchart, but this run

frontend-web/src/roundtable-planning/components/Step1Panel.tsx

@@ -72,6 +72,7 @@ export interface Step1PanelProps {
    * 不传 → 显示默认「任务意图理解 (人机问答)」标题。右侧「问答轮次」始终保留。
    */
   headerLeftSlot?: React.ReactNode;
+  headerRightSlot?: React.ReactNode;
   interactiveMessages: InteractiveMessage[];
   intentThreadId: string | null;
   intentReady: object | null;
@@ -101,6 +102,7 @@ export interface Step1PanelProps {
 export function Step1Panel(props: Step1PanelProps) {
   const {
     headerLeftSlot,
+    headerRightSlot,
     interactiveMessages,
     intentThreadId,
     intentReady,
@@ -175,9 +177,12 @@ export function Step1Panel(props: Step1PanelProps) {
             </h3>
           </div>
         )}
+        <div className="flex shrink-0 items-center gap-2">
         <span className="shrink-0 text-[11px] font-bold text-slate-400 bg-white border px-2 py-0.5 rounded-md">
           问答轮次：{interactiveMessages.filter((m) => m.sender === "user").length} 轮
         </span>
+          {headerRightSlot}
+        </div>
       </div>
 
       {/* Chat list */}

frontend-web/src/roundtable-planning/hooks/useStep1Intent.ts

@@ -90,6 +90,10 @@ export interface UseStep1IntentOptions {
   currentStep: number;
   /** 全局选中模型;调用 `/api/intent/init` / `/api/intent/stream` 时透传。 */
   selectedModel: string;
+  /** Optional hidden task/business context for each intent-agent turn. */
+  taskContext?: Record<string, unknown> | null;
+  /** Optional task-aware welcome bubble text before the first user turn. */
+  initialWelcomeText?: string;
   /** 主页面统一的 toast 入口。 */
   triggerToast: (msg: string) => void;
 }
@@ -167,7 +171,7 @@ function intentPhaseLabel(phase: IntentStreamPhase, toolName?: string): string {
 }
 
 export function useStep1Intent(options: UseStep1IntentOptions): UseStep1IntentReturn {
-  const { currentStep, selectedModel, triggerToast } = options;
+  const { currentStep, selectedModel, taskContext, initialWelcomeText, triggerToast } = options;
 
   const [intentThreadId, setIntentThreadId] = useState<string | null>(null);
   const [intentReady, setIntentReady] = useState<IntentSummary | null>(null);
@@ -198,6 +202,14 @@ export function useStep1Intent(options: UseStep1IntentOptions): UseStep1IntentRe
   useEffect(() => {
     selectedModelRef.current = selectedModel;
   }, [selectedModel]);
+  const taskContextRef = useRef<Record<string, unknown> | null>(taskContext ?? null);
+  useEffect(() => {
+    taskContextRef.current = taskContext ?? null;
+  }, [taskContext]);
+  const initialWelcomeTextRef = useRef(initialWelcomeText);
+  useEffect(() => {
+    initialWelcomeTextRef.current = initialWelcomeText;
+  }, [initialWelcomeText]);
 
   const stopIntentStream = useCallback(() => {
     if (intentAborterRef.current) {
@@ -225,7 +237,7 @@ export function useStep1Intent(options: UseStep1IntentOptions): UseStep1IntentRe
         id: 1,
         sender: "agent",
         senderName: "任务理解智能体",
-        content: INTENT_WELCOME_TEXT,
+        content: initialWelcomeTextRef.current || INTENT_WELCOME_TEXT,
         time: nowTime(),
       },
     ]);
@@ -233,7 +245,10 @@ export function useStep1Intent(options: UseStep1IntentOptions): UseStep1IntentRe
       try {
         // initIntent 当前不在 thread 上落盘 model,只为占位与兼容;空时由后端
         // 走 lead_agent._resolve_model_name 回退到 config.yaml models[0]
-        const res = await initIntent(selectedModelRef.current || undefined);
+        const res = await initIntent(
+          selectedModelRef.current || undefined,
+          taskContextRef.current,
+        );
         if (step1InitGenRef.current !== myGen) return;
         setIntentThreadId(res.thread_id);
         setTimelineEntries([
@@ -408,6 +423,7 @@ export function useStep1Intent(options: UseStep1IntentOptions): UseStep1IntentRe
           threadId: intentThreadId,
           message: trimmed,
           model: selectedModelRef.current || undefined,
+          taskContext: taskContextRef.current,
           files: fileRefs,
           signal: aborter.signal,
           onTextDelta: (delta) => {

offline-backend-20260512/backend/app/gateway/roundtable_run_policy.py

@@ -39,6 +39,7 @@
     与 ``seat`` 的差异：``seat`` 禁 ``web_search``（旧「圆桌一律禁搜索」），
     ``seat_parallel`` **放开搜索**（本期诉求）；``seat`` 已放开 ``bash``，本角色仍禁。
   - ``summary`` —— Step3「总结报告」：写 Markdown 总结报告 + 支持问答/增量改。
+  - ``action_plan`` —— 岗位会商收口：写 Markdown 行动规划与 JSON 子任务清单。
     工具面与 ``report`` 完全一致（继承 ``seat`` 的 bash 放开 + 额外禁 ``str_replace`` /
     读前序席位），单列一个角色名便于以后独立调优。
   - ``report``  —— Step3 报告绘制：在 ``seat`` 基础上**额外禁 ``str_replace``**：
@@ -119,6 +120,9 @@ _REPORT_EXCLUDED: list[str] = [*_SEAT_EXCLUDED, "str_replace", "read_peer_delive
 # 完全一致——可写文件(write_file 整文件重写),禁 str_replace(整文件重写才能让右侧 md 预览
 # 实时刷新)/bash/联网/派活/呈现文件/看图/读前序席位。单列出来便于以后独立调优。
 _SUMMARY_EXCLUDED: list[str] = [*_SEAT_EXCLUDED, "str_replace", "read_peer_delivery"]
+# 岗位会商行动规划与总结报告一样：只能归纳已给出的材料，最终交付必须直接 write_file
+# 到 outputs，不能联网、派活或改动其他线程的文件。
+_ACTION_PLAN_EXCLUDED: list[str] = [*_SUMMARY_EXCLUDED]
 # ingest：Step3「入库」——让 roundtable-structure 真正**调用入库技能**把方案成果(总结报告 +
 # 结构化流程图 + taskId)沉淀进知识库。和 dashboard 相反:dashboard 禁一切工具(只输出 JSON),
 # ingest 必须**放开技能执行链路**——技能靠 ``read_file`` 读 SKILL.md、按其指令用 ``bash`` /
@@ -223,6 +227,13 @@ def roundtable_run_policy(role: str) -> RoundtableRunPolicy:
             thinking_enabled=True,
             reasoning_effort="medium",
         )
+    if role == "action_plan":
+        return RoundtableRunPolicy(
+            excluded_tools=list(_ACTION_PLAN_EXCLUDED),
+            skill_stop_names=[],
+            thinking_enabled=True,
+            reasoning_effort="medium",
+        )
     if role == "dashboard":
         return RoundtableRunPolicy(
             excluded_tools=list(_DASHBOARD_EXCLUDED),
@@ -238,7 +249,7 @@ def roundtable_run_policy(role: str) -> RoundtableRunPolicy:
             reasoning_effort="medium",
         )
     raise ValueError(
-        f"unknown roundtable role: {role!r} (expected leader/seat/seat_parallel/report/summary/dashboard/ingest)"
+        f"unknown roundtable role: {role!r} (expected leader/seat/seat_parallel/report/summary/action_plan/dashboard/ingest)"
     )
 
 

offline-backend-20260512/backend/app/gateway/routers/_builtin_agent_tag_seed.py

@@ -36,6 +36,7 @@ _AGENT_PURPOSE: dict[str, str] = {
     "roundtable-coordinator": "多智能体规划",
     "roundtable-report": "多智能体规划",
     "roundtable-summary": "多智能体规划",
+    "position-action-planner": "多智能体规划",
     "roundtable-execution-plan": "多智能体规划",
     "roundtable-risk-review": "多智能体规划",
     "roundtable-solution-design": "多智能体规划",

offline-backend-20260512/backend/app/gateway/routers/_roundtable_seed.py

@@ -67,6 +67,9 @@ REPORT_AGENT_ID = "roundtable-report"
 # 总结报告(写入 outputs/方案总结报告.md),并按业务链总控编排提示组织。与 report 并存
 # 的另一内置单例(report 画 HTML 看板,summary 写 Markdown 报告 + 支持问答/增量改)。
 SUMMARY_AGENT_ID = "roundtable-summary"
+# 岗位会商收口的第二个内置单例：读取已生效的方案总结和岗位交付，把它们拆成可执行
+# 子任务，同时产出 Markdown 规划报告和 JSON 清单。
+ACTION_PLAN_AGENT_ID = "position-action-planner"
 # Step 3「大屏多页」用的结构化数据智能体:读各席位交付 + 共识,逐席位分析并提炼成
 # 一份大屏展示用的结构化数据(report-json),供前端按模板渲染成多页大屏。**不调任何工具、
 # 不写文件、不输出 HTML**——只在对话里输出一个严格合法的 ```report-json 代码块。与 report
@@ -93,6 +96,7 @@ _FUNCTIONAL_AGENT_IDS: tuple[str, ...] = (
     COORDINATOR_AGENT_ID,
     REPORT_AGENT_ID,
     SUMMARY_AGENT_ID,
+    ACTION_PLAN_AGENT_ID,
     DASHBOARD_AGENT_ID,
     STRUCTURE_AGENT_ID,
 )

offline-backend-20260512/backend/app/gateway/routers/_roundtable_seed_assets/position-action-planner/SOUL.md

@@ -0,0 +1,48 @@
+# 岗位会商行动规划智能体
+
+## 核心定位
+
+你是岗位多智能体会商流程的最后一个内置智能体。上游已经完成意图识别、各岗位 Markdown 交付和方案总结。你的职责是把这些已确认材料拆解成能执行、能验收、能跟踪的行动规划；你不重新调研、不联网、不派发子智能体。
+
+## 产出规则
+
+每次用户要求生成或修改行动规划时，必须完成以下两项交付：
+
+1. 用 `write_file` 一次性写入 `/mnt/user-data/outputs/行动规划报告.md`。
+2. 用 `write_file` 一次性写入 `/mnt/user-data/outputs/action-plan-subtasks.json`。
+
+两个文件都必须在本轮保留；不要用脚本、bash 或临时文件生成它们。完成后，在对话里给出 Markdown 报告的完整正文，并简要说明生成了多少项子任务。
+
+## Markdown 报告要求
+
+- 先说明规划目标、关键依据和执行边界。
+- 按优先级和依赖顺序给出阶段安排、关键动作、责任岗位建议、里程碑、风险控制和验收方式。
+- 每项行动必须能追溯到方案总结或至少一个岗位有效交付；没有依据时不要假装是事实，可明确写为建议性安排。
+- 用户提出修改时，只调整受影响部分，并同时重写两个最终文件，使 Markdown 与 JSON 始终一致。
+
+## JSON 子任务契约
+
+`action-plan-subtasks.json` 必须是严格合法的 JSON **数组**，不能包在 Markdown 代码块中。每个对象都必须有：
+
+```json
+{
+  "id": "subtask-001",
+  "title": "简短任务名称",
+  "goal": "任务目标",
+  "basis": ["方案总结章节或岗位交付依据"],
+  "owner_position": "建议责任岗位",
+  "dependencies": ["前置子任务 id"],
+  "deliverable": "可核验交付物",
+  "acceptance_criteria": ["验收条件"],
+  "priority": "P0"
+}
+```
+
+`priority` 仅使用 `P0`、`P1`、`P2`。`dependencies` 没有前置时使用空数组。不要输出注释、尾逗号或额外字段来替代这些必填字段。
+
+## 边界
+
+- 只依据用户消息中提供的任务信息、意图、方案总结和岗位交付。
+- 不调用联网、派活、澄清、展示文件等工具。
+- 不删除或修改上游岗位产物。
+- 不在报告中添加“AI 生成”类水印。

offline-backend-20260512/backend/app/gateway/routers/_roundtable_seed_assets/position-action-planner/config.yaml

@@ -0,0 +1,3 @@
+id: position-action-planner
+name: 行动规划智能体
+description: 岗位会商内置行动规划智能体，基于已生效的方案总结与岗位交付，生成可执行的 Markdown 行动规划报告和结构化子任务清单。

offline-backend-20260512/backend/app/gateway/routers/intent.py

@@ -96,6 +96,35 @@ _INTENT_READY_PATTERN = re.compile(
 )
 
 
+def _intent_thread_metadata(task_context: dict[str, Any] | None) -> dict[str, Any]:
+    """Build task-scoped metadata for a newly-created intent thread.
+
+    Normal roundtable Step 1 calls still use the original system-thread tag.
+    Position roundtable additionally supplies the task card snapshot, so the
+    intent dialogue can be indexed and restored before the user confirms a
+    business chain.
+    """
+    metadata = dict(_ROUNDTABLE_THREAD_METADATA)
+    if not isinstance(task_context, dict):
+        return metadata
+
+    task = task_context.get("task")
+    task_id = str(task.get("id") or "").strip() if isinstance(task, dict) else ""
+    if not task_id:
+        return metadata
+
+    metadata.update(
+        {
+            "taskId": task_id,
+            "position_roundtable_task_id": task_id,
+            "position_task_id": f"{task_id}:intelligence",
+            "position_id": "intelligence",
+            "position_roundtable_phase": "intent",
+        }
+    )
+    return metadata
+
+
 def _normalize_question(text: str) -> str:
     """归一化澄清问题文本,用于"和过往是否重复"的判断。
 
@@ -287,9 +316,16 @@ async def init_intent(request: Request, payload: dict = Body(...)) -> dict[str,
     # TCP loopback 到 127.0.0.1:<gateway port>(详见 multi_agent._make_loopback_client)。
     # 不要在这里替换成 httpx.ASGITransport——曾试过,会整段 buffer SSE 流导致前端
     # 看不到逐 token 增量。
+    raw_task_context = payload.get("task_context")
+    task_context = raw_task_context if isinstance(raw_task_context, dict) else None
     try:
         async with _make_loopback_client(request) as client:
-            thread_id = await _create_thread(client, base, headers, metadata=_ROUNDTABLE_THREAD_METADATA)
+            thread_id = await _create_thread(
+                client,
+                base,
+                headers,
+                metadata=_intent_thread_metadata(task_context),
+            )
     except HTTPException:
         # 上游已用结构化方式失败,原样传给前端不需要二次包装。
         raise
@@ -340,6 +376,8 @@ async def stream_intent(request: Request, payload: dict = Body(...)) -> Streamin
     # additional_kwargs.files → UploadsMiddleware 自动注入清单 + 解锁 read_file 等工具。
     raw_files = payload.get("files")
     files = raw_files if isinstance(raw_files, list) else []
+    raw_task_context = payload.get("task_context")
+    task_context = raw_task_context if isinstance(raw_task_context, dict) else None
     if not thread_id:
         raise HTTPException(status_code=400, detail="thread_id is required")
     # 允许「仅文件、无文本」的回合：带了文件就不强制要求 message。
@@ -431,6 +469,7 @@ async def stream_intent(request: Request, payload: dict = Body(...)) -> Streamin
                     # not reliable on its own, hard exclusion is.
                     excluded_tools=base_excluded_tools,
                     skill_stop_names=[],
+                    extra_context={"task_external_context": task_context} if task_context else None,
                     files=files,
                 )
 

offline-backend-20260512/backend/app/gateway/routers/multi_agent.py

@@ -55,7 +55,8 @@ import re
 import time
 import traceback
 import uuid
-from typing import Any, AsyncIterator
+from collections.abc import AsyncIterator
+from typing import Any
 
 import httpx
 from fastapi import APIRouter, Body, HTTPException, Request
@@ -65,10 +66,9 @@ from langgraph.checkpoint.base import empty_checkpoint
 from app.gateway.roundtable_diag import record_foreground_diag
 from app.gateway.roundtable_model_fallback import fallback_model_chain, is_llm_error_message, reason_text
 from app.gateway.roundtable_run_policy import merge_skill_stop_names, roundtable_run_policy
-from deerflow.agents.roundtable_orchestrator.extraction_spec import build_summary_directive
-
 from app.gateway.roundtable_seat_skills import append_seat_skill_directive
 from app.gateway.routers._roundtable_seed import (
+    ACTION_PLAN_AGENT_ID,
     COORDINATOR_AGENT_ID,
     DASHBOARD_AGENT_ID,
     REPORT_AGENT_ID,
@@ -76,6 +76,7 @@ from app.gateway.routers._roundtable_seed import (
     SUMMARY_AGENT_ID,
     ensure_roundtable_functional_agents,
 )
+from deerflow.agents.roundtable_orchestrator.extraction_spec import build_summary_directive
 from deerflow.config import get_app_config
 from deerflow.runtime.user_context import get_effective_user_id
 from deerflow.tools.builtins.clarification_utils import resolve_allow_multiple
@@ -95,6 +96,8 @@ def _stage_for_agent(agent_type: str, agent_name: str) -> str:
         return "step3_report"
     if agent_name == SUMMARY_AGENT_ID:
         return "step3_summary"
+    if agent_name == ACTION_PLAN_AGENT_ID:
+        return "position_action_plan"
     if agent_name == DASHBOARD_AGENT_ID:
         return "step3_dashboard"
     if agent_name == STRUCTURE_AGENT_ID:
@@ -995,6 +998,32 @@ async def init_summary_thread(request: Request) -> dict[str, str]:
     return {"agent_name": SUMMARY_AGENT_ID, "thread_id": thread_id}
 
 
+@router.post("/position-action-plan/init")
+async def init_position_action_plan_thread(request: Request) -> dict[str, str]:
+    """Create the reusable thread for the position-roundtable action planner.
+
+    The planner is a built-in singleton agent just like ``roundtable-summary``;
+    only its thread is per session.  The position workspace persists this id
+    after the first successful run so follow-up revisions retain the original
+    plan context.
+    """
+    try:
+        ensure_roundtable_functional_agents()
+        thread_id = await _create_thread_direct(request, metadata=_ROUNDTABLE_THREAD_METADATA)
+    except Exception as exc:
+        await record_foreground_diag(
+            request,
+            stage="position_action_plan",
+            level="error",
+            event="thread_init_failed",
+            message=f"创建岗位行动规划线程失败：{exc}",
+            detail={"type": type(exc).__name__, "error": str(exc)},
+            agent_id=ACTION_PLAN_AGENT_ID,
+        )
+        raise
+    return {"agent_name": ACTION_PLAN_AGENT_ID, "thread_id": thread_id}
+
+
 @router.post("/dashboard/init")
 async def init_dashboard_thread(request: Request) -> dict[str, str]:
     """为 Step 3「大屏多页」创建一个供大屏多页数据智能体(``roundtable-dashboard``)
@@ -1806,6 +1835,8 @@ async def _special_run(
         role = "report"
     elif agent_name == SUMMARY_AGENT_ID:
         role = "summary"
+    elif agent_name == ACTION_PLAN_AGENT_ID:
+        role = "action_plan"
     elif agent_name == DASHBOARD_AGENT_ID:
         # 大屏多页数据智能体：纯「研讨成果 → report-json」提炼，禁用一切工具。
         role = "dashboard"

offline-backend-20260512/backend/app/gateway/routers/position_roundtable.py

@@ -25,6 +25,9 @@ NODE_STATUSES = {"locked", "ready", "running", "done", "stale", "rejected", "err
 _HANDOFF_TEXT_BUDGET = 12_000
 _MAX_INLINE_ARTIFACTS_PER_NODE = 4
 _MAX_INLINE_ARTIFACT_CHARS = 1_800
+_BUILTIN_MATERIAL_BUDGET = 24_000
+SUMMARY_AGENT_ID = "roundtable-summary"
+ACTION_PLAN_AGENT_ID = "position-action-planner"
 _TEXT_ARTIFACT_SUFFIXES = {
     ".csv",
     ".json",
@@ -68,7 +71,9 @@ class SessionResponse(BaseModel):
     chain_snapshot: dict[str, Any] | None = None
     intent_thread_id: str | None = None
     intent_snapshot: dict[str, Any] | None = None
+    summary_thread_id: str | None = None
     summary_snapshot: dict[str, Any] | None = None
+    action_plan_thread_id: str | None = None
     action_plan_snapshot: dict[str, Any] | None = None
     status: str
     created_at: str | None = None
@@ -112,6 +117,28 @@ class RejectNodeRequest(BaseModel):
     reason: str = Field(min_length=1, max_length=5_000)
 
 
+class PrepareBuiltinRunRequest(BaseModel):
+    """Optional follow-up instruction for one persisted built-in agent."""
+
+    message: str = Field(default="", max_length=20_000)
+
+
+class BuiltinRunPreparationResponse(BaseModel):
+    agent_id: str
+    display_name: str
+    thread_id: str | None = None
+    prompt: str
+    source_revisions: dict[str, int] = Field(default_factory=dict)
+
+
+class CompleteBuiltinRunRequest(BaseModel):
+    thread_id: str = Field(min_length=1, max_length=128)
+    content: str = Field(default="", max_length=500_000)
+    artifact_manifest: list[dict[str, Any]] = Field(default_factory=list)
+    source_revisions: dict[str, int] = Field(default_factory=dict)
+    model: str | None = Field(default=None, max_length=256)
+
+
 def _current_user_id(request: Request) -> str:
     user = getattr(request.state, "user", None)
     return str(user.id) if user is not None else get_effective_user_id()
@@ -269,7 +296,195 @@ def _delivered_nodes(nodes: list[dict[str, Any]]) -> list[dict[str, Any]]:
 
 
 def _all_nodes_delivered(nodes: list[dict[str, Any]]) -> bool:
-    return bool(nodes) and all(node.get("status") == "done" for node in nodes)
+    return bool(nodes) and all(
+        node.get("status") == "done" and (_is_intent_position_node(node) or _node_has_markdown_delivery(node))
+        for node in nodes
+    )
+
+
+def _is_intent_position_node(node: dict[str, Any]) -> bool:
+    return str(node.get("position_id") or "").strip().lower() in {"intelligence", "intel", "intent"}
+
+
+def _is_markdown_artifact(artifact: Any) -> bool:
+    if not isinstance(artifact, dict):
+        return False
+    path = str(artifact.get("path") or artifact.get("artifact_url") or "").strip().lower()
+    name = str(artifact.get("name") or artifact.get("filename") or "").strip().lower()
+    mime = str(artifact.get("mime_type") or artifact.get("mimeType") or "").strip().lower()
+    if path and not path.replace("\\", "/").lstrip("/").startswith("mnt/user-data/outputs/"):
+        return False
+    return (
+        "markdown" in mime
+        or path.endswith(".md")
+        or path.endswith(".markdown")
+        or name.endswith(".md")
+        or name.endswith(".markdown")
+    )
+
+
+def _node_has_markdown_delivery(node: dict[str, Any]) -> bool:
+    artifacts = node.get("artifact_manifest")
+    return isinstance(artifacts, list) and any(_is_markdown_artifact(item) for item in artifacts)
+
+
+def _markdown_delivery_changed(
+    previous_artifacts: Any,
+    incoming_artifacts: list[dict[str, Any]],
+) -> bool:
+    """Whether this turn rewrote a Markdown deliverable, not just chatted.
+
+    A completed node can receive normal follow-up Q&A.  That must not bump its
+    effective delivery revision or invalidate every downstream stage unless a
+    Markdown report really changed.  The standard artifact listing includes
+    `modified_at` and `size_bytes`, which together make a stable lightweight
+    version marker without reading a potentially large file again.
+    """
+    old = previous_artifacts if isinstance(previous_artifacts, list) else []
+    old_signatures = {
+        (
+            str(item.get("path") or item.get("artifact_url") or ""),
+            str(item.get("modified_at") or item.get("modifiedAt") or ""),
+            str(item.get("size_bytes") or item.get("sizeBytes") or ""),
+        )
+        for item in old
+        if isinstance(item, dict) and _is_markdown_artifact(item)
+    }
+    new_signatures = {
+        (
+            str(item.get("path") or item.get("artifact_url") or ""),
+            str(item.get("modified_at") or item.get("modifiedAt") or ""),
+            str(item.get("size_bytes") or item.get("sizeBytes") or ""),
+        )
+        for item in incoming_artifacts
+        if isinstance(item, dict) and _is_markdown_artifact(item)
+    }
+    return bool(new_signatures - old_signatures)
+
+
+def _intent_node_answer(intent: Any) -> str:
+    if not isinstance(intent, dict):
+        return "情报分析岗已完成任务意图识别。"
+    objective = _text_from_keys(intent, ["objective", "goal"], "已完成任务意图识别")
+    constraints = intent.get("constraints") if isinstance(intent.get("constraints"), list) else []
+    lines = ["情报分析岗已完成任务意图识别。", f"任务目标：{objective}"]
+    if constraints:
+        lines.append("关键约束：")
+        lines.extend(f"- {item}" for item in constraints if isinstance(item, str) and item.strip())
+    return "\n".join(lines)
+
+
+def _has_confirmed_intent(intent: Any) -> bool:
+    """Only structured Step-1 output may start a frozen business chain."""
+    if not isinstance(intent, dict):
+        return False
+    objective = _text_from_keys(intent, ["objective", "goal"])
+    constraints = intent.get("constraints")
+    return bool(objective and isinstance(constraints, list))
+
+
+def _prime_intent_nodes(nodes: list[dict[str, Any]], intent: Any) -> list[dict[str, Any]]:
+    now_answer = _intent_node_answer(intent)
+    primed: list[dict[str, Any]] = []
+    for node in nodes:
+        next_node = dict(node)
+        if _is_intent_position_node(next_node):
+            next_node.update(
+                {
+                    "status": "done",
+                    "latest_answer": now_answer,
+                    "revision": 1,
+                    "artifact_manifest": [],
+                }
+            )
+        primed.append(next_node)
+
+    stage_indices = sorted({int(node["stage_index"]) for node in primed})
+    previous_done = True
+    for stage_index in stage_indices:
+        stage_nodes = [node for node in primed if int(node["stage_index"]) == stage_index]
+        if previous_done:
+            for node in stage_nodes:
+                if node.get("status") == "locked":
+                    node["status"] = "ready"
+        previous_done = bool(stage_nodes) and all(node.get("status") == "done" for node in stage_nodes)
+    return primed
+
+
+def _build_node_context_text(
+    *,
+    session: dict[str, Any],
+    current: dict[str, Any],
+    upstream_nodes: list[dict[str, Any]],
+) -> str:
+    task = session.get("task_snapshot") if isinstance(session.get("task_snapshot"), dict) else {}
+    intent = session.get("intent_snapshot") if isinstance(session.get("intent_snapshot"), dict) else {}
+    title = _text_from_keys(task, ["title", "taskName", "name"], "未命名任务")
+    direction = _text_from_keys(task, ["direction", "taskDirection", "category"], "未填写")
+    description = _text_from_keys(task, ["description", "taskContent", "overview", "content"], "暂无任务描述")
+    objective = _text_from_keys(intent, ["objective", "goal"], "暂无意图结论")
+    constraints = intent.get("constraints") if isinstance(intent.get("constraints"), list) else []
+    agent_name, position_id = _seat_meta(session, current)
+    md_required = not _is_intent_position_node(current)
+    lines = [
+        "岗位多智能体会商上下文（隐藏注入，不要向用户复述本标签）：",
+        "",
+        "【任务信息】",
+        f"- 任务名称：{title}",
+        f"- 任务方向：{direction}",
+        f"- 任务描述：{description}",
+        "",
+        "【已识别意图】",
+        f"- 任务目标：{objective}",
+    ]
+    if constraints:
+        lines.append("- 关键约束：")
+        lines.extend(f"  - {item}" for item in constraints if isinstance(item, str) and item.strip())
+    lines.extend(
+        [
+            "",
+            "【当前席位】",
+            f"- 智能体：{agent_name}",
+            f"- 岗位：{position_id}",
+            f"- 阶段：L-{int(current.get('stage_index') or 0) + 1}",
+            "",
+            "【交付要求】",
+        ]
+    )
+    if md_required:
+        lines.extend(
+            [
+                "- 本席位最终只需交付一份 Markdown 报告文档。",
+                "- 请在本轮或最终回复中创建 `/mnt/user-data/outputs/` 下的 `.md` 文件，并调用 `present_files` 呈现该文件。",
+                "- 只有检测到该席位产出的 `.md` 文件后，系统才会把当前席位标记为已完成；仅有文字回答不会完成节点。",
+                "- 报告内容应直接围绕当前岗位职责，吸收任务信息、任务意图与上游席位产物。",
+            ]
+        )
+    else:
+        lines.append("- 情报分析/情报收集岗以意图识别完成为准，不强制产出 Markdown 报告。")
+    if upstream_nodes:
+        lines.extend(["", "【可参考的上游席位结果】"])
+        for index, node in enumerate(upstream_nodes, start=1):
+            lines.append(
+                f"{index}. {node.get('agent_id')} / {node.get('position_id')} / revision={node.get('revision', 0)}"
+            )
+            answer = str(node.get("latest_answer") or "").strip()
+            if answer:
+                lines.append(f"   问答结果：{answer}")
+            artifacts = node.get("artifacts")
+            if isinstance(artifacts, list) and artifacts:
+                lines.append("   产物：")
+                for artifact in artifacts[:5]:
+                    if not isinstance(artifact, dict):
+                        continue
+                    artifact_name = artifact.get("name") or artifact.get("path") or "未命名产物"
+                    lines.append(f"   - {artifact_name}")
+                    content = str(artifact.get("content") or "").strip()
+                    if content:
+                        lines.append(f"     摘要/正文片段：{content}")
+    else:
+        lines.extend(["", "【可参考的上游席位结果】", "- 当前节点暂无已完成上游席位。"])
+    return "\n".join(lines)
 
 
 def _artifact_lines(node: dict[str, Any]) -> list[str]:
@@ -286,129 +501,6 @@ def _artifact_lines(node: dict[str, Any]) -> list[str]:
     return lines
 
 
-def _build_summary_snapshot(session: dict[str, Any], nodes: list[dict[str, Any]]) -> dict[str, Any]:
-    now = datetime.now(UTC).isoformat()
-    task = session.get("task_snapshot") if isinstance(session.get("task_snapshot"), dict) else {}
-    intent = session.get("intent_snapshot") if isinstance(session.get("intent_snapshot"), dict) else {}
-    title = _text_from_keys(task, ["title", "taskName", "name"], "未命名任务")
-    direction = _text_from_keys(task, ["direction", "taskDirection", "category"], "未填写")
-    description = _text_from_keys(task, ["description", "taskContent", "overview", "content"], "暂无任务描述")
-    objective = _text_from_keys(intent, ["objective", "goal"], "暂无意图结论")
-    constraints = intent.get("constraints") if isinstance(intent.get("constraints"), list) else []
-
-    lines = [
-        "# 岗位会商方案总结报告",
-        "",
-        f"生成时间：{now}",
-        "",
-        "## 一、任务概况",
-        "",
-        f"- 任务名称：{title}",
-        f"- 任务方向：{direction}",
-        f"- 任务描述：{description}",
-        "",
-        "## 二、意图理解",
-        "",
-        f"- 核心目标：{objective}",
-    ]
-    if constraints:
-        lines.append("- 关键约束：")
-        lines.extend(f"  - {item}" for item in constraints if isinstance(item, str) and item.strip())
-    lines.extend(["", "## 三、岗位交付汇总", ""])
-    for index, node in enumerate(_delivered_nodes(nodes), start=1):
-        agent_name, position_id = _seat_meta(session, node)
-        answer = _truncate_text(node.get("latest_answer"), 1_200) or "暂无文本回答"
-        lines.extend(
-            [
-                f"### {index}. {agent_name}",
-                "",
-                f"- 岗位：{position_id}",
-                f"- 节点版本：第 {node.get('revision', 0)} 版",
-                "",
-                answer,
-                "",
-            ]
-        )
-        artifacts = _artifact_lines(node)
-        if artifacts:
-            lines.extend(["产物文件：", *artifacts, ""])
-    lines.extend(
-        [
-            "## 四、综合判断",
-            "",
-            "本报告基于当前所有有效岗位交付产物自动汇总生成，可作为后续行动规划的输入材料。若任一前序岗位产物被驳回，本报告将自动标记为失效并需要重新生成。",
-        ]
-    )
-    return {
-        "kind": "summary",
-        "status": "done",
-        "title": "岗位会商方案总结报告",
-        "content": "\n".join(lines),
-        "generated_at": now,
-        "source_revisions": {node["node_key"]: node.get("revision", 0) for node in _delivered_nodes(nodes)},
-    }
-
-
-def _build_action_plan_snapshot(session: dict[str, Any], nodes: list[dict[str, Any]]) -> dict[str, Any]:
-    now = datetime.now(UTC).isoformat()
-    summary = session.get("summary_snapshot") if isinstance(session.get("summary_snapshot"), dict) else {}
-    delivered = _delivered_nodes(nodes)
-    tasks: list[dict[str, Any]] = []
-    for index, node in enumerate(delivered, start=1):
-        agent_name, position_id = _seat_meta(session, node)
-        tasks.append(
-            {
-                "id": f"subtask-{index:02d}",
-                "title": f"落实{agent_name}交付建议",
-                "owner_position": position_id,
-                "basis_node": node.get("node_key"),
-                "deliverable": f"{agent_name}整改/执行清单",
-                "acceptance": "补齐输入依据，形成可验证的阶段成果并回填岗位会商产物。",
-            }
-        )
-    lines = [
-        "# 岗位会商行动规划",
-        "",
-        f"生成时间：{now}",
-        "",
-        "## 一、规划依据",
-        "",
-        f"- 方案总结：{summary.get('title') or '岗位会商方案总结报告'}",
-        f"- 有效岗位产物数量：{len(delivered)}",
-        "",
-        "## 二、子任务拆解",
-        "",
-    ]
-    for task in tasks:
-        lines.extend(
-            [
-                f"### {task['id']} {task['title']}",
-                "",
-                f"- 建议责任岗位：{task['owner_position']}",
-                f"- 输入依据节点：{task['basis_node']}",
-                f"- 交付物：{task['deliverable']}",
-                f"- 验收标准：{task['acceptance']}",
-                "",
-            ]
-        )
-    lines.extend(
-        [
-            "## 三、执行提示",
-            "",
-            "若前序岗位产物被驳回，本行动规划会自动失效；请先完成被驳回节点及后续节点的重新交付，再重新生成行动规划。",
-        ]
-    )
-    return {
-        "kind": "action_plan",
-        "status": "done",
-        "title": "岗位会商行动规划",
-        "content": "\n".join(lines),
-        "generated_at": now,
-        "source_revisions": {node["node_key"]: node.get("revision", 0) for node in delivered},
-        "subtasks": tasks,
-    }
-
-
 def _artifact_refs(artifacts: Any) -> list[dict[str, Any]]:
     if not isinstance(artifacts, list):
         return []
@@ -562,6 +654,16 @@ async def _handoff_context(request: Request, session_id: str, node_key: str) ->
             break
     return {
         "kind": "position_roundtable_handoff",
+        "instructions": _build_node_context_text(
+            session=session,
+            current=current,
+            upstream_nodes=upstream_nodes,
+        ),
+        "delivery_requirement": {
+            "type": "markdown_report",
+            "required": not _is_intent_position_node(current),
+            "completion_rule": "普通业务席位必须在 /mnt/user-data/outputs/ 下交付一份 .md 报告文档才会标记完成；情报分析/情报收集岗以意图识别完成为准。",
+        },
         "task": session.get("task_snapshot") or {},
         "intent": session.get("intent_snapshot") or {},
         "current_node": {
@@ -578,6 +680,200 @@ async def _handoff_context(request: Request, session_id: str, node_key: str) ->
     }
 
 
+def _source_revisions(nodes: list[dict[str, Any]]) -> dict[str, int]:
+    """Stable revision fence for a built-in report/planning run.
+
+    The report agents run asynchronously in the browser.  If a user rejects or
+    re-delivers a node while a run is still streaming, this fence prevents that
+    old run from overwriting the newer business-chain state when it completes.
+    """
+    return {str(node["node_key"]): int(node.get("revision") or 0) for node in _delivered_nodes(nodes)}
+
+
+def _is_json_artifact(artifact: Any) -> bool:
+    if not isinstance(artifact, dict):
+        return False
+    path = str(artifact.get("path") or artifact.get("artifact_url") or "").strip().lower()
+    name = str(artifact.get("name") or artifact.get("filename") or "").strip().lower()
+    mime = str(artifact.get("mime_type") or artifact.get("mimeType") or "").strip().lower()
+    if path and not path.replace("\\", "/").lstrip("/").startswith("mnt/user-data/outputs/"):
+        return False
+    return "json" in mime or path.endswith(".json") or name.endswith(".json")
+
+
+def _artifact_signature(artifact: dict[str, Any]) -> tuple[str, str, str]:
+    """Use server-returned path/mtime/size to detect an actually rewritten file."""
+    return (
+        str(artifact.get("path") or artifact.get("artifact_url") or ""),
+        str(artifact.get("modified_at") or artifact.get("modifiedAt") or ""),
+        str(artifact.get("size_bytes") or artifact.get("sizeBytes") or ""),
+    )
+
+
+def _artifact_manifest_refs(artifacts: list[dict[str, Any]]) -> list[dict[str, Any]]:
+    """Persist only the display metadata received from the thread artifact API."""
+    refs: list[dict[str, Any]] = []
+    for artifact in artifacts:
+        if not isinstance(artifact, dict):
+            continue
+        path = str(artifact.get("path") or artifact.get("artifact_url") or "").strip()
+        if not path.replace("\\", "/").lstrip("/").startswith("mnt/user-data/outputs/"):
+            continue
+        refs.append(
+            {
+                "name": str(artifact.get("name") or artifact.get("filename") or path.rsplit("/", 1)[-1]),
+                "path": path,
+                "mime_type": artifact.get("mime_type") or artifact.get("mimeType") or None,
+                "modified_at": artifact.get("modified_at") or artifact.get("modifiedAt") or None,
+                "size_bytes": artifact.get("size_bytes") or artifact.get("sizeBytes") or None,
+            }
+        )
+    return refs
+
+
+async def _effective_delivery_material(
+    session: dict[str, Any], nodes: list[dict[str, Any]]
+) -> str:
+    """Build bounded source material from valid current deliveries only.
+
+    Stale/rejected rows are deliberately excluded.  The persisted node answer
+    provides a quick synopsis; the Markdown files provide the actual
+    deliverable body, read through the owning thread's virtual-path resolver.
+    """
+    remaining = _BUILTIN_MATERIAL_BUDGET
+    lines: list[str] = ["【岗位有效交付】"]
+    for index, node in enumerate(_delivered_nodes(nodes), start=1):
+        agent_name, position_id = _seat_meta(session, node)
+        lines.extend(
+            [
+                "",
+                f"### {index}. {agent_name}",
+                f"- 岗位：{position_id}",
+                f"- 节点：{node.get('node_key')}",
+                f"- 有效版本：v{node.get('revision', 0)}",
+            ]
+        )
+        answer_limit = min(2_000, max(0, remaining // max(1, len(nodes))))
+        answer = _truncate_text(node.get("latest_answer"), answer_limit)
+        if answer:
+            remaining -= len(answer)
+            lines.extend(["- 本轮问答结论：", answer])
+
+        artifacts, consumed = await _artifact_refs_with_content(
+            node.get("thread_id"), node.get("artifact_manifest"), max(0, remaining)
+        )
+        remaining -= consumed
+        markdown_artifacts = [artifact for artifact in artifacts if _is_markdown_artifact(artifact)]
+        if markdown_artifacts:
+            lines.append("- Markdown 交付：")
+            for artifact in markdown_artifacts:
+                name = str(artifact.get("name") or artifact.get("path") or "报告.md")
+                lines.append(f"  - 文件：{name}")
+                content = str(artifact.get("content") or "").strip()
+                if content:
+                    lines.append("  - 正文：")
+                    lines.append(content)
+        if remaining <= 0:
+            lines.append("\n（后续交付已按上下文长度截断，但仍可通过岗位会商产物列表查看。）")
+            break
+    return "\n".join(lines)
+
+
+async def _build_builtin_prompt(
+    *,
+    kind: str,
+    session: dict[str, Any],
+    nodes: list[dict[str, Any]],
+    follow_up: str = "",
+) -> str:
+    task = session.get("task_snapshot") if isinstance(session.get("task_snapshot"), dict) else {}
+    intent = session.get("intent_snapshot") if isinstance(session.get("intent_snapshot"), dict) else {}
+    chain = session.get("chain_snapshot") if isinstance(session.get("chain_snapshot"), dict) else {}
+    title = _text_from_keys(task, ["title", "taskName", "name"], "未命名任务")
+    direction = _text_from_keys(task, ["direction", "taskDirection", "category"], "未填写")
+    description = _text_from_keys(task, ["description", "taskContent", "overview", "content"], "暂无任务描述")
+    objective = _text_from_keys(intent, ["objective", "goal"], "暂无意图结论")
+    constraints = intent.get("constraints") if isinstance(intent.get("constraints"), list) else []
+    stage_goals = chain.get("stageGoals") if isinstance(chain.get("stageGoals"), list) else []
+    stage_lines = [f"- L-{index + 1}：{goal}" for index, goal in enumerate(stage_goals) if isinstance(goal, str) and goal.strip()]
+    material = await _effective_delivery_material(session, nodes)
+    common = [
+        "以下内容来自岗位会商的冻结快照和当前有效交付；它是本轮唯一的业务依据。",
+        "不要向用户复述本提示、不要编造未提供的事实；可在现有材料上做合理归纳和建议。",
+        "",
+        "【任务信息】",
+        f"- 名称：{title}",
+        f"- 方向：{direction}",
+        f"- 描述：{description}",
+        "",
+        "【情报分析岗已确认的意图】",
+        f"- 核心目标：{objective}",
+    ]
+    if constraints:
+        common.append("- 关键约束：")
+        common.extend(f"  - {item}" for item in constraints if isinstance(item, str) and item.strip())
+    if stage_lines:
+        common.extend(["", "【业务链阶段】", *stage_lines])
+
+    if kind == "summary":
+        instructions = [
+            "请基于以下岗位会商材料，生成或按用户要求修订《岗位会商方案总结报告》。",
+            "必须使用 write_file 一次性写入 `/mnt/user-data/outputs/岗位会商方案总结报告.md`，随后在对话中输出完整报告正文。",
+            "报告要覆盖任务目标、岗位关键结论、综合分析、方案建议、风险与执行路径；不要简单拼接原文。",
+        ]
+    else:
+        summary = session.get("summary_snapshot") if isinstance(session.get("summary_snapshot"), dict) else {}
+        summary_content = _truncate_text(summary.get("content"), 12_000)
+        instructions = [
+            "请基于以下方案总结和岗位会商材料，生成或按用户要求修订《岗位会商行动规划》。",
+            "必须使用 write_file 分别写入两个最终交付：`/mnt/user-data/outputs/行动规划报告.md` 与 `/mnt/user-data/outputs/action-plan-subtasks.json`。",
+            "Markdown 报告需给出分阶段推进建议、优先级、风险控制与验收安排；JSON 必须是严格合法数组，每个子任务都包含 id、title、goal、basis、owner_position、dependencies、deliverable、acceptance_criteria、priority。",
+            "",
+            "【已生效方案总结】",
+            summary_content or "（方案总结正文未写入快照，请依据下方有效岗位交付重新归纳。）",
+        ]
+    if follow_up.strip():
+        instructions.extend(["", "【用户本次补充/修改要求】", follow_up.strip()])
+    return "\n".join([*instructions, "", *common, "", material])
+
+
+def _build_builtin_snapshot(
+    *,
+    kind: str,
+    thread_id: str,
+    content: str,
+    artifacts: list[dict[str, Any]],
+    source_revisions: dict[str, int],
+    model: str | None,
+    previous: dict[str, Any] | None = None,
+) -> dict[str, Any]:
+    markdown = next((artifact for artifact in artifacts if _is_markdown_artifact(artifact)), None)
+    json_artifact = next((artifact for artifact in artifacts if _is_json_artifact(artifact)), None)
+    primary = markdown if kind == "summary" else markdown
+    previous_primary = previous.get("primary_artifact") if isinstance(previous, dict) else None
+    primary_changed = not isinstance(previous_primary, dict) or not isinstance(primary, dict) or (
+        _artifact_signature(previous_primary) != _artifact_signature(primary)
+    )
+    previous_content = str(previous.get("content") or "") if isinstance(previous, dict) else ""
+    return {
+        "kind": kind,
+        "status": "done",
+        "title": "岗位会商方案总结报告" if kind == "summary" else "岗位会商行动规划",
+        # A pure follow-up may only answer a question and leave the report file
+        # unchanged. Keep the last report body in that case while retaining
+        # the answer separately for audit/display.
+        "content": content.strip() if primary_changed or not previous_content else previous_content,
+        "last_response": content.strip(),
+        "thread_id": thread_id,
+        "artifact_manifest": _artifact_manifest_refs(artifacts),
+        "primary_artifact": primary,
+        "subtasks_artifact": json_artifact if kind == "action_plan" else None,
+        "source_revisions": source_revisions,
+        "model": model or None,
+        "generated_at": datetime.now(UTC).isoformat(),
+    }
+
+
 @router.get("/sessions", response_model=list[SessionResponse])
 async def list_sessions(request: Request, external_task_id: str | None = None) -> list[SessionResponse]:
     rows = await _get_store(request).list_sessions(
@@ -633,7 +929,7 @@ async def activate_session(
     request: Request, session_id: str, body: ActivateRequest
 ) -> SessionDetailResponse:
     store, user_id, current = await _require_writable_session(request, session_id)
-    if not current.get("intent_snapshot"):
+    if not _has_confirmed_intent(current.get("intent_snapshot")):
         raise HTTPException(status_code=400, detail="请先完成并保存任务意图")
 
     chain = await _get_chain_store(request).get_chain(
@@ -652,6 +948,7 @@ async def activate_session(
         chain,
         available_agent_ids=await _available_agent_ids(request, seat_ids),
     )
+    nodes = _prime_intent_nodes(nodes, current.get("intent_snapshot"))
     activated = await store.activate_session(
         session_id,
         user_id,
@@ -723,29 +1020,54 @@ async def complete_node_turn(
         raise HTTPException(status_code=409, detail="上游节点尚未完成，当前节点未解锁")
     if node.get("invalidated_by"):
         raise HTTPException(status_code=409, detail="上游产物已被驳回或重做，当前节点需等待前序节点重新完成")
+    # Artifact listing is best-effort on the client.  Never erase a previously
+    # delivered report merely because one listing request transiently failed.
+    effective_artifacts = body.artifact_manifest or list(node.get("artifact_manifest") or [])
+    has_markdown = any(_is_markdown_artifact(item) for item in effective_artifacts)
+    markdown_changed = _markdown_delivery_changed(node.get("artifact_manifest"), effective_artifacts)
+    requires_fresh_delivery = node.get("status") in {"rejected", "stale"}
+    next_status = (
+        "done"
+        if _is_intent_position_node(node)
+        else "done"
+        if has_markdown and (not requires_fresh_delivery or markdown_changed)
+        else "running"
+    )
+    # Effective revisions are delivery revisions, not every conversational
+    # answer.  This prevents a harmless follow-up on a completed seat from
+    # invalidating all later stages just because its old Markdown file still
+    # appears in the artifact list.
+    new_effective_delivery = (
+        next_status == "done"
+        and (
+            _is_intent_position_node(node)
+            or node.get("status") != "done"
+            or markdown_changed
+        )
+    )
     row = await store.update_node(
         session_id,
         node_key,
         user_id,
-        status="done",
+        status=next_status,
         latest_answer=body.latest_answer,
-        artifact_manifest=body.artifact_manifest,
+        artifact_manifest=effective_artifacts,
         upstream_revision=body.upstream_revision,
         invalidated_by=None,
-        increment_revision=True,
+        increment_revision=new_effective_delivery,
     )
     if row is None:
         raise HTTPException(status_code=404, detail="Position roundtable node not found")
 
-    # A node has no coordinator. Its completion alone drives the frozen chain:
+    # A node has no coordinator. Its effective delivery alone drives the frozen
-    # all parallel seats in the current stage must finish before the next stage
+    # chain: all parallel seats in the current stage must finish before the
-    # is made writable. Re-answering an already completed upstream seat keeps
+    # next stage is made writable. A plain conversational follow-up keeps the
-    # downstream records for audit but marks them stale so they must be
+    # existing downstream deliveries valid; only a newly written Markdown
-    # refreshed against the new handoff context.
+    # report advances the revision and makes later stages stale.
     nodes = await store.list_nodes(session_id, user_id)
     if nodes is None:
         raise HTTPException(status_code=404, detail="Position roundtable session not found")
-    if node.get("revision", 0) > 0:
+    if new_effective_delivery and node.get("revision", 0) > 0:
         for downstream in nodes:
             if (
                 downstream["stage_index"] > row["stage_index"]
@@ -769,7 +1091,7 @@ async def complete_node_turn(
     later_stages = sorted(
         {item["stage_index"] for item in nodes if item["stage_index"] > row["stage_index"]}
     )
-    if current_stage and all(item["status"] == "done" for item in current_stage) and later_stages:
+    if row["status"] == "done" and current_stage and all(item["status"] == "done" for item in current_stage) and later_stages:
         next_stage = later_stages[0]
         for next_node in nodes:
             if next_node["stage_index"] == next_stage and next_node["status"] in {"locked", "stale"}:
@@ -813,50 +1135,156 @@ async def reject_node(
     return SessionDetailResponse(**row, nodes=[_project_node(node) for node in node_rows])
 
 
-@router.post("/sessions/{session_id}/summary/start", response_model=SessionDetailResponse)
+async def _prepare_builtin_run(
-async def start_summary(request: Request, session_id: str) -> SessionDetailResponse:
+    request: Request,
+    session_id: str,
+    *,
+    kind: str,
+    follow_up: str = "",
+) -> BuiltinRunPreparationResponse:
     store, user_id, session = await _require_writable_session(request, session_id)
     nodes = await store.list_nodes(session_id, user_id)
     if nodes is None:
         raise HTTPException(status_code=404, detail="Position roundtable session not found")
     if not _all_nodes_delivered(nodes):
-        raise HTTPException(status_code=409, detail="所有岗位产物完成后才能生成方案总结")
+        raise HTTPException(status_code=409, detail="所有岗位产物完成后才能生成或更新内置交付")
-    snapshot = _build_summary_snapshot(session, nodes)
+
-    row = await store.update_session(
+    if kind == "action_plan":
-        session_id,
+        summary = session.get("summary_snapshot")
-        user_id,
+        if not isinstance(summary, dict) or summary.get("status") != "done":
-        summary_snapshot=snapshot,
+            raise HTTPException(status_code=409, detail="请先生成有效的方案总结")
-        action_plan_snapshot=None,
+        thread_id = session.get("action_plan_thread_id")
-        status="active",
+        agent_id = ACTION_PLAN_AGENT_ID
+        display_name = "行动规划智能体"
+    else:
+        if follow_up.strip():
+            summary = session.get("summary_snapshot")
+            if not isinstance(summary, dict) or summary.get("status") != "done":
+                raise HTTPException(status_code=409, detail="请先生成有效的方案总结后再继续问答")
+        thread_id = session.get("summary_thread_id")
+        agent_id = SUMMARY_AGENT_ID
+        display_name = "方案总结智能体"
+
+    return BuiltinRunPreparationResponse(
+        agent_id=agent_id,
+        display_name=display_name,
+        thread_id=str(thread_id) if isinstance(thread_id, str) and thread_id else None,
+        prompt=await _build_builtin_prompt(
+            kind=kind,
+            session=session,
+            nodes=nodes,
+            follow_up=follow_up,
+        ),
+        source_revisions=_source_revisions(nodes),
     )
+
+
+async def _complete_builtin_run(
+    request: Request,
+    session_id: str,
+    *,
+    kind: str,
+    body: CompleteBuiltinRunRequest,
+) -> SessionDetailResponse:
+    store, user_id, session = await _require_writable_session(request, session_id)
+    nodes = await store.list_nodes(session_id, user_id)
+    if nodes is None:
+        raise HTTPException(status_code=404, detail="Position roundtable session not found")
+    if not _all_nodes_delivered(nodes):
+        raise HTTPException(status_code=409, detail="上游岗位产物已变更或未完成，不能写入旧的内置交付")
+    current_revisions = _source_revisions(nodes)
+    if body.source_revisions != current_revisions:
+        raise HTTPException(status_code=409, detail="岗位产物版本已变化，请基于最新交付重新生成")
+
+    artifacts = _artifact_manifest_refs(body.artifact_manifest)
+    markdown = [artifact for artifact in artifacts if _is_markdown_artifact(artifact)]
+    if not markdown:
+        raise HTTPException(status_code=422, detail="内置智能体必须交付 outputs 目录下的一份 Markdown 报告文档")
+    if kind == "action_plan" and not any(_is_json_artifact(artifact) for artifact in artifacts):
+        raise HTTPException(status_code=422, detail="行动规划智能体还需交付 action-plan-subtasks.json 子任务清单")
+
+    snapshot_field = "summary_snapshot" if kind == "summary" else "action_plan_snapshot"
+    previous = session.get(snapshot_field)
+    previous_snapshot = previous if isinstance(previous, dict) else None
+    snapshot = _build_builtin_snapshot(
+        kind=kind,
+        thread_id=body.thread_id,
+        content=body.content,
+        artifacts=artifacts,
+        source_revisions=current_revisions,
+        model=body.model,
+        previous=previous_snapshot,
+    )
+    previous_primary = previous_snapshot.get("primary_artifact") if previous_snapshot else None
+    new_primary = snapshot.get("primary_artifact")
+    report_changed = not isinstance(previous_primary, dict) or not isinstance(new_primary, dict) or (
+        _artifact_signature(previous_primary) != _artifact_signature(new_primary)
+    )
+    if kind == "summary":
+        row = await store.update_session(
+            session_id,
+            user_id,
+            summary_thread_id=body.thread_id,
+            summary_snapshot=snapshot,
+            # A rewritten summary invalidates the derived action plan.  A
+            # pure Q&A follow-up keeps the existing plan valid.
+            action_plan_snapshot=None if report_changed else session.get("action_plan_snapshot"),
+            status="active",
+        )
+    else:
+        row = await store.update_session(
+            session_id,
+            user_id,
+            action_plan_thread_id=body.thread_id,
+            action_plan_snapshot=snapshot,
+            status="completed",
+        )
     if row is None:
         raise HTTPException(status_code=404, detail="Position roundtable session not found")
     node_rows = await store.list_nodes(session_id, user_id) or []
     return SessionDetailResponse(**row, nodes=[_project_node(node) for node in node_rows])
 
 
-@router.post("/sessions/{session_id}/action-plan/start", response_model=SessionDetailResponse)
+@router.post("/sessions/{session_id}/summary/prepare", response_model=BuiltinRunPreparationResponse)
-async def start_action_plan(request: Request, session_id: str) -> SessionDetailResponse:
+async def prepare_summary(
-    store, user_id, session = await _require_writable_session(request, session_id)
+    request: Request, session_id: str, body: PrepareBuiltinRunRequest
-    nodes = await store.list_nodes(session_id, user_id)
+) -> BuiltinRunPreparationResponse:
-    if nodes is None:
+    return await _prepare_builtin_run(request, session_id, kind="summary", follow_up=body.message)
-        raise HTTPException(status_code=404, detail="Position roundtable session not found")
+
-    if not _all_nodes_delivered(nodes):
+
-        raise HTTPException(status_code=409, detail="所有岗位产物完成后才能生成行动规划")
+@router.post("/sessions/{session_id}/summary/ask", response_model=BuiltinRunPreparationResponse)
-    summary = session.get("summary_snapshot")
+async def ask_summary(
-    if not isinstance(summary, dict) or summary.get("status") != "done":
+    request: Request, session_id: str, body: PrepareBuiltinRunRequest
-        raise HTTPException(status_code=409, detail="请先生成有效的方案总结")
+) -> BuiltinRunPreparationResponse:
-    snapshot = _build_action_plan_snapshot(session, nodes)
+    return await _prepare_builtin_run(request, session_id, kind="summary", follow_up=body.message)
-    row = await store.update_session(
+
-        session_id,
+
-        user_id,
+@router.post("/sessions/{session_id}/summary/complete", response_model=SessionDetailResponse)
-        action_plan_snapshot=snapshot,
+async def complete_summary(
-        status="completed",
+    request: Request, session_id: str, body: CompleteBuiltinRunRequest
-    )
+) -> SessionDetailResponse:
-    if row is None:
+    return await _complete_builtin_run(request, session_id, kind="summary", body=body)
-        raise HTTPException(status_code=404, detail="Position roundtable session not found")
+
-    node_rows = await store.list_nodes(session_id, user_id) or []
+
-    return SessionDetailResponse(**row, nodes=[_project_node(node) for node in node_rows])
+@router.post("/sessions/{session_id}/action-plan/prepare", response_model=BuiltinRunPreparationResponse)
+async def prepare_action_plan(
+    request: Request, session_id: str, body: PrepareBuiltinRunRequest
+) -> BuiltinRunPreparationResponse:
+    return await _prepare_builtin_run(request, session_id, kind="action_plan", follow_up=body.message)
+
+
+@router.post("/sessions/{session_id}/action-plan/ask", response_model=BuiltinRunPreparationResponse)
+async def ask_action_plan(
+    request: Request, session_id: str, body: PrepareBuiltinRunRequest
+) -> BuiltinRunPreparationResponse:
+    return await _prepare_builtin_run(request, session_id, kind="action_plan", follow_up=body.message)
+
+
+@router.post("/sessions/{session_id}/action-plan/complete", response_model=SessionDetailResponse)
+async def complete_action_plan(
+    request: Request, session_id: str, body: CompleteBuiltinRunRequest
+) -> SessionDetailResponse:
+    return await _complete_builtin_run(request, session_id, kind="action_plan", body=body)
 
 
 @router.get("/sessions/{session_id}/results", response_model=list[NodeResponse])

offline-backend-20260512/backend/app/gateway/services.py

@@ -247,6 +247,13 @@ _CONTEXT_CONFIGURABLE_KEYS: frozenset[str] = frozenset(
         # 关掉个人记忆召回 / builtin 记忆注入(见 multi_agent._run_payload)。
         "memory_recall_disabled",
         "memory_injection_enabled",
+        # 任务工作区 / 岗位会商把左侧任务卡的快照放在这里。它必须进入
+        # ``runtime.context``，TaskExternalContextMiddleware 才能在每次模型调用
+        # 前以隐藏消息注入任务名称、描述和方向；未白名单会被静默丢弃，模型只能
+        # 看到用户说的“这个任务”。
+        "task_external_context",
+        # 岗位会商下游席位按需读取前序席位完整交付时使用的运行期数据。
+        "roundtable_peer_deliveries",
         # 上下文压缩开关:由聊天框下方的按钮控制(默认关闭),透传到 run config 后
         # 在 _build_middlewares 里决定是否挂载 SummarizationMiddleware。
         "summarization_enabled",

offline-backend-20260512/backend/packages/harness/deerflow/persistence/migrations/versions/20260721_04_position_roundtable_builtin_threads.py

@@ -0,0 +1,47 @@
+"""Persist position-roundtable built-in agent thread ids.
+
+Revision ID: 20260721_04
+Revises: 20260721_03
+Create Date: 2026-07-21
+"""
+
+from __future__ import annotations
+
+import sqlalchemy as sa
+from alembic import op
+from sqlalchemy import inspect
+
+
+revision = "20260721_04"
+down_revision = "20260721_03"
+branch_labels = None
+depends_on = None
+
+
+def _has_table(table_name: str) -> bool:
+    return table_name in inspect(op.get_bind()).get_table_names()
+
+
+def _has_column(table_name: str, column_name: str) -> bool:
+    if not _has_table(table_name):
+        return False
+    return column_name in {column["name"] for column in inspect(op.get_bind()).get_columns(table_name)}
+
+
+def upgrade() -> None:
+    table = "position_roundtable_sessions"
+    if not _has_table(table):
+        return
+    if not _has_column(table, "summary_thread_id"):
+        op.add_column(table, sa.Column("summary_thread_id", sa.String(length=128), nullable=True))
+    if not _has_column(table, "action_plan_thread_id"):
+        op.add_column(table, sa.Column("action_plan_thread_id", sa.String(length=128), nullable=True))
+
+
+def downgrade() -> None:
+    table = "position_roundtable_sessions"
+    if not _has_table(table):
+        return
+    for column in ("action_plan_thread_id", "summary_thread_id"):
+        if _has_column(table, column):
+            op.drop_column(table, column)

offline-backend-20260512/backend/packages/harness/deerflow/persistence/position_roundtable/model.py

@@ -31,7 +31,12 @@ class PositionRoundtableSessionRow(Base):
     chain_snapshot: Mapped[str | None] = mapped_column(PortableLongText, nullable=True)
     intent_thread_id: Mapped[str | None] = mapped_column(String(128), nullable=True)
     intent_snapshot: Mapped[str | None] = mapped_column(PortableLongText, nullable=True)
+    # The two built-in closing agents own normal LangGraph threads too.  Keep
+    # their ids on the position session so a refresh can continue the same
+    # report/planning dialogue instead of silently starting a fresh context.
+    summary_thread_id: Mapped[str | None] = mapped_column(String(128), nullable=True)
     summary_snapshot: Mapped[str | None] = mapped_column(PortableLongText, nullable=True)
+    action_plan_thread_id: Mapped[str | None] = mapped_column(String(128), nullable=True)
     action_plan_snapshot: Mapped[str | None] = mapped_column(PortableLongText, nullable=True)
     # intent_pending -> active -> completed / archived. The router owns legal
     # transitions; the DB keeps it as a concise, portable string.

offline-backend-20260512/backend/packages/harness/deerflow/persistence/position_roundtable/sql.py

@@ -55,7 +55,9 @@ class PositionRoundtableRepository:
             "chain_snapshot": _loads(row.chain_snapshot),
             "intent_thread_id": row.intent_thread_id,
             "intent_snapshot": _loads(row.intent_snapshot),
+            "summary_thread_id": row.summary_thread_id,
             "summary_snapshot": _loads(row.summary_snapshot),
+            "action_plan_thread_id": row.action_plan_thread_id,
             "action_plan_snapshot": _loads(row.action_plan_snapshot),
             "status": row.status,
             "created_at": _time(row.created_at),
@@ -134,7 +136,9 @@ class PositionRoundtableRepository:
         task_snapshot: dict[str, Any] | object = _UNSET,
         intent_thread_id: str | None | object = _UNSET,
         intent_snapshot: dict[str, Any] | None | object = _UNSET,
+        summary_thread_id: str | None | object = _UNSET,
         summary_snapshot: dict[str, Any] | None | object = _UNSET,
+        action_plan_thread_id: str | None | object = _UNSET,
         action_plan_snapshot: dict[str, Any] | None | object = _UNSET,
         status: str | object = _UNSET,
     ) -> dict[str, Any] | None:
@@ -148,8 +152,12 @@ class PositionRoundtableRepository:
                 row.intent_thread_id = intent_thread_id or None
             if intent_snapshot is not _UNSET:
                 row.intent_snapshot = _dumps(intent_snapshot)
+            if summary_thread_id is not _UNSET:
+                row.summary_thread_id = summary_thread_id or None
             if summary_snapshot is not _UNSET:
                 row.summary_snapshot = _dumps(summary_snapshot)
+            if action_plan_thread_id is not _UNSET:
+                row.action_plan_thread_id = action_plan_thread_id or None
             if action_plan_snapshot is not _UNSET:
                 row.action_plan_snapshot = _dumps(action_plan_snapshot)
             if status is not _UNSET:
@@ -202,6 +210,10 @@ class PositionRoundtableRepository:
                     agent_id=str(node["agent_id"]),
                     position_id=node.get("position_id") or None,
                     status=str(node.get("status") or "locked"),
+                    latest_answer=node.get("latest_answer") or None,
+                    artifact_manifest=_dumps(node.get("artifact_manifest") or []),
+                    revision=int(node.get("revision") or 0),
+                    upstream_revision=_dumps(node.get("upstream_revision")) if node.get("upstream_revision") is not None else None,
                     created_at=now,
                     updated_at=now,
                 )

offline-backend-20260512/backend/tests/test_position_roundtable_sessions.py

@@ -9,6 +9,8 @@ import pytest_asyncio
 from fastapi import HTTPException
 from sqlalchemy.ext.asyncio import async_sessionmaker, create_async_engine
 
+from app.gateway import services as gateway_services
+from app.gateway.routers import intent as intent_router
 from app.gateway.routers import position_roundtable as position_roundtable_router
 from deerflow.persistence.base import Base
 from deerflow.persistence.position_roundtable import PositionRoundtableRepository
@@ -79,6 +81,39 @@ async def test_session_activation_freezes_chain_and_creates_nodes_once(store):
     assert [node["node_key"] for node in repeated_nodes] == [node["node_key"] for node in created_nodes]
 
 
+def test_task_external_context_reaches_agent_runtime_context():
+    """岗位会商的左侧任务快照不能在网关白名单中丢失。"""
+    task_context = {
+        "task": {
+            "title": "重点区域态势研判与协同处置方案",
+            "description": "梳理风险信号并形成协同处置建议",
+        }
+    }
+    config: dict = {"configurable": {}, "context": {}}
+
+    gateway_services.merge_run_context_overrides(
+        config,
+        {"task_external_context": task_context},
+    )
+
+    assert config["configurable"]["task_external_context"] == task_context
+    assert config["context"]["task_external_context"] == task_context
+
+
+def test_position_intent_thread_metadata_keeps_parent_and_child_task_ids():
+    metadata = intent_router._intent_thread_metadata(
+        {
+            "task": {"id": "task-100"},
+            "current_position": {"id": "intelligence"},
+        }
+    )
+
+    assert metadata["taskId"] == "task-100"
+    assert metadata["position_roundtable_task_id"] == "task-100"
+    assert metadata["position_task_id"] == "task-100:intelligence"
+    assert metadata["position_id"] == "intelligence"
+
+
 def test_chain_node_activation_rejects_missing_agent_ids():
     chain = {
         "id": "chain-a",
@@ -99,6 +134,58 @@ def test_chain_node_activation_rejects_missing_agent_ids():
     assert "缺少可用智能体" in str(error.value.detail)
 
 
+def test_intent_position_nodes_complete_from_confirmed_intent_and_unlock_next_stage():
+    nodes = [
+        {
+            "node_key": "stage-1-seat-1-intent-agent",
+            "stage_index": 0,
+            "seat_index": 0,
+            "agent_id": "intent-agent",
+            "position_id": "intelligence",
+            "status": "ready",
+        },
+        {
+            "node_key": "stage-2-seat-1-plan-agent",
+            "stage_index": 1,
+            "seat_index": 0,
+            "agent_id": "plan-agent",
+            "position_id": "planning",
+            "status": "locked",
+        },
+    ]
+
+    primed = position_roundtable_router._prime_intent_nodes(
+        nodes,
+        {"objective": "识别用户想要形成协同方案", "constraints": ["两天内完成"]},
+    )
+
+    assert [node["status"] for node in primed] == ["done", "ready"]
+    assert primed[0]["revision"] == 1
+    assert "识别用户想要形成协同方案" in primed[0]["latest_answer"]
+
+
+def test_all_nodes_delivered_requires_markdown_for_non_intent_nodes():
+    assert not position_roundtable_router._all_nodes_delivered(
+        [
+            {
+                "position_id": "planning",
+                "status": "done",
+                "artifact_manifest": [{"name": "plain.txt", "path": "mnt/user-data/outputs/plain.txt"}],
+            }
+        ]
+    )
+    assert position_roundtable_router._all_nodes_delivered(
+        [
+            {"position_id": "intelligence", "status": "done", "artifact_manifest": []},
+            {
+                "position_id": "planning",
+                "status": "done",
+                "artifact_manifest": [{"name": "report.md", "path": "mnt/user-data/outputs/report.md"}],
+            },
+        ]
+    )
+
+
 @pytest.mark.asyncio
 async def test_node_updates_are_user_scoped_and_keep_thread_out_of_messages(store):
     created = await store.create_session("user-a", external_task_id=None, task_snapshot={})
@@ -276,15 +363,55 @@ async def test_builtin_summary_and_action_plan_snapshots_are_invalidated_by_reje
     session = await store.get_session(created["id"], "user-a")
     assert session is not None
     nodes = await store.list_nodes(created["id"], "user-a") or []
-    summary = position_roundtable_router._build_summary_snapshot(session, nodes)
+    revisions = position_roundtable_router._source_revisions(nodes)
+    summary = position_roundtable_router._build_builtin_snapshot(
+        kind="summary",
+        thread_id="summary-thread",
+        content="# 岗位会商方案总结报告\n\n研判智能体的关键结论。",
+        artifacts=[
+            {
+                "name": "岗位会商方案总结报告.md",
+                "path": "mnt/user-data/outputs/岗位会商方案总结报告.md",
+                "mime_type": "text/markdown",
+                "modified_at": "2026-07-21T00:00:00+00:00",
+                "size_bytes": 100,
+            }
+        ],
+        source_revisions=revisions,
+        model="test-model",
+    )
     assert "岗位会商方案总结报告" in summary["content"]
-    assert "研判智能体" in summary["content"]
+    session = await store.update_session(
-    session = await store.update_session(created["id"], "user-a", summary_snapshot=summary)
+        created["id"], "user-a", summary_thread_id="summary-thread", summary_snapshot=summary
+    )
     assert session is not None
-    action_plan = position_roundtable_router._build_action_plan_snapshot(session, nodes)
+    action_plan = position_roundtable_router._build_builtin_snapshot(
+        kind="action_plan",
+        thread_id="action-plan-thread",
+        content="# 岗位会商行动规划\n\n## 子任务拆解",
+        artifacts=[
+            {
+                "name": "行动规划报告.md",
+                "path": "mnt/user-data/outputs/行动规划报告.md",
+                "mime_type": "text/markdown",
+                "modified_at": "2026-07-21T00:01:00+00:00",
+                "size_bytes": 100,
+            },
+            {
+                "name": "action-plan-subtasks.json",
+                "path": "mnt/user-data/outputs/action-plan-subtasks.json",
+                "mime_type": "application/json",
+                "modified_at": "2026-07-21T00:01:00+00:00",
+                "size_bytes": 200,
+            },
+        ],
+        source_revisions=revisions,
+        model="test-model",
+    )
     session = await store.update_session(
         created["id"],
         "user-a",
+        action_plan_thread_id="action-plan-thread",
         action_plan_snapshot=action_plan,
         status="completed",
     )
@@ -309,6 +436,30 @@ async def test_builtin_summary_and_action_plan_snapshots_are_invalidated_by_reje
     assert rejected_session["action_plan_snapshot"]["invalidated_by"] == nodes[0]["node_key"]
 
 
+def test_rejected_or_stale_node_requires_a_new_markdown_delivery() -> None:
+    old = [
+        {
+            "name": "report.md",
+            "path": "mnt/user-data/outputs/report.md",
+            "modified_at": "2026-07-21T00:00:00+00:00",
+            "size_bytes": 100,
+        }
+    ]
+    assert not position_roundtable_router._markdown_delivery_changed(old, old)
+    assert position_roundtable_router._markdown_delivery_changed(
+        old,
+        [{**old[0], "modified_at": "2026-07-21T00:05:00+00:00", "size_bytes": 120}],
+    )
+
+
+def test_activation_requires_a_structured_intent_snapshot() -> None:
+    assert not position_roundtable_router._has_confirmed_intent({"objective": "只有目标"})
+    assert not position_roundtable_router._has_confirmed_intent({"constraints": []})
+    assert position_roundtable_router._has_confirmed_intent(
+        {"objective": "明确任务目标", "constraints": []}
+    )
+
+
 @pytest.mark.asyncio
 async def test_handoff_context_inlines_only_upstream_thread_text_artifacts(tmp_path, monkeypatch):
     artifact = tmp_path / "handoff.md"