Merge branch 'dhk-1' of http://192.168.0.200/knowledge-base-bj/deerflow-server into dhk-1

2026-06-05 11:20:22 +08:00
commit 9c9330114f
@@ -0,0 +1,124 @@
+# 多智能体「后台挂起」—— 当前状态、遗留问题与下次继续指南
+
+> 配套设计文档：`multi-agent-background-run-dev.md`（分 Phase 0-9 的总方案）。
+> 本文记录**截至本次会话的实际进展 + 还没解决的遗留问题**，方便下次直接续接。
+
+---
+
+## 一、整体状态：跑通到哪了
+
+「后台挂起」= Step2 把研讨交给后端**进程内驱动真模型**（leader→各席位→共识→报告），
+survive 刷新/关页，对话与报告作为一个 **run 写进草稿**（和手动跑同一存取），可在历史回看。
+
+**真网关已在本地验证可跑通**（2026-06-04 日志确认：roundtable-coordinator + 各席位 +
+roundtable-report 真调 deepseek、真搜索、跑到 report run success）。但还有几个**内容/产物层面的遗留问题**（见第三节）。
+
+### 已完成（可用）
+- ✅ 后端持久化：`roundtable_jobs` 表（迁移 `20260604_01` / `20260605_01` dialogues / `20260605_02` step3）。
+- ✅ 编排引擎 `packages/harness/deerflow/agents/roundtable_orchestrator/`（recommend + chain + 报告，纯逻辑可单测）。
+- ✅ 控制面 `app/gateway/routers/roundtable_jobs.py`：start/get/by-draft/stream(SSE)/resume/cancel/admin-cleanup。
+- ✅ 执行器 `app/gateway/roundtable_job_executor.py`：后台任务 + `set_current_user` + 生命周期 + **每轮把对话作为 `job-<id>` run 写进 `draft.step2.runs[]`（同手动存取）** + done 写 step3。
+- ✅ 真网关 `app/gateway/roundtable_inprocess_gateway.py`：`SimpleNamespace(app)` + `start_run` 进程内驱动 agent；复用 `multi_agent._extract_dispatch` 解析派活；checkpointer 读最终 messages。
+- ✅ 模拟网关 `roundtable_orchestrator/simulated_gateway.py`（不调真模型，前端联调用）。
+- ✅ 前端：`api/roundtable-jobs.ts`、`lib/job-types.ts`、Step2「后台挂起」按钮、历史下拉 loading、`DispatchChainModal`/`DispatchChainJobModal`（流程图 + 对话面板 markdown + 加宽 + 澄清作答 + 取消 + 查看）。
+- ✅ 上下文「广播」（inline）：席位/总控 prompt 带上「任务背景 + 前序各席位交付 + 别提问」约束（修「席位拿不到关键信息、反问用户、总控瞎编」）。
+- ✅ 后端测试：`tests/test_roundtable_jobs.py` / `test_roundtable_orchestrator.py` / `test_roundtable_job_executor.py` / `test_roundtable_jobs_api.py` 全过。
+
+---
+
+## 二、怎么启动 / 验证（每次开工先做）
+
+1. **开真网关**（默认是模拟网关）：已写进 `offline-backend-20260512/.env`：
+   ```
+   ROUNDTABLE_JOB_GATEWAY=inprocess
+   ```
+   启动日志应出现 `[roundtable-jobs] gateway = inprocess (real models)`。
+
+2. **必须用 `make gateway`（不带 --reload）跑后端！**
+   `make dev` 的 `--reload` 监视 `backend/` 目录，agent 运行时往 `.deer-flow/` 写文件
+   （memory.json / checkpoint / 沙箱输出）会触发热重载，**把长时后台任务杀掉**。
+   （改进项：给 uvicorn 加 `--reload-exclude '.deer-flow/*'`，或始终用 `make gateway`。）
+
+3. **直接查库验证**（最有效，PowerShell 控制台是 GBK，导出到文件看中文）：
+   ```bash
+   cd offline-backend-20260512/backend
+   # 查最新作业的 dialogues / step3
+   PYTHONPATH=. uv run python -c "import sqlite3,json; con=sqlite3.connect('.deer-flow/data/deerflow.db'); ..."
+   ```
+   - 作业表：`.deer-flow/data/deerflow.db` → 表 `roundtable_jobs`（`dialogues`/`step3` 列）。
+   - 草稿表：`roundtable_drafts` → `step2` 列里应有 `runs[ {id:"job-<jobId>", jobId, jobStatus, step2RoundtableDialogues:[...], step3:{...}} ]`。
+
+4. **前端验证**：后台挂起 → 等跑完 → 从历史加载该草稿 → Step2 应显示后台研讨完整对话（走现有 `normalizeStep2Runs`/`hydrateStep2Run`）。
+
+---
+
+## 三、遗留问题（下次重点，已按优先级排）
+
+### P0 —— 报告 agent 不写 HTML 文件（结果绘制拿不到产物）
+- **现象**：后台报告 run 跑完，日志 `report html: no .html found in ...threads/<id>/user-data/outputs`；
+  `run.step3.html` 回退成报告 agent 的**纯文字**（不含 `<html>`），第3步画不出东西。
+- **已做**：`build_report_prompt`（`roundtable_orchestrator/engine.py`）已对齐前端
+  `useStep3Report.buildReportPrompt`，明确「用 write_file 写入 /mnt/user-data/outputs/方案总览.html」；
+  `_read_report_html`（`roundtable_inprocess_gateway.py`）已改为扫 outputs 目录任意 `.html`。
+- **仍没解决**：报告 agent（deepseek-chat）还是没调 `write_file`。
+- **下次怎么查**：
+  1. 对比**前端那条能写出 HTML 的报告 run** 的确切调用（`useStep3Report.ts` 里 `streamMultiAgent`
+     调 `/api/multi-agent/run/stream` 的 body：`agent_type`/`excluded_tools`/`context` 等），
+     和后台 `_run_agent_turn`（`roundtable_inprocess_gateway.py`，agent_name=roundtable-report，
+     `excluded_tools=_REPORT_EXCLUDED`）逐项比对，找出差异（怀疑 reasoning_effort / thinking /
+     某个 context 字段 / recursion 步数导致它没走到 write_file）。
+  2. 看报告 run 的最终 messages（checkpoint）里有没有 write_file 的 tool_call（被 dangling 截断？）。
+  3. 报告 agent 的 SOUL：`app/gateway/routers/_roundtable_seed/roundtable-report/SOUL.md`——确认它默认行为。
+
+### P1 —— 内容质量（本次刚加 inline 广播，待真机验证）
+- **已做**：席位/总控 prompt 注入「任务背景 + 前序交付 + 别提问」。
+- **待验证**：真机跑一次，看席位是否基于真实上下文产出（不再反问/凭记忆瞎编）。
+- **若 inline 不够**：考虑实现**真·thread 广播**（像 `multi_agent._append_thread_message` 那样把
+  「用户意图 / 派活 / 各席位交付」追加进各 thread 历史）。进程内要避开 HTTP loopback 鉴权，
+  得用 checkpointer 直接 append human 消息（参考 `threads` 路由 POST `/state` 的内部逻辑）。这块较重，先看 inline 够不够。
+
+### P2 —— 对话「没有步骤（steps）」
+- **现象**：手动跑的气泡带工具调用步骤卡（`onPhaseChange` → `appendStepIfNew` → `Step2Dialogue.steps`）；
+  后台对话只存了最终正文，没有 steps。
+- **下次怎么做**：后台 `_run_agent_turn` 跑完后，从 checkpoint 的 messages 里抽出 tool_calls，
+  映射成 `Step2Dialogue.steps`（`{id,key,label,ts,args}`），在执行器 `_to_step2_dialogue` 里带上。
+  纯展示增强，不影响内容/报告。
+
+### P3 —— 双跑浪费 + --reload 坑（工程健壮性）
+- 进 Step2 时前端会自动跑一遍本地研讨（真模型、真线程）；点后台挂起后前端那条 run 在后端还会跑完
+  （`stopOrchestration` 只断客户端、停不了后端 run）→ 和后台作业短暂双跑、双倍 token。
+  - 优化方向：进 Step2 改成「显式开始」而非自动跑；或后台挂起时真正 cancel 前端那条 run。
+- `make dev` 的 `--reload` 会杀后台任务（见第二节）；改进项：`--reload-exclude '.deer-flow/*'`。
+
+---
+
+## 四、关键文件清单（下次定位用）
+
+**后端**
+| 文件 | 作用 |
+|------|------|
+| `app/gateway/roundtable_inprocess_gateway.py` | 真网关：进程内驱动 leader/seat/re 编排循环 + `build_report_prompt` + inline 广播（`_seat_message`/`_deliveries_block`） |
+| `.../roundtable_orchestrator/progress.py` | 进度落库 + `on_dialogue` 每轮回调 |
+| `.../persistence/roundtable_jobs/` | job 表 model/sql |
+| `app/gateway/deps.py` | `ROUNDTABLE_JOB_GATEWAY` 选网关 + 注入 executor（draft_store） |
+
+**前端**
+| 文件 | 作用 |
+|------|------|
+| `src/roundtable-planning/api/roundtable-jobs.ts` | 作业 API + SSE |
+| `src/roundtable-planning/lib/job-types.ts` | 类型（RoundtableJob/JobDialogue/...） |
+| `src/roundtable-planning/components/DispatchChainModal.tsx` / `DispatchChainJobModal.tsx` | 流程图弹窗 + 对话面板 |
+| `src/roundtable-planning/pages/RoundtablePlanningPage.tsx` | 后台挂起按钮、历史下拉、**runs[] 多运行模型**（`hydrateStep2Run`/`getStep2SnapshotWithRuns`/`normalizeStep2Runs`） |
+| `src/roundtable-planning/api/drafts.ts` | `DraftStep2RunSnapshot`（带 jobId/jobStatus/step3） |
+
+---
+
+## 五、关键架构事实（避免重复踩坑）
+
+1. **真网关进程内驱动**：`start_run(body, thread_id, SimpleNamespace(app=app))` + `set_current_user`，
+   不走 HTTP（loopback 调 `/api/multi-agent/*` 不可行：`_auth_headers` 只转发用户 token + owner check）。
+2. **后台作业首启自建全新线程**（`StartParams.is_resume`），绝不复用前端本地研讨的线程（会撞 409）。
+3. **存取统一**：后台作业 = `draft.step2.runs[]` 里一个带 jobId 的 run，和手动跑同一存取/hydrate 路径，
+   **不要再开 observe/回显分支**（之前试过，被否）。
+4. **聊天记录确实存上了**（job.dialogues + draft 的 run）；之前"没存上"是测了旧代码 / 看的是前端本地残缺数据。
+5. 多 worker / SSE：状态进 DB、SSE 走 DB 轮询；真网关进程内无 loopback。
@@ -15,6 +15,10 @@
        "@codemirror/lang-markdown": "^6.5.0",
        "@codemirror/lang-python": "^6.2.1",
        "@codemirror/language-data": "^6.5.2",
+        "@dnd-kit/core": "^6.3.1",
+        "@dnd-kit/sortable": "^10.0.0",
+        "@dnd-kit/utilities": "^3.2.2",
+        "@fontsource/noto-sans-sc": "^5.2.9",
        "@langchain/core": "^1.1.15",
        "@langchain/langgraph-sdk": "^1.5.3",
        "@radix-ui/react-avatar": "^1.1.11",
@@ -55,9 +59,12 @@
        "cmdk": "^1.1.1",
        "codemirror": "^6.0.2",
        "date-fns": "^4.1.0",
+        "dayjs": "^1.11.21",
+        "dnd-kit": "^0.0.2",
        "docx": "^9.5.1",
        "dotenv": "^17.2.3",
        "embla-carousel-react": "^8.6.0",
+        "file-saver": "^2.0.5",
        "gsap": "^3.13.0",
        "hast": "^1.0.0",
        "katex": "^0.16.28",
@@ -90,6 +97,7 @@
      },
      "devDependencies": {
        "@tailwindcss/postcss": "^4.0.15",
+        "@types/file-saver": "^2.0.7",
        "@types/gsap": "^3.0.0",
        "@types/node": "^20.14.10",
        "@types/react": "^19.0.0",
@@ -858,6 +866,55 @@
        "w3c-keyname": "^2.2.4"
      }
    },
+    "node_modules/@dnd-kit/accessibility": {
+      "version": "3.1.1",
+      "resolved": "https://registry.npmjs.org/@dnd-kit/accessibility/-/accessibility-3.1.1.tgz",
+      "integrity": "sha512-2P+YgaXF+gRsIihwwY1gCsQSYnu9Zyj2py8kY5fFvUM1qm2WA2u639R6YNVfU4GWr+ZM5mqEsfHZZLoRONbemw==",
+      "dependencies": {
+        "tslib": "^2.0.0"
+      },
+      "peerDependencies": {
+        "react": ">=16.8.0"
+      }
+    },
+    "node_modules/@dnd-kit/core": {
+      "version": "6.3.1",
+      "resolved": "https://registry.npmjs.org/@dnd-kit/core/-/core-6.3.1.tgz",
+      "integrity": "sha512-xkGBRQQab4RLwgXxoqETICr6S5JlogafbhNsidmrkVv2YRs5MLwpjoF2qpiGjQt8S9AoxtIV603s0GIUpY5eYQ==",
+      "dependencies": {
+        "@dnd-kit/accessibility": "^3.1.1",
+        "@dnd-kit/utilities": "^3.2.2",
+        "tslib": "^2.0.0"
+      },
+      "peerDependencies": {
+        "react": ">=16.8.0",
+        "react-dom": ">=16.8.0"
+      }
+    },
+    "node_modules/@dnd-kit/sortable": {
+      "version": "10.0.0",
+      "resolved": "https://registry.npmjs.org/@dnd-kit/sortable/-/sortable-10.0.0.tgz",
+      "integrity": "sha512-+xqhmIIzvAYMGfBYYnbKuNicfSsk4RksY2XdmJhT+HAC01nix6fHCztU68jooFiMUB01Ky3F0FyOvhG/BZrWkg==",
+      "dependencies": {
+        "@dnd-kit/utilities": "^3.2.2",
+        "tslib": "^2.0.0"
+      },
+      "peerDependencies": {
+        "@dnd-kit/core": "^6.3.0",
+        "react": ">=16.8.0"
+      }
+    },
+    "node_modules/@dnd-kit/utilities": {
+      "version": "3.2.2",
+      "resolved": "https://registry.npmjs.org/@dnd-kit/utilities/-/utilities-3.2.2.tgz",
+      "integrity": "sha512-+MKAJEOfaBe5SmV6t34p80MMKhjvUz0vRrvVJbPT0WElzaOJ/1xs+D+KDv+tD/NE5ujfrChEcshd4fLn0wpiqg==",
+      "dependencies": {
+        "tslib": "^2.0.0"
+      },
+      "peerDependencies": {
+        "react": ">=16.8.0"
+      }
+    },
    "node_modules/@esbuild/aix-ppc64": {
      "version": "0.25.12",
      "resolved": "https://registry.npmmirror.com/@esbuild/aix-ppc64/-/aix-ppc64-0.25.12.tgz",
@@ -1334,6 +1391,14 @@
      "resolved": "https://registry.npmjs.org/@floating-ui/utils/-/utils-0.2.11.tgz",
      "integrity": "sha512-RiB/yIh78pcIxl6lLMG0CgBXAZ2Y0eVHqMPYugu+9U0AeT6YBeiJpf7lbdJNIugFP5SIjwNRgo4DhR1Qxi26Gg=="
    },
+    "node_modules/@fontsource/noto-sans-sc": {
+      "version": "5.2.9",
+      "resolved": "https://registry.npmjs.org/@fontsource/noto-sans-sc/-/noto-sans-sc-5.2.9.tgz",
+      "integrity": "sha512-bTUIWGBgJDpwi5qAr+x0/lcgv80IHTB9vl6s2f6EymZEa7qYV99yNRBZuKFT+SYDKVunZrjCEhWtpxqmbXWl5Q==",
+      "funding": {
+        "url": "https://github.com/sponsors/ayuhito"
+      }
+    },
    "node_modules/@iconify/types": {
      "version": "2.0.0",
      "resolved": "https://registry.npmjs.org/@iconify/types/-/types-2.0.0.tgz",
@@ -4995,6 +5060,12 @@
        "@types/estree": "*"
      }
    },
+    "node_modules/@types/file-saver": {
+      "version": "2.0.7",
+      "resolved": "https://registry.npmjs.org/@types/file-saver/-/file-saver-2.0.7.tgz",
+      "integrity": "sha512-dNKVfHd/jk0SkR/exKGj2ggkB45MAkzvWCaqLUUgkyjITkGNzH8H+yUwr+BLJUBjZOe9w8X3wgmXhZDRg1ED6A==",
+      "dev": true
+    },
    "node_modules/@types/geojson": {
      "version": "7946.0.16",
      "resolved": "https://registry.npmjs.org/@types/geojson/-/geojson-7946.0.16.tgz",
@@ -6146,9 +6217,9 @@
      }
    },
    "node_modules/dayjs": {
-      "version": "1.11.20",
-      "resolved": "https://registry.npmjs.org/dayjs/-/dayjs-1.11.20.tgz",
-      "integrity": "sha512-YbwwqR/uYpeoP4pu043q+LTDLFBLApUP6VxRihdfNTqu4ubqMlGDLd6ErXhEgsyvY0K6nCs7nggYumAN+9uEuQ=="
+      "version": "1.11.21",
+      "resolved": "https://registry.npmjs.org/dayjs/-/dayjs-1.11.21.tgz",
+      "integrity": "sha512-98IT+HOahAisibz/yjKbzuOBwYcjJ7BCLPzARyHiyEBmRz4fatF+KPJszEHXsGYjUG234aH/cOjW1wwTbKUZlA=="
    },
    "node_modules/debug": {
      "version": "4.4.3",
@@ -6228,6 +6299,12 @@
        "url": "https://github.com/sponsors/wooorm"
      }
    },
+    "node_modules/dnd-kit": {
+      "version": "0.0.2",
+      "resolved": "https://registry.npmjs.org/dnd-kit/-/dnd-kit-0.0.2.tgz",
+      "integrity": "sha512-d8AYd6I7D2b5u882+QNVGw0slBAt851/LWZ2j/pU+onf5/TGEKXeb47sCyhPYKEAUXp4oLfvWfNCqfkU03R1lw==",
+      "deprecated": "Package no longer supported. Contact Support at https://www.npmjs.com/support for more info."
+    },
    "node_modules/docx": {
      "version": "9.6.1",
      "resolved": "https://registry.npmjs.org/docx/-/docx-9.6.1.tgz",
@@ -6470,6 +6547,11 @@
        }
      }
    },
+    "node_modules/file-saver": {
+      "version": "2.0.5",
+      "resolved": "https://registry.npmjs.org/file-saver/-/file-saver-2.0.5.tgz",
+      "integrity": "sha512-P9bmyZ3h/PRG+Nzga+rbdI4OEpNDzAVyy74uVO9ATgzLK6VtAsYybF/+TOCvrc0MO793d6+42lLyZTw7/ArVzA=="
+    },
    "node_modules/format": {
      "version": "0.2.2",
      "resolved": "https://registry.npmjs.org/format/-/format-0.2.2.tgz",
@@ -22,6 +22,7 @@
    "@dnd-kit/core": "^6.3.1",
    "@dnd-kit/sortable": "^10.0.0",
    "@dnd-kit/utilities": "^3.2.2",
+    "@fontsource/noto-sans-sc": "^5.2.9",
    "@langchain/core": "^1.1.15",
    "@langchain/langgraph-sdk": "^1.5.3",
    "@radix-ui/react-avatar": "^1.1.11",
@@ -1,6 +1,10 @@
-import "./styles/globals.css";
-import "tdesign-react/es/style/index.css";
-import "katex/dist/katex.min.css";
+import "@fontsource/noto-sans-sc/chinese-simplified-400.css";
+import "@fontsource/noto-sans-sc/chinese-simplified-500.css";
+import "@fontsource/noto-sans-sc/chinese-simplified-600.css";
+import "@fontsource/noto-sans-sc/chinese-simplified-700.css";
+import "./styles/globals.css";
+import "tdesign-react/es/style/index.css";
+import "katex/dist/katex.min.css";

 import React from "react";
 import ReactDOM from "react-dom/client";
@@ -29,6 +29,7 @@ import {
 import { cn } from "@/lib/utils";

 import { getAvatar } from "./Avatars";
+import { MarkdownContent } from "./MessageBubble";
 import { chainTagClasses } from "../lib/business-chain";
 import { getAvatarTypeFor } from "../lib/roundtable-constants";
 import type {
@@ -140,7 +141,7 @@ export function DispatchChainModal({

  return (
    <Dialog open={open} onOpenChange={onOpenChange}>
-      <DialogContent className="flex max-h-[88vh] flex-col gap-0 sm:max-w-5xl">
+      <DialogContent className="flex max-h-[90vh] w-[94vw] flex-col gap-0 sm:max-w-[1240px]">
        <DialogHeader className="shrink-0">
          <DialogTitle className="flex items-center gap-2">
            {status === "running" && (
@@ -223,9 +224,9 @@ export function DispatchChainModal({
                <div key={d.id} className="flex gap-2.5">
                  <div className="shrink-0">{getAvatar(avatarForDialogue(d), "h-7 w-7")}</div>
                  <div className="min-w-0 flex-1">
-                    <div className="text-xs font-semibold text-foreground/80">{d.name}</div>
-                    <div className="mt-0.5 whitespace-pre-wrap break-words text-sm leading-relaxed text-foreground/70">
-                      {d.content || "…"}
+                    <div className="mb-0.5 text-xs font-semibold text-foreground/80">{d.name}</div>
+                    <div className="prose prose-sm max-w-none break-words text-sm leading-relaxed text-foreground/80 dark:prose-invert">
+                      <MarkdownContent>{d.content || "…"}</MarkdownContent>
                    </div>
                  </div>
                </div>
@@ -289,10 +290,15 @@ export function DispatchChainModal({
                停止
              </Button>
            )}
-            <Button onClick={onView} disabled={!onView} className="gap-1.5">
-              <EyeIcon className="size-4" />
-              {status === "done" ? "查看报告" : "查看"}
-            </Button>
+            {/* 只对「运行中」任务隐藏入口——它正在后台真跑,进入会暂停前端编排、
+                与后台作业双跑或借此重跑。其它状态（排队 / 待澄清 / 已完成 / 已取消 /
+                出错）都不是正在跑的后台作业,可正常查看。 */}
+            {status !== "running" && (
+              <Button onClick={onView} disabled={!onView} className="gap-1.5">
+                <EyeIcon className="size-4" />
+                {status === "done" ? "查看报告" : "查看"}
+              </Button>
+            )}
          </div>
        </DialogFooter>
      </DialogContent>
@@ -1690,32 +1690,39 @@ function IntentSummaryCard({
        {intentReady ? (
          <>
            <ul className="text-xs text-slate-700 space-y-2.5 font-normal">
-              <li className="flex items-baseline gap-2 leading-relaxed">
-                <span className="text-primary text-base leading-none">•</span>
-                <span>
+              <li className="flex items-start gap-2 leading-relaxed">
+                <span className="text-primary text-base leading-none pt-0.5">•</span>
+                <span
+                  className={cn(
+                    "min-w-0 leading-relaxed",
+                    collapsible && !expanded && "line-clamp-3",
+                  )}
+                >
                  <b className="text-slate-900 font-bold">目标</b>: {intentReady.objective}
                </span>
              </li>
            </ul>
            {collapsible && (
-              <button
-                type="button"
-                onClick={() => setExpanded((v) => !v)}
-                className="flex items-center gap-0.5 text-[10px] font-bold text-slate-500 hover:text-primary transition-colors"
-                aria-expanded={expanded}
-              >
-                {expanded ? (
-                  <>
-                    收起
-                    <ChevronUp className="w-3.5 h-3.5" />
-                  </>
-                ) : (
-                  <>
-                    展开
-                    <ChevronDown className="w-3.5 h-3.5" />
-                  </>
-                )}
-              </button>
+              <div className="flex justify-end">
+                <button
+                  type="button"
+                  onClick={() => setExpanded((v) => !v)}
+                  className="inline-flex items-center gap-0.5 text-xs font-bold text-slate-500 hover:text-primary transition-colors"
+                  aria-expanded={expanded}
+                >
+                  {expanded ? (
+                    <>
+                      收起
+                      <ChevronUp className="w-3.5 h-3.5" />
+                    </>
+                  ) : (
+                    <>
+                      展开
+                      <ChevronDown className="w-3.5 h-3.5" />
+                    </>
+                  )}
+                </button>
+              </div>
            )}
            {showFullDetail && (
              <ul className="text-xs text-slate-700 space-y-2.5 font-normal">
@@ -22,6 +22,12 @@ from typing import Any

 from langgraph.checkpoint.base import empty_checkpoint

+from app.gateway.roundtable_run_policy import roundtable_run_policy
+from app.gateway.routers._roundtable_seed import (
+    COORDINATOR_AGENT_ID,
+    REPORT_AGENT_ID,
+    ensure_roundtable_functional_agents,
+)
 from deerflow.agents.roundtable_orchestrator.types import (
    LeaderTurn,
    ReportResult,
@@ -31,22 +37,14 @@ from deerflow.agents.roundtable_orchestrator.types import (
 from deerflow.config.paths import get_paths
 from deerflow.runtime.serialization import serialize_channel_values

-from app.gateway.routers._roundtable_seed import (
-    COORDINATOR_AGENT_ID,
-    REPORT_AGENT_ID,
-    ensure_roundtable_functional_agents,
-)
-
 logger = logging.getLogger(__name__)

 _ROUNDTABLE_THREAD_METADATA: dict[str, Any] = {"thread_type": "roundtable", "system": True}
 _REPORT_FILE = "方案总览.html"

-# 各角色的工具裁剪（镜像 multi_agent 的 _leader_run / _special_run）。
-_LEADER_EXCLUDED = ["web_search"]
-_LEADER_SKILL_STOP = ["agent_orchestration"]
-_SEAT_EXCLUDED = ["ask_clarification", "present_files", "view_image", "agent_orchestration"]
-_REPORT_EXCLUDED = [*_SEAT_EXCLUDED, "bash", "str_replace"]
+# 各角色的工具裁剪 / thinking / reasoning 全部取自单一数据源 roundtable_run_policy
+# （与 multi_agent 的 _leader_run / _special_run 同源），不再在本文件硬编码——历史上
+# 这里漏禁了 web_search，导致后台席位 / 报告能联网搜索而前端不能，现由 policy 统一。


 # ── 进程内底层操作 ───────────────────────────────────────────────────────────
@@ -101,6 +99,56 @@ def _message_text(message: dict[str, Any] | None) -> str:
    return ""


+def _latest_ai_text(messages: list[dict[str, Any]]) -> str:
+    """取**最后一条有正文**的 AI 消息文本。
+
+    席位常以「调一堆工具」的 AIMessage 收尾（正文为空、只有 tool_calls，甚至 dangling），
+    直接取最后一条 AI 消息会拿到空串 → 报告就拿不到席位产物。这里向前找到真正有文字的那条。
+    """
+    for m in reversed(messages):
+        if not isinstance(m, dict):
+            continue
+        if (m.get("type") or m.get("role")) not in {"ai", "assistant"}:
+            continue
+        text = _message_text(m)
+        if text.strip():
+            return text
+    return ""
+
+
+def _normalize_tool_calls(tool_calls: Any) -> list[dict[str, Any]]:
+    """把 checkpoint 里的 tool_calls 归一化成 {name, args, id}（丢掉无名项）。"""
+    out: list[dict[str, Any]] = []
+    for tc in tool_calls or []:
+        if not isinstance(tc, dict):
+            continue
+        name = tc.get("name") or ""
+        if not name:
+            continue
+        out.append({"name": name, "args": tc.get("args") or {}, "id": tc.get("id") or ""})
+    return out
+
+
+def _tool_calls_since_last_human(messages: list[dict[str, Any]]) -> list[dict[str, Any]]:
+    """收集「最后一条 human 消息之后」各 AI 消息的 tool_calls（即**本轮**调用），保序。
+
+    席位 thread 跨轮累积历史，直接扫全量会把往轮的工具调用也算进来；以最后一条 human
+    （本轮 run 的输入）为界，只取其后的 AI 工具调用，对应前端「这一条气泡的步骤卡」。
+    """
+    last_human = -1
+    for i, m in enumerate(messages):
+        if isinstance(m, dict) and (m.get("type") or m.get("role")) in {"human", "user"}:
+            last_human = i
+    out: list[dict[str, Any]] = []
+    for m in messages[last_human + 1 :]:
+        if not isinstance(m, dict):
+            continue
+        if (m.get("type") or m.get("role")) not in {"ai", "assistant"}:
+            continue
+        out.extend(_normalize_tool_calls(m.get("tool_calls")))
+    return out
+
+
 async def _run_agent_turn(
    app,
    *,
@@ -216,27 +264,31 @@ class InProcessRoundtableGateway:
        if self._first_leader and self._seat_intro:
            msg = self._seat_intro + "\n\n" + message
            self._first_leader = False
+        policy = roundtable_run_policy("leader")
        messages = await _run_agent_turn(
            self._app,
            thread_id=thread_ids[coordinator_name],
            agent_name=coordinator_name,
            message=msg,
            model=model or self._model,
-            excluded_tools=_LEADER_EXCLUDED,
-            skill_stop_names=_LEADER_SKILL_STOP,
-            thinking_enabled=False,
-            reasoning_effort="low",
+            excluded_tools=policy.excluded_tools,
+            skill_stop_names=policy.skill_stop_names,
+            thinking_enabled=policy.thinking_enabled,
+            reasoning_effort=policy.reasoning_effort,
        )
        ai = _latest_ai_message(messages)
-        final_text = _message_text(ai)
+        # tool_calls 取最后一条 AI 消息（派活就在这条）；正文取最后一条有文字的 AI 消息。
+        final_text = _latest_ai_text(messages)
        tool_calls = (ai.get("tool_calls") if ai else None) or []
+        # 本轮工具调用（派活 / 澄清），带进 LeaderTurn 供草稿 run 重建步骤卡。
+        turn_tool_calls = _normalize_tool_calls(tool_calls)

        # 澄清优先
        clar = next((tc for tc in tool_calls if (tc.get("name") or "").lower() == "ask_clarification"), None)
        if clar is not None:
            cargs = clar.get("args") or {}
            question = str(cargs.get("question") or "").strip() or "总控需要补充信息，请补充后继续。"
-            return LeaderTurn(final_text=final_text, clarification=question)
+            return LeaderTurn(final_text=final_text, clarification=question, tool_calls=turn_tool_calls)

        # 派活解析（复用 multi_agent 纯函数）
        from app.gateway.routers.multi_agent import _extract_dispatch
@@ -249,38 +301,40 @@ class InProcessRoundtableGateway:
                    dispatched.append((name, task))
            except Exception:  # noqa: BLE001
                logger.warning("dispatch parse failed for tool_call", exc_info=True)
-        return LeaderTurn(final_text=final_text, dispatched=dispatched)
+        return LeaderTurn(final_text=final_text, dispatched=dispatched, tool_calls=turn_tool_calls)

    async def run_seat(self, *, thread_ids, agent_id, task, model) -> SeatTurn:
+        policy = roundtable_run_policy("seat")
        messages = await _run_agent_turn(
            self._app,
            thread_id=thread_ids[agent_id],
            agent_name=agent_id,
            message=task,
            model=model or self._model,
-            excluded_tools=_SEAT_EXCLUDED,
-            skill_stop_names=[],
-            thinking_enabled=True,
-            reasoning_effort="medium",
+            excluded_tools=policy.excluded_tools,
+            skill_stop_names=policy.skill_stop_names,
+            thinking_enabled=policy.thinking_enabled,
+            reasoning_effort=policy.reasoning_effort,
        )
-        return SeatTurn(final_text=_message_text(_latest_ai_message(messages)))
+        return SeatTurn(final_text=_latest_ai_text(messages), tool_calls=_tool_calls_since_last_human(messages))

    async def create_report_thread(self) -> str:
        ensure_roundtable_functional_agents()
        return await _create_thread(self._app)

    async def run_report(self, *, report_thread_id, prompt, model) -> ReportResult:
+        policy = roundtable_run_policy("report")
        messages = await _run_agent_turn(
            self._app,
            thread_id=report_thread_id,
            agent_name=REPORT_AGENT_ID,
            message=prompt,
            model=model or self._model,
-            excluded_tools=_REPORT_EXCLUDED,
-            skill_stop_names=[],
-            thinking_enabled=True,
-            reasoning_effort="medium",
+            excluded_tools=policy.excluded_tools,
+            skill_stop_names=policy.skill_stop_names,
+            thinking_enabled=policy.thinking_enabled,
+            reasoning_effort=policy.reasoning_effort,
        )
-        summary = _message_text(_latest_ai_message(messages))
+        summary = _latest_ai_text(messages)
        html = _read_report_html(self._app, report_thread_id, self._user_id) or summary
        return ReportResult(html=html, summary=summary, model=model or self._model)
@@ -17,7 +17,8 @@ from __future__ import annotations
 import asyncio
 import logging
 from dataclasses import dataclass, field
-from typing import Callable
+from datetime import UTC, datetime
+from typing import Any, Callable

 from deerflow.agents.roundtable_orchestrator import (
    JobProgressWriter,
@@ -36,6 +37,115 @@ class _JobUser:
        self.id = user_id


+# 头像配色：与前端 getAvatarTypeFor 对齐（内置圆桌席位固定色，其余按序轮换）。
+_BUILTIN_AVATAR = {
+    "roundtable-intelligence": "purple",
+    "roundtable-environment": "emerald",
+    "roundtable-solution-design": "human",
+    "roundtable-risk-review": "amber",
+    "roundtable-execution-plan": "pink",
+    "roundtable-summary": "blue",
+}
+_FALLBACK_AVATAR = ["blue", "purple", "emerald", "amber", "pink"]
+
+
+def _avatar_for(agent_id: str, index: int) -> str:
+    return _BUILTIN_AVATAR.get(agent_id) or _FALLBACK_AVATAR[index % len(_FALLBACK_AVATAR)]
+
+
+# 工具步骤卡的回退文案。前端 getStepDisplay 对 tool 步骤是从 key+args 推图标/文案的,
+# 存的 label 基本不展示——这里给个合理回退即可（与前端默认文案对齐）。
+_STEP_LABEL = {
+    "agent_orchestration": "派活给子智能体",
+    "write_file": "写入文件",
+    "str_replace": "写入文件",
+    "read_file": "读取文件",
+    "ls": "列出文件夹",
+    "bash": "执行命令",
+    "web_search": "搜索相关信息",
+    "ask_clarification": "需要你的协助",
+    "present_files": "展示文件",
+}
+
+
+def _steps_from_tool_calls(tool_calls: list[dict], dialogue_id: object, ts: str) -> list[dict]:
+    """把一轮 tool_calls 映射成前端 ``Step2Dialogue.steps``（工具步骤卡）。
+
+    - ``key`` = ``tool_calling:<toolName>``——前端 ``getStepDisplay`` 据此推图标 + 文案;
+    - ``id``  = ``<dialogueId>-<i>``——对齐前端 ``${bubbleId}-${stepIndex}``;
+    - ``args`` 原样带上（write_file 的 path/content、派活的 agent_name/task 等）。
+    """
+    steps: list[dict] = []
+    for i, tc in enumerate(tool_calls or []):
+        name = tc.get("name") or ""
+        if not name:
+            continue
+        steps.append(
+            {
+                "id": f"{dialogue_id}-{i}",
+                "key": f"tool_calling:{name}",
+                "label": _STEP_LABEL.get(name, f"使用 “{name}” 工具"),
+                "ts": ts,
+                "args": tc.get("args") or {},
+            }
+        )
+    return steps
+
+
+def _seat_stub_messages(dialogues: list[dict]) -> dict[str, list[dict]]:
+    """从各对话的 write_file 调用重建 ``seatStubMessages``（threadId → [stub ai message]）。
+
+    对齐前端 ``upsertSeatToolCall`` 的 stub 形状与 id 约定（消息 ``roundtable-stub-<tid>``、
+    工具调用 ``tc-<tid>-<i>``），使加载草稿后点开产物的虚拟 URL 能正确解析到内容。
+    同一 thread 跨轮写多次按 path 去重（后写覆盖），与前端一致。
+    """
+    by_tid: dict[str, dict[str, dict]] = {}
+    for d in dialogues:
+        tid = d.get("threadId")
+        if not tid:
+            continue
+        for tc in d.get("toolCalls") or []:
+            if tc.get("name") != "write_file":
+                continue
+            path = str((tc.get("args") or {}).get("path") or "")
+            by_tid.setdefault(tid, {})[path] = tc  # 后写覆盖
+    out: dict[str, list[dict]] = {}
+    for tid, by_path in by_tid.items():
+        tool_calls = [
+            {"id": f"tc-{tid}-{i}", "name": "write_file", "args": tc.get("args") or {}}
+            for i, tc in enumerate(by_path.values())
+        ]
+        out[tid] = [{"id": f"roundtable-stub-{tid}", "type": "ai", "content": "", "tool_calls": tool_calls}]
+    return out
+
+
+def _to_step2_dialogue(d: dict, idx_by_agent: dict[str, int]) -> dict:
+    """把作业的中性对话记录映射成前端 Step2Dialogue 形状（含 avatarType / steps /
+    agentThreadId / time），使写进草稿 step2.runs 后，前端 hydrateStep2Run 能像手动跑
+    一样直接渲染——步骤卡、时间戳、点开产物全部可用。"""
+    dialogue_id = d.get("id")
+    time_str = d.get("time") or ""
+    base: dict = {
+        "id": dialogue_id,
+        "content": d.get("content") or "",
+        "time": time_str,
+        "streaming": False,
+        "steps": _steps_from_tool_calls(d.get("toolCalls") or [], dialogue_id, time_str),
+    }
+    thread_id = d.get("threadId")
+    if thread_id:
+        base["agentThreadId"] = thread_id
+    if d.get("role") == "leader":
+        return {**base, "sender": "总控协调", "avatarType": "blue", "confidence": "派活说明"}
+    aid = d.get("agentId") or ""
+    return {
+        **base,
+        "sender": d.get("name") or aid,
+        "avatarType": _avatar_for(aid, idx_by_agent.get(aid, 0)),
+        "confidence": "子智能体交付",
+    }
+
+
@dataclass
 class StartParams:
    job_id: str
@@ -52,6 +162,8 @@ class StartParams:
    # 是否为「续跑」（resume 澄清）。首启=False → 真网关建全新线程（不复用前端线程，
    # 避免与前端本地研讨的活跃 run 撞 409）；续跑=True → 复用作业自己的线程。
    is_resume: bool = False
+    # chain 模式来源链条（写进草稿 run 用，展示）。
+    chain: dict[str, Any] | None = None


 # 网关工厂：给定一次启动参数，返回一个 RoundtableGateway 实例。
@@ -94,7 +206,19 @@ class RoundtableJobExecutor:
                except Exception:  # noqa: BLE001
                    logger.warning("persist thread_ids for job %s failed", params.job_id, exc_info=True)

-            progress = JobProgressWriter(self._store, job_id=params.job_id, agents=params.agents)
+            # 每轮模型返回后，把对话作为一个带 jobId 的 run 写进草稿 step2.runs[]
+            # —— 和手动跑同一存取路径（draft_store.update_draft），加载草稿即可显示。
+            async def _on_dialogue(dialogues: list[dict], consensus: int) -> None:
+                await self._write_run_to_draft(
+                    params, dialogues=dialogues, status="running", consensus=consensus, thread_ids=thread_ids
+                )
+
+            progress = JobProgressWriter(
+                self._store,
+                job_id=params.job_id,
+                agents=params.agents,
+                on_dialogue=_on_dialogue if (self._draft_store and params.draft_id) else None,
+            )
            result = await run_orchestration(
                gateway=gateway,
                progress=progress,
@@ -106,26 +230,21 @@ class RoundtableJobExecutor:
                mode=params.mode,
                intent_text=params.intent_text,
            )
-            # 完成且有报告：
-            #  1) 把 step3 挂到 job 上 —— 前端「多运行」模型从单个 job 即可拿到完整结果
-            #     （对话 dialogues + 报告 step3）组装一个 run；
-            #  2) 同时写回关联草稿的顶层 step3 + 推进 furthest_step=3，兼容既有
-            #     「查看报告 → 加载草稿 → Step3」路径。
+            # 终态：把最终对话 + step3 + jobStatus 落进同一个 run（含 furthest_step=3）。
+            await self._write_run_to_draft(
+                params,
+                dialogues=result.dialogues,
+                status=result.status,
+                consensus=100 if result.status == "done" else 95,
+                step3=result.step3,
+                thread_ids=thread_ids,
+            )
+            # step3 也挂 job（API / 弹窗用）。
            if result.status == "done" and result.step3:
                try:
                    await self._store.update_progress(params.job_id, step3=result.step3)
                except Exception:  # noqa: BLE001
                    logger.warning("persist step3 to job %s failed", params.job_id, exc_info=True)
-                if params.draft_id and self._draft_store is not None:
-                    try:
-                        await self._draft_store.update_draft(
-                            params.draft_id,
-                            params.user_id,
-                            step3=result.step3,
-                            furthest_step=3,
-                        )
-                    except Exception:  # noqa: BLE001 — 写报告失败不影响作业本身已 done
-                        logger.warning("persist step3 to draft %s failed", params.draft_id, exc_info=True)
        except asyncio.CancelledError:
            # 干净停止即可，**不在被取消的任务里写 DB**：取消态下二次 await 会再抛
            # CancelledError、并可能打断进行中的 sqlite 连接。``cancelled`` 终态由
@@ -137,6 +256,72 @@ class RoundtableJobExecutor:
        finally:
            reset_current_user(token)

+    async def _write_run_to_draft(
+        self,
+        params: StartParams,
+        *,
+        dialogues: list[dict],
+        status: str,
+        consensus: int,
+        step3: dict | None = None,
+        thread_ids: dict[str, str] | None = None,
+    ) -> None:
+        """把本次后台作业作为一个带 jobId 的 run 写进草稿 step2.runs[]（同手动跑存取）。
+
+        读现有 step2 → 按 run_id=``job-<jobId>`` upsert（保留其它手动 run）→ 设为 activeRun
+        → update_draft。前端加载草稿即可像手动跑一样 hydrate 出对话。
+        """
+        if self._draft_store is None or not params.draft_id:
+            return
+        idx_by_agent = {a.agent_id: i for i, a in enumerate(params.agents)}
+        s2_dialogues = [_to_step2_dialogue(d, idx_by_agent) for d in dialogues]
+        last_leader = next(
+            (d.get("content") for d in reversed(dialogues) if d.get("role") == "leader"), ""
+        )
+        run_id = f"job-{params.job_id}"
+        try:
+            draft = await self._draft_store.get_draft(params.draft_id, params.user_id)
+        except Exception:  # noqa: BLE001
+            draft = None
+        step2 = dict((draft or {}).get("step2") or {})
+        runs = list(step2.get("runs") or [])
+        idx = next((i for i, r in enumerate(runs) if r.get("id") == run_id), -1)
+        created = runs[idx].get("createdAt") if idx >= 0 else datetime.now(UTC).isoformat()
+        prev_step3 = runs[idx].get("step3") if idx >= 0 else None
+        run = {
+            "id": run_id,
+            "title": "后台研讨",
+            "createdAt": created,
+            "source": params.mode,
+            "selectedAgents": [{"agent_id": a.agent_id, "name": a.name} for a in params.agents],
+            # 后台真建的 leader/seat 线程是真实持久的——存进去，前端加载后能像手动跑一样
+            # 按 thread 打开各席位产物（HTTP outputs 兜底）。hydrate 是纯展示、绝不 resume，
+            # 不会触发重跑（init 闸门 + isRoundTableRunning=false 双重拦截）。空则存 None。
+            "threadIds": (thread_ids or None),
+            "coordinatorName": params.coordinator_name,
+            "step2RoundtableDialogues": s2_dialogues,
+            "lastLeaderContent": last_leader or "",
+            "hasConsensus": status == "done",
+            "consensusPercentage": int(consensus),
+            "budgetLimit": 0,
+            # 从各席位 write_file 调用重建 stub，使加载草稿后点开产物的虚拟预览可用。
+            "seatStubMessages": _seat_stub_messages(dialogues),
+            "orchestrationMode": params.mode,
+            "chain": params.chain,
+            "step3": step3 if step3 is not None else prev_step3,
+            "jobId": params.job_id,
+            "jobStatus": status,
+        }
+        runs = [run if i == idx else r for i, r in enumerate(runs)] if idx >= 0 else [*runs, run]
+        step2["runs"] = runs
+        step2["activeRunId"] = run_id
+        try:
+            await self._draft_store.update_draft(
+                params.draft_id, params.user_id, step2=step2, furthest_step=3 if status == "done" else 2
+            )
+        except Exception:  # noqa: BLE001 — 写草稿失败不影响作业本身
+            logger.warning("write run to draft %s failed", params.draft_id, exc_info=True)
+
    def cancel(self, job_id: str) -> bool:
        """请求取消一个在跑的作业。返回是否命中在跑任务。"""
        task = self._tasks.get(job_id)
@@ -0,0 +1,94 @@
+"""圆桌各角色「单轮 run 参数」的**单一数据源**。
+
+历史问题：前端手动路径（``multi_agent.py`` 的 ``_leader_run`` / ``_special_run``）
+和后台进程内网关（``roundtable_inprocess_gateway.py``）各自硬编码了一份「每角色
+工具裁剪 + thinking + reasoning」，结果**漂移**出差异（后台席位 / 报告少禁了
+``web_search``，能联网搜索而前端不能）。
+
+本模块把这套策略收敛成唯一定义，两条路径都从这里取，保证「后台挂起」与「前端
+手动调用」传参**逐字一致**，且以后改一处即可，不再漂移。
+
+角色（``role``）：
+  - ``leader``  —— 总控派活：短决策，关 thinking + low reasoning（首字节快）；
+    只禁 ``web_search``；并配合 ``SkillStopMiddleware`` 在 ``agent_orchestration``
+    调用后停图，便于网关从 AIMessage.tool_calls 解析派活列表。
+  - ``seat``    —— 子智能体交付：开 thinking + medium reasoning；禁派活 /
+    澄清 / 展示文件 / 看图 / 联网搜索（与前端 ``_special_run`` 一致）。
+  - ``report``  —— Step3 报告绘制：在 ``seat`` 基础上**额外禁 ``bash`` /
+    ``str_replace``**：
+      * 禁 ``bash`` → 杜绝它写 Python 脚本再跑出 HTML（脚本产物落在 outputs，
+        但前端虚拟预览只抓 ``write_file`` 的 ``args.content``，抓不到脚本产物 →
+        预览停在半截 ``write_file``，正是「没写完就去生成 py 文件」的现象）；
+      * 禁 ``str_replace`` → 强制返工用整文件 ``write_file`` 重写
+        （``str_replace`` 没有 ``content`` 字段，会让前端虚拟预览读到 undefined
+        而清空沙箱）。
+"""
+
+from __future__ import annotations
+
+from dataclasses import dataclass
+
+# ── 每角色基础工具裁剪（唯一定义处）─────────────────────────────────────────────
+#
+# 注意：``seat`` / ``report`` 都包含 ``web_search`` —— 圆桌研讨里联网搜索一律禁用
+# （与前端 ``_special_run`` 的 excluded_tools 对齐）。leader 也禁 ``web_search``。
+_LEADER_EXCLUDED: list[str] = ["web_search"]
+_SEAT_EXCLUDED: list[str] = ["web_search", "ask_clarification", "present_files", "view_image", "agent_orchestration"]
+_REPORT_EXCLUDED: list[str] = [*_SEAT_EXCLUDED, "bash", "str_replace"]
+
+# leader 内禀的 skill_stop：``agent_orchestration`` 调用后停图。调用方可在此基础上
+# 再叠加自己的 skill_stop（见 ``merge_skill_stop_names``）。
+_LEADER_SKILL_STOP: list[str] = ["agent_orchestration"]
+
+
+@dataclass(frozen=True)
+class RoundtableRunPolicy:
+    """一个角色单轮 run 的归一化参数。
+
+    与 ``multi_agent._run_payload`` 的入参一一对应：``excluded_tools`` /
+    ``skill_stop_names`` / ``thinking_enabled`` / ``reasoning_effort``。
+    """
+
+    excluded_tools: list[str]
+    skill_stop_names: list[str]
+    thinking_enabled: bool
+    reasoning_effort: str
+
+
+def roundtable_run_policy(role: str) -> RoundtableRunPolicy:
+    """返回某角色（``leader`` / ``seat`` / ``report``）的单轮 run 参数。
+
+    返回值里的列表是**新副本**，调用方可安全地再 append/merge 而不污染本模块常量。
+    未知角色抛 ``ValueError``（早失败，避免悄悄退化到错误的工具面）。
+    """
+    if role == "leader":
+        return RoundtableRunPolicy(
+            excluded_tools=list(_LEADER_EXCLUDED),
+            skill_stop_names=list(_LEADER_SKILL_STOP),
+            thinking_enabled=False,
+            reasoning_effort="low",
+        )
+    if role == "seat":
+        return RoundtableRunPolicy(
+            excluded_tools=list(_SEAT_EXCLUDED),
+            skill_stop_names=[],
+            thinking_enabled=True,
+            reasoning_effort="medium",
+        )
+    if role == "report":
+        return RoundtableRunPolicy(
+            excluded_tools=list(_REPORT_EXCLUDED),
+            skill_stop_names=[],
+            thinking_enabled=True,
+            reasoning_effort="medium",
+        )
+    raise ValueError(f"unknown roundtable role: {role!r} (expected leader/seat/report)")
+
+
+def merge_skill_stop_names(policy_names: list[str], caller_names: list[str] | None) -> list[str]:
+    """把角色内禀 skill_stop 与调用方追加的 skill_stop 合并去重（保序）。
+
+    取代历史上 ``list({"agent_orchestration", *skill_stop_names})`` 的 set 写法
+    （set 顺序不确定）；这里用 dict.fromkeys 保序去重，结果确定、便于断言。
+    """
+    return list(dict.fromkeys([*policy_names, *(caller_names or [])]))
@@ -57,18 +57,20 @@ import uuid
 from typing import Any, AsyncIterator

 import httpx
+from fastapi import APIRouter, Body, HTTPException, Request
+from fastapi.responses import StreamingResponse
+from langgraph.checkpoint.base import empty_checkpoint
+
+from app.gateway.roundtable_run_policy import merge_skill_stop_names, roundtable_run_policy
 from app.gateway.routers._roundtable_seed import (
    COORDINATOR_AGENT_ID,
    REPORT_AGENT_ID,
    ensure_roundtable_functional_agents,
 )
 from deerflow.config import get_app_config
-from deerflow.tools.builtins.clarification_utils import resolve_allow_multiple
 from deerflow.runtime.user_context import get_effective_user_id
+from deerflow.tools.builtins.clarification_utils import resolve_allow_multiple
 from deerflow.utils.time import now_iso
-from fastapi import APIRouter, Body, HTTPException, Request
-from fastapi.responses import StreamingResponse
-from langgraph.checkpoint.base import empty_checkpoint

 logger = logging.getLogger(__name__)
 router = APIRouter(prefix="/api/multi-agent", tags=["multi-agent"])
@@ -1008,19 +1010,20 @@ async def _leader_run(
        _broadcast(client, base, headers, "总控智能体收到用户消息:" + new_message, agent_threads, [agent_name])
    )
    thread_id = agent_threads[agent_name]
-    effective_stop_names = list({"agent_orchestration", *skill_stop_names})
+    # 角色参数取自单一数据源 roundtable_run_policy（leader：禁 web_search、停
+    # agent_orchestration 图、关 thinking + low reasoning——短决策、首字节快）。
+    # 调用方可再叠加自己的 skill_stop。后台进程内网关用同一份 policy，保证传参一致。
+    policy = roundtable_run_policy("leader")
+    effective_stop_names = merge_skill_stop_names(policy.skill_stop_names, skill_stop_names)
    body = _run_payload(
        agent_name=agent_name,
        model_name=model_name,
        thread_id=thread_id,
        new_message=new_message,
-        excluded_tools=["web_search"],
+        excluded_tools=policy.excluded_tools,
        skill_stop_names=effective_stop_names,
-        # 总控派活是"下一步派谁、什么任务"的短决策，不需要 CoT。
-        # 关掉 thinking + low reasoning 让"调度中"几乎瞬间转入"正在派活"，
-        # 子智能体仍保留 thinking（在 _special_run 里默认 True），保证交付质量。
-        thinking_enabled=False,
-        reasoning_effort="low",
+        thinking_enabled=policy.thinking_enabled,
+        reasoning_effort=policy.reasoning_effort,
    )
    t_payload_built = time.perf_counter()
    logger.info(
@@ -1163,23 +1166,20 @@ async def _special_run(
    """
    t0 = time.perf_counter()
    thread_id = agent_threads[agent_name]
-    excluded_tools = ["web_search", "ask_clarification", "present_files", "view_image", "agent_orchestration"]
-    if agent_name == REPORT_AGENT_ID:
-        # Step 3「结果绘制」的报告智能体只需**直接 write_file 写出完整 HTML**。
-        # 额外禁用 bash / str_replace：
-        #  - 禁 bash → 杜绝它写 Python/脚本再跑出 HTML（脚本产物落在 outputs 里，
-        #    但前端虚拟预览只抓 write_file 的 args.content，抓不到脚本产物 → 预览停在
-        #    那个半截 write_file 上，正是用户看到的「没写完就去生成 py 文件」）；
-        #  - 禁 str_replace → 强制返工用整文件 write_file 重写（str_replace 没有 content
-        #    字段，会让前端虚拟预览读到 undefined 而清空沙箱）。
-        excluded_tools = [*excluded_tools, "bash", "str_replace"]
+    # 角色参数取自单一数据源 roundtable_run_policy（seat / report）。report 在 seat
+    # 基础上额外禁 bash / str_replace（理由见 roundtable_run_policy 模块注释）。
+    # 后台进程内网关用同一份 policy，保证「后台挂起」与「前端手动调用」传参一致。
+    role = "report" if agent_name == REPORT_AGENT_ID else "seat"
+    policy = roundtable_run_policy(role)
    body = _run_payload(
        agent_name=agent_name,
        model_name=model_name,
        thread_id=thread_id,
        new_message=new_message,
-        excluded_tools=excluded_tools,
-        skill_stop_names=skill_stop_names,
+        excluded_tools=policy.excluded_tools,
+        skill_stop_names=merge_skill_stop_names(policy.skill_stop_names, skill_stop_names),
+        thinking_enabled=policy.thinking_enabled,
+        reasoning_effort=policy.reasoning_effort,
    )
    t_payload_built = time.perf_counter()
    logger.info(
@@ -203,6 +203,7 @@ async def start_job(request: Request, body: StartJobRequest) -> RoundtableJobRes
            thread_ids=body.threadIds,
            coordinator_name=body.coordinatorName,
            draft_id=body.draftId,
+            chain=body.chain,
        )
    )
    return _to_response(row)
@@ -287,6 +288,7 @@ async def resume_job(request: Request, job_id: str, body: ResumeJobRequest) -> R
            coordinator_name=row.get("coordinator_name") or "roundtable-coordinator",
            draft_id=row.get("draft_id"),
            is_resume=True,
+            chain=row.get("chain"),
        )
    )
    return _to_response(updated or row)
@@ -47,6 +47,7 @@ class OrchestrationResult:
    pending_clarification: str | None = None
    error: str | None = None
    last_leader_content: str = ""
+    dialogues: list = field(default_factory=list)


@dataclass
@@ -85,12 +86,13 @@ async def run_orchestration(
      mode: "recommend" | "chain"。
      intent_text: 任务意图，用于报告 prompt（缺省退回 seed_message）。
    """
+    intent = (intent_text or seed_message or "").strip()
    try:
        await progress.begin()
        if mode == "chain":
-            outcome = await _run_chain(gateway, progress, agents, thread_ids, coordinator_name, model, seed_message)
+            outcome = await _run_chain(gateway, progress, agents, thread_ids, coordinator_name, model, seed_message, intent)
        else:
-            outcome = await _run_recommend(gateway, progress, agents, thread_ids, coordinator_name, model, seed_message)
+            outcome = await _run_recommend(gateway, progress, agents, thread_ids, coordinator_name, model, seed_message, intent)

        if outcome.kind == "clarification":
            question = outcome.question or "总控需要补充信息，请补充后继续。"
@@ -99,6 +101,7 @@ async def run_orchestration(
                status="awaiting_input",
                pending_clarification=question,
                last_leader_content=outcome.last_leader_content,
+                dialogues=progress.dialogues(),
            )

        # consensus 或 max_cycles：均收口 → 报告绘制
@@ -118,14 +121,48 @@ async def run_orchestration(
            consensus=True,
            step3=step3,
            last_leader_content=outcome.last_leader_content,
+            dialogues=progress.dialogues(),
        )
    except Exception as exc:  # noqa: BLE001 — 顶层兜底：标记 error 并返回
        logger.exception("roundtable orchestration failed")
        await progress.error(str(exc))
-        return OrchestrationResult(status="error", error=str(exc))
+        return OrchestrationResult(status="error", error=str(exc), dialogues=progress.dialogues())


-# ── recommend 模式 ───────────────────────────────────────────────────────────
+# ── 上下文「广播」（inline 实现）────────────────────────────────────────────────
+#
+# 进程内网关没有 multi_agent 的 thread 广播，席位/总控会看不到「原始任务背景 + 各席位
+# 前序交付」→ 席位反问用户、总控靠记忆瞎编。这里把这些信息**直接拼进 prompt**：
+#   - 每个席位的任务带上：任务背景 + 前序各席位交付 + 「直接产出、别提问」硬约束；
+#   - 总控的续轮 prompt 带上：各席位本轮交付，让它据实派活/收口。
+
+_SEAT_NO_ASK = (
+    "注意：这是自动化圆桌研讨，没有用户在线应答——**请勿向用户提问或等待确认**，"
+    "直接基于已知信息做出合理假设并产出完整的结构化交付（Markdown）。"
+)
+
+
+def _deliveries_block(deliveries: dict[str, str], name_by_id: dict[str, str]) -> str:
+    parts: list[str] = []
+    for aid, text in deliveries.items():
+        t = (text or "").strip()
+        if not t:
+            continue
+        parts.append(f"### {name_by_id.get(aid, aid)}\n{t}")
+    return "\n\n".join(parts)
+
+
+def _seat_message(intent_text: str, deliveries: dict[str, str], name_by_id: dict[str, str], task: str) -> str:
+    """把任务背景 + 前序交付 + 防提问约束拼进席位任务（替代 thread 广播）。"""
+    parts: list[str] = []
+    if intent_text.strip():
+        parts.append(f"【本次研讨任务背景】\n{intent_text.strip()}")
+    block = _deliveries_block(deliveries, name_by_id)
+    if block:
+        parts.append("【前序各席位已完成的交付，请在此基础上承接、补充、不要重复或推翻】\n" + block)
+    parts.append(f"【你的本轮具体任务】\n{task}")
+    parts.append(_SEAT_NO_ASK)
+    return "\n\n---\n\n".join(parts)


 async def _run_recommend(
@@ -136,6 +173,7 @@ async def _run_recommend(
    coordinator_name: str,
    model: str | None,
    seed_message: str,
+    intent_text: str,
 ) -> _LoopOutcome:
    deliveries: dict[str, str] = {}
    last_leader = ""
@@ -152,24 +190,32 @@ async def _run_recommend(
        )
        if turn.final_text:
            last_leader = turn.final_text
-        await progress.add_leader_dialogue(turn.final_text, cycle)
+        await progress.add_leader_dialogue(
+            turn.final_text, cycle, tool_calls=turn.tool_calls, thread_id=thread_ids.get(coordinator_name)
+        )

        if turn.is_clarification:
            return _LoopOutcome("clarification", question=turn.clarification, last_leader_content=last_leader, deliveries=deliveries)
        if turn.is_consensus:
            return _LoopOutcome("consensus", last_leader_content=last_leader, deliveries=deliveries)

-        # **顺序**派活 —— 不要并行
+        # **顺序**派活 —— 不要并行；每个席位带上「背景 + 前序交付」上下文
        await progress.dispatching(cycle)
        for agent_id, task in turn.dispatched:
            await progress.seat_active(agent_id)
-            seat = await gateway.run_seat(thread_ids=thread_ids, agent_id=agent_id, task=task, model=model)
+            seat_msg = _seat_message(intent_text, deliveries, name_by_id, task)
+            seat = await gateway.run_seat(thread_ids=thread_ids, agent_id=agent_id, task=seat_msg, model=model)
            deliveries[agent_id] = seat.final_text
-            await progress.add_seat_dialogue(agent_id, name_by_id.get(agent_id, agent_id), seat.final_text, cycle)
+            await progress.add_seat_dialogue(
+                agent_id, name_by_id.get(agent_id, agent_id), seat.final_text, cycle,
+                tool_calls=seat.tool_calls, thread_id=thread_ids.get(agent_id),
+            )
            await progress.seat_done(agent_id, cycle)

        await progress.bump_consensus()
-        message = CONTINUE_PROMPT
+        # 续轮：把各席位本轮交付喂给总控，让它据实判断（总控 SOUL 依赖看到交付）
+        block = _deliveries_block(deliveries, name_by_id)
+        message = CONTINUE_PROMPT + (f"\n\n各席位最新交付如下：\n{block}" if block else "")

    # 达到最大轮次：强制收口（对齐前端 toast「已达到最大研讨轮次」）
    logger.info("roundtable recommend loop hit MAX_CYCLES=%d, forcing consensus", MAX_CYCLES)
@@ -187,13 +233,15 @@ async def _run_chain(
    coordinator_name: str,
    model: str | None,
    seed_message: str,
+    intent_text: str,
 ) -> _LoopOutcome:
    deliveries: dict[str, str] = {}
    last_leader = ""
    total = len(agents)
+    name_by_id = {a.agent_id: a.name for a in agents}

    for i, seat in enumerate(agents):
-        # (a) 总控按链条顺序只给当前席位派活
+        # (a) 总控按链条顺序只给当前席位派活；带上前序各席位交付，让它据实派活
        await progress.leader_thinking(i + 1)
        leader_task = build_chain_leader_task(
            seed_message=seed_message,
@@ -203,6 +251,9 @@ async def _run_chain(
            seat_name=seat.name,
            is_first=(i == 0),
        )
+        block = _deliveries_block(deliveries, name_by_id)
+        if block:
+            leader_task += f"\n\n前序各席位已完成的交付如下，请据此为本席位分配承接性任务：\n{block}"
        turn = await gateway.run_leader(
            thread_ids=thread_ids,
            coordinator_name=coordinator_name,
@@ -211,34 +262,44 @@ async def _run_chain(
        )
        if turn.final_text:
            last_leader = turn.final_text
-        await progress.add_leader_dialogue(turn.final_text, i + 1)
+        await progress.add_leader_dialogue(
+            turn.final_text, i + 1, tool_calls=turn.tool_calls, thread_id=thread_ids.get(coordinator_name)
+        )
        if turn.is_clarification:
            return _LoopOutcome("clarification", question=turn.clarification, last_leader_content=last_leader, deliveries=deliveries)

        task = pick_chain_task(turn.dispatched, turn.final_text, seat.agent_id, seat.name)

-        # (b) 当前席位交付（顺序，不并行）
+        # (b) 当前席位交付（顺序，不并行）；带上「背景 + 前序交付 + 别提问」上下文
        await progress.seat_active(seat.agent_id)
-        seat_turn = await gateway.run_seat(thread_ids=thread_ids, agent_id=seat.agent_id, task=task, model=model)
+        seat_msg = _seat_message(intent_text, deliveries, name_by_id, task)
+        seat_turn = await gateway.run_seat(thread_ids=thread_ids, agent_id=seat.agent_id, task=seat_msg, model=model)
        deliveries[seat.agent_id] = seat_turn.final_text
-        await progress.add_seat_dialogue(seat.agent_id, seat.name, seat_turn.final_text, i + 1)
+        await progress.add_seat_dialogue(
+            seat.agent_id, seat.name, seat_turn.final_text, i + 1,
+            tool_calls=seat_turn.tool_calls, thread_id=thread_ids.get(seat.agent_id),
+        )
        await progress.seat_done(seat.agent_id, i + 1)

        if total > 0:
            await progress.set_consensus_pct(round((i + 1) / total * 90))

-    # (c) 总控收口：综合全链交付，产出最终结论
+    # (c) 总控收口：综合全链交付，产出最终结论（带上全部席位交付）
    if total > 0:
        await progress.leader_thinking(total + 1)
+        block = _deliveries_block(deliveries, name_by_id)
+        closeout_msg = CHAIN_CLOSEOUT_PROMPT + (f"\n\n各席位完整交付如下：\n{block}" if block else "")
        closeout = await gateway.run_leader(
            thread_ids=thread_ids,
            coordinator_name=coordinator_name,
-            message=CHAIN_CLOSEOUT_PROMPT,
+            message=closeout_msg,
            model=model,
        )
        if closeout.final_text:
            last_leader = closeout.final_text
-        await progress.add_leader_dialogue(closeout.final_text, total + 1)
+        await progress.add_leader_dialogue(
+            closeout.final_text, total + 1, tool_calls=closeout.tool_calls, thread_id=thread_ids.get(coordinator_name)
+        )

    return _LoopOutcome("consensus", last_leader_content=last_leader, deliveries=deliveries)

@@ -279,27 +340,43 @@ def build_report_prompt(
    席位交付按 ``agents`` 顺序列出（缺交付的席位跳过）。报告智能体据此用 write_file
    写出自包含的 ``outputs/方案总览.html``（具体绘制规则在 roundtable-report 的 SOUL）。
    """
+    # 与前端 useStep3Report.buildReportPrompt 对齐 —— 关键是**明确叫报告 agent 用
+    # write_file 写到 /mnt/user-data/outputs/方案总览.html**，否则它只会在对话里回文字、
+    # 不落 HTML 文件（_read_report_html 就读不到）。
+    name_by_id = {a.agent_id: a.name for a in agents}
    lines: list[str] = [
-        "【圆桌最终成果汇总 —— 请据此绘制方案总览 HTML 看板】",
+        "请基于以下圆桌多智能体研讨的成果，撰写一份图文并茂、以文字叙述为主线的 HTML「方案研讨汇报」，"
+        "用 write_file 工具写入 /mnt/user-data/outputs/方案总览.html。",
+        "目标：让用户能轻松读懂这群智能体讨论出了什么方案、为什么是它。请逐个席位提炼其核心观点与交付写成文字，"
+        "再综合成最终方案；图表用来点睛（风险矩阵、时间线、资源占比、贡献雷达等），详略得当——"
+        "缺明确数字时也要据研讨内容合理估算出图并标注「定性/估算」，不要通篇只有文字。",
        "",
-        "## 任务意图",
+        "# 一、任务意图",
        (intent_text or "").strip() or "（未提供）",
        "",
-        "## 各席位交付",
+        "# 二、各席位最终交付",
    ]
-    name_by_id = {a.agent_id: a.name for a in agents}
    has_delivery = False
-    for a in agents:
+    for i, a in enumerate(agents):
        text = (deliveries.get(a.agent_id) or "").strip()
        if not text:
            continue
        has_delivery = True
-        lines.append(f"### {name_by_id.get(a.agent_id, a.agent_id)}")
+        lines.append(f"## {i + 1}. {name_by_id.get(a.agent_id, a.agent_id)}")
        lines.append(text)
        lines.append("")
    if not has_delivery:
-        lines.append("（各席位暂无可用交付，请基于任务意图与总控结论绘制。）")
+        lines.append("（各席位暂无文字交付，请基于任务意图与总控结论合理展开成方案。）")
        lines.append("")
-    lines.append("## 总控最终结论")
+    lines.append("# 三、总控最终共识结论")
    lines.append((consensus_text or "").strip() or "（未提供）")
+    lines.append("")
+    lines.append(
+        "请严格遵循你的职责设定：自包含离线（纯 inline SVG 图表、禁止任何外链 CDN）、研讨叙事型结构"
+        "（结论+KPI → 背景目标 → 各席位观点 → 综合分析 → 最终共识方案 → 执行路线）、"
+        "文字为主图表点睛、浅红主题（主色 #c25555）；写完文件后只用一两句话总结，不要把 HTML 源码贴进对话。"
+    )
+    lines.append(
+        '硬性要求：① 全程使用简体中文；② HTML 必须是完整文档，<head> 最前面声明 <meta charset="utf-8">。'
+    )
    return "\n".join(lines)
@@ -10,12 +10,21 @@
 from __future__ import annotations

 import logging
+from datetime import datetime, timedelta, timezone
 from typing import Any, Awaitable, Callable

 from .types import SeatRef

 logger = logging.getLogger(__name__)

+# 北京时区（UTC+8，固定无 DST）——对话气泡时间戳用，对齐前端 nowTime()
+# = new Date().toLocaleTimeString("zh-CN", {hour12:false}) 的 "HH:MM:SS" 形状。
+_BEIJING_TZ = timezone(timedelta(hours=8))
+
+
+def _now_hms() -> str:
+    return datetime.now(_BEIJING_TZ).strftime("%H:%M:%S")
+
 # 节点状态机：pending → active → done。
 _PENDING = "pending"
 _ACTIVE = "active"
@@ -42,11 +51,15 @@ class JobProgressWriter:
        job_id: str,
        agents: list[SeatRef],
        event_sink: EventSink = None,
+        on_dialogue: Any | None = None,
    ) -> None:
        self._store = store
        self._job_id = job_id
        self._agents = list(agents)
        self._event_sink = event_sink
+        # 每轮 leader/seat 模型返回后回调（dialogues, consensus）—— 执行器据此把对话
+        # 写进草稿（和手动跑同一存取路径）。
+        self._on_dialogue = on_dialogue
        self._state: dict[str, str] = {a.agent_id: _PENDING for a in agents}
        self._spoke_cycle: dict[str, int] = {}
        self._consensus = 0
@@ -147,20 +160,75 @@ class JobProgressWriter:

    # ── 对话记录 ─────────────────────────────────────────────────────────────

-    def _add(self, role: str, agent_id: str | None, name: str, content: str, cycle: int) -> None:
+    def _add(
+        self,
+        role: str,
+        agent_id: str | None,
+        name: str,
+        content: str,
+        cycle: int,
+        *,
+        tool_calls: list[dict] | None = None,
+        thread_id: str | None = None,
+    ) -> None:
        self._dlg_seq += 1
        self._dialogues.append(
-            {"id": self._dlg_seq, "role": role, "agentId": agent_id, "name": name, "content": content, "cycle": cycle}
+            {
+                "id": self._dlg_seq,
+                "role": role,
+                "agentId": agent_id,
+                "name": name,
+                "content": content,
+                "cycle": cycle,
+                # 工具步骤卡数据源（每项 {name, args, id}）；映射成前端 Step2Dialogue.steps。
+                "toolCalls": list(tool_calls or []),
+                # 该气泡对应的 leader/seat 线程——映射成 agentThreadId，供点开产物 / 反查 stub。
+                "threadId": thread_id,
+                # 气泡时间戳，**创建时一次性盖章**（不随后续 flush 抖动），对齐前端 nowTime()。
+                "time": _now_hms(),
+            }
        )

-    async def add_leader_dialogue(self, content: str, cycle: int) -> None:
-        """总控一轮发言（派活说明 / 共识结论）。空内容跳过。"""
+    def dialogues(self) -> list[dict[str, Any]]:
+        return list(self._dialogues)
+
+    async def _flush_dialogue(self) -> None:
+        if self._on_dialogue is None:
+            return
+        try:
+            await self._on_dialogue(self.dialogues(), self._consensus)
+        except Exception:  # noqa: BLE001 — 写草稿失败不打断编排
+            logger.warning("roundtable job dialogue → draft flush failed", exc_info=True)
+
+    async def add_leader_dialogue(
+        self,
+        content: str,
+        cycle: int,
+        *,
+        tool_calls: list[dict] | None = None,
+        thread_id: str | None = None,
+    ) -> None:
+        """总控一轮发言（派活说明 / 共识结论）。空内容跳过。
+
+        ``tool_calls`` / ``thread_id`` 用于重建工具步骤卡（派活 chip）+ agentThreadId。
+        """
        if not (content or "").strip():
            return
-        self._add("leader", None, "总控协调", content, cycle)
+        self._add("leader", None, "总控协调", content, cycle, tool_calls=tool_calls, thread_id=thread_id)
        await self._persist()
+        await self._flush_dialogue()

-    async def add_seat_dialogue(self, agent_id: str, name: str, content: str, cycle: int) -> None:
-        """某席位一轮交付。"""
-        self._add("seat", agent_id, name, content or "", cycle)
+    async def add_seat_dialogue(
+        self,
+        agent_id: str,
+        name: str,
+        content: str,
+        cycle: int,
+        *,
+        tool_calls: list[dict] | None = None,
+        thread_id: str | None = None,
+    ) -> None:
+        """某席位一轮交付。``tool_calls`` / ``thread_id`` 用于步骤卡 + 产物预览。"""
+        self._add("seat", agent_id, name, content or "", cycle, tool_calls=tool_calls, thread_id=thread_id)
        await self._persist()
+        await self._flush_dialogue()
@@ -52,6 +52,9 @@ class LeaderTurn:
    dispatched: list[tuple[str, str]] = field(default_factory=list)
    # 命中澄清时的问题文本；非 None 即 clarification 分支。
    clarification: str | None = None
+    # 本轮总控调用的工具（按出现顺序，每项 {name, args, id}）——通常是 agent_orchestration
+    # 派活 / ask_clarification。用于在草稿 run 里重建「工具步骤卡」，与前端手动跑一致。
+    tool_calls: list[dict] = field(default_factory=list)

    @property
    def is_clarification(self) -> bool:
@@ -67,6 +70,9 @@ class SeatTurn:
    """某席位一轮交付的归一化结果（由 gateway.run_seat 产出）。"""

    final_text: str = ""
+    # 本轮席位调用的工具（按出现顺序，每项 {name, args, id}）——用于在草稿 run 里
+    # 重建「工具步骤卡」+ write_file stub（点开产物），与前端手动跑一致。
+    tool_calls: list[dict] = field(default_factory=list)


@dataclass