diff --git a/capabilities.mjs b/capabilities.mjs
index 6be04ae..8c73781 100644
--- a/capabilities.mjs
+++ b/capabilities.mjs
@@ -233,7 +233,7 @@ export function sandboxDeveloperTools(capabilities) {
 export function sandboxMcpTools(capabilities) {
   const tools = [];
   if (capabilities.static_publish) {
-    tools.push('read_file', 'write_file', 'edit_file', 'create_dir', 'generate_long_image');
+    tools.push('read_file', 'write_file', 'edit_file', 'create_dir', 'generate_docx', 'generate_long_image');
     if (capabilities.shell || capabilities.code_browse) tools.push('list_dir');
   }
   if (capabilities.private_data_space) {
diff --git a/capabilities.test.mjs b/capabilities.test.mjs
index d00d73c..a5a11b9 100644
--- a/capabilities.test.mjs
+++ b/capabilities.test.mjs
@@ -194,6 +194,7 @@ test('sandboxMcpTools returns correct tool list based on capabilities', () => {
     'write_file',
     'edit_file',
     'create_dir',
+    'generate_docx',
     'generate_long_image',
     'private_data_info',
     'private_data_schema',
@@ -257,6 +258,7 @@ test('static_publish with sandboxMcp uses stdio sandbox-fs extension instead of
   assert.equal(sandboxExt.args[1], '/opt/h5/MindSpace/abc123'); // also passed as argv[2]
   assert.ok(sandboxExt.available_tools.includes('write_file'));
   assert.ok(sandboxExt.available_tools.includes('read_file'));
+  assert.ok(sandboxExt.available_tools.includes('generate_docx'));
   assert.ok(sandboxExt.available_tools.includes('generate_long_image'));
 
   // built-in developer extension should only remain for read_image (image_read: true by default)
diff --git a/mindspace-sandbox-mcp.mjs b/mindspace-sandbox-mcp.mjs
index 0aad2f2..8152891 100644
--- a/mindspace-sandbox-mcp.mjs
+++ b/mindspace-sandbox-mcp.mjs
@@ -11,6 +11,7 @@
 import path from 'node:path';
 import fs from 'node:fs';
 import readline from 'node:readline';
+import { fileURLToPath } from 'node:url';
 import { execFileSync } from 'node:child_process';
 import mysql from 'mysql2/promise';
 import { createScheduleService } from './schedule-service.mjs';
@@ -24,6 +25,7 @@ if (!SANDBOX_ROOT) {
 }
 
 const SANDBOX = path.resolve(SANDBOX_ROOT);
+const SANDBOX_MODULE_DIR = path.dirname(fileURLToPath(import.meta.url));
 const PRIVATE_DATA_DIR = path.join(SANDBOX, '.mindspace');
 const PRIVATE_DATA_DB = path.join(PRIVATE_DATA_DIR, 'private-data.sqlite');
 const SQLITE_BIN = process.env.SQLITE_BIN?.trim() || 'sqlite3';
@@ -48,6 +50,70 @@ function resolveSandboxed(p) {
   return resolved;
 }
 
+function resolveDocxGenerateScript() {
+  const candidates = [
+    path.join(SANDBOX, '.agents', 'skills', 'docx-generate', 'generate_docx.py'),
+    path.join(SANDBOX_MODULE_DIR, 'skills', 'docx-generate', 'generate_docx.py'),
+  ];
+  for (const candidate of candidates) {
+    if (fs.existsSync(candidate) && fs.statSync(candidate).isFile()) {
+      return candidate;
+    }
+  }
+  throw new Error(
+    'generate_docx: 未找到 generate_docx.py，请先 load_skill → docx-generate 同步技能到工作区',
+  );
+}
+
+function normalizeDocxSections(sections) {
+  if (!Array.isArray(sections) || sections.length === 0) {
+    throw new Error('generate_docx: sections 必须是非空数组');
+  }
+  return sections.map((section, index) => {
+    if (!section || typeof section !== 'object') {
+      throw new Error(`generate_docx: sections[${index}] 必须是对象`);
+    }
+    const normalized = {
+      heading: section.heading != null ? String(section.heading) : '',
+      paragraphs: Array.isArray(section.paragraphs)
+        ? section.paragraphs.map((paragraph) => String(paragraph))
+        : [],
+    };
+    if (section.table && typeof section.table === 'object') {
+      normalized.table = section.table;
+    }
+    return normalized;
+  });
+}
+
+function runGenerateDocxScript({ outputPath, title, sections }) {
+  const script = resolveDocxGenerateScript();
+  const python = process.env.PYTHON_BIN?.trim() || 'python3';
+  const payload = JSON.stringify({
+    title: String(title ?? ''),
+    sections: normalizeDocxSections(sections),
+  });
+  const stdout = execFileSync(
+    python,
+    [script, '--json', '-', '--output', outputPath],
+    {
+      cwd: SANDBOX,
+      input: payload,
+      encoding: 'utf8',
+      maxBuffer: 10 * 1024 * 1024,
+    },
+  );
+  const abs = resolveSandboxed(outputPath);
+  if (!fs.existsSync(abs) || !fs.statSync(abs).isFile()) {
+    throw new Error(`generate_docx: 脚本执行后未找到 ${outputPath}`);
+  }
+  const size = fs.statSync(abs).size;
+  if (size < 64) {
+    throw new Error(`generate_docx: ${outputPath} 体积异常（${size} 字节）`);
+  }
+  return { bytes: size, stdout: String(stdout ?? '').trim() };
+}
+
 const ALL_TOOLS = [
   {
     name: 'read_file',
@@ -122,6 +188,24 @@ const ALL_TOOLS = [
       required: ['html_path'],
     },
   },
+  {
+    name: 'generate_docx',
+    description:
+      '生成 Word .docx 文件并落盘到工作区（如 public/报告.docx）。公网下载页必须先调用本工具确认文件存在，再写 HTML 相对链接。',
+    inputSchema: {
+      type: 'object',
+      properties: {
+        output_path: { type: 'string', description: '输出 .docx 路径，如 public/协和智慧门诊研究摘要.docx' },
+        title: { type: 'string', description: '文档主标题' },
+        sections: {
+          type: 'array',
+          description: '章节数组；每项可含 heading、paragraphs、可选 table(headers/rows)',
+          items: { type: 'object' },
+        },
+      },
+      required: ['output_path', 'title', 'sections'],
+    },
+  },
   {
     name: 'private_data_info',
     description:
@@ -491,6 +575,25 @@ async function callTool(name, args) {
         },
       ];
     }
+    case 'generate_docx': {
+      const outputPath = String(args.output_path ?? args.path ?? '').trim();
+      if (!outputPath.toLowerCase().endsWith('.docx')) {
+        throw new Error('generate_docx: output_path 必须是 .docx 文件');
+      }
+      resolveSandboxed(outputPath);
+      const title = String(args.title ?? '').trim();
+      if (!title) {
+        throw new Error('generate_docx: title 不能为空');
+      }
+      const sections = args.sections ?? args.payload?.sections;
+      const result = runGenerateDocxScript({ outputPath, title, sections });
+      return [
+        {
+          type: 'text',
+          text: `已生成 ${outputPath}（${result.bytes} 字节）`,
+        },
+      ];
+    }
     case 'private_data_info': {
       ensurePrivateDataDb();
       const size = privateDataSize();
diff --git a/mindspace-sandbox-mcp.test.mjs b/mindspace-sandbox-mcp.test.mjs
index f9676e6..f1e43f2 100644
--- a/mindspace-sandbox-mcp.test.mjs
+++ b/mindspace-sandbox-mcp.test.mjs
@@ -4,6 +4,49 @@ import fs from 'node:fs';
 import os from 'node:os';
 import path from 'node:path';
 import { spawn } from 'node:child_process';
+import { syncPublicDocxDownloads } from './mindspace-public-finish-sync.mjs';
+
+function copyDocxGenerateSkill(root) {
+  const srcDir = path.join(process.cwd(), 'skills', 'docx-generate');
+  const destDir = path.join(root, '.agents', 'skills', 'docx-generate');
+  fs.mkdirSync(destDir, { recursive: true });
+  for (const name of ['generate_docx.py', 'SKILL.md']) {
+    fs.copyFileSync(path.join(srcDir, name), path.join(destDir, name));
+  }
+}
+
+function summarySections(caseName) {
+  return [
+    {
+      heading: '一、项目概述',
+      paragraphs: [
+        `${caseName}项目面向医疗服务流程中的核心痛点，通过数字化与 AI 能力重构诊前、诊中、诊后体验。`,
+        '本摘要基于创新三角评估框架，从创新、成果、市场、团队四个维度提炼关键结论。',
+      ],
+    },
+    {
+      heading: '二、解决方案与技术创新',
+      paragraphs: [
+        '项目以数据驱动和智能辅助决策为主线，形成可复制的技术体系与流程再造方案。',
+        '在技术创新、商业模式与政策吻合度方面均具备行业示范价值。',
+      ],
+    },
+    {
+      heading: '三、成果与价值',
+      paragraphs: [
+        '核心指标显示候诊效率、临床质量或安全水平获得显著提升，患者与医护双侧受益。',
+        '项目已具备向医联体与区域平台输出的标准化能力。',
+      ],
+    },
+    {
+      heading: '四、壁垒、风险与未来展望',
+      paragraphs: [
+        '竞争壁垒来自临床数据积累、流程嵌入深度与专科Know-how。',
+        '未来可在专科深化、基层赋能与支付模式创新方向持续扩展。',
+      ],
+    },
+  ];
+}
 
 function startSandbox(root, envOverrides = {}) {
   const child = spawn(process.execPath, ['mindspace-sandbox-mcp.mjs', root], {
@@ -102,3 +145,75 @@ test('sandbox MCP exposes schedule tools only when schedule env is configured',
     ['schedule_create_item', 'schedule_create_reminder', 'schedule_list_items'],
   );
 });
+
+test('generate_docx writes public Word files for Mark-style summary request', async (t) => {
+  const root = fs.mkdtempSync(path.join(os.tmpdir(), 'mindspace-sandbox-docx-'));
+  fs.mkdirSync(path.join(root, 'public'), { recursive: true });
+  copyDocxGenerateSkill(root);
+
+  const server = startSandbox(root, {
+    ALLOWED_TOOLS: 'generate_docx,write_file,list_dir',
+  });
+  t.after(() => server.child.kill());
+
+  await server.request('initialize');
+  const listed = await server.request('tools/list');
+  assert.ok(listed.result.tools.some((tool) => tool.name === 'generate_docx'));
+
+  const cases = [
+    {
+      docx: 'public/协和智慧门诊研究摘要.docx',
+      html: 'public/协和智慧门诊研究摘要.html',
+      title: '北京协和医院「智慧门诊 + AI 预问诊」研究摘要',
+    },
+    {
+      docx: 'public/广州妇儿CDSS研究摘要.docx',
+      html: 'public/广州妇儿CDSS研究摘要.html',
+      title: '广州妇儿中心「数据驱动 + AI 临床决策」研究摘要',
+    },
+  ];
+
+  for (const item of cases) {
+    const generated = await server.request('tools/call', {
+      name: 'generate_docx',
+      arguments: {
+        output_path: item.docx,
+        title: item.title,
+        sections: summarySections(item.title),
+      },
+    });
+    assert.equal(generated.result.isError, false, generated.result.content?.[0]?.text);
+    assert.match(generated.result.content[0].text, /已生成 public\/.+\.docx/);
+    assert.ok(fs.statSync(path.join(root, item.docx)).size > 500);
+  }
+
+  for (const item of cases) {
+    const docxName = path.basename(item.docx);
+    await server.request('tools/call', {
+      name: 'write_file',
+      arguments: {
+        path: item.html,
+        content: `<!doctype html><html><body><a href="${docxName}" download>下载 Word</a></body></html>`,
+      },
+    });
+  }
+
+  const listedPublic = await server.request('tools/call', {
+    name: 'list_dir',
+    arguments: { path: 'public' },
+  });
+  assert.equal(listedPublic.result.isError, false);
+  assert.match(listedPublic.result.content[0].text, /协和智慧门诊研究摘要\.docx/);
+  assert.match(listedPublic.result.content[0].text, /广州妇儿CDSS研究摘要\.docx/);
+
+  const sync = syncPublicDocxDownloads({ publishDir: root });
+  assert.deepEqual(sync.missing, []);
+  assert.deepEqual(sync.synced, []);
+
+  for (const item of cases) {
+    const abs = path.join(root, item.docx);
+    assert.ok(fs.existsSync(abs));
+    const zipHeader = fs.readFileSync(abs).subarray(0, 2).toString('utf8');
+    assert.equal(zipHeader, 'PK', `${item.docx} should be a valid zip/docx`);
+  }
+});
diff --git a/skills/docx-generate/SKILL.md b/skills/docx-generate/SKILL.md
index ea2a3e6..52e51f5 100644
--- a/skills/docx-generate/SKILL.md
+++ b/skills/docx-generate/SKILL.md
@@ -19,7 +19,17 @@ description: 在工作区内用 Python 标准库（zipfile + XML）生成 Word .
 - 用户要 Word / docx / .doc 文档（输出 `.docx`）
 - 需要保存到 `oa/`、`private/`、`public/` 等分区
 
-## 推荐命令
+## 推荐方式（优先）
+
+**优先调用 sandbox-fs 的 `generate_docx` 工具**（与 `generate_long_image` 同级），直接写入 `public/文件名.docx` 或 `oa/文件名.docx`：
+
+- `output_path`：如 `public/协和智慧门诊研究摘要.docx`
+- `title`：文档标题
+- `sections`：章节数组（`heading`、`paragraphs`、可选 `table`）
+
+生成后必须 `list_dir public/` 确认目标文件已落盘，再写 HTML 下载链接。
+
+## 备选命令（仅当 MCP 不可用时）
 
 技能目录内有 `generate_docx.py`（仅依赖 Python 3 标准库）：