diff --git a/.github/workflows/desktop-packages.yml b/.github/workflows/desktop-packages.yml
index 7aee76e..a44e79b 100644
--- a/.github/workflows/desktop-packages.yml
+++ b/.github/workflows/desktop-packages.yml
@@ -165,6 +165,43 @@ jobs:
             dist/filecollector-windows-*-x64.zip.sha256
           if-no-files-found: error
 
+  # 可选的 LibreOffice 转换器包 (Office→PDF 视觉转换链路): 主包体积不变,
+  # 用户按需下载解压到便携包根目录。LO_VERSION 固定于脚本内, 升级时手动更新。
+  # PR 不构建 (350MB 下载纯耗时); main/tag 上构建, tag 时随 Release 发布。
+  converter-package:
+    name: Build Windows LibreOffice converter package
+    runs-on: windows-latest
+    if: github.event_name != 'pull_request'
+    defaults:
+      run:
+        shell: msys2 {0}
+    steps:
+      - name: Check out source
+        uses: actions/checkout@v4
+
+      - name: Set up MSYS2 UCRT64
+        uses: msys2/setup-msys2@v2
+        with:
+          msystem: UCRT64
+          update: false
+          install: >-
+            unzip
+            zip
+
+      - name: Build LibreOffice converter package
+        run: |
+          chmod +x tools/build-converter-package.sh
+          tools/build-converter-package.sh
+
+      - name: Upload converter package
+        uses: actions/upload-artifact@v4
+        with:
+          name: filecollector-windows-converter-x64
+          path: |
+            dist/filecollector-converter-windows-*-x64.zip
+            dist/filecollector-converter-windows-*-x64.zip.sha256
+          if-no-files-found: error
+
   macos:
     name: Build macOS app bundle (ARM64)
     runs-on: macos-latest
@@ -188,7 +225,7 @@ jobs:
 
   publish:
     name: Publish desktop packages
-    needs: [lint, deb, flatpak, windows, macos]
+    needs: [lint, deb, flatpak, windows, macos, converter-package]
     if: startsWith(github.ref, 'refs/tags/v')
     runs-on: ubuntu-24.04
     permissions:
@@ -216,6 +253,12 @@ jobs:
           name: filecollector-windows-x64
           path: release-assets
 
+      - name: Download Windows converter package
+        uses: actions/download-artifact@v4
+        with:
+          name: filecollector-windows-converter-x64
+          path: release-assets
+
       - name: Download macOS package
         uses: actions/download-artifact@v4
         with:
@@ -228,8 +271,9 @@ jobs:
           test "$(find release-assets -maxdepth 1 -name '*.deb' -type f | wc -l)" -eq 1
           test "$(find release-assets -maxdepth 1 -name '*.flatpak' -type f | wc -l)" -eq 1
           test "$(find release-assets -maxdepth 1 -name 'filecollector-windows-*-x64.zip' -type f | wc -l)" -eq 1
+          test "$(find release-assets -maxdepth 1 -name 'filecollector-converter-windows-*-x64.zip' -type f | wc -l)" -eq 1
           test "$(find release-assets -maxdepth 1 -name 'filecollector-macos-*-arm64.zip' -type f | wc -l)" -eq 1
-          test "$(find release-assets -maxdepth 1 -name '*.sha256' -type f | wc -l)" -eq 3
+          test "$(find release-assets -maxdepth 1 -name '*.sha256' -type f | wc -l)" -eq 4
 
       - name: Create or update GitHub Release
         uses: softprops/action-gh-release@v2
diff --git a/README.md b/README.md
index 372e291..c1c0d46 100644
--- a/README.md
+++ b/README.md
@@ -324,6 +324,14 @@ FileCollector 支持自动将二进制文件转换为 Markdown 格式，无需
 - **文档文件**（PDF、DOCX、PPTX、XLSX、ODT、ODP、ODS、RTF 等）：先通过 LibreOffice 转换为 PDF，再通过 `pdftoppm` 渲染为图片序列，逐页发送给 VLM。
 - **转换缓存**：转换结果缓存在工作目录下的 `.filecollector_cache/` 目录中，基于文件 SHA256 哈希判断是否需要重新转换，避免重复处理。
 - **可配置扩展名**：在 AI 设置对话框中可自定义允许被 VLM 处理的二进制文件扩展名列表，修改后自动重新评估预处理队列。
+- **失败原因透明**：转换失败时队列行与预览区会显示具体原因（如缺少 LibreOffice），不再只提示"转换失败"。
+
+### 处理方式：视觉转换 / 文本抽取
+
+在 **AI 设置 → VLM** 的 **处理方式** 下拉中可选两种二进制文档处理流程：
+
+- **视觉转换**（默认）：文档 → PDF → 页面图片 → VLM/OCR，保留图表、公式、表格等视觉结构。Office 文档需要 LibreOffice —— Windows 用户可下载独立的**转换器包**（Release 附件 `filecollector-converter-windows-x64-*.zip`），解压到便携包根目录（与 `bin/` 同级）即可，应用会自动探测 `tools/LibreOffice/program/soffice.exe`；已安装官方 LibreOffice 的用户无需转换器包，应用同样自动探测。
+- **文本抽取**：纯本地解析 DOCX/XLSX/PPTX/ODT/ODP/ODS（ZIP+XML 结构）与 PDF（内置 `pdftotext`），**无需联网、无需 AI 凭据**，秒级完成；代价是丢失图表、公式与嵌入图片，仅适合纯文字型文档。扫描版 PDF 抽不出文本时自动回退视觉转换流程。两种模式的结果在缓存中分别标记，互不污染，切换模式后对已有条目点"重新转换"即可按新模式重跑。
 
 ### VLM 服务商
 
diff --git a/README_EN.md b/README_EN.md
index 69f81ba..97ad9ae 100644
--- a/README_EN.md
+++ b/README_EN.md
@@ -324,6 +324,14 @@ FileCollector can automatically convert binary files into Markdown format, elimi
 - **Document files** (PDF, DOCX, PPTX, XLSX, ODT, ODP, ODS, RTF, etc.): First converted to PDF via LibreOffice, then rendered as image sequences via `pdftoppm`, and sent page-by-page to VLM.
 - **Conversion cache**: Converted results are cached in the `.filecollector_cache/` directory under the working directory. The system uses SHA256 file hashes to determine whether re-conversion is needed, avoiding redundant processing.
 - **Configurable extensions**: In the AI Settings dialog, you can customize the list of binary file extensions allowed for VLM processing. Changes automatically trigger re-evaluation of the preprocessing queue.
+- **Transparent failure reasons**: When conversion fails, the queue row and preview area show the concrete reason (e.g., missing LibreOffice) instead of a bare "conversion failed".
+
+### Processing mode: Visual conversion / Text extraction
+
+Choose how binary documents are processed via the **Processing Mode** dropdown in **AI Settings → VLM**:
+
+- **Visual conversion** (default): document → PDF → page images → VLM/OCR, preserving charts, formulas, and table layouts. Office documents need LibreOffice — Windows users can download the standalone **converter package** (Release asset `filecollector-converter-windows-x64-*.zip`) and extract it into the portable package root (alongside `bin/`); the app auto-detects `tools/LibreOffice/program/soffice.exe`. If the official LibreOffice is already installed, no converter package is needed — it is auto-detected as well.
+- **Text extraction**: Pure local parsing of DOCX/XLSX/PPTX/ODT/ODP/ODS (ZIP+XML) and PDF (built-in `pdftotext`) — **no network, no AI credentials required**, finishes in seconds; the trade-off is losing charts, formulas, and embedded images, so it suits text-only documents. Scanned PDFs that yield no text automatically fall back to the visual conversion flow. Results of both modes are tagged separately in the cache; after switching modes, click "retry" on existing items to re-run with the new mode.
 
 ### VLM Providers
 
diff --git a/docs/USAGE.md b/docs/USAGE.md
index 5741851..26fca26 100644
--- a/docs/USAGE.md
+++ b/docs/USAGE.md
@@ -121,6 +121,8 @@
 
 「设置 → AI 设置」中的 **AI 助手（侧边栏）** 与 **视觉语言模型（VLM）** 均支持**多模型配置方案（Model Profile）**：可同时预置多套模型配置，各自通过下拉菜单一键切换。
 
+VLM 顶部还有 **处理方式** 下拉，决定二进制文档走哪条管线：**视觉转换**（默认，文档渲染为图片后送 VLM/OCR，保留图表/公式/表格；Office 文件需要 LibreOffice，Windows 可下载 Release 里的转换器包解压到便携包根目录）或 **文本抽取**（本地直接解析 Office/PDF，快速离线且无需 AI 凭据，但丢失图表/公式/嵌入图片；扫描版 PDF 自动回退视觉转换）。两种模式的缓存结果分开标记，切换后对已有条目点"重新转换"即可按新模式重跑。
+
 **配置文件位置**：`~/.config/filecollector/settings.json`
 
 #### 配置文件格式
diff --git a/docs/USAGE_EN.md b/docs/USAGE_EN.md
index db10004..7b383ba 100644
--- a/docs/USAGE_EN.md
+++ b/docs/USAGE_EN.md
@@ -120,6 +120,8 @@ After selecting a working directory, use the search box at the top of the left-s
 
 Both **AI Assistant (Sidebar)** and **Vision-Language Model (VLM)** in "Settings → AI Settings" support **Model Profiles**: you can pre-configure multiple model setups and switch between them instantly from a dropdown menu.
 
+The VLM section also has a **Processing Mode** dropdown that selects the pipeline for binary documents: **Visual conversion** (default — documents are rendered to images and sent to VLM/OCR, keeping charts/formulas/tables; Office files need LibreOffice — Windows users can download the converter package from Release assets and extract it into the portable package root) or **Text extraction** (parses Office/PDF locally — fast, offline, no AI credentials required, but charts/formulas/embedded images are lost; scanned PDFs automatically fall back to visual conversion). Cached results are tagged per mode; after switching, click "retry" on existing items to re-run with the new mode.
+
 **Config file location**: `~/.config/filecollector/settings.json`
 
 #### Config file format
diff --git a/docs/building/windows.md b/docs/building/windows.md
index 8e6ec8f..f5fe29a 100644
--- a/docs/building/windows.md
+++ b/docs/building/windows.md
@@ -34,7 +34,13 @@ chmod +x tools/build-windows-package.sh
 tools/build-windows-package.sh
 ```
 
-脚本会下载并静态编译 cmark-gfm，运行 Meson 单元测试，递归收集 UCRT64 运行时 DLL（含 GTK 图像加载器的间接依赖）、GTK 动态模块、GSettings schema、完整 Adwaita/hicolor 图标主题，以及项目实际使用的两个 XApp symbolic 图标。然后它会验证所有源码引用的 symbolic 图标都能在包内主题中解析，并在打包前后用 `tools/verify_windows_package.py` 对全部 PE 文件做导入表闭包校验；运行时 DLL 数量不足或存在未解析的非系统 DLL 时，构建直接失败。
+脚本会下载并静态编译 cmark-gfm，运行 Meson 单元测试，递归收集 UCRT64 运行时 DLL（含 GTK 图像加载器的间接依赖）、GTK 动态模块、GSettings schema、完整 Adwaita/hicolor 图标主题、zh_CN 翻译文件（.mo）、项目实际使用的两个 XApp symbolic 图标，以及 ZIP 导出依赖的 zip.exe（含 msys-2.0.dll）。然后它会验证所有源码引用的 symbolic 图标都能在包内主题中解析，并在打包前后用 `tools/verify_windows_package.py` 对全部 PE 文件做导入表闭包校验；运行时 DLL 数量不足或存在未解析的非系统 DLL 时，构建直接失败。
+
+> 注意：locale 与 zip.exe 曾因 MSYS2 参数路径转换（`meson --prefix=/usr` 的
+> 参数被改写为 MSYS2 安装根下的路径）而静默缺失，导致 v4.9.1 及之前的
+> Windows 包为纯英文界面且 ZIP 导出不可用。脚本现通过 `MSYS2_ARG_CONV_EXCL`
+> 排除转换，并在打包后硬性校验 `locale/zh_CN/LC_MESSAGES/filecollector.mo`
+> 与 `bin/zip.exe` 存在于 ZIP 内。
 
 排错提示：DLL 收集脚本按 `MSYSTEM_PREFIX` 判定运行时目录，并用 `ldd` 递归解析依赖。若本地构建得到的 `bin/` 里几乎没有 DLL，先确认是在 MSYS2 **UCRT64** 终端中执行（`echo $MSYSTEM_PREFIX` 应为 `/ucrt64`），以及 `ldd staging/bin/filecollector.exe` 能正常解析。
 
@@ -45,6 +51,25 @@ dist/filecollector-windows-<meson.build 中的版本>-x64.zip
 dist/filecollector-windows-<meson.build 中的版本>-x64.zip.sha256
 ```
 
+## LibreOffice 转换器包（可选）
+
+主包不捆绑 LibreOffice（体积原因）。Office 文档的**视觉转换**链路（soffice → PDF → pdftoppm → VLM）依赖 LibreOffice，通过独立的转换器包交付：
+
+```bash
+tools/build-converter-package.sh
+```
+
+脚本下载官方 LibreOffice MSI（版本固定于脚本内 `LO_VERSION`），用 `msiexec /a` 管理员解包（不写注册表、不进 Program Files），打包为：
+
+```text
+dist/filecollector-converter-windows-<LO_VERSION>-x64.zip
+dist/filecollector-converter-windows-<LO_VERSION>-x64.zip.sha256
+```
+
+布局为 `tools/LibreOffice/program/soffice.exe`，解压到便携包根目录（与 `bin/` 同级）即生效；应用通过 `BinaryConverter.resolve_soffice()` 自动探测。CI 中该包在 main/tag 推送时构建，tag 推送时随 Release 发布。
+
+注意：`msiexec` 的 `/a` 参数从 bash 直接传入会被 MSYS2 改写为路径（与 `--prefix=/usr` 同类陷阱），脚本通过 `cmd //c` 复合字符串绕过。
+
 ## 试运行
 
 将 ZIP 解压后，始终通过下列启动器启动：
diff --git a/meson.build b/meson.build
index 262a307..d909bc9 100644
--- a/meson.build
+++ b/meson.build
@@ -128,6 +128,7 @@ sources = files(
   'src/services/binary_converter.vala',
   'src/services/token_estimator.vala',
   'src/services/binary_preprocessor.vala',
+  'src/services/office_text_extractor.vala',
   'src/services/config_manager.vala',
   'src/services/file_content_reader.vala',
   'src/services/file_generator.vala',
@@ -163,8 +164,10 @@ sources = files(
 )
 
 # win32_dpapi_shim.c 仅 Windows 构建需要 (DPAPI 封装), 其它平台不存在该文件。
+# win32_console_shim.c 仅 Windows 构建需要 (GUI 子系统下的 CLI 控制台接管)。
 if host_machine.system() == 'windows'
   sources += files('src/win32_dpapi_shim.c')
+  sources += files('src/win32_console_shim.c')
 endif
 
 # macos_keychain_shim.c 仅 macOS 构建需要 (SecItem* 封装), 其它平台不存在该文件。
diff --git a/po/POTFILES b/po/POTFILES
index 69cdfb9..5a8a1c6 100644
--- a/po/POTFILES
+++ b/po/POTFILES
@@ -13,6 +13,7 @@ src/services/ai_client.vala
 src/services/ai_types.vala
 src/services/binary_converter.vala
 src/services/binary_preprocessor.vala
+src/services/office_text_extractor.vala
 src/services/config_manager.vala
 src/services/file_content_reader.vala
 src/services/file_generator.vala
diff --git a/po/zh_CN.po b/po/zh_CN.po
index 3a9b1eb..9ec27e0 100644
--- a/po/zh_CN.po
+++ b/po/zh_CN.po
@@ -2693,3 +2693,135 @@ msgstr "正在预处理 0/0 个文件..."
 #: src/widgets/vlm_queue_controller.vala:146
 msgid "Cancel All"
 msgstr "取消全部"
+
+# ── 二进制预处理: 双模式 + 失败原因透出 (2026-11 新增) ──
+# 注: 本批次由人工补录 (本机无 msgmerge), 后续重新生成 .pot 时请逐条核对
+
+#: src/services/binary_converter.vala
+msgid "Failed to create tmp dir for PDF conversion"
+msgstr "创建 PDF 转换临时目录失败"
+
+#: src/services/binary_converter.vala
+#, c-format
+msgid "LibreOffice conversion failed (exit %d): %s"
+msgstr "LibreOffice 转换失败 (退出码 %d): %s"
+
+#: src/services/binary_converter.vala
+msgid "LibreOffice is required to convert Office documents. Download the converter package (see README) or install LibreOffice, then retry."
+msgstr "转换 Office 文档需要 LibreOffice：请下载转换器包（见 README）或安装 LibreOffice 后重试。"
+
+#: src/services/binary_converter.vala
+#, c-format
+msgid "pdftoppm failed: %s"
+msgstr "pdftoppm 失败: %s"
+
+#: src/services/binary_preprocessor.vala
+msgid "Multimodal AI is disabled or credential is empty, cannot convert binary files"
+msgstr "多模态 AI 未启用或凭据为空, 无法转换二进制文件"
+
+#: src/services/vlm_task_runner.vala
+msgid "Multimodal AI is disabled in settings"
+msgstr "多模态 AI 在设置中未启用"
+
+#: src/services/vlm_task_runner.vala
+msgid "Multimodal AI credential is empty, cannot convert binary files"
+msgstr "多模态 AI 凭据为空, 无法转换二进制文件"
+
+#: src/services/vlm_task_runner.vala
+#, c-format
+msgid "Image load failed: %s"
+msgstr "图片加载失败: %s"
+
+#: src/services/vlm_task_runner.vala
+#, c-format
+msgid "Document render failed: %s"
+msgstr "文档渲染失败: %s"
+
+#: src/services/office_text_extractor.vala
+#, c-format
+msgid "File too large for text extraction: %s"
+msgstr "文件过大, 无法进行文本抽取: %s"
+
+#: src/services/office_text_extractor.vala
+msgid "Unsupported format for text extraction"
+msgstr "该格式不支持文本抽取"
+
+#: src/services/office_text_extractor.vala
+msgid "Not a valid DOCX (missing word/document.xml)"
+msgstr "不是有效的 DOCX (缺少 word/document.xml)"
+
+#: src/services/office_text_extractor.vala
+msgid "Not a valid XLSX (no worksheet content)"
+msgstr "不是有效的 XLSX (无工作表内容)"
+
+#: src/services/office_text_extractor.vala
+msgid "Not a valid PPTX (no slides)"
+msgstr "不是有效的 PPTX (无幻灯片)"
+
+#: src/services/office_text_extractor.vala
+msgid "Not a valid ODF document (missing content.xml)"
+msgstr "不是有效的 ODF 文档 (缺少 content.xml)"
+
+#: src/services/office_text_extractor.vala
+#, c-format
+msgid "pdftotext failed: %s"
+msgstr "pdftotext 失败: %s"
+
+#: src/services/office_text_extractor.vala
+msgid "PDF appears to be scanned (no extractable text)"
+msgstr "PDF 疑似扫描件 (无可抽取文本)"
+
+#: src/services/office_text_extractor.vala
+msgid "Not a ZIP container (EOCD not found)"
+msgstr "不是 ZIP 容器 (未找到 EOCD)"
+
+#: src/services/office_text_extractor.vala
+msgid "ZIP64 archives are not supported"
+msgstr "不支持 ZIP64 归档"
+
+#: src/services/office_text_extractor.vala
+msgid "Encrypted ZIP entries are not supported"
+msgstr "不支持加密的 ZIP 条目"
+
+#: src/services/office_text_extractor.vala
+#, c-format
+msgid "Unsupported ZIP compression method %u"
+msgstr "不支持的 ZIP 压缩方法 %u"
+
+#: src/services/office_text_extractor.vala
+msgid "ZIP deflate decompression failed"
+msgstr "ZIP deflate 解压失败"
+
+#: src/services/office_text_extractor.vala
+msgid "Corrupted ZIP container"
+msgstr "ZIP 容器已损坏"
+
+#: src/services/office_text_extractor.vala
+#, c-format
+msgid "Text truncated (over %s)"
+msgstr "文本已截断 (超过 %s)"
+
+#: src/cli.vala
+msgid "Text extracted locally"
+msgstr "本地文本抽取完成"
+
+#: src/widgets/preferences_dialog.vala
+msgid "Processing Mode"
+msgstr "处理方式"
+
+#: src/widgets/preferences_dialog.vala
+msgid "Visual conversion (keeps charts, formulas, tables)"
+msgstr "视觉转换 (保留图表/公式/表格)"
+
+#: src/widgets/preferences_dialog.vala
+msgid "Text extraction (fast, offline, text only)"
+msgstr "文本抽取 (快速离线, 仅文本)"
+
+#: src/widgets/preferences_dialog.vala
+msgid "Visual conversion renders documents to images and needs LibreOffice for Office files; text extraction parses Office/PDF locally without AI credentials."
+msgstr "视觉转换把文档渲染为图片, Office 文件需要 LibreOffice; 文本抽取在本地直接解析 Office/PDF, 无需 AI 凭据。"
+
+#: src/window.vala
+#, c-format
+msgid "Reason: %s"
+msgstr "原因: %s"
diff --git a/src/cli.vala b/src/cli.vala
index a1d062e..4b3f8a3 100644
--- a/src/cli.vala
+++ b/src/cli.vala
@@ -322,8 +322,22 @@ public class CliController : GLib.Object {
         return true;
     }
 
+    // 输入路径解析 (--select-file / --load): 相对路径先按进程 cwd 查找
+    // (旧行为, 兼容既有脚本), 找不到且设置了 --work-dir 时再按 work_dir
+    // 解析 (README 完整工作流示例的约定)。输出路径 (--save / --export)
+    // 不在此列, 保持相对 cwd 不变。
+    private string resolve_input_path (string path) {
+        if (Path.is_absolute (path) || work_dir == null) return path;
+        if (FileUtils.test (path, FileTest.EXISTS)) return path;
+        try {
+            var resolved = work_dir.resolve_relative_path (path);
+            if (resolved.query_exists ()) return resolved.get_path ();
+        } catch (Error e) { }
+        return path;
+    }
+
     private bool add_file (string path) {
-        var file = File.new_for_path (path);
+        var file = File.new_for_path (resolve_input_path (path));
         if (!file.query_exists ()) {
             stderr.printf (_("Error: File does not exist: %s\n"), path);
             return false;
@@ -348,13 +362,18 @@ public class CliController : GLib.Object {
                 item.preprocessed_content = md;
                 item.preprocess_status = PreprocessStatus.COMPLETED;
                 item.from_cache = from_cache;
+                item.preprocess_error = null;
                 stdout.printf (_("  ↳ %s: %s\n"),
-                    from_cache ? _("Reused from cache") : _("VLM conversion called"),
+                    from_cache ? _("Reused from cache")
+                    : (ConfigManager.get_binary_process_mode () == ConfigManager.BINARY_MODE_TEXT)
+                        ? _("Text extracted locally")
+                        : _("VLM conversion called"),
                     file.get_basename ());
             } catch (Error e) {
                 stderr.printf (_("  ⚠ Preprocessing %s failed: %s\n"),
                     file.get_basename (), e.message);
                 item.preprocess_status = PreprocessStatus.FAILED;
+                item.preprocess_error = e.message;
             }
         }
         return true;
@@ -441,7 +460,8 @@ public class CliController : GLib.Object {
     }
 
     private bool load_project (string path) {
-        var file = File.new_for_path (path);
+        string resolved = resolve_input_path (path);
+        var file = File.new_for_path (resolved);
         if (!file.query_exists ()) {
             stderr.printf (_("Error: Project file does not exist: %s\n"), path);
             return false;
@@ -450,7 +470,7 @@ public class CliController : GLib.Object {
         try {
             string content;
             size_t len;
-            FileUtils.get_contents (path, out content, out len);
+            FileUtils.get_contents (resolved, out content, out len);
             var parser = new Json.Parser ();
             parser.load_from_data (content);
 
diff --git a/src/controllers/ai_controller.vala b/src/controllers/ai_controller.vala
index aafe517..bd2f161 100644
--- a/src/controllers/ai_controller.vala
+++ b/src/controllers/ai_controller.vala
@@ -903,8 +903,13 @@ public class AIController : GLib.Object {
     // ─── 静态辅助方法 ────────────────────────────────────────────────
 
     private static string rel_path (string root, string full) {
-        if (full.has_prefix (root)) {
-            string r = full.substring (root.length);
+        // 归一分隔符为 "/": Windows 反斜杠路径下 substring 会产出
+        // "\src\main.vala" 形式, 回传给 AI 后再解析会被 Path.is_absolute
+        // 误判为绝对路径 (Windows 下 "\x" 也算绝对)
+        string r = full.replace ("\\", "/");
+        string base_r = root.replace ("\\", "/");
+        if (r.has_prefix (base_r)) {
+            r = r.substring (base_r.length);
             while (r.has_prefix ("/")) r = r.substring (1);
             return r.length > 0 ? r : full;
         }
@@ -912,8 +917,11 @@ public class AIController : GLib.Object {
     }
 
     private static string normalize_path (string path) {
+        // 统一分隔符为 "/": Windows 上 get_path() 返回反斜杠路径, 若不归一,
+        // 沙箱前缀比较 (allowed + "/") 永远不命中, AI 文件工具全部误判越界
+        string p = path.replace ("\\", "/");
         var stack = new Gee.ArrayList<string> ();
-        foreach (unowned string part in path.split ("/")) {
+        foreach (unowned string part in p.split ("/")) {
             if (part == "" || part == ".") continue;
             if (part == "..") {
                 if (stack.size > 0) stack.remove_at (stack.size - 1);
@@ -922,7 +930,7 @@ public class AIController : GLib.Object {
             }
         }
         string joined = string.joinv ("/", (string[]) stack.to_array ());
-        return path.has_prefix ("/") ? "/" + joined : joined;
+        return p.has_prefix ("/") ? "/" + joined : joined;
     }
 
     private string? resolve_ai_path (string path) {
diff --git a/src/main.vala b/src/main.vala
index 2180871..f954029 100644
--- a/src/main.vala
+++ b/src/main.vala
@@ -438,9 +438,19 @@ public class FileCollectorApp : Adw.Application {
     }
 #endif
 
+#if WINDOWS
+    [CCode (cname = "win32_console_init")]
+    private static extern void win32_console_init ();
+#endif
+
     public static int main (string[] args) {
 #if MACOS
         setup_macos_bundle_env ();
+#endif
+#if WINDOWS
+        // GUI 子系统 exe 在 cmd/PowerShell 里跑 CLI 时 stdout 无句柄,
+        // 先检测并接管父进程控制台 (管道环境下为空操作)
+        win32_console_init ();
 #endif
         var app = new FileCollectorApp ();
         int ret = app.run (args);
diff --git a/src/models/item_data.vala b/src/models/item_data.vala
index eeedad0..6c314db 100644
--- a/src/models/item_data.vala
+++ b/src/models/item_data.vala
@@ -18,6 +18,8 @@ public class ItemData : GLib.Object {
     public PreprocessStatus preprocess_status { get; set; default = PreprocessStatus.NONE; }
     public string? preprocessed_content { get; set; }
     public bool from_cache { get; set; default = false; }
+    // 预处理失败原因 (FAILED 状态下展示); 重试/状态重置时必须一并清空
+    public string? preprocess_error { get; set; }
     public int start_line { get; set; default = 0; }
     public int end_line { get; set; default = 0; }
 
diff --git a/src/services/binary_converter.vala b/src/services/binary_converter.vala
index 8cea0c7..04e5fef 100644
--- a/src/services/binary_converter.vala
+++ b/src/services/binary_converter.vala
@@ -1,6 +1,35 @@
 public class BinaryConverter : GLib.Object {
     public const int MAX_IMAGE_DIMENSION = 2048;
 
+    // ─── 捆绑工具解析 ────────────────────────────────────────────────
+    // resolve_bundled_tool / new_poppler_launcher 位于 Platform (无 GDK 依赖,
+    // 便于单元测试链接); 这里仅保留 soffice 探测链 (纯 FileUtils 逻辑)。
+
+    // 解析 soffice 可执行文件。探测顺序:
+    //   1. 转换器包约定布局: <exe>/../tools/LibreOffice/program/soffice.exe
+    //   2. Windows 常见安装路径 (官方安装器默认目录, soffice.exe 不进 PATH)
+    //   3. null → 调用方回退裸名 PATH 搜索 (Linux/macOS 现状)
+    public static string? resolve_soffice () {
+        string exe = Platform.get_executable_path ();
+        if (exe != "." && exe.length > 0) {
+            var portable = Path.build_filename (
+                Path.get_dirname (exe), "..", "tools", "LibreOffice",
+                "program", "soffice.exe");
+            if (FileUtils.test (portable, FileTest.EXISTS)) return portable;
+        }
+        string[] roots = {
+            GLib.Environment.get_variable ("ProgramFiles"),
+            GLib.Environment.get_variable ("ProgramFiles(x86)")
+        };
+        foreach (unowned string? root in roots) {
+            if (root == null || root.length == 0) continue;
+            var candidate = Path.build_filename (
+                root, "LibreOffice", "program", "soffice.exe");
+            if (FileUtils.test (candidate, FileTest.EXISTS)) return candidate;
+        }
+        return null;
+    }
+
     // 复用临时基目录, 避免每次转换都创建/销毁顶层临时目录.
     // 子目录以互斥锁保护的自增计数器命名, 保证 VLM 队列并发调用时不冲突.
     private static string? temp_base_dir = null;
@@ -75,14 +104,15 @@ public class BinaryConverter : GLib.Object {
         return "image/png";
     }
 
-    public static string[]? convert_to_base64_images (string path) {
+    public static string[]? convert_to_base64_images (string path) throws Error {
         string pdf_path = path;
         string lower = path.down ();
         string? tmp_pdf = null;
 
         if (!lower.has_suffix (".pdf")) {
+            // soffice 缺失/转换失败会抛出带指引的错误 (NOT_FOUND/FAILED),
+            // 由调用方捕获后作为失败原因展示
             tmp_pdf = convert_office_to_pdf (path);
-            if (tmp_pdf == null) return null;
             pdf_path = tmp_pdf;
         }
 
@@ -97,7 +127,7 @@ public class BinaryConverter : GLib.Object {
         return result;
     }
 
-    private static string? convert_office_to_pdf (string src) {
+    private static string convert_office_to_pdf (string src) throws Error {
         // 复用基目录, 在其中创建本次任务的子目录, 避免并发冲突.
         // 子目录名使用互斥锁保护的自增计数器, 保证 VLM 队列多线程下唯一.
         string? tmp_base = get_temp_base ();
@@ -109,40 +139,45 @@ public class BinaryConverter : GLib.Object {
                 // 子目录创建失败时回退到独立 mkdtemp, 保证转换流程能继续
                 task_dir = DirUtils.make_tmp ("fc_pdf_XXXXXX");
                 if (task_dir == null) {
-                    warning ("Failed to create tmp dir for PDF");
-                    return null;
+                    throw new IOError.FAILED (_("Failed to create tmp dir for PDF conversion"));
                 }
             }
         } else {
             // 基目录创建失败时回退到独立 mkdtemp
             task_dir = DirUtils.make_tmp ("fc_pdf_XXXXXX");
             if (task_dir == null) {
-                warning ("Failed to create tmp dir for PDF");
-                return null;
+                throw new IOError.FAILED (_("Failed to create tmp dir for PDF conversion"));
             }
         }
 
-        string[] argv = {"soffice", "--headless", "--convert-to", "pdf", "--outdir", task_dir, src};
         try {
+            // 转换器包 / 官方安装路径命中 → 绝对路径; 否则裸名 PATH 搜索
+            string[] argv = { resolve_soffice () ?? "soffice", "--headless",
+                              "--convert-to", "pdf", "--outdir", task_dir, src };
             int status;
             string stdout_str, stderr_str;
             Process.spawn_sync (null, argv, null, SpawnFlags.SEARCH_PATH, null,
                                 out stdout_str, out stderr_str, out status);
-            if (status == 0) {
-                string basename = Path.get_basename (src);
-                int dot = basename.last_index_of (".");
-                string pdf_name = (dot > 0 ? basename.substring (0, dot) : basename) + ".pdf";
-                return Path.build_filename (task_dir, pdf_name);
-            } else {
+            if (status != 0) {
                 // 捕获 soffice 的 stderr 输出, 便于诊断转换失败原因
                 // (如文件损坏/格式不支持/LibreOffice 未安装等)
-                warning ("LibreOffice conversion failed (status %d): %s",
-                         status, (stderr_str ?? "").strip ());
+                throw new IOError.FAILED (_("LibreOffice conversion failed (exit %d): %s").printf (
+                    status, (stderr_str ?? "").strip ()));
             }
+            string basename = Path.get_basename (src);
+            int dot = basename.last_index_of (".");
+            string pdf_name = (dot > 0 ? basename.substring (0, dot) : basename) + ".pdf";
+            return Path.build_filename (task_dir, pdf_name);
+        } catch (SpawnError e) {
+            // 裸名搜索也找不到: Windows 上既没有转换器包也没装 LibreOffice。
+            // 失败路径必须清理本次的 task_dir, 否则每次尝试都泄漏一个目录。
+            cleanup_dir (task_dir);
+            throw new IOError.NOT_FOUND (_(
+                "LibreOffice is required to convert Office documents. Download the converter package (see README) or install LibreOffice, then retry."));
         } catch (Error e) {
-            warning ("LibreOffice conversion failed: %s", e.message);
+            cleanup_dir (task_dir);
+            throw e;
         }
-        return null;
     }
 
     private static string[]? render_pdf_to_base64_images (string pdf_path) {
@@ -153,22 +188,24 @@ public class BinaryConverter : GLib.Object {
         }
 
         string prefix = Path.build_filename (tmp_dir, "page");
-        string[] argv = {"pdftoppm", "-png", "-r", "200", pdf_path, prefix};
         try {
-            int status;
-            string stdout_str, stderr_str;
             // 捕获 stderr: pdftoppm 失败时 (如 PDF 损坏/poppler 未安装) 会把
             // 诊断信息写到 stderr, 原实现丢弃 stderr 只打印退出码, 无法诊断。
-            Process.spawn_sync (null, argv, null, SpawnFlags.SEARCH_PATH, null,
-                                out stdout_str, out stderr_str, out status);
-            if (status != 0) {
-                warning ("pdftoppm failed (status %d): %s",
-                         status, (stderr_str ?? "").strip ());
+            // Windows 便携包用捆绑的 pdftoppm (绝对路径 + POPPLER_DATADIR),
+            // 其余平台回退裸名 PATH 搜索。
+            string? tool = Platform.resolve_bundled_tool ("pdftoppm.exe");
+            string[] argv = { tool ?? "pdftoppm", "-png", "-r", "200", pdf_path, prefix };
+            var launcher = Platform.new_poppler_launcher ();
+            var proc = launcher.spawnv (argv);
+            string out_str, err_str;
+            proc.communicate_utf8 (null, null, out out_str, out err_str);
+            if (!proc.get_successful ()) {
+                warning (_("pdftoppm failed: %s"), (err_str ?? "").strip ());
                 cleanup_dir (tmp_dir);
                 return null;
             }
         } catch (Error e) {
-            warning ("pdftoppm failed: %s", e.message);
+            warning (_("pdftoppm failed: %s"), e.message);
             cleanup_dir (tmp_dir);
             return null;
         }
@@ -216,9 +253,10 @@ public class BinaryConverter : GLib.Object {
      *     返回临时 PDF 路径, is_temp = true. 调用方处理完毕后须调用
      *     cleanup_dir(Path.get_dirname(result)) 清理临时目录.
      *
-     * 返回 null 表示无法解析/转换失败.
+     * soffice 缺失/转换失败时抛出带指引的错误 (NOT_FOUND/FAILED),
+     * 不再返回 null —— 调用方捕获后作为失败原因展示。
      */
-    public static string? resolve_upload_source (string path, out bool is_temp) {
+    public static string resolve_upload_source (string path, out bool is_temp) throws Error {
         is_temp = false;
         string lower = path.down ();
 
@@ -228,13 +266,11 @@ public class BinaryConverter : GLib.Object {
             lower.has_suffix (".jpeg") || lower.has_suffix (".webp") ||
             lower.has_suffix (".bmp") || lower.has_suffix (".tiff") ||
             lower.has_suffix (".tif")) {
-            is_temp = false;
             return path;
         }
 
         // 其余 (Office 等): 转 PDF 后上传
-        string? tmp_pdf = convert_office_to_pdf (path);
-        if (tmp_pdf == null) return null;
+        string tmp_pdf = convert_office_to_pdf (path);
         is_temp = true;
         return tmp_pdf;
     }
diff --git a/src/services/binary_preprocessor.vala b/src/services/binary_preprocessor.vala
index 42463df..735fab6 100644
--- a/src/services/binary_preprocessor.vala
+++ b/src/services/binary_preprocessor.vala
@@ -15,11 +15,15 @@ public class BinaryPreprocessor : GLib.Object {
     ) throws Error {
         from_cache = false;
 
+        // 0. 当前处理模式对应的缓存 flow 标记 (同一文件两种模式产物不同)
+        string flow = (ConfigManager.get_binary_process_mode () == ConfigManager.BINARY_MODE_TEXT)
+            ? "text" : "vlm";
+
         // 1. 先走轻量指纹 (size:mtime) 快速命中缓存, 避免对大文件 (PDF/图片) 计算 SHA256.
         //    命中即返回, 未命中 (含旧版无 quick 字段的缓存) 再回退到完整 SHA256 路径.
         string quick = PreprocessCache.compute_file_hash_fast (item.file_path);
         var cache = new PreprocessCache (work_dir_path);
-        string? cached = cache.get_cached_markdown_quick (item.file_path, quick);
+        string? cached = cache.get_cached_markdown_quick (item.file_path, quick, flow);
         if (cached != null) {
             from_cache = true;
             return cached;
@@ -36,13 +40,23 @@ public class BinaryPreprocessor : GLib.Object {
             warning ("Hash computation failed for %s: %s", item.file_path, e.message);
         }
         if (hash_valid) {
-            cached = cache.get_cached_markdown (item.file_path, hash);
+            cached = cache.get_cached_markdown (item.file_path, hash, flow);
             if (cached != null) {
                 from_cache = true;
                 return cached;
             }
         }
 
+        // 2.5 "文本抽取" 模式: 支持的格式直接本地抽纯文本, 不需要 VLM 凭据与
+        //     网络, 因此放在凭据检查之前。抽取失败 (含扫描版 PDF 抽空) 返回
+        //     null, 继续走下方视觉转换/OCR 流程兜底。vlm 模式下此函数直接
+        //     返回 null, 不改变现有行为。
+        string? text_md = try_extract_text_document (
+            item.file_path, hash_valid ? hash : null, cache);
+        if (text_md != null) {
+            return text_md;
+        }
+
         // 3. Load VLM settings
         var settings = ConfigManager.load_multimodal_ai_settings ();
         bool has_credential = (settings.provider == ConfigManager.PROVIDER_PADDLEOCR)
@@ -50,7 +64,7 @@ public class BinaryPreprocessor : GLib.Object {
             : settings.api_key.length > 0;
         if (!settings.enabled || !has_credential) {
             throw new IOError.FAILED (
-                "多模态 AI 未启用或凭据为空, 无法转换二进制文件"
+                _("Multimodal AI is disabled or credential is empty, cannot convert binary files")
             );
         }
 
@@ -59,10 +73,7 @@ public class BinaryPreprocessor : GLib.Object {
         if (settings.provider == ConfigManager.PROVIDER_PADDLEOCR) {
             // PaddleOCR 云端路径: 上传原文件, 轮询并合并 Markdown
             bool is_temp;
-            string? upload_source = BinaryConverter.resolve_upload_source (item.file_path, out is_temp);
-            if (upload_source == null) {
-                throw new IOError.FAILED ("准备上传源失败: " + item.file_path);
-            }
+            string upload_source = BinaryConverter.resolve_upload_source (item.file_path, out is_temp);
             try {
                 var client = new PaddleOCRClient (settings.paddleocr_token, settings.timeout);
                 // 仅在哈希有效时传入缓存, 使配图落盘到对应缓存子文件夹
@@ -109,7 +120,7 @@ public class BinaryPreprocessor : GLib.Object {
 
         // 7. Save to cache (仅在 hash 有效时保存, 避免空 hash 污染 manifest)
         if (hash_valid) {
-            cache.save_markdown (item.file_path, hash, md);
+            cache.save_markdown (item.file_path, hash, md, flow);
         }
 
         return md;
@@ -122,13 +133,45 @@ public class BinaryPreprocessor : GLib.Object {
         ItemData item, string? work_dir_path
     ) throws Error {
         if (work_dir_path == null) return null;
+        // 按当前处理模式匹配 flow 标记 (两种模式的产物互不命中)
+        string flow = (ConfigManager.get_binary_process_mode () == ConfigManager.BINARY_MODE_TEXT)
+            ? "text" : "vlm";
         // 先轻量指纹命中, 未命中回退 SHA256 路径 (兼容旧缓存)
         string quick = PreprocessCache.compute_file_hash_fast (item.file_path);
         var cache = new PreprocessCache (work_dir_path);
-        string? cached = cache.get_cached_markdown_quick (item.file_path, quick);
+        string? cached = cache.get_cached_markdown_quick (item.file_path, quick, flow);
         if (cached != null) return cached;
         string hash = PreprocessCache.compute_file_hash (item.file_path);
-        return cache.get_cached_markdown (item.file_path, hash);
+        return cache.get_cached_markdown (item.file_path, hash, flow);
+    }
+
+    // "文本抽取" 模式入口: 支持的格式本地抽纯文本并写入缓存 (flow=text)。
+    // 返回 null 表示未抽取成功 —— 模式非 text / 格式不支持 / 抽取失败
+    // (含扫描版 PDF 抽空), 调用方应继续走视觉转换/OCR 流程兜底。
+    // 本分支不需要 VLM 凭据与网络, 调用点必须放在凭据检查之前。
+    public static string? try_extract_text_document (
+        string file_path, string? hash, PreprocessCache? cache
+    ) {
+        if (ConfigManager.get_binary_process_mode () != ConfigManager.BINARY_MODE_TEXT) {
+            return null;
+        }
+        int dot = file_path.last_index_of (".");
+        if (dot < 0) return null;
+        if (!OfficeTextExtractor.supports (file_path.substring (dot).down ())) {
+            return null;
+        }
+        try {
+            string md = OfficeTextExtractor.extract_markdown (file_path);
+            if (cache != null && hash != null) {
+                cache.save_markdown (file_path, hash, md, "text");
+            }
+            return md;
+        } catch (Error e) {
+            // 抽取失败一律回退 (不抛出), 原因留日志供诊断
+            warning ("Text extraction failed for %s, falling back to visual/OCR: %s",
+                     file_path, e.message);
+            return null;
+        }
     }
 
     // 默认提示词 (基于 file_path 字符串, 不依赖 ItemData GObject)
diff --git a/src/services/config_manager.vala b/src/services/config_manager.vala
index d3be58a..e8e400a 100644
--- a/src/services/config_manager.vala
+++ b/src/services/config_manager.vala
@@ -122,6 +122,13 @@ public class ConfigManager : GLib.Object {
     public const string PROVIDER_OPENAI = "openai";
     public const string PROVIDER_PADDLEOCR = "paddleocr";
 
+    // 二进制文件处理模式:
+    //   vlm  = 视觉转换 (soffice → pdftoppm → VLM/OCR, 保图表/公式/表格, 默认)
+    //   text = 文本抽取 (纯本地解析 OOXML/ODF + pdftotext, 离线且无需凭据,
+    //          但丢失图表/公式/嵌入图片)
+    public const string BINARY_MODE_VLM = "vlm";
+    public const string BINARY_MODE_TEXT = "text";
+
     public struct AISettings {
         public bool enabled;
         public string profile_name;  // 当前激活的 ai_models 配置方案名, 空串表示尚未选择
@@ -351,6 +358,39 @@ public class ConfigManager : GLib.Object {
         }
     }
 
+    // ─── 二进制文件处理模式 (视觉转换 / 文本抽取) ────────────────────────
+
+    public static string get_binary_process_mode () {
+        config_mutex.lock ();
+        try {
+            var root = load_settings_root_unlocked ();
+            if (root != null && root.has_member ("binary_process_mode")) {
+                string mode = root.get_string_member ("binary_process_mode");
+                if (mode == BINARY_MODE_TEXT) return BINARY_MODE_TEXT;
+            }
+        } catch (Error e) {
+            warning ("Failed to load binary_process_mode: %s", e.message);
+        } finally {
+            config_mutex.unlock ();
+        }
+        return BINARY_MODE_VLM;
+    }
+
+    public static void save_binary_process_mode (string mode) {
+        config_mutex.lock ();
+        try {
+            Json.Object root = load_settings_root_unlocked () ?? new Json.Object ();
+            // 非法值一律落回默认, 防止手改配置文件后两条管线都走不到
+            root.set_string_member ("binary_process_mode",
+                mode == BINARY_MODE_TEXT ? BINARY_MODE_TEXT : BINARY_MODE_VLM);
+            write_settings_root_unlocked (root);
+        } catch (Error e) {
+            warning ("Failed to save binary_process_mode: %s", e.message);
+        } finally {
+            config_mutex.unlock ();
+        }
+    }
+
     // ─── 多模态 AI 设置 ──────────────────────────────────────────────────
 
     public struct MultimodalAISettings {
diff --git a/src/services/git_service.vala b/src/services/git_service.vala
index db77ce9..1279b7d 100644
--- a/src/services/git_service.vala
+++ b/src/services/git_service.vala
@@ -30,6 +30,11 @@ public class GitService : GLib.Object {
 
     public static string run_git (string work_dir, string[] args) throws GLib.Error {
         string[] argv = { git_executable () };
+        // 非 ASCII 文件名默认被 core.quotePath 转成 "\346\226\207..." 八进制
+        // 转义, status/log 解析端按原样取路径后存在性检查必然失败, 中文
+        // 文件会被静默漏掉。统一关闭转义, 输出保持原始 UTF-8。
+        argv += "-c";
+        argv += "core.quotePath=false";
         foreach (var a in args) argv += a;
 
         var launcher = new SubprocessLauncher (
diff --git a/src/services/office_text_extractor.vala b/src/services/office_text_extractor.vala
new file mode 100644
index 0000000..0cf2bd1
--- /dev/null
+++ b/src/services/office_text_extractor.vala
@@ -0,0 +1,522 @@
+using Gee;
+
+// 纯 Vala 的 Office 文档 / PDF 文本抽取器 (可选的 "文本抽取" 处理模式)。
+//
+// 与视觉转换 (soffice → pdftoppm → VLM) 互补: 文本模式零网络、无需 VLM 凭据、
+// 秒级完成, 但只覆盖文字与表格结构 —— 图表/公式/嵌入图片会丢失, 因此默认
+// 模式仍为视觉转换 (ConfigManager.binary_process_mode)。
+//
+// 覆盖格式:
+//   - OOXML (.docx/.xlsx/.pptx): ZIP 容器 + XML, 内置最小 ZIP 读取器解析
+//   - ODF (.odt/.ods/.odp): 同为 ZIP + XML
+//   - PDF: 调用捆绑的 pdftotext (与 pdftoppm 同属 poppler, Windows 包已捆绑;
+//     其余平台回退 PATH 搜索)。抽不出文本 (扫描版) 抛 NOT_SUPPORTED,
+//     调用方据此回退视觉/OCR 流程。
+//
+// ZIP 解析范围: PKZIP 经典布局 (EOCD + central directory + local header),
+// 支持 STORED 与 DEFLATE; ZIP64 / 加密条目明确报错而不是静默给出错误结果。
+public class OfficeTextExtractor : GLib.Object {
+
+    // 单文件守卫: 防止超大容器把内存吃爆 (文本抽取场景下 50MB 已极宽裕)
+    private const int MAX_FILE_SIZE = 50 * 1024 * 1024;
+    // 抽出文本的上限, 超出截断 (导出与 token 估算不需要无限正文)
+    private const int MAX_TEXT_SIZE = 8 * 1024 * 1024;
+
+    private const string[] SUPPORTED_EXTS = {
+        ".docx", ".xlsx", ".pptx", ".odt", ".ods", ".odp", ".pdf"
+    };
+
+    // ext 为带前导点的小写扩展名 (与 ItemData.is_allowed_binary_target 口径一致)
+    public static bool supports (string ext_lower) {
+        foreach (unowned string e in SUPPORTED_EXTS) {
+            if (ext_lower == e) return true;
+        }
+        return false;
+    }
+
+    public static string extract_markdown (string path) throws Error {
+        var f = File.new_for_path (path);
+        var info = f.query_info (FileAttribute.STANDARD_SIZE, FileQueryInfoFlags.NONE);
+        if (info.get_size () > MAX_FILE_SIZE) {
+            throw new IOError.FAILED (_("File too large for text extraction: %s").printf (
+                GLib.format_size ((int64) info.get_size ())));
+        }
+
+        string md;
+        string lower = path.down ();
+        if (lower.has_suffix (".pdf")) {
+            md = extract_pdf (path);
+        } else {
+            uint8[] data;
+            FileUtils.get_data (path, out data);
+            var zip = new ZipReader (data);
+            if (lower.has_suffix (".docx")) {
+                md = extract_docx (zip);
+            } else if (lower.has_suffix (".xlsx")) {
+                md = extract_xlsx (zip);
+            } else if (lower.has_suffix (".pptx")) {
+                md = extract_pptx (zip);
+            } else if (lower.has_suffix (".ods")) {
+                md = extract_odf (zip, true);
+            } else if (lower.has_suffix (".odt") || lower.has_suffix (".odp")) {
+                md = extract_odf (zip, false);
+            } else {
+                throw new IOError.NOT_SUPPORTED (_("Unsupported format for text extraction"));
+            }
+        }
+
+        if (md.length > MAX_TEXT_SIZE) {
+            md = md.substring (0, MAX_TEXT_SIZE) + "\n\n"
+                + _("Text truncated (over %s)").printf (GLib.format_size (MAX_TEXT_SIZE));
+        }
+        return md;
+    }
+
+    // ─── 各格式抽取 ───────────────────────────────────────────────────
+
+    private static string extract_docx (ZipReader zip) throws Error {
+        var entry = zip.find ("word/document.xml");
+        if (entry == null) {
+            throw new IOError.NOT_SUPPORTED (_("Not a valid DOCX (missing word/document.xml)"));
+        }
+        string xml = bytes_to_utf8 (entry.data);
+        // 表格单元格 → \t, 行/段落 → \n, 软换行/制表符保留; 其余标签剥离
+        string t = xml.replace ("</w:tc>", "\t")
+                      .replace ("</w:tr>", "\n")
+                      .replace ("</w:p>", "\n")
+                      .replace ("<w:tab/>", "\t")
+                      .replace ("<w:br/>", "\n");
+        return tidy_text (decode_entities (strip_tags (t)));
+    }
+
+    private static string extract_xlsx (ZipReader zip) throws Error {
+        var shared = new Gee.ArrayList<string> ();
+        var sst = zip.find ("xl/sharedStrings.xml");
+        if (sst != null) {
+            string xml = bytes_to_utf8 (sst.data);
+            foreach (string si in extract_blocks (xml, "si")) {
+                // <si> 内可能是单个 <t> 或富文本 <r><t>..</t></r> 多段: 拼接全部 <t>
+                var sb2 = new StringBuilder ();
+                foreach (string t in extract_tag_contents (si, "t")) {
+                    sb2.append (t);
+                }
+                shared.add (decode_entities (sb2.str));
+            }
+        }
+
+        // 工作表按文件名中的数字自然排序 (sheet1, sheet2, ... sheet10)
+        var sheets = zip.entries_matching ("xl/worksheets/");
+        sheets.sort ((a, b) => {
+            return num_after (a.name, "sheet") - num_after (b.name, "sheet");
+        });
+        if (sheets.size == 0) {
+            throw new IOError.NOT_SUPPORTED (_("Not a valid XLSX (no worksheet content)"));
+        }
+
+        var sb = new StringBuilder ();
+        int sheet_idx = 0;
+        foreach (var sheet in sheets) {
+            sheet_idx++;
+            if (sb.len > 0) sb.append ("\n");
+            sb.append ("## Sheet ").append (sheet_idx.to_string ()).append ("\n");
+            string xml = bytes_to_utf8 (sheet.data);
+            foreach (string row in extract_blocks (xml, "row")) {
+                var cells = new Gee.ArrayList<string> ();
+                foreach (string cell in extract_blocks (row, "c")) {
+                    // t="s" → sharedStrings 索引; 其余读 <v> 原值
+                    var vs = extract_tag_contents (cell, "v");
+                    string val = vs.size > 0 ? vs.get (0) : "";
+                    if (val.length > 0 && cell.contains ("t=\"s\"")) {
+                        int idx = int.parse (val);
+                        val = (idx >= 0 && idx < shared.size) ? shared.get (idx) : "";
+                    } else {
+                        val = decode_entities (val);
+                    }
+                    cells.add (val);
+                }
+                if (cells.size == 0) continue;
+                sb.append (string.joinv ("\t", cells.to_array ()));
+                sb.append ("\n");
+            }
+        }
+        return sb.str;
+    }
+
+    private static string extract_pptx (ZipReader zip) throws Error {
+        var slides = zip.entries_matching ("ppt/slides/");
+        slides.sort ((a, b) => {
+            return num_after (a.name, "slide") - num_after (b.name, "slide");
+        });
+        if (slides.size == 0) {
+            throw new IOError.NOT_SUPPORTED (_("Not a valid PPTX (no slides)"));
+        }
+        var sb = new StringBuilder ();
+        int n = 0;
+        foreach (var slide in slides) {
+            n++;
+            if (sb.len > 0) sb.append ("\n");
+            sb.append ("## Slide ").append (n.to_string ()).append ("\n");
+            string xml = bytes_to_utf8 (slide.data);
+            string t = xml.replace ("</a:tc>", "\t")
+                          .replace ("</a:p>", "\n")
+                          .replace ("<a:br/>", "\n");
+            sb.append (tidy_text (decode_entities (strip_tags (t))));
+        }
+        return sb.str;
+    }
+
+    // ODF 家族共用 content.xml; 电子表格额外处理单元格/行边界
+    private static string extract_odf (ZipReader zip, bool spreadsheet) throws Error {
+        var entry = zip.find ("content.xml");
+        if (entry == null) {
+            throw new IOError.NOT_SUPPORTED (_("Not a valid ODF document (missing content.xml)"));
+        }
+        string xml = bytes_to_utf8 (entry.data);
+        string t = xml.replace ("</text:p>", "\n")
+                      .replace ("</text:h>", "\n")
+                      .replace ("<text:tab/>", "\t")
+                      .replace ("<text:line-break/>", "\n");
+        if (spreadsheet) {
+            t = t.replace ("</table:table-cell>", "\t")
+                 .replace ("</table:table-row>", "\n");
+        }
+        return tidy_text (decode_entities (strip_tags (t)));
+    }
+
+    private static string extract_pdf (string path) throws Error {
+        string? tool = Platform.resolve_bundled_tool ("pdftotext.exe");
+        string[] argv = { tool ?? "pdftotext", "-layout", "-enc", "UTF-8", path, "-" };
+        var launcher = Platform.new_poppler_launcher ();
+        var proc = launcher.spawnv (argv);
+        string out_str, err_str;
+        proc.communicate_utf8 (null, null, out out_str, out err_str);
+        if (!proc.get_successful ()) {
+            throw new IOError.FAILED (_("pdftotext failed: %s").printf ((err_str ?? "").strip ()));
+        }
+        string text = (out_str ?? "").strip ();
+        if (text.length == 0) {
+            // 扫描版 PDF (纯图): 文本抽取无价值, 调用方据此回退视觉/OCR 流程
+            throw new IOError.NOT_SUPPORTED (_("PDF appears to be scanned (no extractable text)"));
+        }
+        return tidy_text (text);
+    }
+
+    // ─── 文本收尾 ─────────────────────────────────────────────────────
+
+    // 连续空行压缩到最多一个, 行尾空白清理 (顺带处理 CRLF 残留)
+    private static string tidy_text (string text) {
+        var sb = new StringBuilder ();
+        int blank = 0;
+        foreach (unowned string line in text.split ("\n")) {
+            string trimmed = line.strip ();
+            if (trimmed.length == 0) {
+                blank++;
+                if (blank <= 2) sb.append ("\n");
+                continue;
+            }
+            blank = 0;
+            sb.append (trimmed);
+            sb.append ("\n");
+        }
+        return sb.str;
+    }
+
+    // ─── XML 处理 ─────────────────────────────────────────────────────
+
+    // 剥掉全部标签 (含注释); 边界标签已在调用方 replace 为 \n/\t
+    private static string strip_tags (string xml) {
+        var sb = new GLib.StringBuilder.sized (xml.length);
+        int pos = 0;
+        while (pos < xml.length) {
+            int lt = xml.index_of ("<", pos);
+            if (lt < 0) {
+                sb.append (xml.substring (pos));
+                break;
+            }
+            if (lt > pos) sb.append (xml.substring (pos, lt - pos));
+            if (lt + 3 < xml.length && xml[lt + 1] == '!' && xml[lt + 2] == '-'
+                && xml[lt + 3] == '-') {
+                int end = xml.index_of ("-->", lt + 4);
+                pos = (end >= 0) ? end + 3 : xml.length;
+            } else {
+                int gt = xml.index_of (">", lt);
+                pos = (gt >= 0) ? gt + 1 : xml.length;
+            }
+        }
+        return sb.str;
+    }
+
+    private static string decode_entities (string s) {
+        if (!s.contains ("&")) return s;
+        var sb = new GLib.StringBuilder.sized (s.length);
+        int pos = 0;
+        while (pos < s.length) {
+            int amp = s.index_of ("&", pos);
+            if (amp < 0) {
+                sb.append (s.substring (pos));
+                break;
+            }
+            if (amp > pos) sb.append (s.substring (pos, amp - pos));
+            int semi = s.index_of (";", amp);
+            if (semi < 0 || semi - amp > 12) {
+                // 不是实体: 原样输出 '&'
+                sb.append ("&");
+                pos = amp + 1;
+                continue;
+            }
+            string ent = s.substring (amp + 1, semi - amp - 1);
+            string? decoded = null;
+            switch (ent) {
+                case "amp": decoded = "&"; break;
+                case "lt": decoded = "<"; break;
+                case "gt": decoded = ">"; break;
+                case "quot": decoded = "\""; break;
+                case "apos": decoded = "'"; break;
+                default:
+                    if (ent.has_prefix ("#x") || ent.has_prefix ("#X")) {
+                        int v = hex_val (ent.substring (2));
+                        decoded = (v > 0 && v <= 0x10FFFF) ? ((unichar) v).to_string () : null;
+                    } else if (ent.has_prefix ("#")) {
+                        int v = int.parse (ent.substring (1));
+                        decoded = (v > 0 && v <= 0x10FFFF) ? ((unichar) v).to_string () : null;
+                    }
+                    break;
+            }
+            if (decoded != null) {
+                sb.append (decoded);
+            } else {
+                sb.append (s.substring (amp, semi - amp + 1));
+            }
+            pos = semi + 1;
+        }
+        return sb.str;
+    }
+
+    private static int hex_val (string hex) {
+        if (hex.length == 0 || hex.length > 6) return -1;
+        int v = 0;
+        for (int i = 0; i < hex.length; i++) {
+            char c = hex[i];
+            v <<= 4;
+            if (c >= '0' && c <= '9') v += c - '0';
+            else if (c >= 'a' && c <= 'f') v += c - 'a' + 10;
+            else if (c >= 'A' && c <= 'F') v += c - 'A' + 10;
+            else return -1;
+        }
+        return v;
+    }
+
+    // 提取 <tag ...> ... </tag> 整块 (含标签); start 精确匹配 "<tag" 且其后是
+    // '>' / '/' / 空白, 排除 <rowBreaks> 这类前缀碰撞
+    private static Gee.ArrayList<string> extract_blocks (string xml, string tag) {
+        var result = new Gee.ArrayList<string> ();
+        string open = "<" + tag;
+        string close = "</" + tag + ">";
+        int pos = 0;
+        while (true) {
+            int s = find_exact_open (xml, open, pos);
+            if (s < 0) break;
+            int e = xml.index_of (close, s);
+            if (e < 0) break;
+            result.add (xml.substring (s, e - s + close.length));
+            pos = e + close.length;
+        }
+        return result;
+    }
+
+    // 提取 <tag ...>内容</tag> 的内容 (不含标签); 空元素 <tag/> 记为 ""
+    private static Gee.ArrayList<string> extract_tag_contents (string xml, string tag) {
+        var result = new Gee.ArrayList<string> ();
+        string open = "<" + tag;
+        string close = "</" + tag + ">";
+        int pos = 0;
+        while (true) {
+            int s = find_exact_open (xml, open, pos);
+            if (s < 0) break;
+            int gt = xml.index_of (">", s);
+            if (gt < 0) break;
+            if (gt > s && xml[gt - 1] == '/') {
+                result.add ("");
+                pos = gt + 1;
+                continue;
+            }
+            int e = xml.index_of (close, gt);
+            if (e < 0) break;
+            result.add (xml.substring (gt + 1, e - gt - 1));
+            pos = e + close.length;
+        }
+        return result;
+    }
+
+    // 查找 "<tag" 且其后紧跟 '>' / '/' / 空白的出现位置; 找不到返回 -1
+    private static int find_exact_open (string xml, string open, int from) {
+        int s = xml.index_of (open, from);
+        while (s >= 0) {
+            char nxt = (s + open.length < xml.length) ? xml[s + open.length] : '\0';
+            if (nxt == '>' || nxt == '/' || nxt == ' '
+                || nxt == '\t' || nxt == '\n' || nxt == '\r') {
+                return s;
+            }
+            s = xml.index_of (open, s + 1);
+        }
+        return -1;
+    }
+
+    // 文件名中 prefix 之后的连续数字 (sheet1.xml → 1); 无数字排到最后
+    private static int num_after (string name, string prefix) {
+        int p = name.index_of (prefix);
+        if (p < 0) return int.MAX;
+        int v = 0;
+        bool any = false;
+        for (int i = p + prefix.length; i < name.length; i++) {
+            char c = name[i];
+            if (c >= '0' && c <= '9') {
+                v = v * 10 + (c - '0');
+                any = true;
+            } else if (any) {
+                break;
+            }
+        }
+        return any ? v : int.MAX;
+    }
+
+    private static string bytes_to_utf8 (uint8[] data) {
+        var sb = new GLib.StringBuilder.sized (data.length);
+        sb.append_len ((string) data, data.length);
+        return sb.str;
+    }
+
+    // ─── 最小 ZIP 读取器 ──────────────────────────────────────────────
+
+    private class ZipEntry : GLib.Object {
+        public string name;
+        public uint8[] data;
+        public ZipEntry (string name, uint8[] data) {
+            this.name = name;
+            this.data = data;
+        }
+    }
+
+    private class ZipReader : GLib.Object {
+        private Gee.HashMap<string, ZipEntry> map =
+            new Gee.HashMap<string, ZipEntry> ();
+
+        public ZipReader (uint8[] data) throws Error {
+            parse (data);
+        }
+
+        public ZipEntry? find (string name) {
+            return map[name];
+        }
+
+        public Gee.ArrayList<ZipEntry> entries_matching (string prefix) {
+            var result = new Gee.ArrayList<ZipEntry> ();
+            foreach (var e in map.values) {
+                if (e.name.has_prefix (prefix)) result.add (e);
+            }
+            return result;
+        }
+
+        private void parse (uint8[] data) throws Error {
+            // 1. 从尾部扫描 EOCD (0x06054b50): 结构 22 字节 + 注释最长 65535
+            int eocd = -1;
+            int min_scan = int.max (0, data.length - 22 - 65535);
+            for (int i = data.length - 22; i >= min_scan; i--) {
+                if (le32 (data, i) == 0x06054b50u) {
+                    eocd = i;
+                    break;
+                }
+            }
+            if (eocd < 0) {
+                throw new IOError.NOT_SUPPORTED (_("Not a ZIP container (EOCD not found)"));
+            }
+            uint16 total = le16 (data, eocd + 10);
+            uint32 cd_size = le32 (data, eocd + 12);
+            uint32 cd_offset = le32 (data, eocd + 16);
+            if (total == 0xFFFFu || cd_offset == 0xFFFFFFFFu || cd_size == 0xFFFFFFFFu) {
+                throw new IOError.NOT_SUPPORTED (_("ZIP64 archives are not supported"));
+            }
+            if ((int) cd_offset + (int) cd_size > data.length) {
+                throw corrupted ();
+            }
+            // 2. 遍历 central directory
+            int p = (int) cd_offset;
+            for (uint16 i = 0; i < total; i++) {
+                if (p + 46 > data.length || le32 (data, p) != 0x02014b50u) {
+                    throw corrupted ();
+                }
+                uint16 flags = le16 (data, p + 8);
+                uint16 method = le16 (data, p + 10);
+                uint32 comp_size = le32 (data, p + 20);
+                uint16 name_len = le16 (data, p + 28);
+                uint16 extra_len = le16 (data, p + 30);
+                uint16 comment_len = le16 (data, p + 32);
+                uint32 local_offset = le32 (data, p + 42);
+                if ((flags & 0x1u) != 0) {
+                    throw new IOError.NOT_SUPPORTED (_("Encrypted ZIP entries are not supported"));
+                }
+                if (p + 46 + name_len > data.length) {
+                    throw corrupted ();
+                }
+                string name = bytes_to_utf8 (data[p + 46 : p + 46 + name_len]);
+                // 3. local header 定位数据区 (local 与 central 的 name/extra 长度可能不同)
+                if (local_offset + 30 > (uint32) data.length
+                    || le32 (data, (int) local_offset) != 0x04034b50u) {
+                    throw corrupted ();
+                }
+                uint16 lname_len = le16 (data, (int) local_offset + 26);
+                uint16 lextra_len = le16 (data, (int) local_offset + 28);
+                int start = (int) local_offset + 30 + lname_len + lextra_len;
+                if (start < 0 || start + (int) comp_size > data.length) {
+                    throw corrupted ();
+                }
+                uint8[] content;
+                if (method == 0) {
+                    content = data[start : start + (int) comp_size];
+                } else if (method == 8) {
+                    content = inflate (data[start : start + (int) comp_size]);
+                } else {
+                    throw new IOError.NOT_SUPPORTED (
+                        _("Unsupported ZIP compression method %u").printf (method));
+                }
+                map[name] = new ZipEntry (name, content);
+                p += 46 + name_len + extra_len + comment_len;
+            }
+        }
+
+        private uint8[] inflate (uint8[] raw) throws Error {
+            // 注意: vala 0.56 的 glib vapi 没有 GLib.BytesBuilder, 用
+            // GLib.ByteArray 累积; ZlibDecompressor 构造参数复用
+            // ZlibCompressorFormat 枚举 (0.56 vapi 无独立的
+            // ZlibDecompressorFormat 绑定)。
+            var mins = new MemoryInputStream.from_data (raw);
+            var decomp = new GLib.ZlibDecompressor (GLib.ZlibCompressorFormat.RAW);
+            var cis = new GLib.ConverterInputStream (mins, decomp);
+            // Gee 泛型不支持数组类型参数, 二进制累积用 GLib.ByteArray
+            var out_buf = new GLib.ByteArray ();
+            var buf = new uint8[64 * 1024];
+            while (true) {
+                ssize_t n = cis.read (buf);
+                if (n < 0) {
+                    throw new IOError.FAILED (_("ZIP deflate decompression failed"));
+                }
+                if (n == 0) break;
+                out_buf.append (buf[0 : (int) n]);
+            }
+            return out_buf.steal ();
+        }
+
+        private static Error corrupted () {
+            return new IOError.FAILED (_("Corrupted ZIP container"));
+        }
+
+        private static uint16 le16 (uint8[] d, int off) {
+            return (uint16) (d[off] | (d[off + 1] << 8));
+        }
+
+        private static uint32 le32 (uint8[] d, int off) {
+            return ((uint32) d[off]) | ((uint32) d[off + 1] << 8)
+                | ((uint32) d[off + 2] << 16) | ((uint32) d[off + 3] << 24);
+        }
+    }
+}
diff --git a/src/services/preprocess_cache.vala b/src/services/preprocess_cache.vala
index 5bf72aa..b3a72cb 100644
--- a/src/services/preprocess_cache.vala
+++ b/src/services/preprocess_cache.vala
@@ -70,7 +70,10 @@ public class PreprocessCache : GLib.Object {
         }
     }
 
-    public string? get_cached_markdown (string abs_path, string current_hash) {
+    // flow: 产生该结果的处理管线 ("vlm" 视觉转换 / "text" 文本抽取)。
+    // 同一文件两种模式的产物不同, 命中条件必须匹配; 旧缓存条目无 flow 字段
+    // 视为 "vlm", 保证向后兼容。
+    public string? get_cached_markdown (string abs_path, string current_hash, string flow = "vlm") {
         cache_mutex.lock ();
         string? result = null;
         string rel_path = get_rel_path (abs_path);
@@ -78,7 +81,7 @@ public class PreprocessCache : GLib.Object {
         if (manifest.has_member (rel_path)) {
             var entry = manifest.get_object_member (rel_path);
             string cached_hash = entry.get_string_member_with_default ("hash", "");
-            if (cached_hash == current_hash) {
+            if (cached_hash == current_hash && entry_flow (entry) == flow) {
                 string? md_path = resolve_md_path (entry, cached_hash);
                 if (md_path != null && FileUtils.test (md_path, FileTest.EXISTS)) {
                     try {
@@ -112,7 +115,7 @@ public class PreprocessCache : GLib.Object {
      * 旧版缓存条目未记录 quick 字段时会 miss, 此时调用方回退到 get_cached_markdown
      * (走 SHA256 路径) 以兼容既有磁盘缓存.
      */
-    public string? get_cached_markdown_quick (string abs_path, string quick_fp) {
+    public string? get_cached_markdown_quick (string abs_path, string quick_fp, string flow = "vlm") {
         cache_mutex.lock ();
         string? result = null;
         string rel_path = get_rel_path (abs_path);
@@ -120,7 +123,7 @@ public class PreprocessCache : GLib.Object {
         if (manifest.has_member (rel_path)) {
             var entry = manifest.get_object_member (rel_path);
             string cached_quick = entry.get_string_member_with_default ("quick", "");
-            if (cached_quick == quick_fp) {
+            if (cached_quick == quick_fp && entry_flow (entry) == flow) {
                 string? md_path = resolve_md_path (entry, entry.get_string_member_with_default ("hash", ""));
                 if (md_path != null && FileUtils.test (md_path, FileTest.EXISTS)) {
                     try {
@@ -136,7 +139,7 @@ public class PreprocessCache : GLib.Object {
         return result;
     }
 
-    public void save_markdown (string abs_path, string current_hash, string markdown_content) {
+    public void save_markdown (string abs_path, string current_hash, string markdown_content, string flow = "vlm") {
         string rel_path = get_rel_path (abs_path);
         // 扁平结构: 以 hash 命名的独立子文件夹 cache_dir/{hash}/, 内含 content.md 与
         // (可选) imgs/; markdown 内相对路径 "imgs/xxx.jpg" 相对 {hash}/ 基点, 可正确
@@ -154,6 +157,8 @@ public class PreprocessCache : GLib.Object {
             entry.set_string_member ("quick", compute_file_hash_fast (abs_path));
             // 新结构标记: 缓存子文件夹名 (= hash), 读取时据此定位独立目录
             entry.set_string_member ("cache_subdir", current_hash);
+            // 产生该结果的处理管线 ("vlm"/"text"), 命中条件按 flow 匹配
+            entry.set_string_member ("flow", flow);
             entry.set_boolean_member ("has_images", has_imgs_for (current_hash));
             entry.set_int_member ("timestamp", new DateTime.now_utc ().to_unix ());
             manifest.set_member (rel_path, AI.SchemaHelper.obj_to_node (entry));
@@ -164,6 +169,11 @@ public class PreprocessCache : GLib.Object {
         cache_mutex.unlock ();
     }
 
+    // 条目的 flow 归一: 旧条目无该字段视为 "vlm" (历史行为即视觉转换)
+    private static string entry_flow (Json.Object entry) {
+        return entry.get_string_member_with_default ("flow", "vlm");
+    }
+
     // 将缓存图片落盘到 cache_dir/{hash}/imgs/<relpath>, 供 markdown 内相对路径引用.
     // 调用时机: PaddleOCR 解析完成后, 逐张下载并写盘. 同时把 has_images 标记写回 manifest.
     public void save_image (string current_hash, string relpath, uint8[] data) {
diff --git a/src/services/vlm_queue.vala b/src/services/vlm_queue.vala
index fe088c7..38e3757 100644
--- a/src/services/vlm_queue.vala
+++ b/src/services/vlm_queue.vala
@@ -11,11 +11,13 @@ public class VLMQueueManager : GLib.Object {
     public signal void state_changed (bool has_tasks);
     // 工作线程的结果回送信号. 主线程接收后, 在 items 中按 file_path 查找 ItemData
     // 并更新其属性. md 为 null 表示无内容更新 (PROCESSING/FAILED 阶段).
+    // error_message 仅在 status == FAILED 时有意义, 为用户可读的失败原因.
     public signal void task_completed (
         string file_path,
         string? preprocessed_content,
         PreprocessStatus status,
-        bool from_cache
+        bool from_cache,
+        string? error_message
     );
 
     private Gee.LinkedList<string> pending_queue;
@@ -182,8 +184,9 @@ public class VLMQueueManager : GLib.Object {
             }
         } catch (ThreadError e) {
             warning ("Failed to create VLM thread: %s", e.message);
+            string err = e.message;
             Idle.add (() => {
-                task_completed (file_path, null, PreprocessStatus.FAILED, false);
+                task_completed (file_path, null, PreprocessStatus.FAILED, false, err);
                 finish_task (file_path);
                 return Source.REMOVE;
             });
diff --git a/src/services/vlm_task_runner.vala b/src/services/vlm_task_runner.vala
index ecc4632..a4c4cac 100644
--- a/src/services/vlm_task_runner.vala
+++ b/src/services/vlm_task_runner.vala
@@ -104,7 +104,9 @@ public class VLMTaskRunner : GLib.Object {
         // 更新了 vlm_runner.work_dir, local_work_dir 持有的 GFile 也不会被释放.
         File? local_work_dir = work_dir;
 
-        // 1. 检查缓存
+        // 1. 检查缓存 (flow 匹配当前处理模式: vlm/text 两种产物互不命中)
+        string flow = (ConfigManager.get_binary_process_mode () == ConfigManager.BINARY_MODE_TEXT)
+            ? "text" : "vlm";
         string? cached_md = null;
         string hash = "";
         bool hash_valid = false;
@@ -113,11 +115,11 @@ public class VLMTaskRunner : GLib.Object {
             string quick = PreprocessCache.compute_file_hash_fast (file_path);
             if (local_work_dir != null) {
                 var cache = new PreprocessCache (local_work_dir.get_path ());
-                cached_md = cache.get_cached_markdown_quick (file_path, quick);
+                cached_md = cache.get_cached_markdown_quick (file_path, quick, flow);
                 if (cached_md == null) {
                     hash = PreprocessCache.compute_file_hash (file_path);
                     hash_valid = true;
-                    cached_md = cache.get_cached_markdown (file_path, hash);
+                    cached_md = cache.get_cached_markdown (file_path, hash, flow);
                 }
             } else {
                 hash = PreprocessCache.compute_file_hash (file_path);
@@ -135,7 +137,7 @@ public class VLMTaskRunner : GLib.Object {
         if (cached_md != null) {
             string md = cached_md;
             Idle.add (() => {
-                manager.task_completed (file_path, md, PreprocessStatus.COMPLETED, true);
+                manager.task_completed (file_path, md, PreprocessStatus.COMPLETED, true, null);
                 return Source.REMOVE;
             });
             manager.notify_finished (file_path);
@@ -144,10 +146,28 @@ public class VLMTaskRunner : GLib.Object {
 
         // 2. 标记为 PROCESSING (主线程接收信号后更新 UI)
         Idle.add (() => {
-            manager.task_completed (file_path, null, PreprocessStatus.PROCESSING, false);
+            manager.task_completed (file_path, null, PreprocessStatus.PROCESSING, false, null);
             return Source.REMOVE;
         });
 
+        // 1.5 "文本抽取" 模式: 支持的格式直接本地抽纯文本, 不需要 VLM 凭据与
+        //     网络。内部自行判断模式 (vlm 模式直接返回 null); 抽取失败 (含
+        //     扫描版 PDF 抽空) 返回 null, 继续走下方视觉转换/OCR 流程兜底。
+        {
+            PreprocessCache? cache = (local_work_dir != null)
+                ? new PreprocessCache (local_work_dir.get_path ()) : null;
+            string? text_md = BinaryPreprocessor.try_extract_text_document (
+                file_path, hash_valid ? hash : null, cache);
+            if (text_md != null) {
+                Idle.add (() => {
+                    manager.task_completed (file_path, text_md, PreprocessStatus.COMPLETED, false, null);
+                    return Source.REMOVE;
+                });
+                manager.notify_finished (file_path);
+                return;
+            }
+        }
+
         var settings = ConfigManager.load_multimodal_ai_settings ();
 
         // 校验启用状态与对应服务商所需凭据
@@ -155,8 +175,11 @@ public class VLMTaskRunner : GLib.Object {
             ? settings.paddleocr_token.length > 0
             : settings.api_key.length > 0;
         if (!settings.enabled || !has_credential) {
+            string why = (!settings.enabled)
+                ? _("Multimodal AI is disabled in settings")
+                : _("Multimodal AI credential is empty, cannot convert binary files");
             Idle.add (() => {
-                manager.task_completed (file_path, null, PreprocessStatus.FAILED, false);
+                manager.task_completed (file_path, null, PreprocessStatus.FAILED, false, why);
                 return Source.REMOVE;
             });
             manager.notify_finished (file_path);
@@ -170,10 +193,7 @@ public class VLMTaskRunner : GLib.Object {
                 // ── PaddleOCR 云端路径 ──
                 // 上传原文件 (PDF/图片直传, Office 先转 PDF), 轮询并合并 Markdown.
                 bool is_temp;
-                string? upload_source = BinaryConverter.resolve_upload_source (file_path, out is_temp);
-                if (upload_source == null) {
-                    throw new IOError.FAILED (_("Failed to prepare upload source for %s").printf (file_path));
-                }
+                string upload_source = BinaryConverter.resolve_upload_source (file_path, out is_temp);
                 try {
                     // 仅在哈希有效且有工作目录时构造缓存实例, 供配图落盘与 md 保存复用
                     PreprocessCache? pd_cache = (local_work_dir != null && hash_valid)
@@ -197,12 +217,13 @@ public class VLMTaskRunner : GLib.Object {
 
                 if (ItemData.is_image_file (file_path)) {
                     string? b64 = BinaryConverter.convert_image_to_base64 (file_path);
-                    if (b64 == null) throw new IOError.FAILED ("Image load failed");
+                    if (b64 == null) throw new IOError.FAILED (_("Image load failed: %s").printf (file_path));
                     base64_images = { b64 };
                     mime_types = { BinaryConverter.get_output_mime_for_image (file_path) };
                 } else {
+                    // soffice/pdftoppm 缺失或转换失败时抛出带指引的错误
                     string[]? images = BinaryConverter.convert_to_base64_images (file_path);
-                    if (images == null) throw new IOError.FAILED ("Document render failed");
+                    if (images == null) throw new IOError.FAILED (_("Document render failed: %s").printf (file_path));
                     base64_images = images;
                     mime_types = new string[images.length];
                     for (int i = 0; i < images.length; i++) mime_types[i] = "image/png";
@@ -230,13 +251,15 @@ public class VLMTaskRunner : GLib.Object {
             }
 
             Idle.add (() => {
-                manager.task_completed (file_path, md, PreprocessStatus.COMPLETED, false);
+                manager.task_completed (file_path, md, PreprocessStatus.COMPLETED, false, null);
                 return Source.REMOVE;
             });
         } catch (Error e) {
             warning ("VLM Task failed: %s", e.message);
+            // 失败原因透传给 UI: 队列行/预览占位/tooltip 展示, 不再丢弃 e.message
+            string fail_msg = e.message;
             Idle.add (() => {
-                manager.task_completed (file_path, null, PreprocessStatus.FAILED, false);
+                manager.task_completed (file_path, null, PreprocessStatus.FAILED, false, fail_msg);
                 return Source.REMOVE;
             });
         }
diff --git a/src/services/zip_exporter.vala b/src/services/zip_exporter.vala
index 5999cfd..b13a4aa 100644
--- a/src/services/zip_exporter.vala
+++ b/src/services/zip_exporter.vala
@@ -10,7 +10,8 @@ using Gee;
 //
 // 设计上走外部 `zip` 命令而不是 libzip 是因为:
 //   - 项目已有 git/soffice/pdftoppm 等用 Process.spawn_sync 的先例
-//   - 零额外依赖 (系统装 libreoffice 时通常也带了 zip)
+//   - 零额外依赖 (Linux 上系统装 libreoffice 时通常也带了 zip;
+//     Windows 便携包由 tools/build-windows-package.sh 捆绑 zip.exe)
 //   - 不会因为 libzip 版本差异出 ABI 兼容问题
 public class ZipExporter : GLib.Object {
 
@@ -135,16 +136,27 @@ public class ZipExporter : GLib.Object {
     }
 
     private static string compute_relative_path (string abs_path, File? work_dir) {
+        if (work_dir != null && work_dir.get_path () != null) {
+            // GIO 的相对路径计算按规范路径比较, Windows 反斜杠与 Unix 正斜杠
+            // 皆可 (原实现 has_prefix (wd + "/") 在 Windows 上恒不命中,
+            // 导致所有文件都落入 _external/)
+            var rel = work_dir.get_relative_path (File.new_for_path (abs_path));
+            if (rel != null) {
+                return rel.replace ("\\", "/");
+            }
+        }
         if (work_dir == null) {
-            // 没有工作目录: 用去掉前导 / 的全路径, 避免同名冲突
-            var p = abs_path;
+            // 没有工作目录: 用去掉前导分隔符的全路径, 避免同名冲突。
+            // Windows 的盘符冒号必须一并去掉, 否则 "C:/..." 会被
+            // Path.build_filename 当作绝对路径 (最后的绝对分量胜出),
+            // dest 直接指回源文件, 复制自身必然失败。
+            var p = abs_path.replace ("\\", "/");
             while (p.has_prefix ("/")) p = p.substring (1);
+#if WINDOWS
+            p = p.replace (":", "");
+#endif
             return p;
         }
-        var wd = work_dir.get_path ();
-        if (wd != null && abs_path.has_prefix (wd + "/")) {
-            return abs_path.substring (wd.length + 1);
-        }
         // 文件在工作目录外: 用 _external/<basename>-<hash> 隔离
         // 注释之前写"<basename>-<hash>"但代码未实现 hash, 同名外部文件
         // (如 /tmp/notes.txt 与 /home/notes.txt) 会相互覆盖。
diff --git a/src/utils/glob_helper.vala b/src/utils/glob_helper.vala
index 5133afa..08a547e 100644
--- a/src/utils/glob_helper.vala
+++ b/src/utils/glob_helper.vala
@@ -3,11 +3,13 @@ using Gee;
 public class GlobHelper : GLib.Object {
 
     public static bool match_glob (string pattern, string text) {
-        if (!pattern.contains ("**")) {
-            var spec = new PatternSpec (pattern.down ());
+        // 归一分隔符: Windows 上 AI 给出的模式与路径可能是反斜杠形式
+        string norm_pattern = pattern.replace ("\\", "/");
+        if (!norm_pattern.contains ("**")) {
+            var spec = new PatternSpec (norm_pattern.down ());
             return spec.match_string (text.down ());
         }
-        return match_glob_recursive (pattern.down (), text.down ());
+        return match_glob_recursive (norm_pattern.down (), text.down ());
     }
 
     private static bool match_glob_recursive (string pattern, string text) {
@@ -99,7 +101,9 @@ public class GlobHelper : GLib.Object {
                     }
                 }
                 var child_path = Path.build_filename (current_dir, info.get_name ());
-                string rel = child_path.substring (base_dir.length);
+                // 归一分隔符: Windows 上 build_filename 产出反斜杠路径, 前导 "\"
+                // 剥不掉, 且 "**/*.py" 这类按 "/" 分段的模式永不命中
+                string rel = child_path.substring (base_dir.length).replace ("\\", "/");
                 if (rel.has_prefix ("/")) rel = rel.substring (1);
 
                 if (match_glob (pattern, rel)) {
diff --git a/src/utils/platform.vala b/src/utils/platform.vala
index 8c40c7a..ba963c7 100644
--- a/src/utils/platform.vala
+++ b/src/utils/platform.vala
@@ -65,6 +65,39 @@ namespace Platform {
     private extern static int _NSGetExecutablePath ([CCode (array_length = false)] uint8[] buf, ref uint32 bufsize);
 #endif
 
+    // ─── 捆绑外部工具解析 ────────────────────────────────────────────
+    // Windows 便携包把 poppler 工具 (pdftoppm/pdftotext) 放在主程序同目录
+    // (bin/); 绝对路径优先, 找不到回退裸名 PATH 搜索 (Linux/macOS 现状)。
+    // 放在 Platform 是为了让不依赖 GDK 的单元测试 (OfficeTextExtractor)
+    // 也能链接这两个解析逻辑, 避免 binary_converter 的 Gdk.Pixbuf 依赖
+    // 把 gtk4 拉进测试闭包。
+
+    // 解析捆绑的外部工具, 找不到返回 null (调用方回退裸名 + SpawnFlags.SEARCH_PATH)。
+    public static string? resolve_bundled_tool (string exe_name) {
+        string exe = get_executable_path ();
+        if (exe == "." || exe.length == 0) return null;
+        var candidate = Path.build_filename (Path.get_dirname (exe), exe_name);
+        return FileUtils.test (candidate, FileTest.EXISTS) ? candidate : null;
+    }
+
+    // poppler 工具专用 launcher: 中文 PDF 需要 CMap 数据, 运行时把
+    // POPPLER_DATADIR 指向便携包内 share/poppler (打包脚本已捆绑)。
+    public static SubprocessLauncher new_poppler_launcher () throws Error {
+        var launcher = new SubprocessLauncher (
+            SubprocessFlags.STDOUT_PIPE | SubprocessFlags.STDERR_PIPE
+        );
+        string exe = get_executable_path ();
+        if (exe != "." && exe.length > 0) {
+            var data_dir = Path.build_filename (
+                Path.get_dirname (exe), "..", "share", "poppler");
+            if (FileUtils.test (data_dir, FileTest.IS_DIR)) {
+                launcher.setenv ("POPPLER_DATADIR", data_dir, true);
+            }
+        }
+        return launcher;
+    }
+
+
     /**
      * 应用数据目录 (gtksourceview 主题等)。
      * 解析优先级:
diff --git a/src/widgets/context_menus.vala b/src/widgets/context_menus.vala
index 0aab322..9745700 100644
--- a/src/widgets/context_menus.vala
+++ b/src/widgets/context_menus.vala
@@ -83,7 +83,9 @@ public class ContextMenus : GLib.Object {
                 can_retry_vlm = false;
             }
             if (it.item_type == "file" && work_dir != null && it.file_path != null) {
-                if (!it.file_path.has_prefix (work_dir.get_path () + "/")) {
+                // GIO 的 File.has_prefix 按规范路径比较, 跨平台兼容两种分隔符
+                // (原 has_prefix (wd + "/") 字符串手术在 Windows 上恒不命中)
+                if (!File.new_for_path (it.file_path).has_prefix (work_dir)) {
                     has_external = true;
                 }
             }
diff --git a/src/widgets/preferences_dialog.vala b/src/widgets/preferences_dialog.vala
index 99eb5bc..ac53a37 100644
--- a/src/widgets/preferences_dialog.vala
+++ b/src/widgets/preferences_dialog.vala
@@ -28,8 +28,10 @@ public class PreferencesDialog : GLib.Object {
 
     private Gee.ArrayList<ConfigManager.VLMProfile> vlm_profiles;
     private bool switching_vlm_profile = false;
+    private bool switching_mm_mode = false;
 
     private Switch chk_mm_enabled;
+    private Adw.ComboRow combo_mm_process_mode;
     private Adw.ComboRow combo_mm_provider;
     private Adw.ComboRow combo_vlm_profile;
     private EntryRow edit_mm_base_url;
@@ -228,6 +230,31 @@ public class PreferencesDialog : GLib.Object {
         mm_enabled_row.set_activatable_widget (chk_mm_enabled);
         mm_group.add (mm_enabled_row);
 
+        // ── 二进制处理方式下拉: 视觉转换 (默认, 保图表/公式/表格) / 文本抽取
+        //    (快速离线, 无需凭据, 但丢失图表/公式/嵌入图片)。立即持久化,
+        //    两条管线在下次预处理时读取。
+        var process_mode_model = new StringList (new string[] {
+            _("Visual conversion (keeps charts, formulas, tables)"),
+            _("Text extraction (fast, offline, text only)")
+        });
+        combo_mm_process_mode = new Adw.ComboRow ();
+        combo_mm_process_mode.set_title (_("Processing Mode"));
+        combo_mm_process_mode.set_subtitle (_(
+            "Visual conversion renders documents to images and needs LibreOffice for Office files; text extraction parses Office/PDF locally without AI credentials."));
+        combo_mm_process_mode.set_model (process_mode_model);
+        mm_group.add (combo_mm_process_mode);
+        switching_mm_mode = true;
+        combo_mm_process_mode.set_selected (
+            ConfigManager.get_binary_process_mode () == ConfigManager.BINARY_MODE_TEXT ? 1 : 0);
+        switching_mm_mode = false;
+        combo_mm_process_mode.notify["selected"].connect (() => {
+            if (switching_mm_mode) return;
+            ConfigManager.save_binary_process_mode (
+                combo_mm_process_mode.get_selected () == 1
+                    ? ConfigManager.BINARY_MODE_TEXT : ConfigManager.BINARY_MODE_VLM);
+            ai_settings_changed ();
+        });
+
         // ── VLM 模型配置方案下拉: 预置存于配置文件 vlm_models 数组, 双向同步
         var mm_profile_row = new ComboRow ();
         mm_profile_row.set_title (_("Model Profile"));
diff --git a/src/widgets/queue_list_factory.vala b/src/widgets/queue_list_factory.vala
index f752818..8914f6b 100644
--- a/src/widgets/queue_list_factory.vala
+++ b/src/widgets/queue_list_factory.vala
@@ -206,11 +206,19 @@ public class QueueListFactory : GLib.Object {
                 }
             });
 
+            // 监听失败原因变化: FAILED 状态下行内显示简短原因
+            ulong error_handler_id = data.notify["preprocess-error"].connect (() => {
+                if (list_item != null && list_item.get_item () != null) {
+                    render_row (list_item, data, label, icon);
+                }
+            });
+
             // 将句柄 ID 与所监视的数据模型指针弱挂载到 ListItem 容器上,
             // 供 unbind 时双重校验安全剥离信号
             list_item.set_data<ulong> ("content_notify_id", handler_id);
             list_item.set_data<ulong> ("status_notify_id", status_handler_id);
             list_item.set_data<ulong> ("cache_notify_id", cache_handler_id);
+            list_item.set_data<ulong> ("error_notify_id", error_handler_id);
             list_item.set_data<ulong> ("position_notify_id", pos_handler);
             list_item.set_data<ItemData> ("monitored_data_ptr", data);
         });
@@ -222,6 +230,7 @@ public class QueueListFactory : GLib.Object {
             ulong handler_id = list_item.get_data<ulong> ("content_notify_id");
             ulong status_handler_id = list_item.get_data<ulong> ("status_notify_id");
             ulong cache_handler_id = list_item.get_data<ulong> ("cache_notify_id");
+            ulong error_handler_id = list_item.get_data<ulong> ("error_notify_id");
             ulong pos_handler = list_item.get_data<ulong> ("position_notify_id");
             var data = list_item.get_data<ItemData> ("monitored_data_ptr");
 
@@ -235,6 +244,9 @@ public class QueueListFactory : GLib.Object {
             if (cache_handler_id != 0 && data != null && GLib.SignalHandler.is_connected (data, cache_handler_id)) {
                 GLib.SignalHandler.disconnect (data, cache_handler_id);
             }
+            if (error_handler_id != 0 && data != null && GLib.SignalHandler.is_connected (data, error_handler_id)) {
+                GLib.SignalHandler.disconnect (data, error_handler_id);
+            }
             if (pos_handler != 0 && GLib.SignalHandler.is_connected (list_item, pos_handler)) {
                 GLib.SignalHandler.disconnect (list_item, pos_handler);
             }
@@ -243,6 +255,7 @@ public class QueueListFactory : GLib.Object {
             list_item.set_data ("content_notify_id", null);
             list_item.set_data ("status_notify_id", null);
             list_item.set_data ("cache_notify_id", null);
+            list_item.set_data ("error_notify_id", null);
             list_item.set_data ("position_notify_id", null);
             list_item.set_data ("monitored_data_ptr", null);
             var unbind_box = list_item.get_child () as Gtk.Box;
@@ -299,6 +312,12 @@ public class QueueListFactory : GLib.Object {
                             break;
                         case PreprocessStatus.FAILED:
                             display_name += _(" [Conversion failed]");
+                            if (data.preprocess_error != null && data.preprocess_error.length > 0) {
+                                // 队列行空间有限: 单行 + 截断, 完整原因在预览/tooltip
+                                string reason = data.preprocess_error.replace ("\n", " ").strip ();
+                                if (reason.length > 50) reason = reason.substring (0, 50) + "…";
+                                display_name += " " + reason;
+                            }
                             break;
                     }
                 }
diff --git a/src/widgets/vlm_queue_controller.vala b/src/widgets/vlm_queue_controller.vala
index a36783e..6b09cc5 100644
--- a/src/widgets/vlm_queue_controller.vala
+++ b/src/widgets/vlm_queue_controller.vala
@@ -102,7 +102,9 @@ public class VlmQueueController : GLib.Object {
 
     // ─── 内部 ──────────────────────────────────────────────────────────
 
-    private void on_task_completed (string file_path, string? md, PreprocessStatus status, bool from_cache) {
+    private void on_task_completed (
+        string file_path, string? md, PreprocessStatus status, bool from_cache, string? error_message
+    ) {
         var items = app_state.items;
         for (int i = 0; i < items.size; i++) {
             var it = items.get (i);
@@ -110,6 +112,8 @@ public class VlmQueueController : GLib.Object {
                 if (md != null) it.preprocessed_content = md;
                 it.preprocess_status = status;
                 it.from_cache = from_cache;
+                // FAILED 时带用户可读原因; 其它状态清空旧原因防止残留
+                it.preprocess_error = (status == PreprocessStatus.FAILED) ? error_message : null;
                 // 合并刷新: 批量预处理完成时每个文件完成都会触发, 直接 refresh_list
                 // 会形成 N 次 O(n) 扫描 + 末尾预览级联. 委托给 Window 合并到 150ms 窗口
                 // 只刷新一次.
diff --git a/src/win32_console_shim.c b/src/win32_console_shim.c
new file mode 100644
index 0000000..10342f4
--- /dev/null
+++ b/src/win32_console_shim.c
@@ -0,0 +1,24 @@
+#include "win32_console_shim.h"
+
+#include <stdio.h>
+
+void win32_console_init (void) {
+    /* stdout 已有有效句柄时 (Git Bash / CI 的管道、文件重定向、已附加的
+     * 控制台) 不做任何事, 保住管道输出。GUI 子系统直启时 std 句柄为空,
+     * GetFileType 返回 FILE_TYPE_UNKNOWN 并置 ERROR_INVALID_HANDLE。 */
+    SetLastError (0);
+    DWORD out_type = GetFileType (GetStdHandle (STD_OUTPUT_HANDLE));
+    if (!(out_type == FILE_TYPE_UNKNOWN && GetLastError () == ERROR_INVALID_HANDLE)) {
+        return;
+    }
+
+    /* 父进程没有控制台 (双击 / start 启动) 时附加失败, 直接返回, GUI 照常
+     * 运行; 有控制台 (cmd / PowerShell 跑 CLI) 时接管 std 流。 */
+    if (!AttachConsole (ATTACH_PARENT_PROCESS)) {
+        return;
+    }
+    freopen ("CONOUT$", "w", stdout);
+    freopen ("CONOUT$", "w", stderr);
+    /* 程序输出恒为 UTF-8, 控制台默认码页 (中文系统 CP936) 下会乱码 */
+    SetConsoleOutputCP (CP_UTF8);
+}
diff --git a/src/win32_console_shim.h b/src/win32_console_shim.h
new file mode 100644
index 0000000..711f3f4
--- /dev/null
+++ b/src/win32_console_shim.h
@@ -0,0 +1,11 @@
+#ifndef FILECOLLECTOR_WIN32_CONSOLE_SHIM_H
+#define FILECOLLECTOR_WIN32_CONSOLE_SHIM_H
+
+#include <windows.h>
+
+/* CLI 输出接管: exe 为 GUI 子系统 (GUI 启动不弹黑框), 在 cmd/PowerShell
+ * 里直接跑 CLI 时 stdout 无句柄, 输出完全不可见。此函数在检测到该场景时
+ * 附加父进程控制台并接管 std 流。Git Bash / CI 等管道环境不做任何事。 */
+void win32_console_init (void);
+
+#endif
diff --git a/src/window.vala b/src/window.vala
index f85b989..31b86e0 100644
--- a/src/window.vala
+++ b/src/window.vala
@@ -2683,12 +2683,17 @@ public class FileCollectorWindow : Adw.ApplicationWindow {
 
     // 确保指定文件路径的所有父目录都已加载到树中
     private void ensure_path_loaded (string abs_path) {
-        if (work_dir == null || !abs_path.has_prefix (work_dir.get_path () + "/")) return;
+        if (work_dir == null) return;
+        // Windows 上 get_path() 返回反斜杠路径, 归一为 "/" 后再做前缀比较
+        // 与层级切分, 否则懒加载判断恒不命中
+        string wd_path = (work_dir.get_path () ?? "").replace ("\\", "/");
+        string norm = abs_path.replace ("\\", "/");
+        if (!norm.has_prefix (wd_path + "/")) return;
         if (root_store.get_n_items () == 0) return;
         ensure_path_token++;
         uint my_token = ensure_path_token;
 
-        string rel = abs_path.substring (work_dir.get_path ().length + 1);
+        string rel = norm.substring (wd_path.length + 1);
         string[] parts = rel.split ("/");
         var current = (DirectoryItem) root_store.get_item (0);
 
@@ -3552,6 +3557,11 @@ public class FileCollectorWindow : Adw.ApplicationWindow {
                 case PreprocessStatus.FAILED:
                     show_action_bar = true;
                     btn_retry_preprocess.tooltip_text = _("AI conversion failed\nClick to retry conversion");
+                    if (item.preprocess_error != null && item.preprocess_error.length > 0) {
+                        string reason = item.preprocess_error.replace ("\n", " ");
+                        if (reason.length > 120) reason = reason.substring (0, 120) + "…";
+                        btn_retry_preprocess.tooltip_text += "\n\n" + reason;
+                    }
                     break;
                 case PreprocessStatus.PROCESSING:
                     btn_retry_preprocess.tooltip_text = _("Processing...");
@@ -3585,7 +3595,15 @@ public class FileCollectorWindow : Adw.ApplicationWindow {
                     apply_preview_content_full (item, _("Processing..."));
                     break;
                 case PreprocessStatus.FAILED:
-                    apply_preview_content_full (item, _("[AI conversion failed; click the retry button on the toolbar to reconvert]"));
+                    {
+                        string placeholder = _("[AI conversion failed; click the retry button on the toolbar to reconvert]");
+                        if (item.preprocess_error != null && item.preprocess_error.length > 0) {
+                            string reason = item.preprocess_error.replace ("\n", " ");
+                            if (reason.length > 200) reason = reason.substring (0, 200) + "…";
+                            placeholder += "\n" + _("Reason: %s").printf (reason);
+                        }
+                        apply_preview_content_full (item, placeholder);
+                    }
                     break;
                 case PreprocessStatus.CHECKING:
                     apply_preview_content_full (item, _("[Checking cache...]"));
@@ -3986,6 +4004,7 @@ public class FileCollectorWindow : Adw.ApplicationWindow {
 
         item.preprocessed_content = null;
         item.from_cache = false;
+        item.preprocess_error = null;
         item.preprocess_status = PreprocessStatus.PENDING;
 
         refresh_list ();
@@ -4019,6 +4038,7 @@ public class FileCollectorWindow : Adw.ApplicationWindow {
                         item.preprocess_status = PreprocessStatus.NONE;
                         item.preprocessed_content = null;
                         item.from_cache = false;
+                        item.preprocess_error = null;
                     }
                 }
 
@@ -4057,9 +4077,11 @@ public class FileCollectorWindow : Adw.ApplicationWindow {
         if (ctx_item == null || ctx_item.file_path == null) return;
         string path_to_copy = ctx_item.file_path;
         if (!ctx_item.force_absolute && work_dir != null && !use_absolute) {
-            string wd = work_dir.get_path () + "/";
-            if (path_to_copy.has_prefix (wd)) {
-                path_to_copy = path_to_copy.substring (wd.length);
+            // 归一分隔符: Windows 反斜杠路径对 wd + "/" 前缀恒不命中
+            string wd = (work_dir.get_path () ?? "").replace ("\\", "/") + "/";
+            string norm = path_to_copy.replace ("\\", "/");
+            if (norm.has_prefix (wd)) {
+                path_to_copy = norm.substring (wd.length);
             }
         }
         get_clipboard ().set_text (path_to_copy);
@@ -4149,9 +4171,10 @@ public class FileCollectorWindow : Adw.ApplicationWindow {
         if (ctx_tree_item == null) return;
         string path_to_copy = ctx_tree_item.path;
         if (work_dir != null) {
-            string wd = work_dir.get_path () + "/";
-            if (path_to_copy.has_prefix (wd)) {
-                path_to_copy = path_to_copy.substring (wd.length);
+            string wd = (work_dir.get_path () ?? "").replace ("\\", "/") + "/";
+            string norm = path_to_copy.replace ("\\", "/");
+            if (norm.has_prefix (wd)) {
+                path_to_copy = norm.substring (wd.length);
             }
         }
         get_clipboard ().set_text (path_to_copy);
@@ -4329,9 +4352,10 @@ public class FileCollectorWindow : Adw.ApplicationWindow {
         if (ctx_preview_item == null || ctx_preview_item.file_path == null) return;
         string path_to_copy = ctx_preview_item.file_path;
         if (work_dir != null) {
-            string wd = work_dir.get_path () + "/";
-            if (path_to_copy.has_prefix (wd)) {
-                path_to_copy = path_to_copy.substring (wd.length);
+            string wd = (work_dir.get_path () ?? "").replace ("\\", "/") + "/";
+            string norm = path_to_copy.replace ("\\", "/");
+            if (norm.has_prefix (wd)) {
+                path_to_copy = norm.substring (wd.length);
             }
         }
         get_clipboard ().set_text (path_to_copy);
@@ -5100,8 +5124,10 @@ public class FileCollectorWindow : Adw.ApplicationWindow {
 
     private string get_display_path (ItemData data) {
         if (data.force_absolute || use_absolute || work_dir == null) return data.file_path;
-        var wd_path = work_dir.get_path () + "/";
-        if (data.file_path.has_prefix (wd_path)) return data.file_path.substring (wd_path.length);
+        // 归一分隔符, Windows 反斜杠路径下前缀判断恒不命中
+        var wd_path = (work_dir.get_path () ?? "").replace ("\\", "/") + "/";
+        var norm = data.file_path.replace ("\\", "/");
+        if (norm.has_prefix (wd_path)) return norm.substring (wd_path.length);
         return data.file_path;
     }
 
diff --git a/tests/meson.build b/tests/meson.build
index 9990422..3b5d2e5 100644
--- a/tests/meson.build
+++ b/tests/meson.build
@@ -170,3 +170,16 @@ test_search_query_parser_exe = executable('test_search_query_parser',
   dependencies: [glib_dep, gio_dep, gee_dep],
 )
 test('test_search_query_parser', test_search_query_parser_exe)
+
+# ─── 12. OfficeTextExtractor (纯 Vala ZIP/XML 解析; PDF 分支依赖外部
+#     pdftotext, 属集成验证不在此覆盖). poppler 工具解析位于 platform
+#     (无 GDK 依赖), 因此测试闭包不含 binary_converter / gtk4.
+office_text_extractor_src = files(join_paths(src_dir, 'services', 'office_text_extractor.vala'))
+platform_src              = files(join_paths(src_dir, 'utils',    'platform.vala'))
+test_office_text_extractor_exe = executable('test_office_text_extractor',
+  'test_office_text_extractor.vala',
+  office_text_extractor_src,
+  platform_src,
+  dependencies: [glib_dep, gio_dep, gee_dep],
+)
+test('test_office_text_extractor', test_office_text_extractor_exe)
diff --git a/tests/test_office_text_extractor.vala b/tests/test_office_text_extractor.vala
new file mode 100644
index 0000000..d98ed71
--- /dev/null
+++ b/tests/test_office_text_extractor.vala
@@ -0,0 +1,323 @@
+// OfficeTextExtractor 单元测试: 内存构造 ZIP fixture, 覆盖
+// STORED/DEFLATE 两种压缩、docx/xlsx/pptx/odt 抽取、实体解码、
+// 工作表/幻灯片自然排序、EOCD 注释容错、ZIP64/加密/垃圾数据的明确报错。
+// PDF 分支依赖外部 pdftotext, 不在此覆盖 (属于集成验证)。
+//
+// fixture 侧的 ZIP 写入器是测试专用的最小实现: 只写 reader 关心的字段,
+// CRC 一律置 0 (OfficeTextExtractor.ZipReader 不校验 CRC)。
+
+using Gee;
+
+static int failures = 0;
+
+static void check (bool cond, string what) {
+    if (cond) {
+        print ("PASS: %s\n", what);
+    } else {
+        print ("FAIL: %s\n", what);
+        failures++;
+    }
+}
+
+delegate void TrialFunc () throws Error;
+
+static void run (string name, TrialFunc f) {
+    try {
+        f ();
+    } catch (Error e) {
+        print ("FAIL: %s (unexpected throw: %s)\n", name, e.message);
+        failures++;
+    }
+}
+
+// ─── ZIP fixture 写入器 ──────────────────────────────────────────────
+
+class ZipBuilder : GLib.Object {
+    // vala 0.56 的 glib vapi 没有 GLib.BytesBuilder: 用 GLib.ByteArray 累积
+    private GLib.ByteArray body = new GLib.ByteArray ();
+    private uint32 written = 0;
+    private Gee.ArrayList<EntryMeta> entries = new Gee.ArrayList<EntryMeta> ();
+
+    private class EntryMeta : GLib.Object {
+        public string name;
+        public uint16 flags;
+        public uint16 method;
+        public uint32 comp_size;
+        public uint32 uncomp_size;
+        public uint32 offset;
+    }
+
+    public void add_stored (string name, uint8[] content, uint16 flags = 0) {
+        add_entry_raw (name, content, content.length, 0, flags);
+    }
+
+    public void add_deflated (string name, uint8[] content) throws Error {
+        uint8[] comp = deflate_bytes (content);
+        add_entry_raw (name, comp, content.length, 8, 0);
+    }
+
+    private void add_entry_raw (string name, uint8[] stored_data,
+                                uint32 uncomp_size, uint16 method, uint16 flags) {
+        var meta = new EntryMeta ();
+        meta.name = name;
+        meta.flags = flags;
+        meta.method = method;
+        meta.comp_size = (uint32) stored_data.length;
+        meta.uncomp_size = uncomp_size;
+        meta.offset = written;
+        entries.add (meta);
+
+        uint8[] nb = name.data;
+        u32 (0x04034b50u);
+        u16 (20);          // version needed
+        u16 (flags);
+        u16 (method);
+        u16 (0); u16 (0);  // time, date
+        u32 (0);           // crc (reader 不校验)
+        u32 (meta.comp_size);
+        u32 (uncomp_size);
+        u16 ((uint16) nb.length);
+        u16 (0);           // extra len
+        push (nb);
+        push (stored_data);
+    }
+
+    public uint8[] finish () {
+        uint32 cd_offset = written;
+        foreach (var m in entries) {
+            uint8[] nb = m.name.data;
+            u32 (0x02014b50u);
+            u16 (20); u16 (20);  // version made by / needed
+            u16 (m.flags);
+            u16 (m.method);
+            u16 (0); u16 (0);
+            u32 (0);
+            u32 (m.comp_size);
+            u32 (m.uncomp_size);
+            u16 ((uint16) nb.length);
+            u16 (0);             // extra
+            u16 (0);             // comment
+            u16 (0);             // disk start
+            u16 (0);             // internal attrs
+            u32 (0);             // external attrs
+            u32 (m.offset);
+            push (nb);
+        }
+        uint32 cd_size = written - cd_offset;
+        u32 (0x06054b50u);
+        u16 (0); u16 (0);        // disk numbers
+        u16 ((uint16) entries.size);
+        u16 ((uint16) entries.size);
+        u32 (cd_size);
+        u32 (cd_offset);
+        u16 (0);                 // comment len
+        return body.steal ();
+    }
+
+    private void push (uint8[] piece) {
+        body.append (piece);
+        written += (uint32) piece.length;
+    }
+
+    private void u16 (uint16 v) {
+        push (new uint8[] { (uint8) (v & 0xFF), (uint8) ((v >> 8) & 0xFF) });
+    }
+
+    private void u32 (uint32 v) {
+        push (new uint8[] {
+            (uint8) (v & 0xFF), (uint8) ((v >> 8) & 0xFF),
+            (uint8) ((v >> 16) & 0xFF), (uint8) ((v >> 24) & 0xFF)
+        });
+    }
+
+    private static uint8[] deflate_bytes (uint8[] input) throws Error {
+        var comp = new GLib.ZlibCompressor (GLib.ZlibCompressorFormat.RAW);
+        var out_acc = new GLib.ByteArray ();
+        uint8[] out_buf = new uint8[input.length * 2 + 4096];
+        size_t bytes_read;
+        size_t bytes_written;
+        // 第一次: 喂入全部输入
+        comp.convert (input, out_buf, GLib.ConverterFlags.NONE, out bytes_read, out bytes_written);
+        out_acc.append (out_buf[0 : (int) bytes_written]);
+        // INPUT_AT_END 告知输入结束 -> zlib 执行 Z_FINISH 输出终止流并返回 FINISHED
+        // (FLUSH 只是同步刷新内部缓冲, 不会终止 deflate 流)
+        bool finished = false;
+        for (int i = 0; i < 4 && !finished; i++) {
+            var res = comp.convert (new uint8[0], out_buf, GLib.ConverterFlags.INPUT_AT_END,
+                                    out bytes_read, out bytes_written);
+            out_acc.append (out_buf[0 : (int) bytes_written]);
+            finished = (res == GLib.ConverterResult.FINISHED);
+        }
+        if (!finished) {
+            error ("fixture deflate not finished after flush");
+        }
+        return out_acc.steal ();
+    }
+}
+
+// ─── 工具 ────────────────────────────────────────────────────────────
+
+static string write_fixture (string name, uint8[] data) throws Error {
+    string dir = DirUtils.make_tmp ("fc_ext_test_XXXXXX");
+    string path = GLib.Path.build_filename (dir, name);
+    FileUtils.set_data (path, data);
+    return path;
+}
+
+// ─── 测试用例 ────────────────────────────────────────────────────────
+
+static void test_supports () {
+    check (OfficeTextExtractor.supports (".docx"), "supports .docx");
+    check (OfficeTextExtractor.supports (".pdf"), "supports .pdf");
+    check (OfficeTextExtractor.supports (".odp"), "supports .odp");
+    check (!OfficeTextExtractor.supports (".doc"), "rejects legacy .doc");
+    check (!OfficeTextExtractor.supports (".txt"), "rejects .txt");
+}
+
+static void test_docx_stored () throws Error {
+    var xml = """<?xml version="1.0"?><w:document><w:body>
+<w:p><w:r><w:t>标题：项目说明</w:t></w:r></w:p>
+<w:p><w:r><w:t>带有 &amp; 符号与 &lt;标签&gt; 的文本</w:t></w:r></w:p>
+<w:tbl><w:tr><w:tc><w:r><w:t>单元格A</w:t></w:r></w:tc><w:tc><w:r><w:t>单元格B</w:t></w:r></w:tc></w:tr></w:tbl>
+<w:p><w:r><w:t>结尾</w:t></w:r></w:p>
+</w:body></w:document>""";
+    var builder = new ZipBuilder ();
+    builder.add_stored ("word/document.xml", xml.data);
+    builder.add_stored ("[Content_Types].xml", "<Types/>".data);
+    string path = write_fixture ("t.docx", builder.finish ());
+    string md = OfficeTextExtractor.extract_markdown (path);
+    check (md.contains ("标题：项目说明"), "docx: 中文段落提取");
+    check (md.contains ("带有 & 符号与 <标签> 的文本"), "docx: 实体解码");
+    check (md.contains ("单元格A\t单元格B"), "docx: 表格单元格按 \\t 分隔");
+    check (md.contains ("结尾"), "docx: 尾段保留");
+}
+
+static void test_docx_deflate () throws Error {
+    var xml = "<w:document><w:body><w:p><w:r><w:t>压缩条目正文</w:t></w:r></w:p></w:body></w:document>";
+    var builder = new ZipBuilder ();
+    builder.add_deflated ("word/document.xml", xml.data);
+    string path = write_fixture ("t2.docx", builder.finish ());
+    string md = OfficeTextExtractor.extract_markdown (path);
+    check (md.contains ("压缩条目正文"), "docx: DEFLATE 条目可解压");
+}
+
+static void test_xlsx () throws Error {
+    var sst = """<sst><si><t>名称</t></si><si><t>数值列</t></si><si><t>第二张表</t></si><si><t>最后一张</t></si></sst>""";
+    var sheet1 = """<worksheet><sheetData>
+<row r="1"><c r="A1" t="s"><v>0</v></c><c r="B1" t="s"><v>1</v></c></row>
+<row r="2"><c r="A2"><v>42</v></c><c r="B2"><v>3.14</v></c></row>
+</sheetData><rowBreaks/></worksheet>""";
+    var sheet2 = "<worksheet><sheetData><row r=\"1\"><c r=\"A1\" t=\"s\"><v>2</v></c></row></sheetData></worksheet>";
+    var sheet10 = "<worksheet><sheetData><row r=\"1\"><c r=\"A1\" t=\"s\"><v>3</v></c></row></sheetData></worksheet>";
+    var builder = new ZipBuilder ();
+    // 故意乱序加入, 验证 sheet1/sheet2/sheet10 按数字自然排序
+    builder.add_stored ("xl/worksheets/sheet10.xml", sheet10.data);
+    builder.add_stored ("xl/worksheets/sheet2.xml", sheet2.data);
+    builder.add_stored ("xl/worksheets/sheet1.xml", sheet1.data);
+    builder.add_stored ("xl/sharedStrings.xml", sst.data);
+    string path = write_fixture ("t.xlsx", builder.finish ());
+    string md = OfficeTextExtractor.extract_markdown (path);
+    check (md.contains ("名称\t数值列"), "xlsx: sharedStrings 引用解析");
+    check (md.contains ("42\t3.14"), "xlsx: 数值单元格原样输出");
+    check (md.contains ("<rowBreaks>") == false, "xlsx: rowBreaks 未被误当成 row");
+    check (md.contains ("第二张表"), "xlsx: 第二张表存在");
+    int pos_sheet2 = md.index_of ("第二张表");
+    int pos_sheet10 = md.index_of ("最后一张");
+    check (pos_sheet2 >= 0 && pos_sheet10 > pos_sheet2, "xlsx: sheet2 在 sheet10 之前 (自然排序)");
+}
+
+static void test_pptx () throws Error {
+    var s1 = "<p:sld><p:txBody><a:p><a:r><a:t>第一页要点</a:t></a:r></a:p><a:p><a:r><a:t>副要点</a:t></a:r></a:p></p:txBody></p:sld>";
+    var s2 = "<p:sld><p:txBody><a:p><a:r><a:t>第二页内容</a:t></a:r></a:p></p:txBody></p:sld>";
+    var s10 = "<p:sld><p:txBody><a:p><a:r><a:t>第十页内容</a:t></a:r></a:p></p:txBody></p:sld>";
+    var builder = new ZipBuilder ();
+    builder.add_stored ("ppt/slides/slide10.xml", s10.data);
+    builder.add_stored ("ppt/slides/slide1.xml", s1.data);
+    builder.add_stored ("ppt/slides/slide2.xml", s2.data);
+    string path = write_fixture ("t.pptx", builder.finish ());
+    string md = OfficeTextExtractor.extract_markdown (path);
+    check (md.contains ("## Slide 1") && md.contains ("第一页要点"), "pptx: 首页提取");
+    check (md.contains ("副要点"), "pptx: 段落边界换行");
+    int pos2 = md.index_of ("第二页内容");
+    int pos10 = md.index_of ("第十页内容");
+    check (pos2 >= 0 && pos10 > pos2, "pptx: slide2 在 slide10 之前 (自然排序)");
+}
+
+static void test_odt () throws Error {
+    var xml = """<office:document-content><office:body><office:text>
+<text:h>一级标题</text:h><text:p>段落一</text:p>
+<text:p>段落二<text:line-break/>续行</text:p>
+</office:text></office:body></office:document-content>""";
+    var builder = new ZipBuilder ();
+    builder.add_stored ("content.xml", xml.data);
+    string path = write_fixture ("t.odt", builder.finish ());
+    string md = OfficeTextExtractor.extract_markdown (path);
+    check (md.contains ("一级标题"), "odt: 标题提取");
+    check (md.contains ("段落二\n续行"), "odt: line-break 处理");
+}
+
+static void test_eocd_trailing_comment () throws Error {
+    var builder = new ZipBuilder ();
+    builder.add_stored ("word/document.xml",
+        "<w:document><w:body><w:p><w:r><w:t>注释容错</w:t></w:r></w:p></w:body></w:document>".data);
+    uint8[] zip = builder.finish ();
+    // 模拟 ZIP 尾部注释: EOCD 之后附带任意字节
+    uint8[] garbage = "this is a zip comment, not part of the archive".data;
+    uint8[] padded = new uint8[zip.length + garbage.length];
+    for (int i = 0; i < zip.length; i++) padded[i] = zip[i];
+    for (int i = 0; i < garbage.length; i++) padded[zip.length + i] = garbage[i];
+    string path = write_fixture ("t3.docx", padded);
+    string md = OfficeTextExtractor.extract_markdown (path);
+    check (md.contains ("注释容错"), "docx: EOCD 尾部注释容错");
+}
+
+static void expect_not_supported (string name, uint8[] data) throws Error {
+    string path = write_fixture (name, data);
+    bool threw = false;
+    try {
+        OfficeTextExtractor.extract_markdown (path);
+    } catch (Error e) {
+        threw = (e is IOError.NOT_SUPPORTED);
+    }
+    check (threw, name + " → NOT_SUPPORTED");
+}
+
+static void test_error_paths () throws Error {
+    // 完全不是 ZIP 的垃圾数据
+    expect_not_supported ("garbage.docx", "definitely not a zip archive........".data);
+    // ZIP64: EOCD 中 cd_offset = 0xFFFFFFFF
+    expect_not_supported ("zip64.xlsx", new uint8[] {
+        0x50, 0x4b, 0x05, 0x06, 0x00, 0x00, 0x00, 0x00, 0x01, 0x00,
+        0x01, 0x00, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0x00, 0x00
+    });
+    // 加密条目: central directory flags bit0 = 1
+    var builder = new ZipBuilder ();
+    builder.add_stored ("word/document.xml", "<w:document/>".data, 0x0001);
+    expect_not_supported ("enc.docx", builder.finish ());
+}
+
+static void test_txt_rejected () throws Error {
+    string path = write_fixture ("t.txt", "plain text".data);
+    bool threw = false;
+    try {
+        OfficeTextExtractor.extract_markdown (path);
+    } catch (Error e) {
+        threw = (e is IOError.NOT_SUPPORTED);
+    }
+    check (threw, "txt → NOT_SUPPORTED (模式外格式由调用方门禁拦截)");
+}
+
+int main () {
+    run ("supports", () => { test_supports (); });
+    run ("docx stored", () => { test_docx_stored (); });
+    run ("docx deflate", () => { test_docx_deflate (); });
+    run ("xlsx", () => { test_xlsx (); });
+    run ("pptx", () => { test_pptx (); });
+    run ("odt", () => { test_odt (); });
+    run ("eocd comment", () => { test_eocd_trailing_comment (); });
+    run ("error paths", () => { test_error_paths (); });
+    run ("txt rejected", () => { test_txt_rejected (); });
+
+    print ("\n%s\n", failures == 0 ? "ALL TESTS PASSED" : "%d TEST(S) FAILED".printf (failures));
+    return failures == 0 ? 0 : 1;
+}
diff --git a/tests/test_vlm_queue.vala b/tests/test_vlm_queue.vala
index 13cd208..aff9e3e 100644
--- a/tests/test_vlm_queue.vala
+++ b/tests/test_vlm_queue.vala
@@ -7,7 +7,7 @@
 //
 // 注意: 真实 executor 模式 (参考 vlm_task_runner.vala):
 //   1. 在 worker 线程上执行实际工作
-//   2. 通过 Idle.add 在主线程上发射 manager.task_completed (path, md, status, from_cache)
+//   2. 通过 Idle.add 在主线程上发射 manager.task_completed (path, md, status, from_cache, error)
 //   3. 调用 manager.notify_finished (path) 释放并发槽
 // 测试用 executor 必须遵循同样模式。
 
@@ -43,7 +43,7 @@ void pump_loop (uint ms = 200) {
 // 模拟立即完成的 executor: 在 Idle 上发 task_completed, 然后释放槽
 void immediate_executor (string path, VLMQueueManager m) {
     Idle.add (() => {
-        m.task_completed (path, "md-for-" + path, PreprocessStatus.COMPLETED, false);
+        m.task_completed (path, "md-for-" + path, PreprocessStatus.COMPLETED, false, null);
         return Source.REMOVE;
     });
     m.notify_finished (path);
@@ -116,7 +116,7 @@ void test_max_concurrency () {
         lock.unlock ();
 
         Idle.add (() => {
-            m.task_completed (path, "md", PreprocessStatus.COMPLETED, false);
+            m.task_completed (path, "md", PreprocessStatus.COMPLETED, false, null);
             return Source.REMOVE;
         });
         m.notify_finished (path);
@@ -146,7 +146,7 @@ void test_wait_for_completion () {
     mgr.executor = (path, m) => {
         Thread.usleep (100 * 1000); // 100ms
         Idle.add (() => {
-            m.task_completed (path, "md", PreprocessStatus.COMPLETED, false);
+            m.task_completed (path, "md", PreprocessStatus.COMPLETED, false, null);
             return Source.REMOVE;
         });
         m.notify_finished (path);
@@ -190,7 +190,7 @@ void test_wait_for_completion_timeout () {
         Thread.usleep (2000 * 1000); // 2 秒
         // 测试结束后才通知, 避免线程泄漏
         Idle.add (() => {
-            m.task_completed (path, "md", PreprocessStatus.COMPLETED, false);
+            m.task_completed (path, "md", PreprocessStatus.COMPLETED, false, null);
             return Source.REMOVE;
         });
         m.notify_finished (path);
@@ -227,7 +227,7 @@ void test_cancel_clears_pending () {
         lock.unlock ();
         Thread.usleep (100 * 1000);
         Idle.add (() => {
-            m.task_completed (path, "md", PreprocessStatus.COMPLETED, false);
+            m.task_completed (path, "md", PreprocessStatus.COMPLETED, false, null);
             return Source.REMOVE;
         });
         m.notify_finished (path);
@@ -258,7 +258,7 @@ void test_pause_resume () {
     mgr.executor = (path, m) => {
         Thread.usleep (20 * 1000);
         Idle.add (() => {
-            m.task_completed (path, "md", PreprocessStatus.COMPLETED, false);
+            m.task_completed (path, "md", PreprocessStatus.COMPLETED, false, null);
             return Source.REMOVE;
         });
         m.notify_finished (path);
@@ -313,7 +313,7 @@ void test_bg_threads_cleanup () {
     mgr.executor = (path, m) => {
         Thread.usleep (30 * 1000);
         Idle.add (() => {
-            m.task_completed (path, "md", PreprocessStatus.COMPLETED, false);
+            m.task_completed (path, "md", PreprocessStatus.COMPLETED, false, null);
             return Source.REMOVE;
         });
         m.notify_finished (path);
diff --git a/tools/build-converter-package.sh b/tools/build-converter-package.sh
new file mode 100644
index 0000000..f22f58e
--- /dev/null
+++ b/tools/build-converter-package.sh
@@ -0,0 +1,62 @@
+#!/usr/bin/env bash
+# Build the OPTIONAL Windows LibreOffice converter package.
+#
+# 主包不捆绑 LibreOffice (体积原因)。Office 文档视觉转换链路
+# (soffice → PDF → pdftoppm → VLM) 依赖 BinaryConverter.resolve_soffice ()
+# 按以下顺序探测:
+#   1. <应用包根>/tools/LibreOffice/program/soffice.exe   ← 本转换器包的布局
+#   2. C:\Program Files\LibreOffice\program\soffice.exe   ← 官方安装器默认目录
+#   3. PATH 中的 soffice                                   ← Linux/macOS 现状
+#
+# 本脚本下载官方 LibreOffice MSI, 用 msiexec 管理员解包 (不写注册表、
+# 不进 Program Files), 把程序树按 tools/LibreOffice/... 布局打成独立 ZIP。
+# 用户解压到 FileCollector 便携包根目录 (与 bin/ 同级) 即生效。
+#
+# 在 MSYS2 UCRT64 终端执行 (需要 curl/zip/unzip; msiexec 由 Windows 提供)。
+set -euo pipefail
+
+readonly DIST_DIR="dist"
+readonly WORK_DIR="converter-work"
+# 固定 LibreOffice "still" 分支版本; 升级时同步更新, 并确认 URL 有效:
+#   https://download.documentfoundation.org/libreoffice/stable/
+readonly LO_VERSION="25.8.4"
+readonly LO_URL="https://download.documentfoundation.org/libreoffice/stable/${LO_VERSION}/win/x86_64/LibreOffice_${LO_VERSION}_Win_x86-64.msi"
+
+rm -rf "${WORK_DIR}"
+mkdir -p "${DIST_DIR}" "${WORK_DIR}/unpack" "${WORK_DIR}/staging/tools"
+
+echo "Downloading LibreOffice ${LO_VERSION} MSI..."
+curl --fail --location --retry 3 --output "${WORK_DIR}/LibreOffice.msi" "${LO_URL}"
+
+MSI_WIN="$(cygpath -w "${PWD}/${WORK_DIR}/LibreOffice.msi")"
+TARGET_WIN="$(cygpath -w "${PWD}/${WORK_DIR}/unpack")"
+
+echo "Extracting MSI (msiexec /a, administrative extract, no install)..."
+# msiexec 的 /a /qn 参数不能直接从 bash 传入: MSYS2 会把 "/a" 当 POSIX 路径
+# 改写成 MSYS2 安装根路径 (与 meson --prefix=/usr 同一类陷阱), 借 cmd 复合
+# 字符串绕过参数转换。
+cmd //c "msiexec /a \"${MSI_WIN}\" /qn TARGETDIR=\"${TARGET_WIN}\""
+
+# msi 解包的嵌套层级因版本而异, 按 soffice.exe 实际位置反推 LibreOffice 根
+SOFFICE="$(find "${WORK_DIR}/unpack" -name soffice.exe -print -quit)"
+test -n "${SOFFICE}"
+LO_ROOT="$(dirname "$(dirname "${SOFFICE}")")"   # .../LibreOffice/program/soffice.exe → .../LibreOffice
+test -f "${LO_ROOT}/program/soffice.exe"
+test -d "${LO_ROOT}/share"
+
+cp -r "${LO_ROOT}" "${WORK_DIR}/staging/tools/LibreOffice"
+test -f "${WORK_DIR}/staging/tools/LibreOffice/program/soffice.exe"
+test -f "${WORK_DIR}/staging/tools/LibreOffice/program/soffice.bin"
+
+PACKAGE_PATH="${DIST_DIR}/filecollector-converter-windows-x64-${LO_VERSION}.zip"
+(
+  cd "${WORK_DIR}/staging"
+  zip -r "../../${PACKAGE_PATH}" tools
+)
+unzip -l "${PACKAGE_PATH}" | grep -Eq 'tools/LibreOffice/program/soffice\.exe'
+unzip -l "${PACKAGE_PATH}" | grep -Eq 'tools/LibreOffice/program/soffice\.bin'
+sha256sum "${PACKAGE_PATH}" > "${PACKAGE_PATH}.sha256"
+
+du -h "${PACKAGE_PATH}"
+echo "Built ${PACKAGE_PATH}"
+echo "Usage: extract into the FileCollector portable package root (alongside bin/), keeping the tools/ layout."
diff --git a/tools/build-windows-package.sh b/tools/build-windows-package.sh
index f0108a5..7dc53ac 100755
--- a/tools/build-windows-package.sh
+++ b/tools/build-windows-package.sh
@@ -21,6 +21,12 @@ fi
 MSYS_PREFIX="${MSYSTEM_PREFIX:-/ucrt64}"
 export PYTHONUTF8=1
 export LANG=C.UTF-8
+# meson --prefix=/usr 会被 MSYS2 参数转换改写为 MSYS2 安装根下的真实路径
+# (实测 --prefix=/usr → --prefix=D:/<msys2根>/usr), locale 因此被装进
+# install-windows/a/_temp/msys64/usr/... 错误嵌套路径, 打包时静默丢掉全部
+# 翻译 (v4.9.1 及之前的 Windows 包均为纯英文界面)。排除这两类参数;
+# cmark 的 -DCMAKE_INSTALL_PREFIX=/ucrt64 指向真实环境根, 不在排除之列。
+export MSYS2_ARG_CONV_EXCL="/usr;--prefix"
 export PKG_CONFIG_PATH="${MSYS_PREFIX}/lib/pkgconfig:${PKG_CONFIG_PATH:-}"
 export CPATH="${MSYS_PREFIX}/include:${CPATH:-}"
 export LIBRARY_PATH="${MSYS_PREFIX}/lib:${LIBRARY_PATH:-}"
@@ -64,6 +70,9 @@ cp -r data "${STAGING_DIR}/share/data"
 if [[ -d "${INSTALL_ROOT}/usr/share/locale" ]]; then
   cp -r "${INSTALL_ROOT}/usr/share/locale" "${STAGING_DIR}/locale"
 fi
+# 翻译缺失曾静默出包 (v4.9.1 及之前的 Windows 包均为纯英文界面, 见上方
+# MSYS2_ARG_CONV_EXCL 注释), 此处硬性卡关: .mo 不在 staging 直接构建失败。
+test -f "${STAGING_DIR}/locale/zh_CN/LC_MESSAGES/filecollector.mo"
 
 # Dynamically loaded GTK/GIO components are not visible to ldd and must be staged explicitly.
 cp ${MSYS_PREFIX}/lib/gdk-pixbuf-2.0/2.10.0/loaders/*.dll \
@@ -90,6 +99,44 @@ python3 tools/verify_gnome_icon_theme.py "${STAGING_DIR}/share/icons"
 cp ${MSYS_PREFIX}/lib/gio/modules/libgiognutls.dll "${STAGING_DIR}/lib/gio/modules/"
 python3 tools/collect_dlls.py "${STAGING_DIR}/lib/gio/modules/libgiognutls.dll" "${STAGING_DIR}/bin"
 
+# zip_exporter 通过外部 zip 命令打包, 普通用户机器没有 Info-ZIP, 便携包
+# 必须自带。MSYS2 的 zip 位于 msys 层 (/usr/bin), 依赖 msys-2.0.dll;
+# 两者都要进 bin/ (CreateProcess 搜索调用方 exe 所在目录优先于 PATH,
+# zip_exporter 的裸 "zip" 命令无需改 PATH 即可命中)。
+test -f /usr/bin/zip.exe
+test -f /usr/bin/msys-2.0.dll
+cp /usr/bin/zip.exe "${STAGING_DIR}/bin/zip.exe"
+cp /usr/bin/msys-2.0.dll "${STAGING_DIR}/bin/msys-2.0.dll"
+
+# 二进制预处理 (视觉转换模式) 需要 poppler 工具: pdftoppm 把 PDF 渲染成页图
+# 送 VLM, pdftotext 供 "文本抽取" 模式抽纯文本。其依赖 (cairo/freetype/
+# fontconfig/jpeg/png/tiff) GTK 已带, collect_dlls.py 递归收集剩余部分。
+test -f "${MSYS_PREFIX}/bin/pdftoppm.exe"
+test -f "${MSYS_PREFIX}/bin/pdftotext.exe"
+cp "${MSYS_PREFIX}/bin/pdftoppm.exe" "${STAGING_DIR}/bin/pdftoppm.exe"
+cp "${MSYS_PREFIX}/bin/pdftotext.exe" "${STAGING_DIR}/bin/pdftotext.exe"
+python3 tools/collect_dlls.py \
+  "${STAGING_DIR}/bin/pdftoppm.exe" "${STAGING_DIR}/bin/pdftotext.exe" \
+  "${STAGING_DIR}/bin"
+
+# poppler 数据文件 (CJK CID 字体 CMap 等): 中文 PDF 没有它, pdftoppm/pdftotext
+# 会输出乱码或空文本。优先取 MSYS2 环境内的 share/poppler; 没有则下载上游
+# poppler-data 补齐。
+if [[ -d "${MSYS_PREFIX}/share/poppler" ]]; then
+  cp -r "${MSYS_PREFIX}/share/poppler" "${STAGING_DIR}/share/poppler"
+else
+  readonly POPPLER_DATA_VERSION="0.4.12"
+  curl --fail --location --retry 3 --output /tmp/poppler-data.tar.gz \
+    "https://poppler.freedesktop.org/releases/poppler-data-${POPPLER_DATA_VERSION}.tar.gz"
+  tar -xzf /tmp/poppler-data.tar.gz -C /tmp
+  mkdir -p "${STAGING_DIR}/share/poppler"
+  # 上游 tarball 顶层即 cidToUnicode/ cMap/ unicodeMap/ 三目录 (无 share/ 嵌套)
+  for d in cidToUnicode cMap unicodeMap; do
+    cp -r "/tmp/poppler-data-${POPPLER_DATA_VERSION}/${d}" "${STAGING_DIR}/share/poppler/${d}"
+  done
+fi
+test -d "${STAGING_DIR}/share/poppler/cidToUnicode"
+
 cat > "${STAGING_DIR}/bin/filecollector-launch.bat" <<'BAT'
 @echo off
 set "GDK_PIXBUF_MODULEDIR=%~dp0..\lib\gdk-pixbuf-2.0\2.10.0\loaders"
@@ -109,6 +156,10 @@ unzip -l "${PACKAGE_PATH}" | grep -Eq 'bin/filecollector\.exe'
 unzip -l "${PACKAGE_PATH}" | grep -Eq 'bin/filecollector-launch\.bat'
 unzip -l "${PACKAGE_PATH}" | grep -Eq 'lib/gio/modules/libgiognutls\.dll'
 unzip -l "${PACKAGE_PATH}" | grep -Eq 'share/glib-2\.0/schemas/gschemas\.compiled'
+unzip -l "${PACKAGE_PATH}" | grep -Eq 'locale/zh_CN/LC_MESSAGES/filecollector\.mo'
+unzip -l "${PACKAGE_PATH}" | grep -Eq 'bin/zip\.exe'
+unzip -l "${PACKAGE_PATH}" | grep -Eq 'bin/pdftoppm\.exe'
+unzip -l "${PACKAGE_PATH}" | grep -Eq 'bin/pdftotext\.exe'
 # 再校验 ZIP 本体, 防止打包环节漏拷文件。
 python3 tools/verify_windows_package.py "${PACKAGE_PATH}"
 sha256sum "${PACKAGE_PATH}" > "${PACKAGE_PATH}.sha256"
diff --git a/vala-lint.conf b/vala-lint.conf
index 3ea6aca..47f2d4f 100644
--- a/vala-lint.conf
+++ b/vala-lint.conf
@@ -43,7 +43,7 @@ ignore-comments=true
 
 [naming-convention]
 # 全大写标识符 (Vala 常量约定) 不算违反 naming-convention
-exceptions=MAX_FILE_CONTENT_SIZE,MAX_FILE_SIZE,PEEK_SIZE,HICOLOR_RESOURCE_PREFIX,DOCUMENT_EXTENSIONS,IMAGE_EXTENSIONS,SNAPSHOT_ICON_CHOICES
+exceptions=MAX_FILE_CONTENT_SIZE,MAX_FILE_SIZE,MAX_TEXT_SIZE,PEEK_SIZE,HICOLOR_RESOURCE_PREFIX,DOCUMENT_EXTENSIONS,IMAGE_EXTENSIONS,SNAPSHOT_ICON_CHOICES,SUPPORTED_EXTS,BINARY_MODE_VLM,BINARY_MODE_TEXT
 
 [note]
 keywords=TODO,FIXME,XXX,HACK