# dsh-inbox-paste [English](README.md) DeepSeek Harness **Web 插件**:在对话框粘贴/选择文件,落盘到 `inbox/`,待发送与发送后显示缩略图。Agent 用 `list_inbox` / `read_inbox` 读**文本**;**图片理解需自备视觉 skill 或模型**(见下方)。 ![示例:粘贴截图后由 image-reader skill 分析](docs/images/example-chat.png) ## 环境要求 - [DeepSeek Harness](https://github.com/deepseek-ai/DeepSeek-Harness) + **web** profile - Node.js ≥ 22、`dsh` CLI ## 安装 ```powershell dsh plugin --profile web add github:JarszmLovesAI/dsh_upload_file dsh web ``` 卸载:`dsh plugin --profile web remove dsh-inbox-paste` ## 使用 1. 输入框内 **Ctrl+V** 粘贴,或点 **曲别针** 选文件。 2. 上方待发送区出现缩略图;点 × 可移除(发送前不落盘)。 3. 发送后文件写入 `{session.cwd}/inbox/`;气泡可显示缩略图(见配置)。 4. Agent:`list_inbox` → 文本用 `read_inbox`;**图片**请调用 [image-reader](skills/image-reader/README.md) 或你自己的视觉工具。 ### 图片能上传,但模型能「看见」吗? **不能自动。** 本插件不走 Harness 原生多模态附件链,主模型默认只收到路径标记。要分析 inbox 里的图片,请: - 安装仓库附带的 **image-reader** skill,并在 `~/.dsh/.env` 配置 `ARK_API_KEY`;或 - 使用你已有的视觉 MCP / skill / 多模态模型。 ## 配置 文件:`%USERPROFILE%\.dsh\profiles\web\cordis.patch.yml`(按你的 profile 名调整) ```yaml - id: inbox-paste config: uploadMaxBytes: 26214400 inboxDirName: inbox allowedExtensions: [] customizeUserBubble: true # false = 不自定义发送后气泡,可能显示路径原文 ``` | `customizeUserBubble` | 效果 | |----------------------|------| | `true`(默认) | 发送后气泡显示缩略图,隐藏 `[inbox:…](path)` | | `false` | 使用 Harness 内置 user 气泡 | 改配置后重启 `dsh web` 并刷新浏览器。 ## 与 dsh-file-upload 对比 | | dsh-inbox-paste | dsh-file-upload | |--|-----------------|-----------------| | 粘贴 | ✅ | ❌ | | 存储 | `{cwd}/inbox/` | `{cwd}/.dsh-uploads/` | | 内置 PDF 解析 | ❌ | ✅ | ## 测试 ```sh npm test ``` ## License MIT