dsh-web:DeepSeek Harness 的插件生态包,任务看板 + SSH + 移动端 + 图像理解三步装齐
dsh-web-deepseek-harness-web-plugin-sidebar-all-in-one
by Mycelium Protocol
GitHub:zhu1090093659/dsh-web ⭐ 6,031 | Forks 386 | TypeScript | Apache 2.0
创意工坊:https://dsh-market.com
npm:@linxin666/dsh-web-all
创建:2026-08-12
三步装齐
# ① 安装
dsh plugin --profile web add @linxin666/dsh-web-all@latest
# ② 重启
dsh web
# ③ 侧边栏所有功能一键开启
前提:已安装 DeepSeek Harness,dsh web 可正常启动。
是什么
dsh-web 是 DeepSeek Harness(DSH)Web GUI 的插件聚合生态包,理念是「一切皆插件」:每个功能都是独立的 npm 包,可插拔、可替换、可再开发——装全家桶是完整 AI 开发工作台,只挑一两个也能静默融入原生界面。
| 能力 | 原生 dsh web | dsh-web 全家桶 |
|---|---|---|
| 任务看板 | 无 | 多列看板 + cron 定时真实执行 |
| 移动端远程 | 无 | 扫码配对、SSE 实时同步;同链接也可配对 PC 浏览器 |
| SSH 运维 | 无 | 终端 / 文件传输 / 隧道 / 集群 |
| 图像理解 | 无 | describe_image 视觉工具(兼容所有 OpenAI 视觉端点) |
| 右侧面板 | 无 | 资源管理器 / 编辑器 / 终端 / Git / 浏览器 |
| Git 图谱 | 无 | 分支选择器 + 提交历史泳道图 |
| 主题皮肤 | 默认 | Blue Fantasy 内置,其他皮肤从创意工坊按需安装 |
| Agent 预设 | 官方预设 | + 梁神模式(两阶段锚定,实测 98.5 分) |
| 救助模式 | 无 | dsh-doctor:事务式快照/回滚,默认开启 |
各插件功能详解
任务看板(Task Board)
五列看板:待规划 / 待办 / 进行中 / 已完成 / 已失败。点卡片上的「执行」,任务交给真实的 DSH Agent 会话跑,跑完状态自动回写。
支持 cron 定时执行:详情里配表达式(比如每天 23:00 自动升级 DSH、每周一 09:00 生成周报),关闭浏览器后仍按时跑。附可选的空闲睡眠保护,支持 Windows / macOS / Linux(systemd-logind),允许屏幕熄灭,阻止整机因空闲睡眠中断任务。
移动端远程控制(Mobile Remote)
侧边栏底部的手机图标打开配对面板。扫二维码(或复制链接)配对后,手机进入独立移动端界面:看会话、开新会话、收发消息、切模型和思考强度、调权限预设,与桌面端实时同步(SSE)。
同一份配对链接也能配对 PC 浏览器,让另一台电脑运行完整 Web GUI。配对令牌一次性限时,「停止」随时吊销所有设备。开 cloudflared 公网隧道后可跨网络使用。
Cloudflare quick tunnel(trycloudflare.com)和 Tailscale Serve 不透传 SSE,实时推送会降级为轮询;需即时推送则用 named tunnel。
远程连接(SSH Ops)
- Web 终端:xterm.js,实时输出,窗口大小自适应
- 文件传输:SFTP 上传 / 下载,进度条,可浏览远程目录
- 端口转发:本地隧道直连远程内网服务(数据库、API、管理后台),仅监听 127.0.0.1
- 集群执行:一条命令并发多台主机,按别名 / 环境 / 标签过滤
- Agent 直连:Agent 和面板共用同一份主机配置,对话说”连一下 xxx 看状态”,Agent 即去远程执行
主机配置存在 ~/.dsh/dsh-ssh.json,支持从 ~/.ssh/config 一键导入,密钥 / 密码认证均可。
图像理解(Image Understanding)
给纯文本模型补视觉通道:describe_image 工具检测到图片(本地路径 / https URL / 会话附件)后,发给配置好的 OpenAI 兼容视觉端点(Qwen-VL、GLM-4V、GPT-4o、本地 Ollama 均可),只有返回文本进入会话记录,图片本身不进。
输入框加了图片按钮,选图后生成附件引用插入草稿,模型用 describe_image 分析。支持 prompt 参数自定义指令(OCR / UI 诊断 / 翻译),比默认描述准。配置在设置 → 插件配置 → Image understanding。
梁神模式(LiangShen Mode)
两阶段锚定的 Agent 预设(dsh-liangshen):
- 第一阶段:首轮模型请求只看到精确双工具(持久
bash+str_replace_editor)和一行 persona,无运行时上下文注入 - 第二阶段:首次工具调用后,检测到首个 minimal-like 推理块时切换为 PTC Mode,恢复完整工具注册表和所有 prompt section
效果:Standard/PTC 模式约 91/92 分,Minimal 模式约 99/96 分,梁神模式 Windows 原生实测均值 98.5,在不牺牲完整工具能力的前提下维持高质量输出。Resume 不丢状态,支持 plan mode。
救助模式(dsh-doctor)
事务式救援体系,默认开启:
- Doctor Supervisor 后台服务检测启动失败、崩溃、心跳丢失、Web 故障、白屏
- 每次修复是一个事务:快照当前 profile → 在候选环境应用确定性规则 → 经隔离 dump-config + Web 健康门禁 → 原子提升,失败按字节回滚
- Web 控制台(设置 → 插件配置 → Doctor 卡片)展示故障事件,提供诊断、修复、回滚动作
- 「发送给 Harness」把最近故障摘要和错误堆栈投回当前会话,让 Agent 就地诊断
创意工坊(dsh-market.com)
对标 Steam Workshop:社区统一分发皮肤、宠物与插件。按设备点赞热度排序,前三名登首页颁奖台;皮肤支持实时试穿预览,插件提供一键复制安装命令。
Blue Fantasy 蓝色幻想皮肤随皮肤插件内置:鲸鱼插画垫在半透明面板下,靛蓝色调贯穿全局。鲸鱼娘宠物和其他皮肤从工坊浏览、试穿后按需安装。
站点本身纯静态构建,由脚本从三类真值源(skin.json / pet.json / community.json)确定性生成,点赞等动态能力由 Cloudflare Workers + D1 承载,push 到 main 自动部署。
单独安装某个插件
不需要全家桶时,按需安装:
dsh plugin --profile web add @linxin666/dsh-client-ui-task-board@latest # 任务看板
dsh plugin --profile web add @linxin666/dsh-ssh@latest # SSH 运维
dsh plugin --profile web add @linxin666/dsh-tool-describe-image@latest # 图像理解
dsh plugin --profile web add @linxin666/dsh-pet@latest # 宠物
dsh plugin --profile web add @linxin666/dsh-liangshen@latest # 梁神模式
dsh plugin --profile web add @linxin666/dsh-doctor@latest # 救助模式
dsh plugin --profile web add dsh-better-sidebar@latest # 右侧面板
为什么”一切皆插件”
DSH 本身已经是成熟的 AI 编程 harness。dsh-web 解决的是 Web GUI 功能碎片化的问题:每个团队需要的组合不同(有人要 SSH,有人要移动端,有人只要皮肤),但过去每个功能都需要单独寻找、单独配置、手动维护兼容性。
插件体系把这个问题收拢进统一的安装机制(官方 profile 机制,不改 DSH 源码),创意工坊解决发现和分发,救助模式解决升级和故障,三者合起来构成一个可持续维护的生态,而不是一堆散包。
相关链接
- GitHub:https://github.com/zhu1090093659/dsh-web
- 创意工坊:https://dsh-market.com
- npm(聚合包):https://www.npmjs.com/package/@linxin666/dsh-web-all
Mycelium Protocol — 追踪 AI 系统的底层演化
关于 Mycelium
菌丝协议。持续追踪 AI 工具、系统和实验的内容节点。
dsh-web: The DeepSeek Harness Plugin Ecosystem — Task Board, SSH, Mobile, and Image Understanding in Three Steps
by Mycelium Protocol
GitHub: zhu1090093659/dsh-web ⭐ 6,031 | Forks 386 | TypeScript | Apache 2.0
Creative Workshop: https://dsh-market.com
npm: @linxin666/dsh-web-all
Created: 2026-08-12
Three Steps
# ① Install
dsh plugin --profile web add @linxin666/dsh-web-all@latest
# ② Restart
dsh web
# ③ Enable everything from the sidebar
Prerequisite: DeepSeek Harness installed and dsh web running normally.
What It Is
dsh-web is the plugin ecosystem bundle for DeepSeek Harness (DSH) Web GUI. The philosophy: everything is a plugin. Each feature is an independent npm package — composable, replaceable, and re-developable. Install the full bundle for a complete AI development workstation, or pick one or two and they quietly blend into the native interface.
| Capability | Native dsh web | dsh-web bundle |
|---|---|---|
| Task board | None | Multi-column kanban + real cron execution |
| Mobile remote | None | QR code pairing, SSE sync; same link pairs PC browsers too |
| SSH ops | None | Terminal / file transfer / tunnels / cluster |
| Image understanding | None | describe_image tool (any OpenAI-compatible vision endpoint) |
| Right panel | None | File explorer / editor / terminal / Git / browser |
| Git graph | None | Branch selector + commit history swim-lane |
| Themes | Default | Blue Fantasy built-in, others from the Creative Workshop |
| Agent presets | Official | + LiangShen mode (two-phase anchored, 98.5 avg score) |
| Rescue mode | None | dsh-doctor: transactional snapshot/rollback, on by default |
Plugin Breakdown
Task Board — Five columns: backlog / todo / in-progress / done / failed. “Execute” sends the task to a real DSH agent session; status updates automatically. Cron scheduling works after the browser closes. Optional idle-sleep protection on Windows, macOS, and Linux.
Mobile Remote — Pair via QR code. Phone gets a dedicated mobile UI: view sessions, start new ones, send messages, switch model and thinking intensity, adjust permission presets — all in real-time sync with desktop (SSE). The same pairing link works for a PC browser too. Tokens are one-use and time-limited. Cloudflare named tunnels enable cross-network pairing.
Note: Cloudflare quick tunnel (trycloudflare.com) and Tailscale Serve don’t forward SSE — real-time push downgrades to polling. For instant push, use a named tunnel.
SSH Ops — xterm.js terminal, SFTP file transfer with progress, port forwarding (loopback-only), cluster-wide command execution with tag/environment filters, and agent integration (say “check the status on xxx” in chat; the agent runs it remotely). Configuration in ~/.dsh/dsh-ssh.json, imports from ~/.ssh/config.
Image Understanding — Adds vision to text-only models. describe_image detects images (local path / https URL / session attachment), sends them to any OpenAI-compatible vision endpoint (Qwen-VL, GLM-4V, GPT-4o, local Ollama), and only the returned text enters the session — the image itself doesn’t. Custom prompt param supports OCR, UI diagnosis, translation.
LiangShen Mode — Two-phase anchored agent preset:
- Phase 1: model sees only two tools (persistent
bash+str_replace_editor) and one persona line — no runtime context injection - Phase 2: after first tool call, when a minimal-like reasoning block appears, switches to PTC Mode with full tool registry and all prompt sections
Standard/PTC scores ~91/92; Minimal ~99/96; LiangShen averages 98.5 on Windows native — maintains full tool capability without sacrificing output quality. Resume-safe, plan mode supported.
Rescue Mode (dsh-doctor) — Transactional rescue, on by default. Detects crashes, heartbeat loss, web failures, white screens. Each fix is a transaction: snapshot → apply deterministic rules → isolated dump-config + web health gate → atomic promotion. Failure rolls back byte-for-byte. Web console shows fault events with diagnose/fix/rollback actions; “Send to Harness” formats the latest fault summary into a troubleshooting prompt in the current session.
Creative Workshop (dsh-market.com)
Positioned like Steam Workshop for DSH. Community distributes skins, pets, and plugins. Rankings by device upvotes; top three on the front page podium. Skins support live preview before install; plugins get one-click install commands.
Blue Fantasy is the built-in default: whale illustration under translucent panels, indigo tones throughout. Other skins and the whale-girl pet are available on the Workshop.
The site itself is a pure static build, generated deterministically from three source-of-truth files (skin.json / pet.json / community.json). Dynamic features (likes, per-device votes) run on Cloudflare Workers + D1. Push to main deploys automatically.
Individual Plugin Install
When the full bundle is more than you need:
dsh plugin --profile web add @linxin666/dsh-client-ui-task-board@latest # Task board
dsh plugin --profile web add @linxin666/dsh-ssh@latest # SSH ops
dsh plugin --profile web add @linxin666/dsh-tool-describe-image@latest # Image understanding
dsh plugin --profile web add @linxin666/dsh-liangshen@latest # LiangShen mode
dsh plugin --profile web add @linxin666/dsh-doctor@latest # Rescue mode
dsh plugin --profile web add dsh-better-sidebar@latest # Right panel
Why “Everything Is a Plugin”
DSH is already a mature AI coding harness. dsh-web solves Web GUI feature fragmentation: every team needs a different combination (some need SSH, some need mobile, some just want a skin), but until now each feature required finding separately, configuring separately, and manually managing compatibility.
The plugin architecture routes this through DSH’s official profile mechanism — no source code changes to DSH itself. The Creative Workshop handles discovery and distribution. Rescue mode handles upgrades and failures. Together they form a maintainable ecosystem, not a pile of loose packages.
Links
- GitHub: https://github.com/zhu1090093659/dsh-web
- Creative Workshop: https://dsh-market.com
- npm: https://www.npmjs.com/package/@linxin666/dsh-web-all
Mycelium Protocol — tracking the deep evolution of AI systems
© 2026 Mycelium Protocol. All rights reserved.
关于本站 · 免责声明
🍄 Mushroom Research Blog 是非营利、免费公开的个人科技观察博客与公众号 XStack18,不接受商业合作、不代表任何企业或机构立场,也不谋求商业利益。我们以个人视角客观中立地记录和分析 AI、Web3 等领域的最新模型发布与技术动态——不止转述新闻标题或二手信息,而是给出有独立思考的深入分析,希望帮更多人获得有价值的一手科技认知。
⚠️ 文中介绍的开源代码与模型,仅供学习交流与技术借鉴。它们大多仍处于早期阶段,有待进一步研究和验证,请勿直接用于工作或生产环境;如需采用,请先自行充分测试,并核实其许可证与安全性。
Open-source code and models featured here are shared for learning and reference only. Most are early-stage and still need further study and verification — please don't use them directly in your work or in production. Test them thoroughly and check their licenses and security first.
- 本站文章均为作者基于公开信息的个人研究与观点整理,不代表文中提及的任何公司、产品、模型的官方立场,未与其构成商业关联或合作关系。
- 科技行业信息更新极快,我们尽力保证内容准确、及时,但不对完整性、实时性做绝对保证,具体请以相关企业/项目官方公告为准。
- 文中引用的第三方商标、产品名称、图片、数据等版权归原权利人所有,我们会尽量注明来源;如你认为存在版权疑问或侵权,请通过下方邮箱联系我们,收到通知后会尽快核实处理(更正、加注来源或删除)。
- 文章内容仅为技术科普与个人观点,不构成投资、法律或其他专业建议,据此进行任何决策的后果需自行判断和承担。
📮 侵权 / 勘误 / 合作咨询:hello@mushroom.cv
💬 评论与讨论
使用 GitHub 账号登录后发表评论