返回目录

dsh-tianshu-tui

编辑精选维护状态: 活跃

huiliyi37/dsh-tianshu-tui

官方 DeepSeek Harness 的交互式终端 UI 插件:自研 ANSI 极简交互渲染、流式 Markdown/工具卡、16+ 主题、slash 命令与选择器、输入历史与本地偏好持久化、LSP 诊断、memory记忆,很丝滑的开发体验。

前往 GitHub项目主页
$ npx -y @deepseek-ai/dsh@0.1.0-rc.8 plugin --profile tui add @huiliyi37/dsh-tianshu-tui

259

星标

9

Fork

TypeScript

语言

Apache-2.0

许可证

2026-08-13

创建于

2026-08-29

最近推送

Apache-2.0 交互式终端 UI 插件,装入官方 DeepSeek Harness(dsh)tui profile 即用:会话工作区、图片端到端、推理可视化,纯展示层不注入模型面。

DSH 适配

兼容

作者声明

安全审计

未审计

最后核验

2026-08-21

许可证

Apache-2.0

01它能帮你完成什么?

  • Run DeepSeek Harness coding sessions in a full interactive terminal workspace

    Real-time session rendering with resume, /fork branching, /rewind rollback, /export Markdown transcripts, slash menus, approval cards and live status/config/tasks/workflow panels

    Developers using the official dsh CLI who want a rich in-terminal coding-agent experience instead of the plain CLI

  • Send images to the model end-to-end from the terminal

    Clipboard-pasted images rendered inline (kitty / iTerm2), delivered via the harness attachment service, auto-described by a vision-bridge model when the main model cannot see

    Users who paste screenshots or designs into coding sessions and need vision-capable models to actually see them

  • Watch and control the agent's reasoning and approvals in real time

    Collapsible thinking blocks (Ctrl+O), pending approval cards with inline diff preview, mode cycling (normal → plan → always-approve)

    Developers who want visible reasoning traces and structured approval flow during dsh coding sessions

02如何接入 DeepSeek Harness?

前置条件

  • Official dsh CLI `@deepseek-ai/dsh` at `0.1.0-rc.8` — the package is not a standalone program and cannot run with just `npm i`
  • Node.js `^22.19 || >=24`
  • `pnpm` on PATH (the `dsh plugin` command forwards to it)

安装步骤

  1. 01

    Install the plugin into the tui profile: `npx -y @deepseek-ai/dsh@0.1.0-rc.8 plugin --profile tui add @huiliyi37/dsh-tianshu-tui` (or from Git: `... add github:huiliyi37/dsh-tianshu-tui`)

    $ npx -y @deepseek-ai/dsh@0.1.0-rc.8 plugin --profile tui add @huiliyi37/dsh-tianshu-tui

  2. 02

    Launch: `npx -y @deepseek-ai/dsh@0.1.0-rc.8 --profile tui` (if the official CLI is globally installed at 0.1.0-rc.8, plain `dsh` works too)

    $ npx -y @deepseek-ai/dsh@0.1.0-rc.8 --profile tui

验证接入成功

  • Success when the welcome page shows the dsh-tianshu-tui brand; exit with `Ctrl+Q` or `/exit`

03DSH 适配与能力边界

DSH 适配兼容

TUI plugin installed into the official dsh CLI's tui profile via `dsh plugin --profile tui add`; pure presentation layer over the official DeepSeek Harness session event stream

  • Full session workspace in the terminal

    dsh coding sessionsreal-time rendering, append-only scroll transcript, session resume on startup, /fork branching, /rewind rollback (session truncation + optional file restore), /export Markdown transcript

    /rewind truncates the session event log and can roll files back to a pre-boundary snapshot/export writes a Markdown transcript file
  • End-to-end image pipeline with vision bridge

    clipboard-pasted or attached imagesinline rendering via kitty / iTerm2 graphics protocols, delivery through the harness attachment service, auto vision-model description when the main model is text-only

    oversized images are compressed before sending (1568px long-edge cap, JPEG 0.82 → 0.55 → 1024px + 0.55)
  • Complete input surface

    user typing in the terminalslash dropdown (fuzzy prefix match, MRU sort, ghost preview), @-path Tab completion, @mention expansion, bracketed paste, optional vim keys, external editor (Ctrl+E), history search (Ctrl+F), Ctrl+. keymap overlay

    local preferences persisted to ~/.dsh-tui/prefs.jsoninput history persisted to ~/.dsh-tui/input-history.json (1000-entry cap; delete the file to clear)
  • Auditable pure presentation layer

    session event stream from the host harnessall rendered state derived from session events; user input becomes ordinary logged messages

04适合谁?何时不该用?

适合

  • Developers using the official dsh CLI who want a rich in-terminal coding-agent experience instead of the plain CLI
  • Users who paste screenshots or designs into coding sessions and need vision-capable models to actually see them
  • Developers who want visible reasoning traces and structured approval flow during dsh coding sessions

不适合

  • Not a standalone program: requires the official `@deepseek-ai/dsh` CLI pinned at `0.1.0-rc.8`, Node.js `^22.19 || >=24`, and pnpm on PATH. A stale `dsh` binary on PATH can cause ERR_FS_EISDIR errors.
  • Must not be installed into oh-my-tianshu's (@huiliyi37 ecosystem) tui profile — mixing ecosystems creates fragile cross-ecosystem resolution of @deepseek-ai/* via stale symlinks in ~/.dsh/profiles/node_modules.

05兼容性、维护与安全提示

  • Not a standalone program: requires the official `@deepseek-ai/dsh` CLI pinned at `0.1.0-rc.8`, Node.js `^22.19 || >=24`, and pnpm on PATH. A stale `dsh` binary on PATH can cause ERR_FS_EISDIR errors.
  • Must not be installed into oh-my-tianshu's (@huiliyi37 ecosystem) tui profile — mixing ecosystems creates fragile cross-ecosystem resolution of @deepseek-ai/* via stale symlinks in ~/.dsh/profiles/node_modules.
  • On every launch the plugin checks npm `latest` for self-updates (network call; disable with DSH_TUI_SKIP_UPDATE=1). Image re-asking additionally requires the companion @deepseek-ai/dsh-vision-ask package, which the TUI bundle does not ship.
2026-08-132026-08-18v0.1.2-rc.12

Apache-2.0 · actively maintained (latest release v0.1.2-rc.12, 2026-08-20)

06常见问题

如何把 dsh-tianshu-tui 装进 DeepSeek Harness?

它是 dsh 插件,不是独立程序。先装官方 CLI @deepseek-ai/dsh(0.1.0-rc.8),再运行 `npx -y @deepseek-ai/dsh@0.1.0-rc.8 plugin --profile tui add @huiliyi37/dsh-tianshu-tui`,然后用 `npx -y @deepseek-ai/dsh@0.1.0-rc.8 --profile tui` 启动。

是原生运行还是 MCP?

都不是:它作为 TUI 插件运行在官方 dsh 的 tui profile 内。UI 是纯展示层——所有 agent 状态来自会话事件流,自身不注册任何 prompt、工具或上下文面。

需要什么前置条件?

官方 @deepseek-ai/dsh 0.1.0-rc.8、Node.js ^22.19 || >=24、PATH 上有 pnpm(dsh plugin 会转发给它)。只 npm i 本包跑不起来。

会自动更新或联网吗?

从 npm 安装后,每次启动会对照 npm latest,有新版本就写入 profile(重启生效)。设 DSH_TUI_SKIP_UPDATE=1 可跳过联网检查;github:/link: 安装不会被改写。

有哪些已知限制?

图片再询问(ask_image)需另装同仓伴生插件 @deepseek-ai/dsh-vision-ask;LSP 诊断只进 TUI 本地展示(工具卡徽标 + /lsp 面板),不提供模型工具面;不要把它装配进 oh-my-tianshu 的 tui profile。

08数据与来源

  • 作者声明github.com1ff44e3dea8c…

    **dsh-tianshu-tui**(`@huiliyi37/dsh-tianshu-tui`)是官方 [DeepSeek Harness](https://github.com/deepseek-ai/deepseek-harness)…

  • 作者声明github.com1ff44e3dea8c…

    npx -y @deepseek-ai/dsh@0.1.0-rc.8 plugin --profile tui add @huiliyi37/dsh-tianshu-tui

  • 作者声明github.com1ff44e3dea8c…

    UI 是纯展示层:所有 agent 状态都来自会话事件流。

页面基于项目公开文档、仓库元数据和 DSH Plugins 的结构化解析生成;最后核验于 2026-08-21。发现错误?提交更正。

🏆

最佳 DeepSeek Harness 插件

从全目录挑出的 12 个值得优先安装的插件,覆盖各个分类。

DSH Plugins 是独立的 DeepSeek Harness 插件市场,与 DeepSeek 官方无关,也不代表官方背书。第三方插件未经安全审计,安装前请审查源码。

每周获取最新的 DeepSeek Harness 插件,绝不滥发。