DSH_Graph-Horses

Composable Digital Workforce for DeepSeek Harness · 可组装的专属数字员工团队

在同一个平台 DeepSeek Harness（DSH）中，打造属于你所在行业的专属 AI 员工、团队与工作流，让每个岗位拥有自己的专业能力、认知与记忆。

Build the AI employees, teams, and workflows your industry needs—all within DeepSeek Harness (DSH), with role-specific capabilities, professional knowledge, and memory.

从一个岗位，到一支分工协作的团队，再到可执行、可验收的工作流。自带软件开发专业团队，也支持你围绕 AI 视频制作、内容策划、研究分析等场景，自定义岗位、组织能力与专业知识，让不同员工各司其职，在同一平台完成任务与交付。

Start with one specialist, assemble a team, and turn its responsibilities into a workflow with clear deliverables and review criteria. A software development team is included; you can also configure teams for AI video production, content planning, research, and other domains using the models and tools available in your DSH environment.

简体中文 · English · 安装 / Installation · 使用 / Usage · English quick start

下载插件 / Download plugin · SHA-256 · 版本说明 / Release notes

下载 .tgz 安装插件；GitHub 的 Source code ZIP 是开发源码。当前版本为带已知限制的预发布。 Install the .tgz plugin package. GitHub’s Source code ZIP contains development sources. This is a prerelease with documented limitations.

flowchart LR
    R["岗位定义 / Role definitions"] --> E["专业员工 / Specialists"]
    K["专业认知与记忆 / Knowledge & memory"] <--> E
    E --> T["协作团队 / Teams"]
    T --> W["任务工作流 / Workflows"]
    W --> D["成果与验收 / Deliverables & review"]
    D -->|"归档与核验 / Archive & verify"| K

简体中文

产品能力 · 内置软件团队 · 安装 · 完整使用流程 · 认知与记忆 · 卸载与数据 · 开发与贡献

一个平台，组织你需要的专业分工

graph_horses 是一个独立的 DSH 第三方插件。你可以为员工定义“负责什么、擅长什么、使用什么模型和工具、参考什么专业知识”，将员工组织为团队，再用工作流明确每一步由谁负责、交付什么、如何验收。

能力	你可以做什么
自定义专业员工	设置岗位使命、职责、工作规范与边界；按岗位选择模型、Skill、插件和 MCP 偏好。
组建协作团队	复用已有员工定义，设置团队内名称、上下级关系与职责分工；同一岗位可在不同团队中复用。
编排工作流	为步骤指定负责人、输入、交付物、验收标准和不合格退回目标，形成清晰的任务推进路径。
在原生对话中工作	在 DSH 项目对话选择工作流，当前负责人直接处理你的消息，沿用原生附件、历史、工具与审批。
积累专业认知与记忆	上传资料，浏览概念与实体关系、重组文章和原始证据，分别管理员工私库与团队共享知识。
跟踪交付与验收	查看阶段成果、任务与交付历史；验收后推进，退回后返工，保留原记录与修改依据。
分享与复用配置	使用 JSON 导入、导出员工、团队和工作流定义及必要依赖。

例如：打造一支 AI 视频制作团队

你可以根据自己的制作方式配置以下岗位，并为每个岗位补充专业资料与可用工具：

岗位示例	主要职责	专业认知与记忆示例
制片／策划	明确受众、主题、风格、周期与交付要求	品牌规范、受众资料、已确认的创作要求
编剧	设计叙事结构、脚本、对白与节奏	人物设定、世界观、脚本规范与已确认修改
分镜／视觉设计	把脚本转为镜头、画面和视觉要求	镜头语言、角色外观、色彩和构图规范
影像／音频制作	使用已接入的工具生成或处理素材	模型使用方法、素材规范、已验证的制作经验
剪辑／交付	组织镜头、音频、字幕及输出版本	剪辑规范、字幕样式、画幅与输出要求
质量审核	核对叙事、人物、视觉一致性及技术指标	审核清单、历史缺陷及验收标准

工作流可以是：创作简报 → 脚本 → 分镜 → 素材制作 → 剪辑 → 质量审核 → 交付。同样的方法也适用于其他行业：改变岗位、工具、知识和验收条件，即可建立自己的专业协作方式。

这是自定义配置示例。仓库当前预置的是软件开发团队；视频生成、配音、剪辑等模型或服务需要在 DSH 中另行接入。员工的能力取决于实际可用的模型、Skill、插件、MCP 和工具。

开箱提供软件开发专业团队

全新 profile 首次启动时提供六个完整岗位，以及对应团队和两条工作流：

岗位	主要职责	必备 Skill
产品负责人	需求、范围、优先级、验收与需求覆盖	requirements-acceptance
技术负责人／架构师	架构、接口、技术取舍、工作分解与独立审查	architecture-design
UX/UI 设计师	用户路径、交互状态、视觉与可访问性	ux-design-review
开发人员	增量实现、问题定位与回归测试	incremental-implementation
测试负责人	功能、边界、性能、回归与质量证据	quality-verification
运维负责人	打包、安装升级、运行交接、回滚与卸载验证	release-readiness

• 软件工程 · 10 步：需求与验收基线 → 交互与体验定义 → 架构与实施契约 → 测试与性能计划 → 增量开发 → 独立代码审查 → 功能与性能复测 → 打包与运行交接 → 需求覆盖与用户验收 → 复盘与知识交接。
• 缺陷修复 · 5 步：复现与根因定位 → 最小修复 → 独立审查 → 功能与性能回归 → 修复交付。

每个岗位包含职责与边界、模型继承、必备／优先技能、工作方法、证据与记忆规范、协作及交付定义；每个步骤包含负责人、输入、产物、验收与退回规则。插件共提供 22 项内置技能，覆盖需求、架构、设计、实现、测试、代码审查、发布和工程收尾。详见软件岗位技能。

默认内容可编辑、复用或替换，不包含【演示】工作流、模拟完成记录或伪造的已确认记忆。安装不自动执行业务任务；升级已有 profile 不静默覆盖用户配置，也不自动补入默认对象。

安装与第一次使用

准备条件

项目	当前范围
DSH	官方 DeepSeek Harness 0.2.1-rc.3；其他版本需要重新验证。
Node.js	声明范围 `^22.19.0
界面	DSH Web 和官方 macOS Apple Silicon Desktop 0.2.1-rc.3；插件界面跟随 DSH 的中文／English 设置，业务内容保持原文。
模型	在 DSH 中配置可用模型；员工可继承 DSH 默认值或选择单独的岗位模型。
安装包	维护者提供的 dsh-plugin-composable-workforce-0.28.0.tgz 及对应 SHA-256。

截至 2026-10-26，官方最新 GitHub 发行标签和 npm next 为 0.2.1-rc.3，npm latest 仍为 0.2.1-rc.2。本版本按 rc.3 验证；使用 npm 时可显式安装 @deepseek-ai/dsh@0.2.1-rc.3。

正式安装只需要受支持的官方 DSH 和插件 .tgz，无需克隆、修改或重建 Harness。先核对归档校验和并停止目标宿主，再使用自己的 DSH_HOME 与 profile：

# 替换为你的 DSH 数据目录、profile 和插件归档绝对路径。
DSH_HOME=/absolute/dsh-home dsh plugin --profile web add \
  /absolute/dsh-plugin-composable-workforce-0.28.0.tgz \
  --config.auto-install-peers=false \
  --config.strict-peer-dependencies=false \
  --ignore-scripts

安装或更新后，重新启动目标 Harness，并刷新已打开的页面。版本兼容性以实际官方发行版和同一归档哈希的验收为准；安装命令的包管理参数不代表其他版本已兼容。

最快体验路径： 打开左侧 graph_horses → 查看内置六岗位与软件开发团 → 在 DSH 中打开一个项目及其原生对话 → 选择 工作流：软件工程 → 向当前负责人描述任务 → 从顶部 详情 查看交付并验收。

安装、更新、数据归属与完整清理说明见接入与生命周期指南。

官方 Desktop 桌面版

DSH 已提供官方 Desktop。请从 DSH 官网下载。桌面版共用 Web 插件机制，本插件使用同一个 .tgz；Web 与 Desktop 各有自己的 profile，安装到 Web 不会自动安装到 Desktop。当前实测范围为 macOS Apple Silicon 0.2.1-rc.3；Windows 桌面版和其他架构尚未在本项目验收。

首次打开 Desktop 完成原生初始化，然后退出应用。请使用 Desktop 自带的 dsh 命令安装到 desktop profile；npm 安装的 dsh 不允许管理该保留 profile。macOS 安装在 ~/Applications 时，例如：

"$HOME/Applications/DeepSeek Harness.app/Contents/Resources/runtime/cli/bin/dsh" \
  plugin --profile desktop add /absolute/dsh-plugin-composable-workforce-0.28.0.tgz \
  --config.auto-install-peers=false --config.strict-peer-dependencies=false --ignore-scripts

若装在 /Applications，相应替换路径。使用自定义 DSH_HOME 时，应用和上述命令必须使用同一数据目录。重新打开 Desktop 后，左侧出现 graph_horses。桌面端的保存操作兼容官方 dsh-app://app 鉴权代理。普通卸载用同一条 Desktop 命令执行 plugin --profile desktop remove dsh-plugin-composable-workforce；完整清理的脚本用法见Desktop 接入说明。

从员工到交付：完整使用流程

1. 创建或调整专业员工

在 graph_horses → 自由职业人库 中选择已有岗位，或新建一个岗位。这里的“自由职业人”是可复用的员工定义；项目执行时才建立具体员工实例。

配置入口	使用方法
模型配置	跟随 DSH 当前默认模型，或指定岗位的 provider、model 与推理力度。账号和凭据仍由 DSH 管理。
岗位定义	填写岗位使命、职责、工作规范和边界。可使用“AI 智填”生成可编辑草稿，检查后保存。
能力组合	添加必备／优先 Skill、优先插件与优先 MCP。必备 Skill 每份工作单开始时加载；优先 Skill 在任务匹配时按需读取。
认知记忆	上传专业资料，浏览知识关系、文章和原文，管理自动归档设置。

“AI 智填”使用当前草稿的姓名和模型，生成结果不会自动保存为正式配置。选中某个插件或 MCP 只表达使用偏好，不会自动安装、连接或授予权限；需要先在 DSH 中准备好相应能力。

2. 组建团队

在 团队 中引入已配置的员工，设置团队内名称及上下级关系。员工定义可被多个团队复用，同一岗位也可作为不同成员加入团队。

团队的 资料管理员 开关控制团队共享及成员长期知识的使用与整理。共享资料管理员跟随团队唯一负责人的模型；员工私库使用该员工的模型和独立整理提示，不额外占用业务岗位。

3. 定义工作流

在 工作流 中新建或编辑流程，先引入团队，再为每一步选择该团队内的负责人。当前编辑器提供步骤说明、交付内容、验收标准和退回目标；输入资料与依赖可写入步骤说明：

• 步骤说明与输入：这一阶段要解决什么问题、依据什么资料。
• 交付内容：应产生的文件、结果或说明。
• 验收标准：如何判断完成，哪些情况需要退回。
• 不合格退回目标：本步骤或之前的有效步骤。

保存后，顺序与返工路径会在流程图中显示。步骤尝试次数和运行参数保持有界，避免没有结束条件的重试。

员工和团队引入项目时保存独立配置。当前工作流库则是流程定义的来源：编辑通过版本校验同步到项目引用，项目继续独立保存人员绑定、进度与交付历史。涉及已有交付的破坏性修改会被限制；契约变更可能使后续结果需要重新验证。不要把“编辑模板”理解为自动重做已执行任务。

4. 在项目对话中选择工作流

先在 DSH 中打开项目，再打开该项目的原生对话。输入区初始显示 工作流：无，菜单首项也是“无”。

首次选择某个工作流时，会创建其项目团队、员工和人员关系，不执行任务。该项目首次非空选择会成为项目默认，并清除之前的空覆盖，让项目对话统一继承；之后在某个对话改选其他工作流或“无”，只改变该对话的选择。同项目、同工作流共享步骤进度与验收记录。

直接输入任务即可由当前步骤负责人接手，无需 @。需要单独派工时，可在行首输入 @，选择当前项目已启用的员工；输入中间的 @ 继续使用 DSH 原生引用功能。

顶部步骤条以 灰色 表示未完成、蓝色 表示当前步骤、白色 表示已验收。点击步骤只切换当前步骤与负责人，不会把之前步骤标成完成。执行期间会限制身份与流程切换。原生空白对话的顶部区域在首次发言后显示。

5. 查看成果、验收与返工

点击顶部 详情，左侧查看完整步骤和负责人，右侧查看步骤要求、阶段文件、交付记录、任务摘要及验收结果。工作流 / 项目知识库 页签支持只读查看当前流程团队的跨阶段知识与来源。

查看其他步骤不会激活它，明确点击 设为当前 才会切换。通过当前负责人真实交付的人工验收后，流程才推进；退回需要填写意见，并按配置返回相应步骤。旧交付与验收记录保留，失效结果不再作为本轮有效成果。

选择 无 后恢复 DSH 原生对话行为，移除工作流负责人上下文和模型覆盖；它不会清空历史、重置原生最近使用模型，或改写全局模型默认。更多规则见工作流详情与验收。

专业认知与记忆如何工作

每位员工可以拥有自己的专业资料和可见记忆，团队也可以形成共享知识。专业认知不只是一段角色提示词，还包含能回到原文核对的资料、知识重组和关系结构。

层次	内容	用途
A · 原始资料	上传原件、阅读文本、版本、哈希与来源	核实原文、精确引用、比较变化与冲突。
B · 知识重组	基于来源整理的文章、解释及引用	按任务读取专业细节；模型综合保持候选状态。
C · 导航与关系	概览、概念／实体及有证据的语义关系	定位相关知识，再按需读取 B 或 A。

图谱支持搜索、拖动、缩放、关联展开与同屏阅读。连线来自有证据的语义关系，不因两个页面共用来源就虚构关系。员工私库、团队库和不同项目之间保持作用域隔离；共享同一套知识程序不等于共享全部数据。

资料管理： 在“专业知识”中上传 PDF、DOCX、UTF-8 文本、Markdown、JSON、CSV、YAML 或日志。单文件最多 6 MiB，每次最多 8 个文件／12 MiB。原件与提取文本分别保存；扫描 PDF 和图片文字不做隐式 OCR。抽取或整理失败会保留已归档资料，并提供状态与重试。

四类知识： 专业知识、任务档案、复盘经验、工作约束。后三类有独立自动归档开关；关闭分类开关停止后续自动收录，已有资料仍可阅读，重新开启不会立即补扫。团队资料管理员总开关关闭时，则停止该团队及成员长期知识的读取、写入和整理，保留数据。

增量整理： 来源更新后只处理变更及必要依赖，保留未受影响内容；重复上传或无变化重试不产生新的模型编纂。自动资料按任务稳定阶段合批，首次启动只本地归档。需要编纂时会通过 DSH 模型服务调用模型，可能产生模型费用；打开已有图谱或阅读文章不会因此发起编纂。

事实与候选分开： 模型提案、临时聊天和未验收产物不会自动变成已确认长期事实。失败复盘保留现象、证据、原因假设、建议与验证步骤；不会自行确认原因、修改 Skill 或执行建议。出现冲突时保留版本、适用条件和来源，自然语言冲突识别仍受模型能力与证据范围限制。

详见团队知识库、严格增量规则及共享 LLM Wiki。

分享、迁移与配置复用

员工、团队和工作流标题栏提供导入／导出；使用兼容 JSON 格式交换当前对象及必要依赖。导入先更新草稿，检查并保存后才写入，保留当前对象身份；依赖重新映射为新 ID，不覆盖无关对象。

• 格式标识为 graph_horses，formatVersion: 1，最大 1 MiB；JSON 配置文件不是插件安装包。
• 不导出项目路径、私有记忆、凭据、执行日志或项目工作台任务／成果历史；自由文本仍可能包含你填写的信息，分享前应检查。
• Skill、插件、MCP 和工具保存的是引用，接收环境仍需具备实际能力；导入不会自动安装或扩权。
• 兼容的项目实例配置导入默认暂停，不恢复或重放历史任务。

详见对象设置交换。

官方扩展边界、数据与卸载

插件不修改 DSH 源码、构建产物、内置插件或安装目录，只使用目标官方版本已有的 Cordis、bundle/profile、公开服务及 UI 插槽。岗位、Skill、知识和员工之间的消息不会扩大权限；执行继续遵循 DSH 原生预设、沙箱与审批。插件是第三方项目，不代表 DeepSeek 官方出品或背书。

默认专属数据位置为：

$DSH_HOME/composable-workforce/state.json
$DSH_HOME/composable-workforce/knowledge/state.json/

状态文件保存员工与团队定义、项目实例、记忆、运行及兼容业务记录；知识目录保存资料、原件、版本和派生知识。自定义 statePath 会改变对应位置，完整归属清单见生命周期指南。

保留数据卸载： 先停止目标宿主，执行官方命令，之后重启并刷新页面。专属配置与数据保留，便于重装恢复。

DSH_HOME=/absolute/dsh-home dsh plugin --profile web remove \
  dsh-plugin-composable-workforce --config.auto-install-peers=false

插件专属文件与配置清理： 使用随包脚本先预览，再在目标宿主已停止的条件下执行。请在移除插件前保留该脚本的可用位置；purge 会自行调用官方 remove。

# 指向实际官方 npm 安装和目标数据目录，不是 Harness 源码目录。
export DSH_INSTALLATION_MANIFEST=/official/install/node_modules/@deepseek-ai/dsh/package.json

node /installed/plugin/scripts/plugin.mjs purge \
  --home /absolute/dsh-home --profile web --dry-run

node /installed/plugin/scripts/plugin.mjs purge \
  --home /absolute/dsh-home --profile web --host-stopped

清理只处理经归属校验的插件配置、状态、知识、空专属目录及显式指定的自建服务日志。自定义外部状态需要 --owned-state-path；本仓库开发服务清理需要按指南指定 --service-root。用户项目、已交付成果、共享凭据、其他插件数据及不明文件保留。

当前未满足项： 官方 DSH 0.2.1-rc.3 没有会话持久记录删除接口，员工创建的专属原生会话历史仍会保留。因此，目前通过的是插件运行能力、专属文件和配置级清理，尚未满足全部插件专属持久记录清除的严格要求；0.28.0 是带保留项的验收候选，不能标记为无保留合规成品。 详见ADR-001和版本验收记录。

验证状态与已知限制

最近一次代码审计与归档复测基于官方 DSH、独立临时 profile 和固定归档 SHA-256：

检查	0.28.0 验收结果
类型检查与构建	通过
自动化测试	698 项：697 通过、0 失败、1 项可选私有数据测试跳过
真实归档安装、移除、重装与清理	15 项生命周期检查通过；原生会话持久记录保留项见上文
跨版本升级	官方 alpha.2 / 插件 0.27.0 → rc.3 / 插件 0.28.0，保留自定义岗位定义与库对象
官方原生界面与对话工作流	Web 29 项检查通过；macOS Desktop 实机安装、保存、卸载和重装通过
宿主程序完整性	27,607 个程序文件前后总哈希一致

界面与业务执行测试使用临时离线模型适配器，不等同于实际供应商的回答质量、费用、延迟或生产容量验收。当前仅实测 macOS arm64、Node 24.21.0 和指定 DSH 版本；本地 JSON 全量事务仍随历史规模增长。性能负载、原始证据、根因与后续计划见完整审计报告。

当前资源库界面只保留自由职业人、团队和工作流；工作入口是原生项目对话及顶部详情。项目目标、任务依赖、成果版本与并行调度等旧工作台后端能力和记录仍保留兼容，旧工作台编辑器并非当前公开页面入口。

开发与贡献

以下面向插件开发者，普通使用者按上面的 .tgz 安装即可。

先在本插件开发环境准备 package.json 列明的官方运行时 peer 依赖。现有 setup 脚本可从受支持版本的 Harness SDK 源码目录建立开发链接，只写本插件的 node_modules，不会构建或修改宿主；它使用 DSH_SOURCE_PATH，不是 DSH_INSTALLATION_MANIFEST。

npm ci --legacy-peer-deps
npm --prefix third_party/llm-wiki ci --ignore-scripts

# 可选开发链接方式：已有兼容 SDK 源码目录时使用。
DSH_SOURCE_PATH=/absolute/deepseek-harness-0.2.1-rc.3 npm run setup

npm run typecheck
npm test
npm run build
mkdir -p artifacts
npm pack --pack-destination artifacts

若已通过官方 npm 包在本插件环境准备好全部 peer，无需运行 setup。启动与生命周期脚本可用 DSH_INSTALLATION_MANIFEST 指向实际官方 npm 安装；它不替代本插件自身的 SDK 依赖准备。

命令	用途
npm start	启动本仓库独立开发 profile。
npm run service:start / service:status / service:stop / service:open	macOS 开发后台服务管理；其他系统使用独立终端 npm start。
npm run test:lifecycle -- artifacts/dsh-plugin-composable-workforce-0.28.0.tgz	在官方 npm 安装和临时 profile 验证真实归档生命周期。
npm run test:performance	本地合成负载下的状态、持久化、知识包与布局测量。

发布前需重新生成归档、SHA-256 和绑定该哈希的验收记录；改动随包文件后不能沿用旧归档验收。提交问题时请附版本、复现步骤、预期／实际行为与脱敏日志。新增能力需遵守 AGENTS.md 的官方扩展、数据归属和清理要求。

src/domain/   员工、团队、配置交换、状态与原子存储
src/harness/  官方服务、Agent、工具、模型及执行接入
src/client/   原生界面插槽、对话工作流与客户端通信
src/ui/       资源库、编辑器、知识阅读与交付详情
src/knowledge/ 资料归档、增量编纂、证据与知识范围
src/workflow/ 顺序流程状态机、验收、返工与恢复
src/work/     兼容项目目标、任务、成果版本及调度服务
skills/       22 项随包技能与软件岗位预设
integration/  官方接入、安装卸载及版本验收记录

常见问题

• 必须使用内置软件团队吗？ 不需要。它是可编辑的起点，你可以创建适合自己行业的员工、团队和工作流。
• 为什么升级后没有新增六岗位？ 默认对象只为新状态创建，已有 profile 不被静默覆盖或补种；可手动配置或导入经过检查的模板。
• 为什么选择某个 MCP 后仍不能调用？ 选择仅保存偏好，请先在 DSH 中安装、配置并连接实际服务，再检查原生权限。
• 为什么知识库显示待整理或失败？ 资料归档与模型编纂分开；检查负责人／员工模型是否可用、团队资料管理员开关及资料处理状态，再使用重试。
• 移除员工会删除项目历史吗？ 资源库移除会归档并隐藏定义，保留已有项目实例和记录；目前没有一键恢复入口。卸载插件与删除模板也不是同一操作。

文档、许可与致谢

更新记录 · 术语表 · 业务架构 · 工作流编辑 · 原生能力与 MCP · 配置交换

插件原创代码采用 MIT License。共享 LLM Wiki 程序及其对应源码采用 GPL-3.0，不受根目录 MIT 声明重新许可；各依赖保留自身许可证。完整范围见 LICENSES.md 与第三方声明。

感谢 DeepSeek Harness 提供插件与 Agent 扩展基础，以及 nashsu/llm_wiki 提供本项目知识程序所采用的上游模块。

↑ 返回顶部 · English ↓

---

English

Features · Included software team · Installation · Usage · Knowledge & memory · Uninstall & data · Development

One platform for the specialists your work needs

graph_horses is an independent third-party plugin for DSH. Define what each employee is responsible for, which capabilities and models they use, and what professional knowledge they can reference. Assemble employees into teams, then turn responsibilities into workflows with explicit owners, deliverables, and review criteria.

Capability	What you can do
Custom specialists	Define missions, duties, working standards, boundaries, models, Skills, and plugin/MCP preferences.
Collaborative teams	Reuse employee definitions, name team members, and configure reporting relationships and responsibilities.
Workflows	Specify an owner, inputs, outputs, acceptance criteria, and a rejection target for every step.
Native project conversations	Select a workflow in DSH and talk directly to its current owner, retaining native attachments, history, tools, and approvals.
Professional knowledge and memory	Upload references and inspect concepts, entities, articles, and original evidence in private or shared libraries.
Delivery and review	Inspect stage outputs and history, accept deliveries to advance, or return them for rework with an audit trail.
Portable configurations	Import and export employee, team, and workflow definitions with their required dependencies as JSON.

Example: configure an AI video production team

Example role	Responsibility	Example knowledge and memory
Producer / planner	Audience, concept, style, schedule, and delivery requirements	Brand guidelines, audience research, approved creative briefs
Screenwriter	Story structure, scripts, dialogue, and pacing	Character definitions, story world, script standards, approved revisions
Storyboard / visual designer	Shots, composition, and visual direction	Camera language, character appearance, color and composition rules
Image / audio producer	Generate or process assets with connected tools	Model guides, asset specifications, verified production experience
Editor / delivery specialist	Edit shots, sound, subtitles, and output versions	Editing conventions, subtitle styles, aspect ratios and export requirements
Quality reviewer	Narrative, character and visual consistency, technical checks	Review checklists, past defects, acceptance criteria

A workflow might be brief → script → storyboard → asset production → editing → quality review → delivery. Other industries use the same configuration model with different roles, capabilities, references, and acceptance criteria.

This is a configuration example. The included preset is the software development team. Video generation, speech, editing models, and external services must be connected separately through DSH; an employee can only use capabilities actually available in its environment.

A software development team is included

A fresh profile starts with six complete role definitions, a team, and two workflows:

Role	Focus	Required Skill
Product owner	Requirements, scope, priorities, acceptance, and coverage	requirements-acceptance
Technical lead / architect	Architecture, contracts, trade-offs, work breakdown, independent review	architecture-design
UX/UI designer	User journeys, interaction states, visual design, accessibility	ux-design-review
Developer	Incremental implementation, debugging, regression tests	incremental-implementation
QA lead	Functional, boundary, performance, and regression evidence	quality-verification
Operations lead	Packaging, installation, upgrades, handover, rollback, removal checks	release-readiness

• Software engineering · 10 steps: requirements and acceptance baseline → interaction design → architecture and implementation contracts → test and performance planning → incremental implementation → independent code review → functional and performance verification → packaging and handover → requirements coverage and user acceptance → retrospective and knowledge handover.
• Defect repair · 5 steps: reproduction and root-cause analysis → minimal repair → independent review → functional and performance regression → repair delivery.

Roles include responsibilities, boundaries, model inheritance, required/preferred Skills, working methods, evidence and memory policies, collaboration, and delivery definitions. Steps include owners, inputs, outputs, acceptance criteria, and rework rules. 22 bundled Skills cover the engineering lifecycle; see software role Skills.

You can edit, reuse, or replace these defaults. They contain no demo workflows, simulated completions, or fabricated confirmed memories. Installation does not execute business tasks. Upgrades preserve existing profiles instead of silently replacing definitions or adding new defaults.

Installation and first run

Requirement	Current scope
DSH	Official DeepSeek Harness 0.2.1-rc.3; other versions require separate validation.
Node.js	Declared range: `^22.19.0
Interface	DSH Web and official macOS Apple Silicon Desktop 0.2.1-rc.3. UI language follows DSH Chinese/English settings; user content stays in its original language.
Models	Configure a working model in DSH. Employees can inherit the default or use an explicit role model.
Distribution	The maintainer-provided dsh-plugin-composable-workforce-0.28.0.tgz and matching SHA-256 checksum.

As checked on 2026-10-26, the newest official GitHub tag and npm next are 0.2.1-rc.3; npm latest remains 0.2.1-rc.2. Select @deepseek-ai/dsh@0.2.1-rc.3 explicitly for this release.

End users need the supported official DSH installation and the plugin .tgz. There is no requirement to clone, patch, or rebuild Harness. Verify the checksum, stop the target host, and install into your intended home and profile:

# Replace the data directory, profile, and absolute archive path.
DSH_HOME=/absolute/dsh-home dsh plugin --profile web add \
  /absolute/dsh-plugin-composable-workforce-0.28.0.tgz \
  --config.auto-install-peers=false \
  --config.strict-peer-dependencies=false \
  --ignore-scripts

Restart Harness and refresh open pages after installation or an update. Compatibility is established by testing the exact official release and archive hash; the package-manager flags do not establish compatibility with other versions.

Fastest path: open graph_horses in the sidebar → inspect the six roles and software team → open a project and its native DSH conversation → select the included 软件工程 workflow → describe the task → inspect and review its delivery through Details in the conversation header. Preset names are user-editable business content and are not automatically translated.

See the installation and lifecycle guide for updates, ownership rules, and cleanup details.

Official Desktop app

DSH provides an official Desktop app from the DSH website. It shares the Web plugin system and uses the same .tgz. Web and Desktop have separate profiles; installing in Web does not install in Desktop. This project validates macOS Apple Silicon Desktop 0.2.1-rc.3; Windows and other architectures remain untested.

Open Desktop once to initialize it, then quit. Install into desktop with the command bundled inside Desktop; the npm CLI cannot manage this reserved profile. For a macOS app installed in ~/Applications:

"$HOME/Applications/DeepSeek Harness.app/Contents/Resources/runtime/cli/bin/dsh" \
  plugin --profile desktop add /absolute/dsh-plugin-composable-workforce-0.28.0.tgz \
  --config.auto-install-peers=false --config.strict-peer-dependencies=false --ignore-scripts

Adjust the path for /Applications. If using a custom DSH_HOME, launch the app and its command with the same home. Reopen Desktop to see graph_horses. Saving supports the official dsh-app://app authentication proxy. Remove with the same Desktop command and plugin --profile desktop remove dsh-plugin-composable-workforce; see Desktop integration for full cleanup.

From a role to a reviewed delivery

1. Define an employee

Open graph_horses → Freelancer library and select an existing role or create one. A library entry is a reusable definition; a project receives its own employee instance when the workflow is selected.

Setting	How to use it
Model	Inherit the current DSH default, or choose a provider, model, and reasoning effort. DSH manages accounts and credentials.
Role definition	Describe the mission, duties, working standards, and boundaries. AI autofill can produce an editable draft for review and saving.
Capabilities	Select required/preferred Skills and preferred plugins or MCP servers. Required Skills load once per work order; preferred Skills load when relevant.
Knowledge and memory	Upload references, read their knowledge graph and source evidence, and manage automatic archiving.

AI autofill uses the current draft name and model; its result is not automatically saved. Selecting a plugin or MCP server records a preference. It does not install, connect, or grant access to that integration—prepare the capability in DSH first.

2. Assemble a team

Add configured employees under Teams, assign team-specific names, and define reporting relationships. Employee definitions can be reused across teams, including multiple members based on the same role.

The team librarian switch controls shared and member long-term knowledge use and organization. Shared compilation uses the team's unique lead model; a private library uses its employee's model with a separate organization prompt. Neither creates an extra business employee.

3. Design the workflow

Create or edit a workflow, adopt a team, and choose each step's owner from that roster. The editor provides instructions, required outputs, acceptance criteria, and a rejection target at the current or an earlier step; describe input references and dependencies in the instructions. The diagram shows sequence and rework paths. Attempts and runtime settings are bounded.

Employee and team installations retain independent project configurations. In the current version, the workflow library is the definition source: validated edits synchronize to project references, while projects retain their own personnel bindings, progress, and delivery history. Destructive changes involving existing deliveries are restricted; contract changes can require later results to be revalidated. Editing a definition never automatically reruns completed work.

4. Select it in a native project conversation

Open your project in DSH, then open one of its conversations. The input initially displays Workflow: None, with None first in the menu.

Selecting a workflow creates its project team, employees, and reporting relationships without executing a task. The project's first non-empty selection becomes its default and clears earlier empty overrides so conversations inherit it. Later selections of a different workflow or None affect only that conversation. Conversations using the same workflow in the same project share step progress and review records.

Type normally to work with the current step owner; no @ is needed. For a separate work order, type @ at the beginning of a message and choose an enabled project employee. An @ elsewhere retains DSH's native reference behavior.

The header uses gray for pending, blue for current, and white for accepted steps. Selecting a step changes the current step and owner; it does not mark earlier steps complete. Identity/workflow changes are restricted during execution. The native header appears after the first message in an empty conversation.

5. Inspect, accept, or return the delivery

Open Details in the header to inspect the full step list, owners, requirements, stage files, deliveries, task summaries, and review history. The Workflow / Project knowledge tabs provide read-only access to cross-stage team knowledge and its sources.

Inspecting another step does not activate it; use Set as current explicitly. Only human acceptance of an actual owner delivery advances the workflow. Rejection requires feedback and returns to the configured step. Earlier deliveries remain in history, and invalidated results stop counting as current accepted output.

None removes the workflow owner context and model override and restores native DSH conversation behavior. It does not erase history, reset the last-used native model, or change the global model default. See delivery and acceptance.

Professional knowledge and memory

Employees can maintain their own references and visible memory, while teams build shared knowledge. Professional context includes source-backed information and relationships, alongside role instructions.

Layer	Content	Purpose
A · Sources	Original uploads, readable text, versions, hashes, provenance	Verify details, cite exact evidence, compare changes and conflicts.
B · Synthesis	Organized articles and explanations with source references	Read relevant professional detail; model synthesis remains a candidate.
C · Navigation and relationships	Overviews, concepts, entities, evidenced semantic relationships	Find relevant material, then read B or A as needed.

Search, pan, zoom, expand related concepts, and read articles and original evidence alongside the graph. Edges represent evidenced semantic relationships; sharing a source does not create a relationship. Private, shared, and project libraries retain their scope boundaries even though they use one shared knowledge program.

Uploads: PDF, DOCX, UTF-8 text, Markdown, JSON, CSV, YAML, and logs. Limits are 6 MiB per file, 8 files / 12 MiB per batch. Originals and extracted text are stored separately. Scanned PDFs and image text do not receive implicit OCR. Failed extraction or organization preserves archived material with status and retry support.

Categories: professional knowledge, task archives, retrospective experience, and working constraints. The last three have independent automatic archive switches. Disabling a category stops future automatic admission while existing knowledge remains readable; reenabling does not immediately backfill. Disabling the team librarian stops shared and member long-term reads, writes, and organization while preserving data.

Incremental organization: changes process only affected sources and necessary dependencies, preserving unaffected content. Duplicate uploads or unchanged retries do not generate fresh model compilation. Automatic material is batched at stable task stages; initial startup performs local archiving only. Compilation uses DSH model services and may incur model costs. Browsing existing graphs or articles does not itself trigger compilation.

Evidence and confirmation: model proposals, temporary chat, and unaccepted artifacts do not automatically become confirmed long-term facts. Retrospectives distinguish observations, evidence, causal hypotheses, recommendations, and verification steps; they do not confirm causes, modify Skills, or execute recommendations on their own. Conflicting versions and sources remain traceable; semantic conflict detection is limited by the supplied evidence and model capabilities.

See team knowledge, incremental processing, and shared LLM Wiki.

Share and reuse configurations

Employee, team, and workflow title bars provide JSON import/export for the current object and its required dependencies. Import updates a draft first; review and save to persist it. The current object's identity stays intact, dependencies receive remapped IDs, and unrelated objects are not overwritten.

• Format: graph_horses, formatVersion: 1, maximum 1 MiB. Configuration JSON is separate from the plugin installation archive.
• Exports exclude project paths, private memories, credentials, execution logs, and workbench task/delivery history. Free text may still contain information you entered; inspect it before sharing.
• Skills, plugins, MCP servers, and tools are references. The destination still needs the actual capabilities; import neither installs them nor grants permissions.
• Compatible project-instance imports start paused and do not restore or replay past tasks.

See configuration exchange.

Official extension boundary, data, and removal

The plugin uses the supported release's existing Cordis, bundle/profile, public services, and UI slots. It does not modify Harness source, builds, built-in plugins, or installation files. Roles, Skills, knowledge, and employee messages cannot elevate permissions; execution follows native DSH presets, sandboxing, and approval rules. This is a third-party project, without a claim of official DeepSeek authorship or endorsement.

Default plugin-owned data locations:

$DSH_HOME/composable-workforce/state.json
$DSH_HOME/composable-workforce/knowledge/state.json/

The state file contains role/team definitions, project instances, memories, runs, and compatible business records. The knowledge directory contains sources, originals, versions, and derived knowledge. A custom statePath changes these locations; see the lifecycle guide for the ownership inventory.

Remove while retaining data: stop the target host, run the official command, then restart and refresh. Configuration and data remain available for reinstallation.

DSH_HOME=/absolute/dsh-home dsh plugin --profile web remove \
  dsh-plugin-composable-workforce --config.auto-install-peers=false

Clean plugin-owned files and configuration: preview first, then execute with the target host stopped. Keep the packaged script accessible before removing the package; purge invokes official removal itself.

# The actual official npm installation, not a Harness source checkout.
export DSH_INSTALLATION_MANIFEST=/official/install/node_modules/@deepseek-ai/dsh/package.json

node /installed/plugin/scripts/plugin.mjs purge \
  --home /absolute/dsh-home --profile web --dry-run

node /installed/plugin/scripts/plugin.mjs purge \
  --home /absolute/dsh-home --profile web --host-stopped

Cleanup verifies ownership and removes this plugin's configuration, state, knowledge, empty dedicated parents, and explicitly identified custom service logs. External custom state requires --owned-state-path; repository development-service cleanup requires --service-root as described in the guide. User projects, delivered files, shared credentials, other plugins' data, and unknown files are preserved.

Open requirement: official DSH 0.2.1-rc.3 has no API for deleting persistent session records. Dedicated native employee-session history remains. Runtime, plugin-owned file, and configuration cleanup pass, but removal of every plugin-specific persistent record is still unmet. Version 0.28.0 is a validation candidate with this exception, not an unqualified compliant release. See ADR-001 and the version acceptance record.

Verification status and limitations

The latest code audit and archive verification used an official release, isolated temporary profiles, and one fixed archive SHA-256:

Check	0.28.0 result
Type checking and build	Passed
Automated tests	698 total: 697 passed, 0 failed, 1 optional private-data test skipped
Real archive installation, removal, reinstallation, cleanup	15 lifecycle checks passed; session-persistence exception above remains
Official native UI and conversation workflows	29 checks passed
Host program integrity	Identical aggregate hash across 27,607 program files before and after

UI and execution tests use temporary offline model adapters. They do not establish real provider answer quality, cost, latency, or production capacity. The verified environment is macOS arm64, Node 24.21.0, and the specified DSH release. Full-file local JSON transactions still grow with history. Loads, evidence, root causes, and follow-up plans are in the audit report.

The current library UI exposes employees, teams, and workflows; work happens through native project conversations and header details. Backend services and records for legacy project goals, task dependencies, deliverable versions, and parallel scheduling remain compatible. The old workbench editor is not a current public page entry.

Development and contributions

This section is for plugin development. End users install the .tgz as described above.

Prepare the official runtime peer dependencies declared in package.json inside this plugin's development environment. The existing setup script can link a compatible Harness SDK source directory into this plugin's node_modules; it neither builds nor modifies the host. It uses DSH_SOURCE_PATH, not DSH_INSTALLATION_MANIFEST.

npm ci --legacy-peer-deps
npm --prefix third_party/llm-wiki ci --ignore-scripts

# Optional development links when a compatible SDK checkout is already available.
DSH_SOURCE_PATH=/absolute/deepseek-harness-0.2.1-rc.3 npm run setup

npm run typecheck
npm test
npm run build
mkdir -p artifacts
npm pack --pack-destination artifacts

Skip setup if all peers are already prepared from official npm packages. Launch and lifecycle scripts can use DSH_INSTALLATION_MANIFEST to locate an actual official npm installation; this does not prepare the plugin's own SDK dependencies.

Command	Purpose
npm start	Launch the repository's separate development profile.
npm run service:start / service:status / service:stop / service:open	Manage the macOS development service; on other systems use npm start in a separate terminal.
npm run test:lifecycle -- artifacts/dsh-plugin-composable-workforce-0.28.0.tgz	Verify the actual archive using an official npm installation and a temporary profile.
npm run test:performance	Measure local synthetic state, persistence, knowledge-packet, and layout loads.

Before distribution, generate a new archive, checksum, and acceptance record bound to that hash. Changes to packaged files cannot reuse an older archive's acceptance. Report issues with versions, reproduction steps, expected/actual behavior, and redacted logs. Contributions must follow the public extension, ownership, and cleanup requirements in AGENTS.md.

src/domain/    Employees, teams, configuration exchange, atomic state
src/harness/   Official services, Agent, tools, models, execution
src/client/    Native UI slots, conversation workflows, client transport
src/ui/        Libraries, editors, knowledge reader, delivery details
src/knowledge/ Archiving, incremental synthesis, evidence, scope
src/workflow/  Sequential workflows, review, rework, recovery
src/work/      Compatible project goals, tasks, outputs, scheduling
skills/        22 bundled Skills and software role presets
integration/   Official integration, installation, removal, acceptance

Frequently asked questions

• Do I have to use the software team? No. It is an editable starting point. Create employees, teams, and workflows for your own domain.
• Why did upgrading not add six roles? Defaults are created only for new state. Existing profiles are neither overwritten nor silently reseeded. Configure roles manually or import reviewed templates.
• Why does selecting an MCP server not make it callable? Selection is a preference. Install, configure, and connect the actual service in DSH and check native permissions.
• Why is knowledge pending or failed? Archiving and model compilation are separate. Check the lead/employee model, librarian switch, and processing status, then retry.
• Does removing an employee delete project history? Library removal archives and hides the definition, retaining existing instances and records; there is currently no one-click restore. Removing a template is separate from uninstalling the plugin.

Documentation, licenses, and acknowledgments

Changelog · Glossary · Architecture · Workflow editor · Native capabilities and MCP · Configuration exchange

Original plugin code is licensed under the MIT License. The shared LLM Wiki program and its corresponding source are licensed under GPL-3.0; the root MIT notice does not relicense them. Dependencies retain their own licenses. See LICENSES.md and third-party notices for the complete scope.

Thanks to DeepSeek Harness for its plugin and Agent extension foundations, and nashsu/llm_wiki for the upstream modules used in the shared knowledge program.

↑ Back to top · 简体中文