通过两个受限的 DSH 工具和网页工作区卡片,创建并检查 RecoWork local-agent-project 工作区。
安装
# npm 包(预构建)
dsh plugin --profile web add recowork-dsh
# GitHub 源码(首次需按提示配置 allowBuilds 构建授权后重试)
dsh plugin --profile web add github:recoluan/recowork#path:/packages/recowork-dsh
装任何插件都等于在你的机器上跑第三方代码,权限和你本人一样大——能读你的文件、用你的凭据、访问网络,工具审批管不到它。GitHub 来源的插件还会在安装时执行构建脚本——pnpm 默认拦截,所以安装可能停在 ERR_PNPM_GIT_DEP_PREPARE_NOT_ALLOWED 或 ERR_PNPM_IGNORED_BUILDS;dsh 会打印出需要添加的确切键名,把它加进该 profile 的 pnpm-workspace.yaml 的 allowBuilds 下,重跑一次即可装上。放行构建本身就是一次信任判断:请只安装可信来源,并尽量锁定 commit(github:owner/repo#sha)。
README
该插件的 README 只有英文版本。
recowork-dsh is an experimental DSH Bundle that gives an agent three narrowly scoped tools:
recowork_init: initializes one new RecoWorklocal-agent-projectworkflow.recowork_status: reads its manifest and a small, fixed set of current-work documents.recowork_import: safely adds RecoWork to one existing non-empty project without overwriting existing paths.
It also contributes a model-facing workflow protocol: use initialization only on an explicit request, inspect a workspace before durable work, then read its AGENTS.md and localized working methods before planning or editing. It intentionally does not expose a shell, arbitrary file reads, upgrades, deletion, or renaming.
In DSH Web it adds a collapsible RecoWork project cockpit:
- Selecting a recognized workspace automatically loads its authorized read-only status. The main dashboard shows a full-context continuation action, blocking decisions from open questions, the current project stage, and stable project-memory items extracted from the project brief.
- A guarded action can be queued into the current DSH conversation or copied as the same task package for another conversation. The package includes the workspace path, status read, root
AGENTS.mdprotocol, and the selected objective. - Technical metadata, document health, current-document names, and manual refresh live under Status & settings. Add project separates explicitly confirmed New workspace and Import existing project flows.
The cockpit does not maintain a second task database and cannot edit, delete, repair, or upgrade an initialized workspace. Its generated prompts instruct the agent to read AGENTS.md and the localized working methods, treat workspace documents as authoritative, and honor the project's confirmation rules before changing direction, scope, or files. New-workspace initialization rejects a non-empty destination. Existing-project import lists only non-empty direct children without rw-manifest.json, requires separate confirmation, adds only missing RecoWork files, preserves every matching path, and confines any existing root AGENTS.md change to its marker-bounded RecoWork block.
Its interface follows DSH's active zh or en preference live, including when the user changes the language in DSH Settings. Workspace document names and action text remain in the language of the selected workspace.
The Bundle has no runtime npm dependency on DSH internals. This keeps a locally linked checkout resolvable from its own source directory; DSH accepts the plugin's raw JSON-Schema tool definitions directly.
Install
- Install DeepSeek Harness and start its Web profile once so its profile exists.
- Install the published Bundle into the Web profile, then explicitly authorize one or more existing directories:
dsh plugin --profile web add recowork-dsh@latest
npx --yes recowork-dsh@latest setup --root /absolute/path/to/your/recowork-lab
dsh web
Repeat --root to authorize more than one directory. If your DSH home is not the default ~/.dsh, either set DSH_HOME or pass it explicitly:
DSH_HOME=/path/to/dsh-home npx --yes recowork-dsh@latest setup --root /absolute/path/to/your/recowork-lab
The setup command creates a timestamped backup of cordis.patch.yml, handles DSH's default [] placeholder as a single YAML document, and validates the resulting YAML after writing. If validation fails, it restores the original profile patch before reporting the error. If you previously configured recowork-dsh by hand, review that entry and rerun the command with --adopt-existing; it will back up the file before replacing only the old RecoWork entry. The default initializer and importer run npx --yes recowork@4.1.0.
For local development, replace the first command with dsh plugin --profile web add ./packages/recowork-dsh, then run node ./packages/recowork-dsh/bin/recowork-dsh.js setup --root /absolute/path/to/your/recowork-lab --dsh-home "$DSH_HOME".
Safety contract
rootmust exactly match an entry inallowedRoots;destinationmust be a descendant of that root.- Initialization requires a missing or empty directory and only permits the three supported templates and
zh/enlocales. - Import requires an existing non-empty project without
rw-manifest.json, preserves same-name paths, and only permits the same templates and locales. - Status requires
rw-manifest.jsonand reads only fixed, localized current-work files, capped at 12,000 characters each. - Existing RecoWork workspaces remain user-owned. Use
rw statusorrw upgradedirectly for upgrade review.
The only DSH Web write actions are explicitly confirmed new-workspace initialization and safe existing-project import. Both stay under the approved-root boundary and delegate lifecycle behavior to the RecoWork CLI; neither can bypass existing confirmation gates or workspace-ownership rules.
链接
同类插件
loopx-project/loopx#dsh-loopx-plugin★ 6233
LoopX——面向长周期 Agent 的提供商中立、本地优先状态内核与控制平面:在 DeepSeek Harness 执行层之上持久化 Goal、Todo、门禁、证据、配额、恢复与交接状态;插件负责引导安装 CLI 与技能、准入有界的同会话续跑,并为精确绑定的工作循环提供本地 GoalBar。
Q00/ouroboros#integrations/dsh-plugin★ 6196
通过 DSH MCP 客户端挂载 Ouroboros 的纯配置包,在 DSH 中提供 36 个涵盖需求访谈、Seed、执行、评估与演化流程的工具。
chuspeeism/dashi-taskboard#deepseek-harness★ 3329
把当前已安装并运行中的 Codex Taskboard 嵌入 DeepSeek Harness 侧边栏,并通过 Launcher 运行时描述文件连接,而不是使用固定端口。
NanmiCoder/dsh-agent-teams★ 2012
AgentTeams 多智能体团队。
EthanYoQ/AI-Novel-Writer#dsh-ai-novel-writer★ 1393
安装专用 AI 小说创作预设与工作台:提供带修订号的本地项目资产、紧凑侧边工作台,以及需要原生审批的逐文件变更。
tong-io/tongflow#dsh-tongflow★ 1042
基于 TongFlow 的“片场”插件,用于图片、配音、音乐与视频制作:agent 为每个资产生成 TongFlow 工作流文件(.tongflow.json)并通过 TongFlow 插件执行,内嵌工作流画布,按镜头/角色/take 组织项目,附漫剧模板;以 @tongflow 开头的会话进入 Studio 界面。
社区评论
评论公开保存在 GitHub Discussions。加载评论会连接 GitHub 和 Giscus;发表内容需要 GitHub 账号。