dsh-plugins-plus

中文

这是由 SparkElf 维护的 DeepSeek Harness（dsh）独立插件仓库。这里的插件通过 profile composition bundle 安装到上游 dsh（dsh plugin --profile <name> add <package-or-git-spec>；每个包自行声明 Cordis 插件入口），并且不依赖 deepseek-harness-plus 产品分支。

仓库边界

• deepseek-harness-plus 跟踪 dsh 与 Plus 产品，并通过固定版本的 manifest 组合第三方插件。
• 本仓库保存 SparkElf 自有插件的源码，使这些插件也能直接用于原版 dsh。所有权与分发规则记录在 Plus 插件所有权与分发决策 中。

插件

• @sparkelf/dsh-plugin-ping：提供 /ping 连通性命令；无需模型调用即可回复 pong。
• @sparkelf/dsh-mobile-bridge：完整的 Host 与 Client 插件，提供出站 E2EE 隧道、配对二维码和一等公民的“移动连接”设置页。手机端只注入连接 facade；Harness 各功能包继续拥有自己的响应式界面。
• @sparkelf/dsh-mobile-bridge-server：多用户盲中继，支持邮箱验证码、可选的微信身份通道、页面内相机扫码和持久手机登录，让手机通过自有服务器访问本地 Harness。
• @sparkelf/dsh-dataops-integration：独立DSH的完整DataOps Host + Web Client连接插件，拥有Authorization Code + PKCE、持久target、session-lifetime access credential、Settings和generic MCP组合；不修改DSH core，也不在DataOps托管容器中挂载。
• @sparkelf/dsh-chart：完整的 Host + Web Client 交互图表 capability。Host 记录可 durable replay 的完整 ECharts option，Client 在 Harness 消息流中渲染并负责 resize、主题切换和 disposal。当前为 opt-in 发布候选，不默认挂载，也不在正式发布前声明生产 npm pin。
• @sparkelf/dsh-query-result-analysis：有界分页、checkpoint和hierarchical reduction的独立DataOps query-result分析插件；每个模型批次和归并进入持久DSH子会话，由普通AgentLoop记录模型可见内容并拥有重试。

CI / CD

• CI（ci.yml）：每次 push 和 PR 都执行 pnpm install、tsc --noEmit、Vitest 单元测试和可发布制品构建。
• CD（publish.yml）：v* tag 触发后，重新执行 typecheck、测试和构建，再使用 NPM_TOKEN secret 将包发布到 npm。
• UI 系统测试只断言可操作交互和可见状态变化；间距、对齐、像素几何、计算样式和截图由人在目标视口审查，截图不作通过条件。真实DataOps取数、分析、图表和刷新回放用例要求已配置且已连接的profile及显式业务prompt/marker；缺少输入时只记为skip，不作为验收通过。

添加插件

1. 创建 packages/<name>，按照 Cordis 插件模式实现 name、inject 和 apply；上游 dsh 的 @deepseek-ai/dsh-command-compact 可作为参考。
2. 在包旁添加不需要密钥的单元测试。
3. 插件不得依赖 deepseek-harness-plus；只能依赖已发布的上游包（@deepseek-ai/cordis、@deepseek-ai/dsh-*）。
4. 更新包版本，并通过 v<repo-version> tag 发布。
5. 根文档和面向用户的包 README 必须在同一 README.md 中维护“中文”和“English”两个对应章节，内容变更时同步更新。

English

This repository contains independent DeepSeek Harness (dsh) plugins maintained by SparkElf. The plugins install into upstream dsh through profile composition bundles (dsh plugin --profile <name> add <package-or-git-spec>; each package declares its own Cordis plugin entry) and never depend on the deepseek-harness-plus product fork.

Repository boundaries

• deepseek-harness-plus tracks dsh and the Plus product, and composes third-party plugins through a version-pinned manifest.
• This repository owns the source of SparkElf plugins so they remain usable with vanilla dsh. The ownership and distribution rules are recorded in the Plus plugin ownership and distribution decision.

Plugins

• @sparkelf/dsh-plugin-ping: provides the /ping connectivity command and replies with pong without a model call.
• @sparkelf/dsh-mobile-bridge: a complete Host and Client plugin providing an outbound E2EE tunnel, pairing QR, and first-class Mobile Bridge settings section. The phone injects only the connection facade; each Harness package retains ownership of its responsive UI.
• @sparkelf/dsh-mobile-bridge-server: a multi-user blind relay with email-code and optional WeChat identity channels, in-page camera scanning, and persistent phone sign-in, allowing phones to reach a local Harness through a self-hosted server.
• @sparkelf/dsh-dataops-integration: the complete Host + Web Client DataOps connection plugin for standalone DSH. It owns Authorization Code + PKCE, a persistent target, a session-lifetime access credential, Settings, and generic MCP composition; it changes no DSH core and is not mounted in DataOps-managed containers.
• @sparkelf/dsh-chart: a complete Host + Web Client interactive-chart capability. The Host records a complete ECharts option for durable replay; the Client renders it in the Harness conversation and owns resize, theme changes, and disposal. It is currently an opt-in release candidate, is not mounted by default, and does not claim a production npm pin before an actual release exists.
• @sparkelf/dsh-query-result-analysis: an independent DataOps query-result analysis plugin with bounded paging, checkpoints, and hierarchical reduction. Each model batch and reduction runs in a durable DSH child session whose normal AgentLoop logs model-visible content and owns retry.

CI / CD

• CI (ci.yml): runs pnpm install, tsc --noEmit, Vitest unit tests, and publishable artifact builds on every push and PR.
• CD (publish.yml): on v* tags, repeats typecheck, tests, and builds before publishing packages to npm with the NPM_TOKEN secret.
• UI system tests assert only operable interactions and visible state changes. Humans review spacing, alignment, pixel geometry, computed CSS, and screenshots at target viewports; screenshots are not pass criteria. The real DataOps query, analysis, chart, and reload-replay case requires a configured connected profile plus explicit business prompt/markers; missing inputs produce only a skip and never acceptance evidence.

Adding a plugin

1. Create packages/<name> and implement name, inject, and apply following the Cordis plugin pattern; upstream dsh's @deepseek-ai/dsh-command-compact is the reference.
2. Add keyless unit tests beside the package.
3. Keep the plugin independent of deepseek-harness-plus; depend only on published upstream packages (@deepseek-ai/cordis, @deepseek-ai/dsh-*).
4. Bump the package version and publish through a v<repo-version> tag.
5. Root-facing and user-facing package documentation must keep matching “中文” and “English” sections in the same README.md, updated together whenever the content changes.