dsh-harness-panel

dsh-harness-panel

A DeepSeek Harness plugin that adds a Harness page to Settings — one place to inspect and start/stop MCP servers, enable/disable skills, and start/stop hook bridges, styled to match the native settings UI.

English · 中文

Features

MCP

• Every configured @deepseek-ai/dsh-mcp-client row: connection status (observed from the upstream mcp/status seam, honestly unknown when unobserved), transport, sanitized target, tool inventory
• A native start/stop switch on every addressable row, through the same pluginManager Remote the built-in Plugins page uses; disabling stops the mcp-client row and its tools
• One-shot connectivity probe per server (streamable-http initialize POST / stdio spawn handshake, 10s deadline)
• Env/header values never leave the host (keys only, URLs credential-redacted)

Skills

• The current session's user-invocable skill catalog — the same seam the composer / menu reads (remote.skills), so the panel always agrees with the composer
• A native enable/disable switch per skill. Disabling sets the provider's own frontmatter keys (user-invocable: false, disable-model-invocation: true) in the skill's SKILL.md, so it leaves the model catalog and the / menu in every registry scope; the panel keeps listing it with an off switch, and re-enabling restores the original frontmatter lines
• Search filter, model-invocable badge, source path

Hooks

• Mounted hook bridges (@deepseek-ai/dsh-hooks-claude-code / @deepseek-ai/dsh-hooks-codex): each row's configPath file parsed into event → matcher → command views
• A per-command switch inside each expanded group: turning one hook off flips its type to the sentinel the bridges skip, records the original value, and reloads the bridge row so it takes effect immediately. The bridge row itself deliberately carries no whole-row switch — the loader's live re-enable of a row is unreliable, so the page stays at the granularity the file edit serves
• Conventional on-disk locations (~/.claude/settings.json, ~/.codex/hooks.json, project-level, …) scanned read-only and clearly marked not mounted

Design

The visual language follows the shipped settings sections rather than inventing its own: --dsw-* design tokens only (no hard-coded colors), native corner radii (--dsw-radius-xl cards, --dsw-radius-md controls), 0.5px hairline strokes, the underline tab strip used by the built-in Plugins page, and the shipped Switch atom from @deepseek-ai/dsh-client-ui-primitives — the exact control the Plugins page uses. Light and dark themes both work.

Install

Desktop profile (~/.dsh/profiles/desktop):

cd ~/.dsh/profiles/desktop
# package.json → dependencies: "dsh-harness-panel": "link:
"
# package.json → dsh.profile.bundles: add "dsh-harness-panel"
pnpm install
# restart DSH (the desktop profile has no live reload)

Or install by local path from Settings → Plugins.

Security boundary

• MCP and hook bridge switches call the shipped remote.pluginManager.setPluginEnabled: the official manager writes the profile's cordis.patch.yml and reloads the row, exactly as the built-in Plugins page does. This host half registers no write route for them
• The host half adds one write route, POST /harness-panel/skills/toggle, for skill toggles. It requires content-type: application/json, an x-harness-panel: 1 header, and a same-origin Origin when the browser sends one; the body is bounded to 16 KiB. A toggle edits only the two invocation keys of the named skill's own SKILL.md (verified against the file's declared name) and records the original lines for restore
• Beside those two frontmatter lines, the only file the host writes is $DSH_HOME/harness-panel/state.json (default ~/.dsh/harness-panel/state.json): skill names, descriptions, paths, disable times, and the original frontmatter lines, mode 0600. No instruction bodies, no credentials
• Credentials never leave the host: env/header maps surface as key lists, URLs and error text are redacted
• Routes bind to loopback only (the dsh webserver default)

Development

pnpm install
pnpm build       # emits lib/index.js (host ESM) + lib/client.js (browser bundle)
pnpm typecheck
pnpm test        # vitest: mcp aggregation, hooks parsing, sanitization, skill file/state, route gate, switch models

The scripts prefer bunx and fall back to the local Node. On the macOS desktop install the DSH-bundled Node is hardened and cannot load the ad-hoc native rollup/rolldown bindings, so bunx tsdown / bunx vitest run is the path that works there.

Compatibility: DSH >=0.2.0 <0.3.0 (developed against desktop 0.2.1-rc.1).

---

中文

DeepSeek Harness 设置面板插件：在设置导航中新增 Harness 入口，集中查看并启停 MCP 服务器、Skills 与 Hooks，视觉与原生设置页一致。

• MCP：已配置服务器的连接状态（来自上游 mcp/status 事件，无观察时如实显示未知）、传输方式、脱敏目标、工具清单，一键探活；每个可寻址行右侧有原生启停开关，停用后 mcp-client 行与其工具一起下线
• Skills：当前会话可见的 skill 目录（与 / 菜单同源），每个 skill 右侧有原生启用/停用开关。停用直接写入该 skill SKILL.md 的官方 frontmatter 键（user-invocable: false、disable-model-invocation: true），在任何 registry scope 都生效，skill 从模型目录与 / 菜单消失；面板仍以关闭状态列出，重新启用时按记录还原原始行
• Hooks：已挂载的 hook 桥配置解析（事件点 → 匹配器 → 命令），展开后每条 command 有独立开关——关闭时把该 hook 的 type 改为 bridge 跳过的哨兵值（provider 自身的机制），记录原始值并重载 bridge 行立即生效。bridge 整行不设总开关：loader 对整行的热恢复不可靠，面板停留在文件编辑能服务的粒度。本机约定路径上未挂载的配置文件继续只读扫描

安全边界：MCP 与 hook 开关调用 shipped remote.pluginManager.setPluginEnabled，由官方 manager 写 profile 的 cordis.patch.yml 并 reload，与内置 Plugins 页一致；host 只新增一条 skill 开关写入路由 POST /harness-panel/skills/toggle，要求 JSON、x-harness-panel: 1 与同源 Origin，body 上限 16 KiB，且只改目标 SKILL.md 里两个 invocation 键（先校验文件声明的 skill name）。除这两行外，host 唯一直接写出的文件是 $DSH_HOME/harness-panel/state.json（默认 ~/.dsh/harness-panel/state.json，0600），只含 skill 名称、描述、路径、停用时间与原始行记录，不含正文与凭据。路由仅监听 loopback。

开发命令见上文；安装或重新构建后需重启 DSH（desktop profile 无 live reload）。