dsh-control-center
DeepSeek Harness 的三个设置页面：管理 MCP 服务器、从 GitHub 编写或拉取代理技能，以及维护全局角色库。
插件会安装到这里；不确定时保持 web。
npx -y @deepseek-ai/dsh plugin --profile web add github:MYanhui/DSH-plug-in#08205bb9b77b13169d2e7a60f0d4783ac3955378Control Center 以 dsh-control-center 发布，当前版本为 1.0.0。Plugin Hub 会校验它的 manifest，并保存精确安装来源，便于复现安装结果。
点击下方 Badge 复制 Markdown，粘贴到 README 即可。
验证 package.json 声明的 GitHub 仓库，即可管理这个公开页面。认领后，Hub 会优先安排当前版本的安全扫描，并在通过后公开展示结果。
认领这个 Plugin →English | 中文
Three visual Settings pages for the DeepSeek Harness: MCP servers, skills and global personas. One dual-face plugin — a host half that owns the data and the one model-facing contribution, and a browser half that renders the pages into the settings shell.
The Chinese document (README.zh.md) is the fuller one: it adds the verification log, troubleshooting, and the one-click script reference.
Settings
├── …
├── MCP 管理 MCP server CRUD
├── Skill 管理 author or pull skills, inspect and delete them
└── 全局人设 a library of persona documents, one active
The package is a normal DSH bundle. Any channel works.
One script (Windows). Copies the package, backs the patch up once, and
writes the marked mount block — idempotent, -DryRun to preview, -Profile /
-DshHome to target another profile:
powershell -NoProfile -ExecutionPolicy Bypass -File .\scripts\install.ps1
powershell -NoProfile -ExecutionPolicy Bypass -File .\scripts\uninstall.ps1 # one-key restore
Official CLI, which also records the bundle in the profile manifest:
dsh plugin --profile <name> add dsh-control-center
By hand, for a profile whose node_modules is managed manually: copy this
directory to <profile>/node_modules/dsh-control-center, then append the mount
row to the profile's own cordis.patch.yml:
- insert:
- id: control-center
name: 'dsh-control-center'
dsh-hmr watches that file, so the plugin activates on save; the browser half
is part of the client boot graph and needs one page refresh the first time it
appears. After the first installation the script leaves
cordis.patch.yml.bak-dsh-control-center as the final rollback point.
An MCP server is a Loader row, not a settings value, so the page writes a
delimited managed block into the active profile's own patch layer and lets
dsh-hmr reload the composition:
# >>> dsh-control-center:mcp — managed, edit it from Settings instead
- insert: [{"id":"cc-mcp-github","name":"@deepseek-ai/dsh-mcp-client","config":{…}}]
# <<< dsh-control-center:mcp
Everything outside the two markers is preserved byte for byte, and the payload
is JSON — a strict subset of YAML 1.2 — so no URL, argument or environment value
can change the document's shape. The definitions live in
$DSH_HOME/control-center/mcp-servers.json; the block is regenerated from them
on every write, so disabling a server removes its row without losing it.
The row reports its live Loader state, so the list distinguishes configured, enabled and connected.
Skills are ordinary files, so this page is a filesystem editor over the roots
dsh-skill-filesystem already scans — it keeps no database:
| Root | Source |
|---|---|
$DSH_HOME/skills | user-dsh, provider rank 400 — where installs go |
$DSH_AGENTS_HOME/skills | user-agents, rank 500 |
any customSkillDirs | from a live skill-filesystem row |
Project roots (.dsh/skills, .agents/skills) depend on the session's working
directory rather than on this global page and are deliberately left out.
New skill writes <root>/<name>/SKILL.md with name, description,
whenToUse and the body.
Raw editor opens the current file verbatim; saving re-validates the frontmatter.
Pull from GitHub downloads one source archive, finds every SKILL.md at a
plausible depth, and installs each bundle. Three URL shapes are accepted, plus
the git@github.com:owner/repo.git spelling:
https://github.com/owner/repo
https://github.com/owner/repo/tree/<branch>
https://github.com/owner/repo/tree/<branch>/<subpath>
No git binary is needed: the archive is fetched from codeload and unpacked
by a small built-in tar reader. An existing skill is only replaced when
overwrite is on.
The filesystem provider watches its roots, so a created, edited or deleted skill reaches the next model step without a restart.
Each persona is a Markdown document under
$DSH_HOME/control-center/personas/, with its name and description in
personas.json. Exactly one is active, and the page can clear the seat.
The active document is registered as one system-prompt section:
ctx.systemPrompt.section({
name: 'control-center:persona',
order: 5,
text: activeDocument,
interpolate: false,
})
Three properties follow, and all are intended:
deployment:persona-prefix for one agent
does not remove it, so every session is covered;interpolate: false keeps {{…}} as written, so a stray brace
in prose can never fail prompt assembly.Switching the active document disposes and re-registers the section; the next assembly carries the new text.
One POST-only route, /control-center/api, with a method field:
| Method | Purpose |
|---|---|
system.info | paths and version |
mcp.state / mcp.save / mcp.remove / mcp.toggle | MCP servers |
skill.state / skill.read / skill.save / skill.write / skill.remove / skill.pull | skills |
persona.state / persona.read / persona.save / persona.remove / persona.activate | personas |
Responses are { ok: true, value } or { ok: false, error: { code, message } }.
The route is fenced against cross-origin browser callers: an Origin header must
name a loopback host, and a present Content-Type must be JSON. A caller with no
Origin — the Electron IPC bridge, a shell tool — is already running with the
user's own authority.
Both scripts need no dependencies and no harness:
node scripts/selfcheck.mjs # host half: frontmatter, tar, the managed block, all three stores over real files
node scripts/clientsmoke.mjs # browser half: bundle, slot registration, every section rendered with stubbed host replies
skill.pull's network path is not exercised by either; verify it against a real
repository from the running harness.
index.js host half: route, API table, persona section, MCP reconciliation
client.js browser half: the three Settings sections
lib/wire.js body reader, JSON responder, request fence
lib/paths.js DSH_HOME / profile locations, atomic read-write
lib/mcp.js MCP store + the managed patch block
lib/skills.js skill roots, frontmatter, CRUD, GitHub install
lib/github.js repository reference parsing + archive download
lib/tar.js minimal tar reader
lib/personas.js persona store + the prompt section lifecycle
scripts/install.ps1 one-key install (UTF-8 with BOM — see README.zh.md)
scripts/uninstall.ps1 one-key restore
scripts/selfcheck.mjs host-half checks
scripts/clientsmoke.mjs browser-half checks
package.json is already shaped for it (no private, license, repository
with directory, keywords, engines.dsh, peerDependencies,
dsh.manifestVersion) and locale/{en,zh}.json supplies the store card copy.
The catalog is the curated list awesome-dsh-plugin:
add ONE file under data/plugins/, run node scripts/generate-readme.mjs, open
a PR. The step-by-step checklist — plus the entry YAML, a pre-flight validator
and the screenshot template — lives in the repository's publish/ directory.