dsh-skill-mcp-manager

Manage skills and MCP servers from the DSH web UI — a dual-half plugin that adds a "Skills & MCP" section to the Settings page.

• Skills tab: list the user's skills, create new SKILL.md bundles — by typing the instructions or uploading a .md file — and toggle each skill's model / user invocation visibility. By default, the plugin scans $DSH_HOME/skills followed by ~/.agents/skills, matching the filesystem provider's user-root precedence. New skills are written to $DSH_HOME/skills, which @deepseek-ai/dsh-skill-filesystem already watches — a created or toggled skill appears in the next catalog observation automatically. Removal is disable-only in v1 (frontmatter switch), no file deletion.
• MCP Servers tab: list configured servers with live connection state, add (stdio or streamable-http), and remove. Servers persist to the machine-global ~/.dsh/cordis.patch.yml as @deepseek-ai/dsh-mcp-client instances; DSH's own user-patch HMR watcher hot-applies the file, so adds and removes go live without restarting the GUI — verified end-to-end against a real harness instance.

Install

dsh plugin --profile web add /path/to/dsh-skill-mcp-manager

or, file-level (no pnpm needed because the profile's shared module fallback provides all @deepseek-ai runtime deps):

1. Add the bundle to $DSH_HOME/profiles/web/package.json:

   {
     "dependencies": { "dsh-skill-mcp-manager": "link:/path/to/dsh-skill-mcp-manager" },
     "dsh": { "profile": { "bundles": [ /* ...existing..., "dsh-skill-mcp-manager" ] } }
   }
   

2. Symlink it into the profile so the loader resolves it:

   ln -s /path/to/dsh-skill-mcp-manager "$HOME/.dsh/profiles/web/node_modules/dsh-skill-mcp-manager"
   

3. Restart the GUI (dsh web). The new "Skills & MCP" section appears in Settings.

Architecture

Browser (Settings page: Skills & MCP)
   │  Typert RPC  (5 methods over ctx.remote.skillMcpManager)
   ▼
Host half (dsh-skill-mcp-manager)
   ├─ Skills  →  $DSH_HOME/skills/<name>/SKILL.md   (filesystem provider watches)
   └─ MCP     →  settings ns "skill-mcp-manager.mcpServers"  (durable truth)
                 └─ reconcile → ~/.dsh/cordis.patch.yml  ← insert rows
                                   └─ profile-boot's watchUserPatches (HMR)
                                         → loader mounts/disposes dsh-mcp-client

Skills

• Add (typed): validates kebab-case name, writes <root>/<name>/SKILL.md in the standard frontmatter format (name, description, optional whenToUse).
• Add (upload): pick a .md file and every editable field is pre-filled from the file's frontmatter — name (kebab-suggested; title/H1 fallback), description, whenToUse, and the show-to-model/user toggles — so no retyping is needed (you can still edit them). The host keeps the file's body verbatim and preserves its remaining frontmatter keys (custom keys such as metadata); form fields take precedence when edited. Add without a name also derives it from the file. Uploads are capped at 1 MB.
• Toggle: rewrites only disable-model-invocation / user-invocable, body and every other frontmatter key preserved byte-for-byte.
• List: scans the configured roots ($DSH_HOME/skills and ~/.agents/skills by default) one level deep — bundle dirs and flat .md files — parsing the same format dsh-skill-filesystem uses. Duplicate names use the first root, matching the provider's precedence.

Skill review

The upload button appears first in the Add skill form. Expanding a listed skill renders its complete Markdown body with the harness renderer. The source toggle shows the original file, including all frontmatter.

MCP servers

• Each connection shows its registered tools. Docker gateways using --profile also show that profile's child servers, including children with no registered tools. Discovery uses the read-only Docker CLI and matches catalog tool names against the harness registry. Catalog snapshots may omit newer tools; the connection's full registered-tool list remains available.

• A loaded plugin is labeled loaded, since loader activation alone does not prove a successful MCP connection. Child tool counts indicate detected tools, not a live health check. Docker child servers are managed through Docker.

• Remove is available only for connections owned by this panel. Optional connection fields are grouped under Advanced options.

• Docker gateways with --servers show the selected names. Other gateway types and Docker gateways without either selection show their registered tools; Docker discovery limitations and CLI failures appear as warnings.

• Settings is the source of truth. On every change the manager rewrites $DSH_HOME/cordis.patch.yml, preserving every row it does not own (including !!js expressions, which round-trip verbatim through the same js-yaml dialect the loader include uses). It never touches a file that is already in sync, so manual comments survive until an actual change.

• Live status comes from ctx.loader entries filtered to @deepseek-ai/dsh-mcp-client, so a server added outside the manager shows up too (marked external), and connect failures surface as failed/pending.

• The @deepseek-ai/dsh-mcp-client bridge is imported only as a resolvability probe; the GUI shows explicit install guidance when it is missing.

Verification

• node tests/discovery.mjs — Docker profile selection, child tool matching, and failure containment.
• node tests/smoke.mjs — host-half smoke test (stubbed context, temp dirs; never touches $DSH_HOME).
• Second-instance integration (temp DSH_HOME, dsh --profile <it> on a free port): the manager's patch format is applied by the real loader, a fixture stdio MCP server handshakes, and live add + live remove via home-patch rewrites were observed (zero restart). The browser boot manifest includes the package and its /plugins/dsh-skill-mcp-manager/client.js serves 200.

Config (plugin-level)

Field	Default	Meaning
skillRoots	[] → $DSH_HOME/skills, ~/.agents/skills	Absolute or ~ roots to list/manage; the first root wins duplicate names and receives new skills
mcpPatchTarget	'' → $DSH_HOME/cordis.patch.yml	Patch file rows are projected into

Known limitations

• Disable-only skill removal in v1 (per scope decision); deleting files is a follow-up. ~/.agents/.skill-lock.json (a third-party installer's manifest) is never touched and may not reflect changes made here.
• The patch file is re-serialized on real changes, which drops its comments (the "managed by" header is regenerated). Only rewrite when content drifts.
• A serverName you also hand-maintain in another patch layer is flagged in the UI as a collision; the manager never deletes rows it does not own.
• The first boot writes pre-existing settings into the patch file; on surfaces where no HMR watcher is active yet that picks it up on the next run — the GUI flow (save from the UI) happens after boot, when the watcher is live.