ouli-1242/dsh-plugin-tool-management

MCP servers, skills, subagents, prompts, memories and archived sessions for DeepSeek Harness — all in one panel. Bundle them into a scene, swap the whole environment with one click, and roll back to the pre-entry snapshot when you leave.

dsh plugin --profile web add dsh-plugin-tool-management@latest      # web
dsh plugin --profile desktop add dsh-plugin-tool-management@latest  # desktop

Hard-refresh the browser after installing for web (Cmd/Ctrl+Shift+R), or restart the desktop app — then open Settings → Tools Management. For desktop, open the app once and quit it completely before running the command. An older version usually means registry lag or pnpm's minimumReleaseAge — install an exact version (@0.X.X) instead.

---

Highlights

Demo

Player refuses to start (browsers with a stricter CSP sometimes block the inline embed)? Watch the same file directly.

Every interface shown is a real screenshot of the running product.

Music: Kevin MacLeod「Limit 70」, CC BY 4.0.

Screenshots

Scenes	MCP
Skills	Subagents
Prompts	Memories
Sessions	Host

Three scopes

Every conversation gets its own settings by default. Switch once on the Host page → Scope, and the whole panel plus every model tool follows that layer:

	All global	Per workspace	Per conversation (default)
Who shares one set	Every conversation	Conversations in the same workspace	Each conversation separately
Best for	One person, one machine, consistent everywhere	One project, one environment	Parallel tasks in one project
Lost when switching?	—	Nothing is deleted; switch back to restore	Nothing is deleted; switch back to restore

The selector at the top of the panel jumps to a specific conversation or workspace and shows its own copy; when that layer is locked, the unlock button sits on the same row. Within a layer, every switch resolves top-down: this layer's edit > entered scene > creation snapshot > global.

Scenes work in all three scopes. Under per-conversation, conversation A can sit in “Coding” while B stays in “Writing”, each with its own MCP servers, skills and personas.

What each tab does

Let the agent drive it

Six domains, 26 model tools: list, read, create, update, delete, toggle, enter or leave a scene, apply a prompt, run a persona.

• Only 6 ship enabled (~1.6k tokens/turn); the other 20 stay off. Off means not in the schema at all — the model cannot see them, let alone call them.
• Scoped writes: the model writes to the current layer by default and the receipt says which one; scope: "global" is required to touch everything.
• Confirmation cards: stopping servers, deleting, machine-wide toggles and similar writes ask first; recovery actions (restart, restore, undo apply) go straight through.
• Per-tool control: the model-tool table on the Host page adjusts visibility per layer. All 26 on is ~6.2k tokens/turn; profiles can be saved per task and switched together with a scene.

The Host page

More than a version number:

• Checks: host version, the plugin build actually loaded this run, and whether each action is native / adapted / unavailable, with reasons. If an upgrade quietly breaks something, it shows up here.
• Access token: set / disable / change / delete; writes and plaintext secrets go through it.
• Scope: all global / per workspace / per conversation.
• Injection: per-domain switches, minimal-preset fallback, the two slash-menu entries, sidebar entry.
• Model tools: per-tool visibility for all 26, saved per layer.
• Maintenance: old-backup cleanup, feature overview, recent changes.

Content	Location
MCP definitions	~/.dsh/cordis.patch.yml (the only file the plugin writes; backed up first)
Skill policy / custom directories	~/.dsh/tool-management/skills-state.json
Skills / memories / personas / presets	~/.dsh/tool-management/{skills,memories,subagents,prompts}/
Memory index / scenes / profiles	~/.dsh/tool-management/memories-index.json
Recycle bins (skills / memories / personas / presets / scenes)	~/.dsh/tool-management/trash/<domain>-trash/
Archive ledger / retention	~/.dsh/tool-management/history-*.json
MCP sidecars (disabled tools / known tools / notes / settings)	~/.dsh/tool-management/mcp-*.json
Injection settings (six domain switches)	~/.dsh/tool-management/inject-settings.json
Model-tool table / scope overrides	~/.dsh/tool-management/tool-table.json · scope-overrides.json
Runtime log / patch backups	~/.dsh/tool-management/tool-management.log · backups/

• Only its own files: skill sources are never touched; MCP changes go to cordis.patch.yml, with five backups per patch file.
• Token: once set, every write and every plaintext secret requires x-dsh-token; browser reads and writes use cookies, but plaintext still needs the token.
• Process-level boundary is the token: writes normally sit behind the host's browser auth; when the host fence is absent the route falls back to a loopback-only check, which any local process can pass — only the access token gives you a process-level boundary.
• Plaintext on disk: masking happens in the UI only. MCP env / headers and the plugin token stay plaintext in the patch file, and backups are neither encrypted nor rotated. The token gate cannot stop disk reads — file permissions are the boundary.
• Deletes are recoverable: skills / memories / personas / presets / scenes go to the recycle bin (permanent session deletion excepted), with an optional retention period (keep forever by default).

FAQ

Symptom	Fix
No panel after installing	Hard-refresh the browser; if that fails, restart DSH.
The model cannot call a tool	Host page → “Model tools”, make sure the right conversation / workspace is selected (20 tools are off by default). The panel and scripts are unaffected.
Actions unavailable after a DSH upgrade	Check the reason on the Host page; run node scripts/doctor.mjs, then node scripts/host-deps.mjs --fix.
Version mismatch	Supports DSH 0.2.1 only (>=0.2.1-rc.2 <0.2.2); cross-generation upgrades wait for the plugin to catch up.
Duplicate MCP tab	Remove the old loader line from cordis.patch.yml and restart.
DSH won't start after a bad edit	Restore the newest cordis.patch.yml.*.bak-* from ~/.dsh/tool-management/backups/.

Full troubleshooting guide: docs/TROUBLESHOOTING.md.

Docs

Doc	When to read it
Troubleshooting	Missing panel, duplicate tab, actions unavailable after an upgrade, broken sidecars
Architecture	How content reaches the model, why one toggle differs between conversations, which file a change lands in
Compatibility matrix	Capability matrix per DSH version (verified / untested / unavailable)
Changelog · Release overview	What changed in each version

Development

npm install
npm run build        # tsc + sync client
npm test             # build + i18n + contract tests
npm run doctor       # host compatibility check

lib/ is not committed — run npm run build after cloning, and restart dsh web to pick up changes. Runtime dependencies are only fflate and js-yaml (lazy-loaded); every @deepseek-ai/* package uses the host's copy.

License

MIT