✦ dsh-constellation

A live, self-organizing constellation map of your DeepSeek Harness plugin universe.

DeepSeek Harness runs on the idea that everything is a plugin — but once your profile carries a hundred fibers, nothing tells you what your agent is actually made of. Which fiber is stuck pending? Who provides llm? What does dsh-session-title-first-prompt-llm even do?

Constellation answers all three: a floating ✦ button in the DSH Web UI opens a real-time map of every plugin and service fiber in your running profile — organized into balanced capability domains, labeled in plain language, searchable, and operable.

English | 中文

Features

🌌 Graph — your plugin universe, at three zoom levels

The map starts collapsed: one pill per capability domain, sized by member count, with aggregated dependency edges between domains.

• Click a domain to expand its member plugins in a radial layout around the pill, inside a dashed territory boundary — click again to collapse
• Big domains cluster first: domains beyond 12 members show service clusters (llm ×8) — click a cluster to see just those plugins
• Hover or select a plugin to reveal its live dependency edges (providers → it → consumers) while the rest of the map dims — the on-demand "internal logic" of your composition
• Search anything: press / and type a plugin name, service, keyword, or domain — results ranked, keyboard-navigable, and clicking one flies the camera to the node with a pulse highlight
• Nodes are colored by fiber phase (active / pending / failed / disabled…); the legend doubles as a phase filter

✦ AI-native organization

Plugin ecosystems grow faster than any hand-written category list. When your profile has an LLM configured, Constellation puts it to work:

• AI taxonomy — the model proposes a balanced capability taxonomy (no 34-member mega-domain, no singleton domains) with bilingual labels, then assigns every plugin and writes it a short human label: dsh-session-title-first-prompt-llm becomes 标题生成 / Session title LLM
• AI descriptions — plugins whose package ships no description get a one-line summary written by the profile's own model
• One-click re-run — the ✦ AI classify button re-organizes the map whenever your plugin set changes

Everything is cached under $DSH_HOME/storages/ — one classification pass survives restarts. Profiles without an LLM simply fall back to the built-in heuristic classifier; the graph never depends on model access.

🗂 Detail drawer — inspect and operate

Every plugin opens a side drawer with its label, description, phase, and provided/consumed services, plus live entry operations:

• Enable / Disable — flips the entry through the cordis-plugin-loader tree API and persists to the profile (no restart needed)
• Remove — stops the entry and removes it from the loader tree (confirmation required)

📋 Inventory — every entry at a glance

Cards grouped by capability domain: phase, label, description, and services. Click a card for the detail drawer.

🩺 Doctor — the plugin physician

Automatic diagnostics over the live snapshot (click a finding to inspect):

• ⏳ Stuck pending — a fiber waiting on a service no active plugin provides (names the missing service)
• ❌ Failed fibers — plugins that crashed on load
• 🧟 Zombie entries — enabled entries whose fiber is disposed
• ⚠️ Duplicate providers — two plugins providing the same service; only the last one wins

🌏 Bilingual UI

The whole interface speaks Chinese and English with a one-click switch — a natural fit for a harness whose plugin names are English but whose users often aren't.

Install

dsh plugin --profile web add dsh-constellation

Then start the Web UI (dsh --profile web) and click the ✦ button in the bottom-right corner.

How it works

┌─ Browser ──────────────────────────────────────┐
│  client bundle (React, ~90 KB)                 │
│  body-portal panel · canvas constellation map  │
│  polls GET /constellation/graph                │
└───────────────────┬────────────────────────────┘
                    │ same-origin, loopback-fenced
┌───────────────────┴────────────────────────────┐
│  host half (Node)                              │
│  walks ctx.loader.entries() per request:       │
│    fiber.inject keys  → consumed services      │
│    fiber.store  keys  → provided services      │
│    fiber.state        → phase                  │
│  optional llm sub-fiber: taxonomy + labels     │
└────────────────────────────────────────────────┘

The graph projection reads the same runtime records the cordis context proxy consults — no kernel-internal APIs, so it survives minor version drift. Data never leaves localhost: routes reject non-loopback Host headers, and the LLM calls go through the profile's own configured model.

Requirements

• DeepSeek Harness >= 0.1.0-rc.8 with a Web composition (e.g. the web profile)
• Node >= 20
• (optional) a configured default model, for the AI taxonomy / labels / descriptions

Development

npm install
npm run build        # tsc (types) + tsdown (lib/index.js + lib/client.js)

The client bundle follows the official DSH client-bundle preset: a CJS closure factory registered through window.__ModuleLoader__, with React / react-dom kept external against the shell's module table.

License

MIT