Features

• Delegation on demand — the Leader prompt tells the model when splitting a task pays off (independent sub-tasks, distinct expertise, independent review, large reading jobs) and when to just do the work itself.
• Task graphs — acyclic dependency graphs run ready nodes in parallel; a node receives its dependencies' results. Parallel, pipeline, execute-and-review, and multi-level team patterns are supported.
• Roles that map to real tools — 13 built-in roles (explore, librarian, worker, code-reviewer, qa-executor, …). Each child gets only its role's dsh tools (read, grep, edit, bash, web_search, …); reviewers are read-only and return a structured verdict.
• Per-node model and reasoning effort — route by node, role, or task category. Efforts are validated against what each model actually offers.
• Settings page — health check, model selection, routing table, skills, run history with a task-graph view, and a preview of the exact prompts.
• Global and per-conversation switches — turn orchestration off everywhere, or per conversation from the composer.
• Custom connections (advanced) — call an OpenAI-compatible or Anthropic endpoint directly, registered as a dsh model route.
• Recoverable runs — plans, assignments, results, and events are persisted; runs can be inspected and cancelled from the UI.

Screenshots

Overview	Routing
Models and reasoning effort	Per-conversation switch

Install

DeepSeek Harness Desktop

Open Plugins → Add plugin, enter dsh-dynamic-multiagent, and confirm. The plugin page appears under Installed plugins.

Installing with dsh plugin --profile web add … on a machine with Desktop installs into the CLI's web profile, not into Desktop. Use the in-app installer for Desktop.

dsh CLI (web profile)

dsh plugin --profile web add dsh-dynamic-multiagent
dsh web

Open Settings → Multi-agent (多智能体).

From source

git clone https://github.com/Silasvanceneo/dsh-dynamic-multiagent.git
cd dsh-dynamic-multiagent
npm install && npm run build
dsh plugin --profile web add /absolute/path/to/dsh-dynamic-multiagent

Quick start

1. Open the settings page. Overview should show the subagent provider (spawn) as ready.
2. On Models, switch on the dsh models children may use and tag their capabilities. Skip this step and every child inherits the conversation's model.
3. On Routing, optionally pick a model and reasoning effort per role or category, for example a cheap model with effort off for explore and a strong one with max for code-reviewer.
4. Give the conversation a task that benefits from splitting, for example "Survey how errors are handled across this repository, fix the three worst cases, and have the fixes reviewed independently."

The Leader reads the catalog, submits a plan, tells you the plan in one or two sentences, runs it, verifies the results, and reports back. Progress shows in the composer chip and on the Runs tab.

How it works

flowchart LR
    U[You] --> L[Leader<br>conversation model]
    L -->|dynamic_multiagent_status| C[(Models · roles<br>categories · skills)]
    L -->|dynamic_multiagent_plan| G{{Task graph}}
    G -->|ready nodes in parallel| A1[explore]
    G --> A2[worker]
    A1 -->|result| A2
    A2 -->|result| R[code-reviewer]
    R -->|verdict| L
    L -->|verified answer| U

Tool	Purpose
dynamic_multiagent_status	Without a run id: available models (with efforts), routing defaults, roles, categories, skills. With a run id: graph, results, events.
dynamic_multiagent_start	Opens a run and records the user's constraints.
dynamic_multiagent_plan	Validates the graph (ids, acyclic dependencies, roles, models, efforts) and returns each node's model and effort.
dynamic_multiagent_execute	Runs the graph through the host's ctx.subagents provider.
dynamic_multiagent_derive_role	Derives a specialized role that may only narrow its base role's tools, write access, and delegation.
dynamic_multiagent_cancel	Cancels a run and every unfinished child.

Each child is a regular dsh subagent, so it shows up in dsh's own subagent views. It receives a role-specific prompt — not the Leader's — with its task, acceptance criteria, real tool names, write scope, and a fixed report format.

Reasoning effort

A node's effort is resolved in this order: the node's effort → the role default → the category default → the model's own default.

Source	Where the efforts come from	On the wire
dsh model	ctx.llm.resolveModelInfo(), for example DeepSeek off · high · max	the dsh adapter's own mapping
Custom connection, OpenAI-compatible	declared per model	reasoning_effort
Custom connection, Claude 4.6+ / Fable	declared per model (Suggest from id)	thinking: {type: "adaptive"} + output_config.effort
Custom connection, Claude 4.5 and older	declared per model	thinking.budget_tokens

An explicit effort the model does not offer is rejected when the plan is submitted; an unsupported default is dropped with a note. dsh 0.2.x applies a child's reasoningEffort natively. dsh 0.1.x children only inherit provider, model, and output cap, so the plugin re-applies the effort on every request of that child.

Switches

Switch	Where	Effect
Enable orchestration	top of the settings page	Off: no conversation gets the Leader prompt or the six tools.
On for new conversations	top of the settings page	Default for conversations without their own setting.
This conversation	Multi-agent chip next to the composer	Per-conversation override; Follow default clears it.

When orchestration is off for a conversation, the Leader prompt section is not rendered and the tools are masked for that agent, so the model never sees them. Overrides are stored in the plugin's data directory, not in the dsh session log, so logs stay readable if the plugin is removed. Child agents never see the orchestration tools.

Configuration

Settings are edited in the UI and stored in ${DSH_HOME}/dynamic-multiagent/config.json (~/.dsh by default). Deployment options go in the profile's cordis.patch.yml:

- id: dynamic-multiagent
  config:
    dataDir: ${DSH_HOME}/dynamic-multiagent   # where config, runs, and switches live
    subagentProvider: spawn                    # host subagent provider used for children
    exposeTools: true                          # false hides the six tools entirely
    autoInstallSkills: false

Custom connections never store secrets: credentialRef names an environment variable or dsh credential, such as RELAY_API_KEY.

Compatibility

Host	Settings page	Browser ↔ host channel
dsh 0.1.0-rc.6 (CLI)	Settings → Multi-agent	loopback /dynamic-multiagent/*
dsh 0.2.1-rc.3 (Desktop)	Plugins → Dynamic multi-agent	authenticated /api/dynamic-multiagent/*

The settings page, switches, model catalog, and saving were exercised in the real Web UI of both versions. A full end-to-end delegation against a live model provider is not yet covered by automated tests.

Development

npm install
npm run build    # client/ → lib/client.js (React, slots, and primitives come from the dsh shell)
npm test         # node --test
npm run check    # build + syntax check + tests

Path	Contents
src/index.js	Host plugin: prompt section, tools, switches, effort bridge, settings channel
src/core/	Runtime, model routing, roles, prompts, task graph, persistence, connection adapter
client/	Browser half: settings page, composer chip, tool cards
docs/design/	Original design documents (Chinese)

License

MIT