积木 · Jimu

A domain-based task orchestration plugin for DeepSeek Harness.

This plugin decomposes complex multi-step tasks with interdependent stages into an ordered set of independent execution units (domains). After human confirmation, each domain runs in sequence with automatic audit and output hand-off, keeping the entire process visible, traceable, and interruptible.

Background

A general-purpose agent typically completes a complex task in a single pass, leaving intermediate steps opaque and hard to control. Jimu provides a structured execution model:

• Decomposition: the task is split into an ordered domain plan; each domain declares its scope and acceptance criteria;
• Human confirmation: no domain runs before the user confirms the plan, aligning the decomposition with user intent;
• Independent execution: each domain runs in a fresh sub-agent that receives only its own task and relevant upstream output;
• Automatic audit: on completion, each domain's output is verified against its acceptance criteria before the next domain is released;
• Interruptible and traceable: domain progress and audit results are persisted as session events, reviewable at any time; execution can be paused and resumed from a given domain.

Components

Component	Description
domain-plan	Provides the domain_plan tool: decomposes the task into a domain plan and submits it for user confirmation; the verdict is recorded as the domain/plan event
orchestrator	Listens for confirmed plans, starts each domain's sub-agent in order, performs output hand-off and automatic audit, and provides the domain_resume tool to continue from a paused domain
auto-approver	Test helper that answers confirmation panels automatically; for automated tests only, never mount in production

Install

# via npm
dsh plugin --profile web add @spechenxin/jimu

# or via GitHub
dsh plugin --profile web add github:spechenxin/jimu

Restart dsh web after installing. The plugins appear in Settings → Plugins as jimu-domain and jimu-orchestrator.

Usage

Give the model a multi-step task. The flow is:

1. The model calls domain_plan and submits a domain plan for user confirmation;
2. After confirmation, the orchestrator starts sub-agents by domain sequence, handing upstream output between domains;
3. Each domain is audited on completion; on pass the next domain starts, on failure the run pauses;
4. After a fix, instruct the model to call domain_resume to continue from the paused domain.

Event model

Event	Description
domain/plan	The domain plan and its human confirmation verdict
domain/progress	Per-domain execution state (running / passed / paused / failed)
domain/audit	Per-domain audit conclusion and rationale

These events are persisted with the session and support replay and audit.

Development

pnpm install        # install dependencies
pnpm build          # build to lib/
pnpm typecheck      # type check

License

MIT