@goodandready/dsh-agent-orchestrator

Agent Orchestrator

用于 DeepSeek Harness 的多智能体任务分解、DAG 工作流编排和提示缓存优化器。

目标 Profile

插件会安装到这里；不确定时保持 web。

npx -y @deepseek-ai/dsh plugin --profile web add github:GooDAnDReaDY/dsh-agent-orchestrator#cae4c17bf337396bb26e0e770f82c4f03a9bdf56

兼容性与来源证明

Agent Orchestrator 以 @goodandready/dsh-agent-orchestrator 发布，当前版本为 0.1.9。Plugin Hub 会校验它的 manifest，并保存精确安装来源，便于复现安装结果。

DSH 兼容范围: *
运行环境: web
发布来源: github
Registry 更新时间: 2026/9/19

版本

0.1.9stable

2026/9/19

0.1.6stable

2026/9/18

README

📦 @goodandready/dsh-agent-orchestrator

Multi-Agent Task Decomposition, DAG Workflow Orchestration & Prompt Caching Engine for DeepSeek Harness

🇬🇧 English • 🇷🇺 Русский • 🇨🇳 中文说明

⭐ If you like this plugin, please star it on GitHub — it shows me that the plugin is useful to you and motivates me to keep developing it.

🐛 If you find a bug or would like to request a feature, open a GitHub issue in any language — I will review your proposal and implement useful suggestions in a future plugin version.

⚡ Overview & The Problem

Single-agent software engineering architectures suffer from cognitive overload when tasked with complex multi-stage projects: monolithic prompts conflate architecture, styling, core business logic, testing, and documentation into a single generation pass, leading to hallucinated contracts, regression bugs, and excessive token expenditure.

Furthermore, executing multiple subagents independently often resets the LLM KV-cache on each turn, forfeiting prefix cache reuse and incurring substantial latency and financial overhead.

@goodandready/dsh-agent-orchestrator introduces an autonomous multi-agent orchestration framework to DeepSeek Harness:

Intelligent Triage & Decomposition: Analyzes high-level objectives from chat or Kanban cards and breaks them down into fine-grained stages across 12 specialized agent roles.
DAG Execution Engine: Schedules tasks based on Directed Acyclic Graph dependencies, running independent stages concurrently while strictly enforcing blocker gates.
KV-Cache / Prompt Caching Optimizer: Guarantees byte-level prefix invariance for subagents sharing identical models, unlocking 80–90% prompt token cache hits and near-instant TTFT.
Strict Separation of Duties: Backend implementation, UI interface design, and frontend client assembly are strictly isolated into distinct personas and execution stages.
Dual Surface Integration: Dispatched natively via /orchestrate in DSH chat (with a sticky header milestone card) or via @goodandready/dsh-kanban task boards.

🏗️ Architecture

graph TD
    Trigger["Input Task<br/>(/orchestrate in Chat or Kanban Card)"] --> Main["Lead Orchestrator (Triage)"]
    
    subgraph Engine ["DAG Engine & Prompt Caching"]
        L1["Layer 1: Canonical Base Anchor (>1024 tokens)"]
        L2["Layer 2: Shared Task Anchor"]
        L3["Layer 3: Cumulative Artifacts (Append-Only)"]
        L4["Layer 4: Role Suffix Directive"]
    end

    Main --> Engine

    subgraph AgentPool ["Configured Agent Personas (Self-Contained in Settings)"]
        R1["Technical Spec Analyst"]
        R2["System Architect"]
        R3["UI/UX Interface Designer"]
        R4["Backend Developer"]
        R5["Frontend Developer"]
        R6["QA Automation Specialist"]
        R7["Documentation Specialist"]
    end

    Engine --> AgentPool
    AgentPool --> Delivery["Orchestrated Delivery<br/>(Header Utility Card & Kanban Sync)"]

👥 12 Built-In Specialized Agent Roles

All agent profiles are completely self-contained within plugin settings (no external file dependencies):

Role ID	Title	Specialization	Strict Boundaries
`spec`	Technical Spec Analyst	Requirements, Acceptance Criteria (DoD), Schemas	Never writes implementation or styling
`architecture`	System Architect	System Design, DESIGN.md, ADR, Modular Contracts	Never implements production code or deploys
`ui_design`	UI/UX Interface Designer	Layouts, Theme Tokens (`--dsw-alias-*`), Slots	Never writes backend Cordis services
`frontend`	Frontend Developer	React Components, Client Hooks, DOM Events	Never alters backend routes or DB schemas
`backend`	Backend Developer	Cordis Services, WebServer Routes, Data Store	Never writes client React JSX or styles
`fullstack`	Fullstack Integrator	Client-Server Contract Wiring, End-to-End Flow	Adheres strictly to modular limits
`qa_tests`	QA Automation Specialist	Unit Tests (`node:test`), Boundary Verification	Verifies without external network calls
`bugfix`	Hotfix & Triage Engineer	Root-Cause Diagnosis, Minimal Blast Radius Fixes	Never refactors unrelated code
`docs`	Documentation Specialist	Trilingual Documentation (en/ru/zh), Releases	Never overwrites previous documentation
`refactoring`	Refactoring Specialist	Complexity Reduction (YAGNI), Bundle Compression	Preserves backwards compatibility
`research`	Research & Spike Engineer	Technology Evaluation, Library Trade-Offs	Delivers analysis; never merges spike code
`devops`	DevOps & Tooling Engineer	Package Manifests, Build Verification, Systemd	Never exposes private network credentials

🔄 Complexity Scenarios

Hotfix / Trivial (1 Stage): Instant defect elimination or single-parameter tweak.
Simple (2 Stages): Discussion & Spec $\rightarrow$ Targeted Execution.
Medium (3–4 Stages): Spec $\rightarrow$ UI Design $\rightarrow$ Frontend Code $\rightarrow$ QA Tests.
Complex (5–6 Stages): Spec $\rightarrow$ Architecture $\rightarrow$ UI Design $\rightarrow$ Implementation $\rightarrow$ QA $\rightarrow$ Trilingual Docs.
Enterprise / Deep R&D (7 Stages): Spike Research $\rightarrow$ Spec $\rightarrow$ Architecture $\rightarrow$ Parallel Backend & UI Design $\rightarrow$ Frontend Assembly $\rightarrow$ Comprehensive QA $\rightarrow$ Documentation Gate.
Custom DAG Scenarios: Fully configurable in plugin settings with custom stages and blocker checkboxes.

⚡ Prompt Caching Mechanics

Modern LLMs (DeepSeek-V3, Claude 3.5 Sonnet, vLLM) cache prompt KV states strictly from the first token forward. If non-deterministic timestamps or random IDs are placed in the prompt header, cache hit rate drops to 0%.

dsh-agent-orchestrator enforces a 4-layer canonical layout:

Layer 1: Static Base Anchor (>1024 tokens): Byte-identical guidelines and tools definition common across all agents.
Layer 2: Shared Task Anchor: Stable description of user objective and target repository.
Layer 3: Cumulative Context (Append-Only): Outputs of predecessor stages appended in a deterministic sequence, preserving 100% of the preceding KV-cache.
Layer 4: Role Directive (Suffix): Role persona prompt, skills, and subtask-specific scope appended at the end.

This architecture delivers 80–95% cache hits across subagents utilizing the same model, reducing TTFT and cutting token costs by ~90%.

💻 Usage

1. In DSH Chat via Slash Command

/orchestrate Design and build a settings card for the finance plugin

Explicit scenario selection:

/orchestrate complex Build a multi-tenant authentication provider
/orchestrate hotfix Fix null reference in store.js

Short alias:

/orc Refactor state management

2. In @goodandready/dsh-kanban

Open any card on the board.
Click [Собрать пайплайн / Assemble Pipeline].
Select the complexity preset or allow auto-triage.
The card automatically reflects stage transitions and advances to Review upon completion.

🧪 Verification & Automated Testing

Execute the native test suite (121 passing tests across 53 suites, zero network dependencies):

node --test test/*.test.mjs

Verify npm package bundle size compliance (<256 KiB threshold):

npm pack --dry-run --json

Visual verification

Production acceptance of v0.1.6 — Settings card, Dark and Light themes side by side:

dsh-agent-orchestrator v0.1.6 visual verification

📄 License

MIT © GooDAnDReaDY

README

📦 @goodandready/dsh-agent-orchestrator

Multi-Agent Task Decomposition, DAG Workflow Orchestration & Prompt Caching Engine for DeepSeek Harness

🇬🇧 English • 🇷🇺 Русский • 🇨🇳 中文说明

⚡ Overview & The Problem

Furthermore, executing multiple subagents independently often resets the LLM KV-cache on each turn, forfeiting prefix cache reuse and incurring substantial latency and financial overhead.

@goodandready/dsh-agent-orchestrator introduces an autonomous multi-agent orchestration framework to DeepSeek Harness:

Intelligent Triage & Decomposition: Analyzes high-level objectives from chat or Kanban cards and breaks them down into fine-grained stages across 12 specialized agent roles.
DAG Execution Engine: Schedules tasks based on Directed Acyclic Graph dependencies, running independent stages concurrently while strictly enforcing blocker gates.
KV-Cache / Prompt Caching Optimizer: Guarantees byte-level prefix invariance for subagents sharing identical models, unlocking 80–90% prompt token cache hits and near-instant TTFT.
Strict Separation of Duties: Backend implementation, UI interface design, and frontend client assembly are strictly isolated into distinct personas and execution stages.
Dual Surface Integration: Dispatched natively via /orchestrate in DSH chat (with a sticky header milestone card) or via @goodandready/dsh-kanban task boards.

🏗️ Architecture

graph TD
    Trigger["Input Task<br/>(/orchestrate in Chat or Kanban Card)"] --> Main["Lead Orchestrator (Triage)"]
    
    subgraph Engine ["DAG Engine & Prompt Caching"]
        L1["Layer 1: Canonical Base Anchor (>1024 tokens)"]
        L2["Layer 2: Shared Task Anchor"]
        L3["Layer 3: Cumulative Artifacts (Append-Only)"]
        L4["Layer 4: Role Suffix Directive"]
    end

    Main --> Engine

    subgraph AgentPool ["Configured Agent Personas (Self-Contained in Settings)"]
        R1["Technical Spec Analyst"]
        R2["System Architect"]
        R3["UI/UX Interface Designer"]
        R4["Backend Developer"]
        R5["Frontend Developer"]
        R6["QA Automation Specialist"]
        R7["Documentation Specialist"]
    end

    Engine --> AgentPool
    AgentPool --> Delivery["Orchestrated Delivery<br/>(Header Utility Card & Kanban Sync)"]

👥 12 Built-In Specialized Agent Roles

All agent profiles are completely self-contained within plugin settings (no external file dependencies):

Role ID	Title	Specialization	Strict Boundaries
`spec`	Technical Spec Analyst	Requirements, Acceptance Criteria (DoD), Schemas	Never writes implementation or styling
`architecture`	System Architect	System Design, DESIGN.md, ADR, Modular Contracts	Never implements production code or deploys
`ui_design`	UI/UX Interface Designer	Layouts, Theme Tokens (`--dsw-alias-*`), Slots	Never writes backend Cordis services
`frontend`	Frontend Developer	React Components, Client Hooks, DOM Events	Never alters backend routes or DB schemas
`backend`	Backend Developer	Cordis Services, WebServer Routes, Data Store	Never writes client React JSX or styles
`fullstack`	Fullstack Integrator	Client-Server Contract Wiring, End-to-End Flow	Adheres strictly to modular limits
`qa_tests`	QA Automation Specialist	Unit Tests (`node:test`), Boundary Verification	Verifies without external network calls
`bugfix`	Hotfix & Triage Engineer	Root-Cause Diagnosis, Minimal Blast Radius Fixes	Never refactors unrelated code
`docs`	Documentation Specialist	Trilingual Documentation (en/ru/zh), Releases	Never overwrites previous documentation
`refactoring`	Refactoring Specialist	Complexity Reduction (YAGNI), Bundle Compression	Preserves backwards compatibility
`research`	Research & Spike Engineer	Technology Evaluation, Library Trade-Offs	Delivers analysis; never merges spike code
`devops`	DevOps & Tooling Engineer	Package Manifests, Build Verification, Systemd	Never exposes private network credentials

🔄 Complexity Scenarios

Hotfix / Trivial (1 Stage): Instant defect elimination or single-parameter tweak.
Simple (2 Stages): Discussion & Spec $\rightarrow$ Targeted Execution.
Medium (3–4 Stages): Spec $\rightarrow$ UI Design $\rightarrow$ Frontend Code $\rightarrow$ QA Tests.
Complex (5–6 Stages): Spec $\rightarrow$ Architecture $\rightarrow$ UI Design $\rightarrow$ Implementation $\rightarrow$ QA $\rightarrow$ Trilingual Docs.
Enterprise / Deep R&D (7 Stages): Spike Research $\rightarrow$ Spec $\rightarrow$ Architecture $\rightarrow$ Parallel Backend & UI Design $\rightarrow$ Frontend Assembly $\rightarrow$ Comprehensive QA $\rightarrow$ Documentation Gate.
Custom DAG Scenarios: Fully configurable in plugin settings with custom stages and blocker checkboxes.

⚡ Prompt Caching Mechanics

dsh-agent-orchestrator enforces a 4-layer canonical layout:

Layer 1: Static Base Anchor (>1024 tokens): Byte-identical guidelines and tools definition common across all agents.
Layer 2: Shared Task Anchor: Stable description of user objective and target repository.
Layer 3: Cumulative Context (Append-Only): Outputs of predecessor stages appended in a deterministic sequence, preserving 100% of the preceding KV-cache.
Layer 4: Role Directive (Suffix): Role persona prompt, skills, and subtask-specific scope appended at the end.

This architecture delivers 80–95% cache hits across subagents utilizing the same model, reducing TTFT and cutting token costs by ~90%.

💻 Usage

1. In DSH Chat via Slash Command

/orchestrate Design and build a settings card for the finance plugin

Explicit scenario selection:

/orchestrate complex Build a multi-tenant authentication provider
/orchestrate hotfix Fix null reference in store.js

Short alias:

/orc Refactor state management

2. In @goodandready/dsh-kanban

Open any card on the board.
Click [Собрать пайплайн / Assemble Pipeline].
Select the complexity preset or allow auto-triage.
The card automatically reflects stage transitions and advances to Review upon completion.

🧪 Verification & Automated Testing

Execute the native test suite (121 passing tests across 53 suites, zero network dependencies):

node --test test/*.test.mjs

Verify npm package bundle size compliance (<256 KiB threshold):

npm pack --dry-run --json

Visual verification

Production acceptance of v0.1.6 — Settings card, Dark and Light themes side by side:

dsh-agent-orchestrator v0.1.6 visual verification

Agent Orchestrator

兼容性与来源证明

版本

相关插件

Agent Orchestrator

README

📦 @goodandready/dsh-agent-orchestrator

Multi-Agent Task Decomposition, DAG Workflow Orchestration & Prompt Caching Engine for DeepSeek Harness

⚡ Overview & The Problem

🏗️ Architecture

👥 12 Built-In Specialized Agent Roles

🔄 Complexity Scenarios

⚡ Prompt Caching Mechanics

💻 Usage

1. In DSH Chat via Slash Command

2. In @goodandready/dsh-kanban

🧪 Verification & Automated Testing

Visual verification

📄 License

兼容性与来源证明

版本

相关插件

相关插件

相关插件

README

📦 @goodandready/dsh-agent-orchestrator

Multi-Agent Task Decomposition, DAG Workflow Orchestration & Prompt Caching Engine for DeepSeek Harness

⚡ Overview & The Problem

🏗️ Architecture

👥 12 Built-In Specialized Agent Roles

🔄 Complexity Scenarios

⚡ Prompt Caching Mechanics

💻 Usage

1. In DSH Chat via Slash Command

2. In @goodandready/dsh-kanban

🧪 Verification & Automated Testing

Visual verification

📄 License