Skip to content

About

A multi-agent architecture proposal for AI Agents

Resources

Stars

1 star

Watchers

0 watching

Forks

Repository files navigation

Astra Agent Constellation

多智能体编排架构蓝图 —— 单一编排者、位置化执行者、看护者与工具门禁。 An agent-orchestration architecture blueprint: one orchestrator, positional executors, a guardian and a tool gate.

License: CC BY-SA 4.0 GitHub stars Last commit


English

Astra Agent Constellation is an agent-orchestration architecture blueprint — not a software product, plugin or CLI. It defines how one orchestrator + positional executors + a guardian + a tool gate are organised, kept in sync, disciplined and audited. You adopt it by reading the specification, copying the templates, and adapting them to your own topology.

Core Design

Role Responsibility Deployment
Orchestrator Sole user-facing entry point; task breakdown; acceptance Main work machine
Executors (on demand) Complete coding / build / training / testing on a designated platform; never talk directly to the user Machines that carry development / compile / training / test load
Guardian (0–1) Maintains update and recovery of every agent except itself; reachable by the user over an independent channel Separate physical machine (single-point-of-failure isolation)
Tool Gate Unified authentication and audit for all shared tool services Infrastructure layer

Conventions: the model layer reuses the existing gateway (zero new subscriptions); the tool layer is network-ised with credentials held server-side; sync is layered as "knowledge follows code, config follows git, memory stays local, credentials are centralised"; discipline is layered as "gates stay at orchestration, rules live in AGENTS.md, permissions live in tool configuration".

Repository Layout

astra-agent-constellation/
├── DESIGN.md            ← Doc-site visual design spec (IBM Carbon style)
├── mkdocs.yml           ← Doc-site config (Simplified Chinese)
├── docs/                ← Volume 1: normative specification (Simplified Chinese, RFC 2119)
│   ├── 00-overview.md … 09-game-day.md
│   └── references/      ← Volume 3: decision records (ADR 0001–0005)
├── templates/           ← Volume 2: reference implementations (British English, copy & adapt)
│   ├── AGENTS.md / opencode.json / dotfiles/ / agent-registry/ / task-brief/
└── scripts/             ← Companion tooling (health-check.sh, registry-check.py)

Quick Start

  1. Read the adoption path on the doc site, starting from Phase 0 (sync foundation) — the linchpin of the whole scheme.
  2. Copy the templates under templates/ into your repository / machine and adapt as needed.
  3. Work through Phase 1→3 step by step, passing acceptance before each next phase.

Building the Doc Site Locally

python3 -m venv .venv && . .venv/bin/activate
pip install mkdocs-material
mkdocs serve   # or mkdocs build

The interface follows the IBM Carbon design spec (DESIGN.md); fonts use the system font stack and do not rely on the Google Fonts CDN (loads fine domestically in China).

Versioning

Three-layer SemVer (consistent with the astra ecosystem):

Layer Version Purpose
official v0.2.2 GitHub public copy / registry.yaml / git tag (current release)
personal v0.1.0+alrcatraz.Y Private copy personalised revision (after real-topology override)
local v0.1.0+alrcatraz.Y.<variant>.Z Machine-local variant

v1.0.0 (formal release) waits for the AI Gate development to complete and for verification against a real deployment.

Licence

  • Specification text and decision records (docs/, README, DESIGN.md): CC BY-SA 4.0
  • Templates and scripts (templates/, scripts/): MIT (LICENSE-MIT in each directory)

Related Projects


中文

Astra Agent Constellation(阿斯特拉智能体星座) 是一份智能体编排架构蓝图 (蓝图,非软件/插件/CLI):定义「一个编排者 + 若干个位置化执行者 + 一个看护者

  • 一个工具门禁」如何组织、如何同步、如何传导纪律、如何被审计。采纳方式:阅读 规范 → 复制模板 → 按自家拓扑适配。

核心设计

角色 职责 部署
编排者 Orchestrator 唯一用户交互入口;任务拆解;验收 主工作机
执行者 Executors(按需) 在指定平台完成编码/构建/训练/测试;永不直接与用户对话 承担开发/编译/训练/测试功能的机器
看护者 Guardian(0–1) 维护除自己外所有智能体的更新与恢复;经独立渠道可被用户直接访问 独立物理机(防单点)
工具门禁 Tool Gate 所有智能体共用工具服务的统一鉴权与审计 基础设施层

配套约定:模型层共用现有网关(零新订阅);工具层网络化 + 凭证服务端持有; 同步分层「知识跟代码走、配置跟 git 走、记忆留本地、凭证集中保管」;纪律分层 「门留在编排层、规则进 AGENTS.md、权限进工具配置」。

仓库结构

astra-agent-constellation/
├── DESIGN.md            ← 文档站界面设计规范(IBM Carbon 风格)
├── mkdocs.yml           ← 文档站配置(简体中文)
├── docs/                ← 卷一:规范正文(简体中文,RFC 2119)
│   ├── 00-overview.md … 09-game-day.md
│   └── references/      ← 卷三:决策记录(ADR 0001–0005)
├── templates/           ← 卷二:参考实现(英式英语,可复制适配)
│   ├── AGENTS.md / opencode.json / dotfiles/ / agent-registry/ / task-brief/
└── scripts/             ← 配套脚本(health-check.sh、registry-check.py)

快速开始

  1. 阅读 文档站 的采纳路径,从 Phase 0(同步地基) 开始—— 这是整个方案的命门。
  2. 复制 templates/ 下的模板到你的仓库/机器,按需适配。
  3. 按 Phase 1→3 逐步落地,每阶段验收通过再进下一阶段。

本地构建文档站

python3 -m venv .venv && . .venv/bin/activate
pip install mkdocs-material
mkdocs serve   # 或 mkdocs build

注意:界面按 IBM Carbon 规范设计(DESIGN.md),字体走系统字体栈, 不依赖 Google Fonts CDN(国内可正常加载)。

版本

遵循三层 SemVer(与 astra 生态一致):

层级 版本 用途
official v0.2.2 GitHub 公开副本 / registry.yaml / git tag(当前发布)
personal v0.1.0+alrcatraz.Y 私有副本个性化修订(真实拓扑覆盖后)
local v0.1.0+alrcatraz.Y.<variant>.Z 本机变体

正式版 v1.0.0:待 AIGate 开发完成后对应更新,并经过我们实际部署验证后再出。

许可

  • 规范正文与决策记录(docs/、README、DESIGN.md):CC BY-SA 4.0
  • 模板与脚本(templates/、scripts/):MIT(各目录含 LICENSE-MIT)

相关项目

About

A multi-agent architecture proposal for AI Agents

Resources

Stars

1 star

Watchers

0 watching

Forks

Releases

Contributors

Languages