multi-agent-os

内容来源：README.md（说明文档） · 原始地址 · 查看安装指南

原始内容

# Multi-Agent OS

[![License: MIT](https://img.shields.io/badge/License-MIT-yellow.svg)](https://opensource.org/licenses/MIT)
[![Claude Code](https://img.shields.io/badge/Claude%20Code-Plugin-blueviolet)](https://claude.ai/code)
[![Version](https://img.shields.io/badge/Version-1.21.0-blue)](https://github.com/ekson73/multi-agent-os)
[![Sentinel](https://img.shields.io/badge/Sentinel-Protocol-green)](https://github.com/ekson73/multi-agent-os/tree/main/sentinel)
[![Branching: GitHub Flow](https://img.shields.io/badge/branching-GitHub%20Flow%20(Class%20B)-0a7bbb)](./AGENTS.md)

A comprehensive Claude Code plugin for orchestrating AI agents in software development workflows.

## Features

### Sentinel Protocol
- 10 detection rules for anomaly prevention
- Loop detection, scope creep, stagnation alerts
- Error cascade prevention
- Health score calculation

### Status Map System
- 10 template types for different contexts
- Human-centric observability
- Automatic template inference

### Anti-Conflict Protocol v3.2
- 7-phase workflow with mandatory QA
- Git worktree integration
- Lock file coordination

### Worktree Policy v1.1
- Multi-agent branch isolation
- Catch-22 scenario handling
- Orphan branch cleanup

### Statusline Auto-install
- Automatic statusline template installation
- Settings.json auto-configuration
- Checksum-based update detection

### Hierarchical Merge Protocol v1.0
- Parent-child branch convergence
- Child Completion Constraint
- Exception prefixes for hotfix/emergency

### TTL Policy
- Content freshness management
- PROV tags for provenance tracking
- Automatic expiration alerts

### MVV Generator (v1.1.0)
- 8-dimension ontological analysis (Context, Purpose, Taxonomy, Semantics, Lineage, Epistemology, Ontology, Aesthetics)
- Mission/Vision/Values synthesis pipeline
- Automated organizational identity extraction
- `/mvv` command for one-shot execution

## Installation

> **Plugin name vs repo name**: the repository is `multi-agent-os`; the **plugin is named `maos`** (see `.claude-plugin/plugin.json`). Install and reference it as `maos`. Skills/commands surface namespaced as **`/maos:<name>`** (e.g. `/maos:orchestrator`, `/maos:auto-pilot`).

### From Marketplace (recommended)

Plugin management runs **inside a Claude Code session** via the `/plugin` command (not a shell `claude plugins` subcommand):

```text
/plugin marketplace add ekson73/eko-claude-plugins
/plugin install maos@eko-claude-plugins
/reload-plugins
```

Then run `/plugin` (Installed tab) to confirm, and `/help` to see the `/maos:*` skills. Need to update later? `/plugin marketplace update eko-claude-plugins`.

### From source (local dev / self-use)

```bash
git clone https://github.com/ekson73/multi-agent-os.git
claude --plugin-dir ./multi-agent-os     # load for this session
claude plugin validate ./multi-agent-os  # validate the manifest
```

`--plugin-dir` may be repeated to load several plugins and takes precedence over an installed copy of the same name (handy for testing local changes). See [Self-Referential Usage](#self-referential-usage).

### Project / team scope

To make `maos` available to **everyone on a repository**, add to that project's `.claude/settings.json` (collaborators are prompted to install when they trust the folder):

```json
{
  "extraKnownMarketplaces": {
    "eko-claude-plugins": {
      "source": { "source": "github", "repo": "ekson73/eko-claude-plugins" }
    }
  },
  "enabledPlugins": ["maos@eko-claude-plugins"]
}
```

CLI equivalent (writes project scope): `claude plugin install maos@eko-claude-plugins --scope project`. Full schema: [Plugin settings](https://code.claude.com/docs/en/settings#plugin-settings).

## Plugin Structure

```
multi-agent-os/
├── .claude-plugin/
│   └── plugin.json           ← Plugin manifest
├── hooks/
│   └── hooks.json            ← Hook configuration
├── plugin-scripts/           ← Hook executables
│   ├── session-start.sh
│   ├── pre-delegate.sh
│   ├── post-delegate.sh
│   ├── session-end.sh
│   └── governance/
│       ├── worktree-gate.sh
│       ├── auto-name-session.sh
│       └── token-budget-gate.sh  ← GaaS token bloat detection
├── commands/                 ← Slash commands (~15; representative subset shown)
│   ├── sync.md
│   ├── audit.md
│   ├── agentic-status.md
│   ├── worktree.md
│   └── delegate.md
├── agents/                   ← Agent definitions (20+; representative subset shown)
│   ├── orchestrator.md
│   ├── sentinel-monitor.md
│   ├── qa-validator.md
│   ├── consolidator.md
│   ├── forge.md              ← Meta-agent creator (Goldilocks + RBAD + 33 Socratic Questions)
│   ├── governance-auditor.md ← Standards & compliance guardian
│   ├── naming-organizer.md   ← Digital organization & taxonomy
│   ├── data-validator.md     ← Truth & evidence verification (11 validation types)
│   └── validation-auditor.md ← Second-line active auditing
├── skills/                   ← Skills (40+, subdirectory format; representative subset shown)
│   ├── audit/SKILL.md
│   ├── agent-select/SKILL.md
│   ├── context-prep/SKILL.md
│   ├── hierarchical-merge/SKILL.md
│   ├── worktree-policy/SKILL.md
│   ├── anti-conflict/SKILL.md
│   ├── status-map/SKILL.md
│   ├── ttl-policy/SKILL.md
│   ├── ontological-analysis/SKILL.md  ← 8-dimension analysis
│   ├── mvv-synthesis/SKILL.md         ← MVV generation
│   └── response-compression/SKILL.md  ← Output verbosity control
├── protocols/                ← Governance protocols
│   ├── exit-hygiene.md       ← Session exit hygiene (zero loose ends)
│   ├── agent-delegation.md   ← Delegation chain & Forge bootstrap
│   ├── rbad.md               ← Role-Based Agent Design (6-category taxonomy)
│   └── action-priority.md    ← Eisenhower Matrix for task prioritization
├── sentinel/                 ← Sentinel Protocol files
├── statusmap/                ← Status Map templates
└── docs/                     ← Additional documentation
```

## Available Commands

> Namespaced as `/maos:<name>`. This table is **representative** — run `/help` for the full live list, or `/maos:maos-concierge` to be guided to the right command/agent/skill for your intent.

| Command | Description |
|---------|-------------|
| `/maos:agentic-status` | Human-readable status of the agentic system (git + agents + sentinel + locks). *(Was `/status` — renamed to avoid colliding with Claude Code's built-in `/status`.)* |
| `/maos:sync` | Sync content from the framework to a consumer project |
| `/maos:audit` | On-demand audit of sessions, tasks, and orchestration flows |
| `/maos:worktree` | Manage git worktrees for multi-agent isolation |
| `/maos:delegate` | Delegate a task to a specialized sub-agent with context prep |
| `/maos:mvv` | Generate/update Mission, Vision, Values for a repository |
| `/maos:auto-pilot` | Drive an operator goal end-to-end across sub-agents (GaaS/GaaC) |
| `/maos:quiesce` | Drive the session to quiescence — no open ticket/gap/fix/PR, all PRs green |
| `/maos:enhance-pipeline` | Run one feature through the full divergent→convergent→deliver lifecycle |
| `/maos:founder-playbook` | Diagnose an AI-native startup's lifecycle stage and route to the right discipline |
| `/maos:agentic-tool-forge` | Forge a raw intent into the optimal reusable agentic-tool |
| `/maos:session-fission` | Non-destructively split a tangled session into focused new ones |

Plus namespaced utilities (`/maos:auto-shard`, `/maos:code:*`, `/maos:analyze:*`) — see `/help`.

## Available Skills

> **40+ skills** ship with MAOS — the table below is a **representative subset**. Model-invoked (Claude auto-selects by task) or explicit via `/maos:<skill>`. For the full live catalog and guidance on which to use, invoke `/maos:maos-concierge`.

| Skill | Description |
|-------|-------------|
| `audit` | Sentinel Protocol auditing |
| `agent-select` | Agent selection algorithm |
| `context-prep` | Pre-delegation context |
| `hierarchical-merge` | Branch merge rules |
| `worktree-policy` | Worktree enforcement |
| `anti-conflict` | Conflict prevention |
| `status-map` | Status visualization |
| `ttl-policy` | Content freshness |
| `ontological-analysis` | 8-dimension philosophical analysis |
| `mvv-synthesis` | Mission/Vision/Values synthesis |
| `response-compression` | Output verbosity control (none/lite/full/ultra) with role-based profiles |
| `maos-concierge` | Onboarding/router/anchor for the **entire** MAOS framework (start here) |
| `auto-pilot` / `quiesce` / `converge` / `convergence-engine` | Autonomous orchestration, session quiescence, and multi-proposal convergence |
| `enhance-pipeline` | Divergent→convergent→deliver feature lifecycle |
| `agentic-delegation` / `delegate-governance` | Delegation criteria + GaaS/GaaC governance prompts |
| `agentic-tool-forge` / `-evaluator` / `-trainer` | Create, evaluate, and improve reusable agentic-tools |
| `anima` | Sovereign precision-naming engine — names anything (12 correctness + 4 resonance aspects, register-aware); the forge's Phase-4 naming delegate + standalone |
| `founder-playbook` / `founder-stage-*` | AI-native startup lifecycle disciplines (Idea/MVP/Launch/Scale) |
| `morning-briefing` / `pulse` / `session-fission` | Context restoration, re-orientation, session splitting |
| `rule-quality-tests` / `operator-quote-capture` / `pii-masking` / `slm-routing` | Governance, capture, PII detection, model routing |

## Available Agents

> **20+ agents** ship with MAOS — representative subset below. Invoke via the `Task` tool (`subagent_type: maos:<agent>`) or see `/agents`.

| Agent | Description |
|-------|-------------|
| `orchestrator` | Master coordinator |
| `sentinel-monitor` | Anomaly detection |
| `qa-validator` | Quality assurance |
| `consolidator` | Output synthesis |
| `forge` | Meta-agent creator — specialized agents via Goldilocks Principle + RBAD |
| `governance-auditor` | Standards governance, compliance auditing, pattern enforcement |
| `naming-organizer` | Digital organization, taxonomy, naming conventions |
| `data-validator` | Data validation, evidence capture, truth verification (11 types) |
| `validation-auditor` | Second-line active auditing, drift detection, integrity verification |
| `cascade-resolver` / `perspective-trio` / `persona-pipeline` | Convergence engine primitives — score-uplift, orthogonal-trio, review-board |
| `code-reviewer` / `debugger` | Code review and systematic debugging specialists |
| `data-analyst` / `legacy-archaeologist` / `memory-curator` | Analysis, legacy reverse-engineering, memory hygiene |
| `gitops-engineer` / `founder-coach` / `consultants` | GitOps/K8s, startup coaching, expert thinking-archetypes |

## Hooks

The plugin automatically hooks into Claude Code lifecycle:

| Hook | Trigger | Scripts |
|------|---------|---------|
| `SessionStart` | Session initialization | session-start.sh, auto-name-session.sh |
| `PreToolUse[Task]` | Before delegation | pre-delegate.sh, token-budget-gate.sh |
| `PreToolUse[Bash]` | Before shell commands | worktree-gate.sh |
| `PostToolUse[Task]` | After delegation | post-delegate.sh |
| `Stop` | Session end | session-end.sh |

## Documentation

| Document | Description |
|----------|-------------|
| [Worktrees Guide](https://github.com/ekson73/multi-agent-os/blob/HEAD/docs/worktrees-guide.md) | Multi-agent worktree coordination |
| [Hierarchical Merge Protocol](https://github.com/ekson73/multi-agent-os/blob/HEAD/protocols/hierarchical-merge-protocol.md) | Branch convergence rules |
| [Framework Consumption](https://github.com/ekson73/multi-agent-os/blob/HEAD/docs/framework-consumption.md) | Consumer project integration |
| [Agent Format](https://github.com/ekson73/multi-agent-os/blob/HEAD/docs/agent-format.md) | YAML frontmatter specification for agents |
| [Exit Hygiene](https://github.com/ekson73/multi-agent-os/blob/HEAD/protocols/exit-hygiene.md) | Session exit protocol (zero loose ends) |
| [Agent Delegation](https://github.com/ekson73/multi-agent-os/blob/HEAD/protocols/agent-delegation.md) | Delegation chain & Forge bootstrap |
| [RBAD](https://github.com/ekson73/multi-agent-os/blob/HEAD/protocols/rbad.md) | Role-Based Agent Design taxonomy |
| [Action Priority](https://github.com/ekson73/multi-agent-os/blob/HEAD/protocols/action-priority.md) | Eisenhower Matrix for prioritization |

## Self-Referential Usage

This plugin can use itself during development:

```bash
# In multi-agent-os directory
claude --plugin-dir .
```

## Interoperability

Skills in this plugin follow the [Agent Skills open standard](https://agentskills.io) and are compatible with 30+ AI coding tools including Claude Code, Cursor, OpenAI Codex, Gemini CLI, Kiro, VS Code, GitHub Copilot, Goose, Windsurf, and others. Install with `npx skills add` or native plugin mechanisms.

This repository includes an [AGENTS.md](https://github.com/ekson73/multi-agent-os/blob/HEAD/AGENTS.md) file following the [AAIF standard](https://agents.md) (60k+ projects) to guide any AI coding agent working on this codebase.

## License

MIT License - See LICENSE file for details.

---

*Multi-Agent OS v1.21.0 | Created by Emilson Moraes | Powered by Claude Code*

Multi-Agent OS

A comprehensive Claude Code plugin for orchestrating AI agents in software development workflows.

Features

Sentinel Protocol

10 detection rules for anomaly prevention
Loop detection, scope creep, stagnation alerts
Error cascade prevention
Health score calculation

Status Map System

10 template types for different contexts
Human-centric observability
Automatic template inference

Anti-Conflict Protocol v3.2

7-phase workflow with mandatory QA
Git worktree integration
Lock file coordination

Worktree Policy v1.1

Multi-agent branch isolation
Catch-22 scenario handling
Orphan branch cleanup

Statusline Auto-install

Automatic statusline template installation
Settings.json auto-configuration
Checksum-based update detection

Hierarchical Merge Protocol v1.0

Parent-child branch convergence
Child Completion Constraint
Exception prefixes for hotfix/emergency

TTL Policy

Content freshness management
PROV tags for provenance tracking
Automatic expiration alerts

MVV Generator (v1.1.0)

8-dimension ontological analysis (Context, Purpose, Taxonomy, Semantics, Lineage, Epistemology, Ontology, Aesthetics)
Mission/Vision/Values synthesis pipeline
Automated organizational identity extraction
/mvv command for one-shot execution

Installation

Plugin name vs repo name: the repository is multi-agent-os; the plugin is named maos (see .claude-plugin/plugin.json). Install and reference it as maos. Skills/commands surface namespaced as /maos:<name> (e.g. /maos:orchestrator, /maos:auto-pilot).

From Marketplace (recommended)

Plugin management runs inside a Claude Code session via the /plugin command (not a shell claude plugins subcommand):

/plugin marketplace add ekson73/eko-claude-plugins
/plugin install maos@eko-claude-plugins
/reload-plugins

Then run /plugin (Installed tab) to confirm, and /help to see the /maos:* skills. Need to update later? /plugin marketplace update eko-claude-plugins.

From source (local dev / self-use)

git clone https://github.com/ekson73/multi-agent-os.git
claude --plugin-dir ./multi-agent-os     # load for this session
claude plugin validate ./multi-agent-os  # validate the manifest

--plugin-dir may be repeated to load several plugins and takes precedence over an installed copy of the same name (handy for testing local changes). See Self-Referential Usage.

Project / team scope

To make maos available to everyone on a repository, add to that project's .claude/settings.json (collaborators are prompted to install when they trust the folder):

{
  "extraKnownMarketplaces": {
    "eko-claude-plugins": {
      "source": { "source": "github", "repo": "ekson73/eko-claude-plugins" }
    }
  },
  "enabledPlugins": ["maos@eko-claude-plugins"]
}

CLI equivalent (writes project scope): claude plugin install maos@eko-claude-plugins --scope project. Full schema: Plugin settings.

Plugin Structure

multi-agent-os/
├── .claude-plugin/
│   └── plugin.json           ← Plugin manifest
├── hooks/
│   └── hooks.json            ← Hook configuration
├── plugin-scripts/           ← Hook executables
│   ├── session-start.sh
│   ├── pre-delegate.sh
│   ├── post-delegate.sh
│   ├── session-end.sh
│   └── governance/
│       ├── worktree-gate.sh
│       ├── auto-name-session.sh
│       └── token-budget-gate.sh  ← GaaS token bloat detection
├── commands/                 ← Slash commands (~15; representative subset shown)
│   ├── sync.md
│   ├── audit.md
│   ├── agentic-status.md
│   ├── worktree.md
│   └── delegate.md
├── agents/                   ← Agent definitions (20+; representative subset shown)
│   ├── orchestrator.md
│   ├── sentinel-monitor.md
│   ├── qa-validator.md
│   ├── consolidator.md
│   ├── forge.md              ← Meta-agent creator (Goldilocks + RBAD + 33 Socratic Questions)
│   ├── governance-auditor.md ← Standards & compliance guardian
│   ├── naming-organizer.md   ← Digital organization & taxonomy
│   ├── data-validator.md     ← Truth & evidence verification (11 validation types)
│   └── validation-auditor.md ← Second-line active auditing
├── skills/                   ← Skills (40+, subdirectory format; representative subset shown)
│   ├── audit/SKILL.md
│   ├── agent-select/SKILL.md
│   ├── context-prep/SKILL.md
│   ├── hierarchical-merge/SKILL.md
│   ├── worktree-policy/SKILL.md
│   ├── anti-conflict/SKILL.md
│   ├── status-map/SKILL.md
│   ├── ttl-policy/SKILL.md
│   ├── ontological-analysis/SKILL.md  ← 8-dimension analysis
│   ├── mvv-synthesis/SKILL.md         ← MVV generation
│   └── response-compression/SKILL.md  ← Output verbosity control
├── protocols/                ← Governance protocols
│   ├── exit-hygiene.md       ← Session exit hygiene (zero loose ends)
│   ├── agent-delegation.md   ← Delegation chain & Forge bootstrap
│   ├── rbad.md               ← Role-Based Agent Design (6-category taxonomy)
│   └── action-priority.md    ← Eisenhower Matrix for task prioritization
├── sentinel/                 ← Sentinel Protocol files
├── statusmap/                ← Status Map templates
└── docs/                     ← Additional documentation

Available Commands

Namespaced as /maos:<name>. This table is representative — run /help for the full live list, or /maos:maos-concierge to be guided to the right command/agent/skill for your intent.

Command	Description
`/maos:agentic-status`	Human-readable status of the agentic system (git + agents + sentinel + locks). (Was `/status` — renamed to avoid colliding with Claude Code's built-in `/status`.)
`/maos:sync`	Sync content from the framework to a consumer project
`/maos:audit`	On-demand audit of sessions, tasks, and orchestration flows
`/maos:worktree`	Manage git worktrees for multi-agent isolation
`/maos:delegate`	Delegate a task to a specialized sub-agent with context prep
`/maos:mvv`	Generate/update Mission, Vision, Values for a repository
`/maos:auto-pilot`	Drive an operator goal end-to-end across sub-agents (GaaS/GaaC)
`/maos:quiesce`	Drive the session to quiescence — no open ticket/gap/fix/PR, all PRs green
`/maos:enhance-pipeline`	Run one feature through the full divergent→convergent→deliver lifecycle
`/maos:founder-playbook`	Diagnose an AI-native startup's lifecycle stage and route to the right discipline
`/maos:agentic-tool-forge`	Forge a raw intent into the optimal reusable agentic-tool
`/maos:session-fission`	Non-destructively split a tangled session into focused new ones

Plus namespaced utilities (/maos:auto-shard, /maos:code:*, /maos:analyze:*) — see /help.

Available Skills

40+ skills ship with MAOS — the table below is a representative subset. Model-invoked (Claude auto-selects by task) or explicit via /maos:<skill>. For the full live catalog and guidance on which to use, invoke /maos:maos-concierge.

Skill	Description
`audit`	Sentinel Protocol auditing
`agent-select`	Agent selection algorithm
`context-prep`	Pre-delegation context
`hierarchical-merge`	Branch merge rules
`worktree-policy`	Worktree enforcement
`anti-conflict`	Conflict prevention
`status-map`	Status visualization
`ttl-policy`	Content freshness
`ontological-analysis`	8-dimension philosophical analysis
`mvv-synthesis`	Mission/Vision/Values synthesis
`response-compression`	Output verbosity control (none/lite/full/ultra) with role-based profiles
`maos-concierge`	Onboarding/router/anchor for the entire MAOS framework (start here)
`auto-pilot` / `quiesce` / `converge` / `convergence-engine`	Autonomous orchestration, session quiescence, and multi-proposal convergence
`enhance-pipeline`	Divergent→convergent→deliver feature lifecycle
`agentic-delegation` / `delegate-governance`	Delegation criteria + GaaS/GaaC governance prompts
`agentic-tool-forge` / `-evaluator` / `-trainer`	Create, evaluate, and improve reusable agentic-tools
`anima`	Sovereign precision-naming engine — names anything (12 correctness + 4 resonance aspects, register-aware); the forge's Phase-4 naming delegate + standalone
`founder-playbook` / `founder-stage-*`	AI-native startup lifecycle disciplines (Idea/MVP/Launch/Scale)
`morning-briefing` / `pulse` / `session-fission`	Context restoration, re-orientation, session splitting
`rule-quality-tests` / `operator-quote-capture` / `pii-masking` / `slm-routing`	Governance, capture, PII detection, model routing

Available Agents

20+ agents ship with MAOS — representative subset below. Invoke via the Task tool (subagent_type: maos:<agent>) or see /agents.

Agent	Description
`orchestrator`	Master coordinator
`sentinel-monitor`	Anomaly detection
`qa-validator`	Quality assurance
`consolidator`	Output synthesis
`forge`	Meta-agent creator — specialized agents via Goldilocks Principle + RBAD
`governance-auditor`	Standards governance, compliance auditing, pattern enforcement
`naming-organizer`	Digital organization, taxonomy, naming conventions
`data-validator`	Data validation, evidence capture, truth verification (11 types)
`validation-auditor`	Second-line active auditing, drift detection, integrity verification
`cascade-resolver` / `perspective-trio` / `persona-pipeline`	Convergence engine primitives — score-uplift, orthogonal-trio, review-board
`code-reviewer` / `debugger`	Code review and systematic debugging specialists
`data-analyst` / `legacy-archaeologist` / `memory-curator`	Analysis, legacy reverse-engineering, memory hygiene
`gitops-engineer` / `founder-coach` / `consultants`	GitOps/K8s, startup coaching, expert thinking-archetypes

Hooks

The plugin automatically hooks into Claude Code lifecycle:

Hook	Trigger	Scripts
`SessionStart`	Session initialization	session-start.sh, auto-name-session.sh
`PreToolUse[Task]`	Before delegation	pre-delegate.sh, token-budget-gate.sh
`PreToolUse[Bash]`	Before shell commands	worktree-gate.sh
`PostToolUse[Task]`	After delegation	post-delegate.sh
`Stop`	Session end	session-end.sh

Documentation

Document	Description
Worktrees Guide	Multi-agent worktree coordination
Hierarchical Merge Protocol	Branch convergence rules
Framework Consumption	Consumer project integration
Agent Format	YAML frontmatter specification for agents
Exit Hygiene	Session exit protocol (zero loose ends)
Agent Delegation	Delegation chain & Forge bootstrap
RBAD	Role-Based Agent Design taxonomy
Action Priority	Eisenhower Matrix for prioritization

Self-Referential Usage

This plugin can use itself during development:

# In multi-agent-os directory
claude --plugin-dir .

Interoperability

Skills in this plugin follow the Agent Skills open standard and are compatible with 30+ AI coding tools including Claude Code, Cursor, OpenAI Codex, Gemini CLI, Kiro, VS Code, GitHub Copilot, Goose, Windsurf, and others. Install with npx skills add or native plugin mechanisms.

This repository includes an AGENTS.md file following the AAIF standard (60k+ projects) to guide any AI coding agent working on this codebase.

License

MIT License - See LICENSE file for details.

Multi-Agent OS v1.21.0 | Created by Emilson Moraes | Powered by Claude Code