convez.ai
AgenteAutomaçãoAvançado
andywxy1

CEO Plugin — Orquestrador de Agentes Especializados

andywxy1/ceo-plugin

Orquestra 170+ agentes especializados com coordenação de equipes e loops verify-fix

Ver no GitHub
InicialPluginScore · de 0 a 100

Plugin para Claude Code que coordena 170+ agentes especializados em 13 domínios diferentes. Implementa coordenação de equipes, loops verify-fix, roteamento de modelos e execução controlada para projetos complexos. Suporta reuniões de produtos com discussão inter-departamental e relatórios estruturados.

Instalação

$npx convez-ai andywxy1/ceo-plugin

1 comando no terminal: detecta o formato do repositório (marketplace ou plugin avulso) e instala via Claude Code. Requer Node 18+ e o Claude Code instalado.

Métricas

Estrelas
6
Forks
1
Acompanhando
1
Saúde
42%
Último commit
há 7 meses
Versão
—
Issues abertas
0
Contribuidores
2

Casos de uso

  • Coordenar equipes de agentes para projetos complexos multi-domínio
  • Realizar reuniões de produto com discussão inter-departamental entre agentes
  • Implementar ciclos de verificação e correção automáticos com escalação

Veredito

Solução robusta para orquestração em larga escala de agentes especializados. Destacam-se a coordenação lateral entre equipes, loops verify-fix e roteamento adaptativo de modelos. Requer familiaridade técnica avançada com Claude Code e gestão de contexto compartilhado.

README

A tradução para o português deste README ainda está na fila.

English | 简体中文

What's New in v3.1.0

Product Meeting — /ceo:meeting assembles a cross-functional team of agents acting as company staff. They discuss your product freely via SendMessage — proposing bugs, debating improvements, asking each other questions across departments. An independent Meeting Observer then reads the full transcript and writes a comprehensive report with per-person narratives, severity-rated findings, blind-spot analysis, and prioritized action items.

Previously in v3.0.0

Team Coordination — Coupled workstreams use TeamCreate + SendMessage so agents can negotiate laterally (API contracts, shared state) without routing everything through the orchestrator.

Verify-Fix Loop — Every implementation task is reviewed by a read-only agent (disallowedTools: Edit, Write). Failed tasks loop back with specific feedback. Max 3 retries before user escalation.

Onboarding & Settings — /ceo:setup walks you through model tier, verify-fix preferences, team mode, and checkpoint frequency. Writes settings.json so the CEO remembers your choices.

Model Routing — Every agent now has a model: field in its frontmatter. Architects, reviewers, and strategists default to Opus. Implementers and developers default to Sonnet. Override globally via settings.json:

TierBehavior
max_qualityAll agents on Opus
balancedOpus + Sonnet split (default)
max_speedAll agents on Sonnet

Shared Context Protocol — Parallel agents read and append to a shared context file (ceo-projects/<name>/context/<workstream>.md), preventing duplicate discovery work. Read the protocol docs →

What's Inside

  • 3 skills — /ceo:ceo (the meta-orchestrator) + /ceo:meeting (product meeting simulation) + /ceo:setup (onboarding wizard)
  • 170+ specialized agents spanning 13 domains
  • 19 reference docs — NEXUS framework, phase playbooks, scenario runbooks, handoff templates, anti-patterns guide
  • Settings system — settings.json + JSON Schema for full configurability
  • Session-start hook — automatically suggests /ceo when multi-domain tasks are detected

Installation

Prerequisites

  • Claude Code v1.0.33 or later
  • Run claude --version to check

Step 1: Add the marketplace

/plugin marketplace add andywxy1/ceo-plugin

Step 2: Install the plugin

/plugin install ceo@ceo-plugin

Step 3: Run setup (recommended)

/ceo:setup

This walks you through model tier, verify-fix preferences, team mode, and checkpoint frequency. Takes about 1 minute.

Step 4: Reload and verify

/reload-plugins

Run /agents to see the 170+ agents loaded, or /help to see ceo:ceo and ceo:setup listed under available skills.

Updating

/plugin install ceo@ceo-plugin

Then reload:

/reload-plugins

Your settings.json is preserved across updates.

Usage

/ceo:ceo

The CEO follows a strict protocol with hard gates at every transition:

  1. Phase 0 — Settings — loads settings.json (model tier, verify-fix, team mode, checkpoints)
  2. Phase 1 — Discovery — asks questions to understand your project scope, domains, and constraints
    • Hard Gate: user must confirm project brief before advancing
  3. Phase 2 — Planning — matches agents to needs, builds an execution plan with workstreams and dependencies
    • Hard Gate: user must explicitly approve the plan before any agents spawn
  4. Phase 3 — Pre-flight — spawns key agents in review-only mode to surface ambiguities (mandatory for ALL project scales)
    • Hard Gate: all ambiguities must be resolved before execution begins
  5. Phase 4 — Execution — orchestrates agents via teams or standalone dispatch, runs verify-fix loops, tracks progress
    • Hard Gate: every task must pass read-only review before marking complete

The CEO uses multiple enforcement layers to prevent common orchestration failures:

LayerMechanism
Hard Gates<HARD-GATE> blocks at every phase transition — non-negotiable barriers
Verify-Fix LoopEvery task reviewed by a read-only agent; feedback loops until pass or escalation
Team CoordinationCoupled agents use SendMessage for lateral negotiation within TeamCreate groups
Model RoutingPer-agent model assignment via frontmatter, overridable by settings.json
Shared ContextParallel agents share discoveries via context files to prevent duplicate work
Rationalization Prevention15-entry table of common CEO shortcuts with rebuttals
Red Flag Callouts13 internal thoughts that trigger immediate re-evaluation
Anti-Pattern Guide10 documented orchestration failure modes with fixes

Rigid Protocols (never bend)

Phase sequence, hard gates, Tier 1 discovery questions, verify-fix loop, reviewer is read-only, quality gate checklists, 3-retry escalation limit, CEO-never-implements rule, handoff template format, plan approval gate, mandatory pre-flight, shared context initialization.

Flexible Protocols (adapt to context)

Number of agents per phase, sprint duration, parallel tracks, scale classification, scenario runbook selection, Tier 2/3 question selection, checkpoint frequency, team vs. standalone dispatch, reviewer choice, shared context granularity.

Run /ceo:setup to configure interactively, or edit settings.json directly.

SettingValuesDefaultDescription
model_tiermax_quality / balanced / max_speedbalancedWhich model runs each agent type
verify_fix.enabledtrue / falsetrueRun read-only reviewer after each task
verify_fix.max_retries1-53Verify-fix cycles before user escalation
verify_fix.reviewer_modelopus / sonnetopusWhich model runs the reviewer
team_modeauto / always / neverautoWhen to use TeamCreate/SendMessage
checkpointtask / workstream / phaseworkstreamHow often CEO pauses for approval
preflight_agents1-53Agents consulted during pre-flight
project_dirpath./ceo-projectsWhere project files are stored
shared_contexttrue / falsetrueCreate shared context files for workstreams

Full schema: settings.schema.json

For larger projects, execution maps to the 7-phase NEXUS pipeline:

Phase 0: Intelligence & Discovery (3-7d)     -> Gate: Executive Summary Generator
Phase 1: Strategy & Architecture (5-10d)      -> Gate: Studio Producer + Reality Checker
Phase 2: Foundation & Scaffolding (3-5d)       -> Gate: DevOps + Evidence Collector
Phase 3: Build & Iterate (2-12wk)             -> Gate: Agents Orchestrator
Phase 4: Quality & Hardening (3-7d)           -> Gate: Reality Checker (sole authority)
Phase 5: Launch & Growth (2-4wk)              -> Gate: Studio Producer + Analytics Reporter
Phase 6: Operate & Evolve (ongoing)           -> Governance: Studio Producer

Every NEXUS phase has a hard gate, mandatory checklist-to-task conversion, and evidence requirements.

Scenario Runbooks

Pre-built activation templates for common project types:

ScenarioTimelineDetails
Startup MVP4-6 weeks18-22 agents, compressed discovery through launch
Enterprise Feature8-12 weeksFull compliance and multi-team coordination
Marketing Campaign2-4 weeksMulti-channel content production
Incident Response1-5 daysP0/P1 emergency response

Agent Domains

DomainAgentsExamples
Engineering23Backend Architect, Frontend Developer, DevOps, Security Engineer, SRE
Marketing26SEO, TikTok, Xiaohongshu, Content Creator, Growth Hacker
Game Dev19Unity, Unreal, Godot, Roblox, Narrative Designer, Level Designer
Sales9Deal Strategist, Pipeline Analyst, Sales Coach, Proposal Strategist
Design8UX Architect, UI Designer, Brand Guardian, Visual Storyteller
Testing8API Tester, Performance Benchmarker, Accessibility Auditor
Paid Media7PPC, Programmatic, Paid Social, Tracking Specialist
Support6Analytics Reporter, Finance Tracker, Infrastructure Maintainer
Project Mgmt6Project Shepherd, Studio Producer, Jira Workflow Steward
Product5Product Manager, Sprint Prioritizer, Trend Researcher
Spatial/XR5visionOS, WebXR, Metal Engineer, XR Interface Architect
Specialized29MCP Builder, Workflow Architect, Document Generator, ZK Steward
Strategy1Agents Orchestrator

All agents now include model: frontmatter (Opus or Sonnet). Read-only agents (reviewers, auditors, checkers) include disallowedTools: Edit, Write.

Local Development

To test changes locally without installing:

claude --plugin-dir /path/to/ceo-plugin

Run /reload-plugins after making changes to pick them up without restarting.

License

GPL-3.0

Ver o original no GitHub

Bom para

Relacionados