convez.ai
AgenteAutomaçãoAvançado
bluminal

LumenAI

bluminal/lumenai

Organização de IA que coordena descoberta, desenvolvimento, entrega e operações

Ver no GitHub
BomPluginScore · de 0 a 100

Marketplace de plugins Claude Code que estrutura agentes de IA para simular uma organização de engenharia completa. Implementa 15 agentes especializados em descoberta de produto, desenvolvimento, entrega, operações e aprendizado, orquestrando tarefas do ciclo de vida de software via 12 comandos.

Instalação

$npx convez-ai bluminal/lumenai

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
3
Forks
1
Acompanhando
0
Saúde
87%
Último commit
há 1 semana
Versão
v2.7.1
Issues abertas
5
Contribuidores
3

Casos de uso

  • Estruturar fluxo completo de desenvolvimento de software com agentes especializados
  • Orquestrar revisões de código, segurança e infraestrutura em paralelo
  • Automatizar planejamento de requisitos e escrita de planos de implementação
  • Coordenar análise de performance, observabilidade e melhorias operacionais

Veredito

Plugin sofisticado que vai além de automação simples: modela dinâmica de organização de engenharia com agentes hierárquicos (orquestração, especialistas, pesquisa). Exige entendimento profundo de workflows de software, mas oferece orquestração enterprise genuína com múltiplos quality gates.

README

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

LumenAI

A Claude Code plugin marketplace by Bluminal Labs.

What is this?

LumenAI is a structured registry of Claude Code plugins — collections of AI agents and commands that work together to accomplish complex software delivery tasks.

/plugin marketplace add bluminal/lumenai

Plugins

Synthex

The first plugin in the marketplace. Synthex models a software startup's org chart as a collection of AI agents that synthesize to deliver complete, production-quality software.

/plugin install synthex

The organization spans the full software lifecycle: discover, build, ship, operate, and learn — with 15 agents organized into three layers and 11 commands that orchestrate them.

Agents (15)

Orchestration Layer — Lead roles that coordinate specialists and drive execution.

AgentRoleType
Tech LeadFull-stack orchestrator, primary coding agentExecution + Orchestration
Lead Frontend EngineerFrontend tech lead, delegates to framework specialistsExecution + Delegation
Product ManagerRequirements gathering, implementation planning, product strategyPlanning + Strategy

Specialist Layer — Domain experts invoked by leads or commands for focused work.

AgentRoleType
ArchitectSystem architecture guidance, ADRs, plan feasibility reviewAdvisory + Planning
Code ReviewerCraftsmanship review, specification compliance, convention adherenceAdvisory (PASS/WARN/FAIL)
Security ReviewerSecurity review quality gate (vulnerabilities, secrets, access control)Advisory (PASS/WARN/FAIL)
Terraform Plan ReviewerInfrastructure-as-code review (cost, risk, security)Advisory (PASS/WARN/FAIL)
Quality EngineerTest strategy, coverage analysis, test writingExecution + Advisory
Design System AgentDesign tokens, component governance, compliance auditsExecution + Advisory
Performance EngineerFull-stack performance analysis (Core Web Vitals, queries, bundles)Advisory
SRE AgentSLOs/SLIs, observability, runbooks, blameless postmortemsAdvisory + Execution
Technical WriterAPI docs, user guides, migration guides, changelogsExecution

Research & Analysis Layer — Agents focused on understanding users, measuring outcomes, and driving improvement.

AgentRoleType
UX ResearcherResearch plans, personas, journey maps, Opportunity Solution TreesPlanning + Advisory
Metrics AnalystDORA metrics, HEART/AARRR frameworks, OKR trackingAdvisory
Retrospective FacilitatorStructured retrospectives, improvement item trackingPlanning + Advisory

Commands (12)

CommandPurposeAgents Orchestrated
initInitialize project configuration and directories--
next-priorityExecute next highest-priority tasksTech Lead
refine-requirementsImprove PRD clarity through multi-agent reviewPM + Tech Lead + Lead Frontend Engineer
write-implementation-planTransform PRD into implementation planPM + Architect + Design System Agent + Tech Lead
review-codeMulti-perspective code reviewCode Reviewer + Security Reviewer + Performance Engineer (opt.)
write-adrCreate Architecture Decision RecordArchitect (interactive)
write-rfcCreate Request for CommentsArchitect + PM + Tech Lead + Security Reviewer
test-coverage-analysisAnalyze test gaps, optionally write testsQuality Engineer
design-system-auditAudit frontend for design system complianceDesign System Agent
retrospectiveStructured cycle retrospectiveMetrics Analyst + Retrospective Facilitator
reliability-reviewOperational readiness assessmentSRE Agent + Terraform Plan Reviewer (opt.)
performance-auditFull-stack performance analysisPerformance Engineer

Synthex+ (Beta)

A companion plugin to Synthex that adds persistent team orchestration via Claude Code's beta Agent Teams API. Synthex+ reuses Synthex agent definitions — it does not duplicate or modify them.

BETA — Requires the CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS=1 feature flag. Both the Agent Teams API and this plugin are under active development.

/plugin install synthex-plus

Where standard Synthex spawns ephemeral subagents (each unaware of the others), Synthex+ creates persistent teams where agents share a task list, exchange messages via mailboxes, and coordinate autonomously.

Synthex CommandSynthex+ EquivalentWhat Changes
next-priorityteam-implementPersistent team with real-time coordination instead of sequential subagent invocations
review-codeteam-reviewCross-domain messaging between reviewers (e.g., code reviewer alerts security reviewer)
write-implementation-planteam-planReviewers persist across review cycles, retaining full context
refine-requirementsteam-refinePersistent reviewers with cross-perspective messaging during PRD refinement

Commands (5)

CommandPurposeSynthex Agents Used
team-initInitialize Synthex+ configuration--
team-implementSustained multi-agent implementationTech Lead + Frontend Engineer + Quality Engineer + Code Reviewer + Security Reviewer
team-reviewMulti-perspective code review with cross-domain communicationCode Reviewer + Security Reviewer + Performance Engineer (opt.) + Design System Agent (opt.)
team-planCollaborative implementation planning with persistent reviewersProduct Manager + Architect + Design System Agent + Tech Lead
team-refineCollaborative PRD refinement with persistent reviewersProduct Manager + Tech Lead + Lead Frontend Engineer

When to use Synthex+ over Synthex: Multi-component work spanning 3+ files across 2+ system layers, large code reviews (500+ LOC), security-sensitive changes, planning for 10+ requirements, or refining large PRDs (20+ requirements). For quick, focused tasks, standard Synthex is lighter and more cost-effective.

See the Synthex+ README for full documentation.

Multi-Model Review

Multi-model review fans review prompts out to multiple LLM-family proposers (OpenAI, Google, local-Ollama) via CLI adapters and consolidates findings into a single deduplicated, severity-reconciled, attributed list.

The primary benefit: catching correlated-error blind spots — bugs and issues that any single LLM family would miss but that show up when multiple families review independently.

Off by default. Opt in via /synthex:init (interactive prompt) or by editing .synthex/config.yaml. CLI-only — Synthex does not store API keys.

See docs/specs/multi-model-review/architecture.md for the full design and docs/specs/multi-model-review/adapter-recipes.md for per-adapter setup.

Native Looping

Synthex 0.8+ ships a native --loop flag on iteration-friendly commands (next-priority, write-implementation-plan, refine-requirements, review-code, and all four Synthex+ team commands), plus a generic /synthex:loop for arbitrary prompts. Loops iterate in the same agent thread by default (auto-compaction handles the context window) and persist per-session state at .synthex/loops/<loop-id>.json for resume across sessions. See plugins/synthex/docs/native-looping.md for the full framework spec.

Automated Testing

All agents are tested using a three-layer testing pyramid. Since agents are pure markdown (no runtime code), testing works by invoking agents with synthetic fixtures and validating their outputs.

LayerWhatCostWhen
1 - SchemaValidates markdown structure, sections, tables, verdict format$0Every PR
2 - BehavioralRegex/JS assertions against cached agent outputs~$3/run (cached)Manual trigger
3 - SemanticLLM-as-judge evaluates accuracy and quality~$8/runManual trigger

Current coverage: 404 tests across 18 test suites — 206 for Synthex agents + 131 for Synthex+ templates, hooks, and command outputs + 67 for shared infrastructure. See CLAUDE.md for full details.

cd tests && npx vitest run schemas/   # Layer 1: instant, free

Project Structure

lumenai/
├── .claude-plugin/marketplace.json     # Marketplace registry
├── plugins/
│   ├── synthex/                        # Synthex plugin
│   │   ├── .claude-plugin/plugin.json  # Plugin manifest (15 agents, 12 commands)
│   │   ├── agents/                     # Agent definitions (.md files)
│   │   ├── commands/                   # Command definitions (.md files)
│   │   └── config/defaults.yaml        # Default project configuration
│   └── synthex-plus/                   # Synthex+ plugin (BETA)
│       ├── .claude-plugin/plugin.json  # Plugin manifest (5 commands)
│       ├── commands/                   # Team command definitions (.md files)
│       ├── templates/                  # Team composition templates
│       ├── hooks/                      # Hook behavioral specs + hooks.json
│       ├── scripts/                    # Thin shell shims for hook events
│       ├── config/defaults.yaml        # Default configuration
│       └── docs/                       # Decision guide, context management, output formats
├── tests/                              # Automated agent testing framework
│   ├── schemas/                        # Layer 1: Schema validators + Vitest tests
│   │   └── synthex-plus/               # Synthex+ validators (templates, hooks, outputs)
│   ├── helpers/                        # Invocation wrapper, cache, parser, snapshots
│   ├── fixtures/                       # Synthetic test inputs with planted issues
│   │   └── synthex-plus/               # Synthex+ fixtures
│   └── promptfoo.config.yaml           # Layer 2+3: Behavioral + semantic tests
├── docs/
│   ├── reqs/main.md                    # Product requirements
│   ├── plans/main.md                   # Implementation plan
│   ├── agent-interactions.md           # Agent interaction map and orchestration flows
│   └── research-sources.md             # Research behind each agent's design
├── CLAUDE.md                           # Developer instructions
└── README.md                           # This file

How to Extend

See CLAUDE.md for instructions on adding new agents, commands, and plugins.

License

Apache 2.0 — See LICENSE for details.

Ver o original no GitHub

Bom para

Relacionados