convez.ai
SkillAutomaçãoAvançado
hornosg

Team AI Harness

hornosg/team-ai-harness

32 agentes IA organizados em 3 times com roteamento automático e cerimônias de risco

Ver no GitHub
InicialPluginScore · de 0 a 100

Harness multi-team de 32 agentes IA especializados (Dev, Product, Marketing) com roteamento central, níveis de cerimônia L1-L4 e rastreabilidade FILE-ID/TEST-ID. Funciona com Claude Code, OpenCode, Cursor e Copilot, persistindo contexto arquitetural via Engram.

Instalação

$npx convez-ai hornosg/team-ai-harness

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
1
Forks
1
Acompanhando
0
Saúde
42%
Último commit
há 3 meses
Versão
—
Issues abertas
3
Contribuidores
1

Casos de uso

  • Orquestração de desenvolvimento multi-camadas com controle de risco por cerimônia
  • Pipelines cross-domain: launches de feature, growth initiatives e incident management
  • Rastreabilidade spec-to-code-to-test com FILE-ID e TEST-ID automáticos
  • Persistência de decisões arquitetônicas e contexto entre sessões

Veredito

Skill robusto para orquestração de múltiplos agentes IA em equipes grandes. Exige Node.js 20+, Engram e compreensão de cerimônias/DDD; ideal para organizações que precisam governança de IA estruturada, não para projetos simples ou iniciantes.

README

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

Team AI Harness

A multi-team AI agent harness for development, product, and marketing — built for Claude Code, OpenCode, Cursor, and GitHub Copilot.

License: MIT Claude Code OpenCode


What is this?

A ready-to-use harness of 32 specialized AI agents organized across three teams (Dev, Product, Marketing) with a central routing layer. You install it once into any project's management/ folder and get:

  • A single entry point (@meta-router) that classifies every request and routes it to the right agent chain
  • Ceremony levels (L1–L4) that enforce appropriate rigor based on risk — L4 for anything touching money, auth, or compliance is non-negotiable
  • Spec-driven development with FILE-ID / TEST-ID traceability from plan to code to test
  • Persistent memory across sessions via Engram
  • Multi-adapter sync: same canonical agents generate Claude Code, OpenCode, Cursor, and Copilot formats

Architecture

Owner
  └── @meta-router                  ← single entry point
        ├── @dev-orchestrator        ← L1–L4 ceremony levels
        │     ├── @dev-architect
        │     ├── @dev-technical-leader
        │     ├── @dev-senior-backend / @dev-senior-frontend
        │     ├── @dev-qa
        │     └── @dev-security      ← mandatory on L4
        ├── @product-orchestrator
        │     ├── @product-owner
        │     ├── @product-strategist
        │     └── @product-analyst
        └── @marketing-orchestrator
              ├── @marketing-senior-copywriter
              ├── @marketing-growth-marketer
              └── @marketing-brand-strategist

32 agents total. See AGENTS.md for the full roster with roles, models, and tools.


Key Features

  • Ceremony levels — L1 (quick fix) to L4 (money/auth/compliance). Each level defines which agents are required, what planning is mandatory, and what can't be skipped.
  • FILE-ID / TEST-ID system — every planned file gets a traceable ID. Every test maps to a FILE-ID. The code reviewer verifies this mechanically before merge.
  • Cross-domain pipelines — launch a feature, run a growth initiative, or manage an incident across Product → Dev → Marketing with a single @meta-router command.
  • Roadmap integration — agents read roadmap/roadmap.yaml to contextualize requests. @meta-router status crosses roadmap + proposals + OpenSpec specs into one report.
  • Multi-adapter sync — one canonical source in agents/ generates outputs for Claude Code (.claude/agents/), OpenCode (.opencode/agents/), Cursor (.cursor/rules/), and Copilot (.github/agents/).
  • Persistent memory — all senior agents use Engram to persist architectural decisions, bug resolutions, and session context across conversations.

Prerequisites

ToolVersionInstall
Node.js≥ 20.19.0nodejs.org
Claude Codelatestnpm install -g @anthropic-ai/claude-code
OpenCodelatestnpm install -g opencode-ai
OpenSpeclatestnpm install -g @fission-ai/openspec@latest
Engramlatestbrew install gentleman-programming/tap/engram

Install Engram MCP plugin

# Claude Code
claude plugin marketplace add Gentleman-Programming/engram
claude plugin install engram

# OpenCode
engram setup opencode

Quick Start

Option A — Install into an existing project

curl -fsSL https://raw.githubusercontent.com/<your-org>/team-ai-harness/main/scripts/install-management.sh | bash

Or clone and run locally:

git clone https://github.com/<your-org>/team-ai-harness
cd team-ai-harness
./scripts/install-management.sh /path/to/your-project

This creates a management/ folder inside your project with all agents, skills, rules, roadmap templates, and a root CLAUDE.md that points to management/CLAUDE.md.

Option B — Use this repo directly

git clone https://github.com/<your-org>/team-ai-harness
cd team-ai-harness
./scripts/sync-agents.sh   # generates .claude/agents/ and .opencode/agents/

Then open Claude Code or OpenCode in the repo root:

claude
@meta-router I need to add payment processing to the checkout flow

First Steps After Install

  1. Fill in management/PROJECT.md — project identity, stack, services, principles. Agents read this as ground truth.
  2. Complete management/roadmap/roadmap.yaml — milestones and epics. Ask @meta-router crear epica [X] if you're starting fresh.
  3. Check the status — @meta-router status gives a full report crossing roadmap + proposals + specs.

Ceremony Levels

LevelWhenRequired agents
L1Quick fix, reversible, < 30 minJunior + TL (optional)
L2Standard feature, known patternsTL + Senior + QA
L3Architectural change, DB migrationArchitect + TL + Senior + QA + Monitoring
L4Money / Auth / ComplianceArchitect + Security (both, non-negotiable) + TL + Senior + QA + Monitoring

L4 triggers (hardcoded, can't be downgraded): payment, auth, token, wallet, session, PCI, BCRA, compliance.


Document Flow

Each document has exactly one owner:

DocumentOwner agentWhen
PROP-NNN.md (proposal)product-ownerNew idea / discovery
roadmap/epicas/ENN-*.mdorchestratorsProposal approved
docs/adr/ADR-XX.mdarchitectStructural decision L3/L4
openspec/changes/[name]/tasks.mdarchitect (L3/L4) / technical-leader (L2/L3)Before coding
FILE-IDs (F-NNN) + TEST-IDs (T-NNN)same as tasks.md ownerInside tasks.md
Codesenior-backend / senior-frontendImplementation
QA reportqaBefore sign-off
Code reviewtechnical-leaderEvery PR L2+
Security reportsecurityL4 mandatory

See AGENTS.md for the full flow diagram.


FILE-ID / TEST-ID System

Every planned file gets a unique FILE-ID. Every test maps to a FILE-ID. The code reviewer verifies mechanically before merge.

## FILE-ID Table
| FILE-ID | Path | Action | Layer | Description |
|---------|------|--------|-------|-------------|
| F-001   | src/domain/model/order.ts | CREATE | Domain | Order aggregate |
| F-002   | src/application/usecase/place-order.ts | CREATE | Application | Place order use case |

## TEST-ID Table
| TEST-ID | FILE-ID | Type | Scenario | Expected |
|---------|---------|------|----------|----------|
| T-001 | F-001 | Unit | Create order with valid data | Order with UUID |
| T-002 | F-001 | Unit | Create with invalid amount | ValidationError |
| T-003 | F-002 | Unit | Execute with duplicate | ConflictError |

If T-NNN doesn't exist in code → Critical finding → no merge.


Cross-Domain Pipelines

@meta-router launch feature [X]

Runs automatically: Product (define) → Dev (implement) → Marketing (launch) → Product (measure)

PipelineSequence
Feature launchProduct → Dev → Product (validation) → Marketing → Dev (deploy) → Product (metrics)
Growth initiativeProduct → Dev → Marketing → Product (measure)
Production incidentDev (fix) → Product (impact) → Marketing (comms if needed)
New segmentProduct (research) → Marketing (messaging) → Dev (gaps) → Marketing (campaign)

Modifying Agents

Canonical sources live in agents/. Generated files in .claude/agents/ and .opencode/agents/ are never edited directly.

# Edit canonical source
vim agents/dev/architect.md

# Sync to all configured adapters
./scripts/sync-agents.sh

# Sync to all adapters (Claude, OpenCode, Cursor, Copilot)
./scripts/sync-agents.sh --all

# Dry run
./scripts/sync-agents.sh --dry-run

Repository Structure

team-ai-harness/
├── agents/                        ← canonical source (edit here)
│   ├── orchestrators/             ← meta-router, dev/product/marketing orchestrators
│   ├── dev/                       ← 11 dev agents
│   ├── product/                   ← 7 product agents
│   └── marketing/                 ← 10 marketing agents
├── skills/                        ← cada skill es un directorio con SKILL.md
│   ├── dev/                       ← code-reviewer/, planner/, owasp-top10/, hexagonal-{go,python,flutter}/,
│   │                                 kong/, prometheus/, grafana/, loki/, tracing/, observability-stack/, digital-ocean/, ...
│   ├── marketing/                 ← market-audit/, market-seo/, market-copy/, ...
│   └── shared/                    ← bmad-*/, roadmap-management/, roadmap-status/
├── rules/                         ← architecture, api-standards, security (adapter-ready)
├── adapters/                      ← format transforms per AI tool
├── roadmap/
│   ├── roadmap.yaml               ← milestones + epics
│   ├── epicas/                    ← ENN-*.md templates
│   └── propuestas/                ← PROP-NNN-*.md templates
├── config/
│   ├── routing-rules.yaml
│   ├── ceremony-levels.yaml
│   └── cross-domain-pipelines.yaml
├── scripts/
│   ├── sync-agents.sh             ← generates adapter outputs
│   └── install-management.sh      ← installer for any project
├── .claude/agents/                ← generated for Claude Code
├── .opencode/agents/              ← generated for OpenCode
├── AGENTS.md                      ← full agent roster + flow diagram
└── PROJECT.md                     ← project identity template

Skills Reference

CategorySkillPurpose
DevplannerGenerate FILE-IDs + TEST-IDs before coding
Devcode-reviewer7-dimension PR review with score /35
Devconventional-commitCommit format and enforcement
Devpr-workflowPR best practices and size limits
Devowasp-top10OWASP Top 10:2021 gate + audit
Devmemory-protocolWhen and how to use Engram
Marketingmarket-auditFull market analysis via WebFetch
Marketingmarket-competitorsCompetitor scanning + comparison matrix
Marketingmarket-seoSEO checklist + local optimization
Marketingmarket-copyCopy review and improvement
Marketingmarket-landing-croLanding page conversion optimization
Sharedbmad-brainstorming100+ ideas, anti-bias rotation
Sharedbmad-distillatorLossless document compression
Sharedbmad-party-modeMulti-agent discussion simulation
Sharedbmad-editorial-review-structureDocument structure analysis
Sharedbmad-editorial-review-proseClinical copy editing
Sharedbmad-review-adversarialCynical reviewer, minimum 10 findings
Sharedroadmap-managementCreate proposals, epics, update roadmap
Sharedroadmap-statusFull status report: roadmap + proposals + specs
Hexagonalhexagonal-goHexagonal + DDD para Go (Chi/pgx/Wire)
Hexagonalhexagonal-pythonHexagonal + DDD para Python (FastAPI/SQLAlchemy async)
Hexagonalhexagonal-flutterClean Architecture + DDD para Flutter (BLoC/Riverpod)
StackkongAPI Gateway: routing declarativo (decK), plugins, DB-less, seguridad del borde
StackprometheusMétricas, golden signals, PromQL, alerting
StackgrafanaDashboards, SLOs, dashboards-as-code, correlación de señales
StacklokiLogging estructurado (JSON), LogQL, pipeline de ingestión
Stackobservability-stackVisión integral Prometheus + Grafana + Loki; flujo de diagnóstico
StacktracingTracing distribuido: OpenTelemetry + Grafana Tempo, propagación W3C
Stackdigital-oceanDeploy e infra: App Platform/Droplets, doctl, VPC, managed DB, Spaces, costos

Credits

Engram — Persistent Memory for AI Agents

This harness relies on Engram by Gentleman Programming for persistent memory across agent sessions and chat compactions.

Engram provides a single binary (zero dependencies) backed by SQLite + FTS5, exposed as an MCP server. It enables any MCP-compatible agent — Claude Code, OpenCode, Gemini CLI, VS Code Copilot — to remember architectural decisions, bug resolutions, and session context indefinitely.

brew install gentleman-programming/tap/engram

Give them a star: github.com/Gentleman-Programming/engram

OpenSpec

Spec-driven development framework by Fission AI. Provides the openspec/changes/ structure used by the planner skill and all L2+ ceremony workflows.

npm install -g @fission-ai/openspec@latest

Claude Code

Agent runtime by Anthropic. The .claude/agents/ format and subagent invocation pattern (@agent-name) are Claude Code conventions.


Contributing

Contributions welcome. A few conventions to keep the harness portable:

  • No project-specific content in agents — agents must work for any project. Project identity lives in PROJECT.md, not in agent files.
  • Skills are documents, not commands — skills live in skills/<name>/SKILL.md and are read by agents. They don't go in .claude/commands/.
  • Canonical source in agents/ — never edit generated files in .claude/agents/ or .opencode/agents/ directly.
  • L4 is non-negotiable — don't add exceptions or workarounds for money/auth/compliance ceremony requirements.

To add a new agent:

# 1. Create canonical file
vim agents/dev/my-new-agent.md   # with proper frontmatter

# 2. Sync
./scripts/sync-agents.sh

# 3. Add to AGENTS.md roster table

License

MIT © 2025 hornosg

Permission is hereby granted, free of charge, to any person obtaining a copy of this software and associated documentation files (the "Software"), to deal in the Software without restriction, including without limitation the rights to use, copy, modify, merge, publish, distribute, sublicense, and/or sell copies of the Software, and to permit persons to whom the Software is furnished to do so, subject to the following conditions:

The above copyright notice and this permission notice shall be included in all copies or substantial portions of the Software.

THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.

Ver o original no GitHub

Bom para

Relacionados