convez.ai
AgenteAutomaçãoAvançado
lucasrp

Edge of Chaos — Framework de Agentes IA Autônomos

lucasrp/edge-of-chaos

Framework completo para agentes IA autônomos com skills, dashboard e heartbeat

Ver no GitHub
InicialPluginScore · de 0 a 100

Framework Python para implantar agentes IA autônomos baseados em Claude Code, com identidade própria, dashboard, 22 skills pré-instaladas e ciclo heartbeat. Inclui renderização de configuração, gerenciamento de skills, serviços systemd e CLI tools para orquestração completa.

Instalação

$npx convez-ai lucasrp/edge-of-chaos

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
3
Acompanhando
0
Saúde
28%
Último commit
há 2 semanas
Versão
—
Issues abertas
28
Contribuidores
6

Casos de uso

  • Implantar agentes IA com identidade e autonomia definida
  • Automatizar ciclos de pesquisa, análise e publicação de conteúdo
  • Gerenciar múltiplos skills e protocolos pre/post-flight para agentes

Veredito

Ferramenta sólida e bem-estruturada para quem precisa orquestrar agentes IA em produção com autonomia real. Exige familiaridade com systemd, Python e arquitetura de skills, mas oferece abstração valiosa para render/deploy/heartbeat.

README

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

edge-of-chaos — Autonomous AI Agent Framework

Framework for deploying autonomous AI agents based on Claude Code. Each agent has its own identity, dashboard, skills, and heartbeat cycle.

Quick Start (5 min)

# 1. Clone
git clone https://github.com/lucasrp/edge-of-chaos.git my-agent
cd my-agent

# 2. Configure (edit 5 required fields)
cp agent.yaml.example agent.yaml
nano agent.yaml

# 3. Render + Install
python3 tools/edge-render
python3 tools/edge-apply

# 4. Validate
python3 tools/edge-doctor

Done. Dashboard running, 22 skills installed, heartbeat ready.

agent.yaml

Bootstrap spec. Used at render/install time to create runtime config. Five required fields:

name: my-agent                    # unique name (lowercase, hyphens)
codename: ma                      # prefix for skills (/ma-pesquisa, /ma-heartbeat)
missao: "What this agent does"    # 1-2 sentences
persona: "How it communicates"    # tone and style
dominio: "government"             # work domain

Everything else has smart defaults. See agent.yaml.example for all options.

What edge-apply Does (8 phases)

  1. Render — generates all files from agent.yaml + templates
  2. Directories — creates content dirs (blog/, reports/, logs/, etc.)
  3. Skills — installs 22 skills with your prefix to ~/.claude/skills/
  4. Identity — CLAUDE.md, memory files, config, onboarding templates
  5. Dashboard venv — Flask operator surface with FTS5 search
  6. Tools venv — edge-consult, review-gate, edge-deepresearch
  7. Systemd — heartbeat timer + dashboard server service (blog-server name kept for compatibility)
  8. Tools — CLI tools + symlinks to ~/.local/bin/

After Install

# Start dashboard (service name kept as blog-server for compatibility)
systemctl --user enable --now blog-server

# Start autonomous heartbeat (every 2h)
systemctl --user enable --now agent-heartbeat.timer

# Or run manually
~/.local/bin/heartbeat.sh

The first heartbeat publishes a self-introduction and delivers the first useful content about the domain. No warmup phase — the agent produces from day one.

Structure

my-agent/
├── agent.yaml              ← your config (gitignored)
├── agent.yaml.example      ← template with all fields
├── config/
│   ├── preflight.yaml      ← canonical pre-skill protocol source (compiled and executed by CLI)
│   ├── postflight.yaml     ← canonical post-skill protocol source (compiled and executed by CLI)
│   ├── strategy.md         ← operator direction (agent reads, proposes)
│   ├── interests.md        ← shared interests (guides exploration)
│   ├── branding.yaml       ← agent phenotype (name, colors, dashboard config via legacy `blog` key)
│   └── runtime-routers.yaml ← rendered LLM router config used at runtime
├── skills/                 ← 22 core skills (genotype)
├── tools/                  ← CLI tools (edge-consult, edge-fontes, etc.)
├── blog/                   ← Flask + htmx dashboard server (legacy package name)
├── search/                 ← FTS5 + vector search engine
├── templates/              ← .tpl files rendered by edge-render
├── memory/                 ← personality, rules, method (genotype)
├── autonomy/               ← autonomy policy, capabilities, frontier
└── systemd/                ← service + timer templates

Key Concepts

Genotype / Phenotype / Epigenetics

Every change goes through one question: is this genotype or phenotype?

  • Genotype — shared code (skills, tools, dashboard server). Lives in the repo. Propagates via git pull.
  • Phenotype — rendered runtime config + local state. Per-agent. Generated at install and evolves in place.
  • Epigenetics — runtime state (feed entries, reports, memory). Never replicates.

Heartbeat

The agent wakes every 2h via systemd timer. It runs internal heartbeat curation over sessions, threads, tasks, and health, then dispatches one action skill: research, discovery, report, planner, or autonomy.

Adversarial Review

The agent never evaluates its own output. Before publishing, it submits conclusions to GPT/Grok via edge-consult. This creates a cross-model review loop.

Onboarding

New agents produce from the first heartbeat. A checklist tracks progress (identity, production, recognition, calibration) and completes organically as the agent delivers real content. No sequential phases — onboarding is concurrent with production.

Publication Pipeline (consolidate-state)

Atomic publication: state snapshot → adversarial review → quality gate → entry publish → HTML report → state commit → git commit.

Tools

ToolPurpose
edge-renderGenerate files from agent.yaml + templates
edge-applyProvision host (idempotent, 8 phases)
edge-doctorValidate installation (30 checks)
edge-consultCross-model adversarial review
edge-fontesUnified external source search
edge-indexIndex content into FTS5 + vectors
edge-searchHybrid search (semantic + keyword)
edge-eventStructured event logging
edge-dispatchShadow dispatch-cycle envelope for heartbeat and operator runs
review-gateLLM-as-judge quality gate

Requirements

  • Claude Code CLI (npm install -g @anthropic-ai/claude-code)
  • Python 3.10+ with venv
  • Git
  • Linux with systemd (or macOS with launchd)

API keys (optional, degrade gracefully):

  • OPENAI_API_KEY — enables adversarial review and quality gates
  • EXA_API_KEY — enables semantic search via Exa
  • XAI_API_KEY — enables Grok as second reviewer

License

MIT

Ver o original no GitHub

Bom para

Relacionados