convez.ai
AgenteAutomaçãoAvançado
GaspardCoche

Agent System – Orquestração Multi-agente com Obsidian

GaspardCoche/agent-system

Orquestração de agentes IA com memória persistente em Obsidian e GitHub Actions

Ver no GitHub
InicialPluginScore · de 0 a 100

Sistema multi-agentes autônomo orquestrado via GitHub Actions e Claude Code, onde cada agente especializado (Pesquisador, Analisador, Forjador, Sentinel) possui memória persistente em Obsidian Vault e comunica via protocolo padronizado. Inclui 10 agentes com papéis específicos (lideração, análise, automação, auditoria) e knowledge graph integrado para captura de aprendizados.

Instalação

$npx convez-ai GaspardCoche/agent-system

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
0
Forks
0
Acompanhando
0
Saúde
28%
Último commit
há 8 horas
Versão
—
Issues abertas
29
Contribuidores
3

Casos de uso

  • Automação de workflows complexos com múltiplos agentes paralelos
  • Gestão de conhecimento empresarial com knowledge graph interativo
  • Auditoria, triage de dados e geração de insights automatizados
  • CI/CD avançado com validação, testes e agregação de relatórios

Veredito

Arquitetura empresarial robusta para orquestração de agentes com persistência de conhecimento. Excelente para times que precisam coordenar múltiplos bots autônomos com contexto compartilhado; requer experiência em GitHub Actions, Python e design de sistemas distribuídos.

README

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

Agent System

Systeme multi-agents autonome orchestre via GitHub Actions et Claude Code. Chaque agent a un role specialise, une memoire persistante (Obsidian Vault), et un protocole de communication standardise.

Architecture

GitHub Issues / Cron / Manual
         |
    Orchestrator ──── decompose la tache
         |
    +----+----+
    |         |
Researcher  Analyzer    (parallele)
    |         |
    +----+----+
         |
       Forge ──── implemente
         |
      Sentinel ──── teste & valide
         |
      Aggregate ──── rapport + notification

Agents

AgentRoleTrigger
OrchestratorDecompose, route, superviseIssues, PRs, cron, manual
ForgeImplementation codeDispatch
SentinelTests & QADispatch
IrisEmail triage & digestCron 7h30, manual
ScoutVeille web, enrichissementDispatch, manual
AriaLead generation, CRMDispatch, manual
NexusAudit Google AdsCron Lundi, manual
SageAuto-amelioration systemeCron Dimanche, manual
RalphAutomatisation, webhooksrepository_dispatch
LumenAnalyse donnees & insightsDispatch

Vault (Knowledge Base)

Le vault Obsidian dans docs/vault/ est la memoire persistante du systeme. Chaque agent :

  1. Lit INDEX.md + son fichier memoire au debut de chaque run
  2. Ecrit ses apprentissages apres chaque run
  3. Le knowledge graph est automatiquement reconstruit via vault-sync.yml

Structure du Vault

docs/vault/
  INDEX.md                    -- Hub central, point d'entree
  business/
    vision.md                 -- Vision & mission
    strategy.md               -- Strategie business
    roadmap.md                -- Roadmap & priorites
  prospects/
    pipeline.md               -- Pipeline leads (Scout + Aria)
  campaigns/
    google-ads.md             -- Etat campagnes (Nexus)
  operations/
    daily-digest.md           -- Digest email (Iris)
    decisions.md              -- Decisions d'architecture
    runbooks.md               -- Procedures operationnelles
    kpis.md                   -- Metriques systeme
  content/
    social-media.md           -- Strategie contenu
    brand-voice.md            -- Ton editorial
    editorial-calendar.md     -- Calendrier editorial
  tech/
    infrastructure.md         -- Architecture technique
    integrations.md           -- APIs & MCP servers
  security/
    access-control.md         -- Controle d'acces
  agents/
    dispatch-log.md           -- Journal orchestrations
    nexus-memory.md           -- Memoire Nexus
    iris-memory.md            -- Memoire Iris
    scout-memory.md           -- Memoire Scout
    sage-memory.md            -- Memoire Sage
    forge-memory.md           -- Memoire Forge
    sentinel-memory.md        -- Memoire Sentinel
    ralph-memory.md           -- Memoire Ralph
    lumen-memory.md           -- Memoire Lumen
    aria-memory.md            -- Memoire Aria
  templates/
    vault-note.md             -- Template note generique
    agent-memory.md           -- Template memoire agent

Ouvrir dans Obsidian

open -a Obsidian docs/vault

Obsidian affiche le graph interactif des connexions entre les notes (Cmd+Shift+G).

Dashboard

Dashboard web deploye sur GitHub Pages (https://gaspardcoche.github.io/agent-system/) avec 4 onglets :

  • Runs -- Historique des runs, lancement d'agents, rapports inline
  • Status -- Sante des agents, secrets configures/manquants
  • Agents -- Documentation, inputs, instructions de setup
  • Graph -- Knowledge graph interactif (D3.js)

Skills

Systeme de skills qui transforme les patterns MCP frequents en scripts Python autonomes :

SkillDescription
firecrawl_scrapeScrape URL -> markdown
github_create_issueCreer/commenter issues
gemini_analyzeAnalyse gros fichiers via Gemini
slack_notifyNotification Slack formatee

Les skills sont geres par Sage qui analyse les retrospectives et propose de nouveaux skills.

Setup

Prerequis

  • GitHub repo avec Actions activees
  • Secrets GitHub configures (voir ci-dessous)
  • GitHub Pages active sur le repo (deploy depuis docs/)

Secrets GitHub requis

SecretAgent(s)Notes
CLAUDE_CODE_OAUTH_TOKENTousclaude setup-token en local
GEMINI_API_KEYLumen, SageGoogle AI Studio
GMAIL_TOKEN_JSONIrisOAuth Gmail
GMAIL_USER_EMAILIrisAdresse Gmail
FIRECRAWL_API_KEYScoutfirecrawl.dev
HUBSPOT_API_KEYAriaHubSpot private app
FULLENRICH_API_KEYAriafullenrich.com
GOOGLE_ADS_DEVELOPER_TOKENNexusGoogle Ads API Center
GOOGLE_ADS_CLIENT_IDNexusGoogle Cloud OAuth
GOOGLE_ADS_CLIENT_SECRETNexusGoogle Cloud OAuth
GOOGLE_ADS_REFRESH_TOKENNexusOAuth flow
SLACK_WEBHOOK_URLNotificationsSlack Incoming Webhook

Lancer un agent

# Via GitHub CLI
gh workflow run orchestrator.yml -f task_description="Description de la tache"

# Agent specifique
gh workflow run nexus.yml -f report_type=weekly_audit -f dry_run=true

# Via le dashboard GitHub Pages
# Ouvrir https://gaspardcoche.github.io/agent-system/ -> Config (PAT) -> Run agent

Rebuild le knowledge graph

python3 .github/scripts/vault_builder.py --vault-dir docs/vault --output docs/data/graph.json

Protocole de communication inter-agents

Chaque agent lit sa tache depuis /tmp/agent_task.json et ecrit ses resultats dans /tmp/agent_result.json :

{
  "agent": "<nom>",
  "task_id": "<id>",
  "status": "complete|failed|needs_retry|pending_approval",
  "summary": "<max 150 mots>",
  "artifacts": ["<filepath>"],
  "next_agent": "<nom ou null>",
  "retrospective": {
    "what_worked": "...",
    "what_failed": "...",
    "mcp_patterns": ["<tool:pattern:count>"],
    "improvement_suggestion": "..."
  }
}

Self-Improvement

  1. Chaque agent remplit retrospective dans son resultat JSON
  2. Sage lit toutes les retrospectives chaque dimanche
  3. Sage propose des ameliorations de prompts via PR
  4. Les erreurs repetees vont dans memory/lessons_learned.md
  5. Les patterns MCP frequents deviennent des skills valides

Fichiers cles

FichierRole
CLAUDE.mdInstructions globales pour tous les agents
.github/workflows/_reusable-claude.ymlBase workflow commune
.github/scripts/vault_builder.pyGenere le knowledge graph
.github/scripts/generate_report.pyGenere les rapports
docs/index.htmlDashboard complet
docs/data/agents.jsonConfig agents pour le dashboard
skills/registry.jsonRegistry des skills
memory/lessons_learned.mdErreurs a ne pas repeter

Licence

Projet prive.

Ver o original no GitHub

Bom para

Relacionados