Agent Daily Publisher
quarktimes/agent-daily-publisher
Multi-agent que transforma logs de código em artigos técnicos publicados
Ver no GitHubSistema multi-agent em produção que captura sessões de desenvolvimento do Claude Code, analisa logs técnicos e publica artigos automaticamente em plataformas como Dev.to, CSDN e WeChat. Implementa arquitetura completa com padrões ReAct, loops de auto-avaliação, recuperação de erros e observabilidade via MCP.
Instalação
npx convez-ai quarktimes/agent-daily-publisher1 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
Casos de uso
- Automatizar publicação de artigos técnicos a partir de sessões de desenvolvimento
- Demonstrar arquitetura avançada de sistemas multi-agent em produção
- Implementar pipelines com avaliação de qualidade e loops de revisão automática
Veredito

Projeto exemplar de engenharia de agentes em Python, que vai além de wrappers de frameworks ao demonstrar padrões sólidos: ReAct loops, contratos JSON Schema, auto-avaliação e degradação graciosa. Valioso tanto para portfólio técnico quanto para aprender arquitetura multi-agent em escala.
README
A tradução para o português deste README ainda está na fila.
🎯 Why This Exists
Two audiences, one solution:
- For me: Automatically capture daily Claude Code sessions and publish technical blog articles to 掘金, Dev.to, Medium, etc.
- For interviewers: A living portfolio piece that proves I understand production multi-agent systems at depth.
"Anyone can chain LangChain calls. Few can design agent systems with evaluation loops, error recovery, and protocol-level integration."
🏛 Architecture at a Glance
┌──────────┐ ┌──────────┐ ┌──────────┐
│ Capture │→ │ Analyze │→ │ Write │→ ─ ─ ─ ┐
│ (Tool- │ │ (CoT) │ │ (Gen w/ │ │
│ Use) │ │ │ │ Const.) │ │
└──────────┘ └──────────┘ └──────────┘ │
┌──────▼──────┐
│ Judge │
│ (Self-Eval) │
└──────┬──────┘
│ pass
┌──────────┐ ┌──────────┐ │
│ Publish │← │ Adapt │← ─ ─ ─ ─ ─ ─ ─ ─ ─ ─┘
│ (ErrRec) │ │ (Transf) │
└──────────┘ └──────────┘
┌─────────────────────────┐
│ Observability (MCP) │
└─────────────────────────┘
Each stage is an autonomous agent with its own toolset, decision-making loop, and structured output contract.
🧠 Six Agents, Six Patterns
| Agent | Pattern | What It Shows |
|---|---|---|
| Capture | Tool-Use | Dynamic tool discovery, data aggregation |
| Analyze | Chain-of-Thought | Multi-step reasoning, pattern extraction |
| Write | Generation w/ Constraints | Controlled generation, template adherence |
| Judge | Self-Evaluation | Quality scoring, revision feedback loop |
| Adapt | Transformation | Content style transfer, multi-language |
| Publish | Error Recovery | Retry strategies, partial failure handling |
✨ What Makes This Different
Not another LangChain wrapper. This is a from-first-principles agent framework:
- 🔄 ReAct Loop — agents think-act-observe, not call-return
- 📋 JSON Schema Contracts — every agent<->agent handshake is validated
- ⚖️ Self-Evaluation Loop — Judge Agent rejects, Write Agent revises, loop until quality
- 🔧 Dynamic Tool Registry — agents discover available tools at runtime
- 🔭 Full Observability — every decision, every tool call, every token is traceable
- 📡 MCP Protocol — external systems can query pipeline state
- 🛡 Graceful Degradation — one platform failure doesn't crash the pipeline
🚀 Quick Start
# Install
git clone https://github.com/yourname/agent-daily-publisher
cd agent-daily-publisher
bash scripts/install.sh
# Dry run (no publishing)
python -m workflows.daily_pipeline
# Live publish
python -m workflows.daily_pipeline --publish
# Configure platforms
# Edit config/platforms.yaml (set enabled: true + API keys)
Prerequisites
- Python 3.12+
ANTHROPIC_API_KEYenvironment variable- Platform API keys (for publishing)
📦 Project Structure
agent-daily-publisher/
├── core/ # Agent framework (from scratch)
│ ├── agent.py # BaseAgent — ReAct loop
│ ├── pipeline.py # PipelineOrchestrator + JudgeLoop
│ ├── tool_registry.py # Dynamic tool discovery
│ ├── structured_output.py # JSON Schema enforcement
│ ├── memory.py # Hierarchical memory management
│ └── observer.py # Observability protocol
├── agents/ # 6 specialized agents
├── tools/ # Tools + platform connectors
├── mcp/ # MCP protocol server
├── workflows/ # Pipeline orchestration
├── templates/ # Platform-specific templates
├── config/ # YAML configuration
└── scripts/ # Install, hooks, cron
📊 Production Features
| Feature | Implementation |
|---|---|
| Error Recovery | Exponential backoff, 3 retries, graceful degradation |
| Observability | JSONL traces per agent call, tool call, pipeline run |
| Configuration | YAML-based, no code changes for platform config |
| Testing | Mock Claude client for dry runs |
| Scheduling | cron/launchd integration via daily-run.sh |
| Extensibility | Add new agents via BaseAgent subclass, new platforms via BasePublisher |
🛠 Platform Support
| Platform | Status | Required |
|---|---|---|
| Dev.to | ✅ Ready | DEVTO_API_KEY |
| 掘金 (Juejin) | ✅ Ready | JUEGIN_TOKEN |
| Medium | 🚧 Planned | MEDIUM_TOKEN |
| CSDN | 🚧 Planned | CSDN_TOKEN |
| 知乎 (Zhihu) | 🔮 Future | Cookie/API |
| 微信公众号 | 🔮 Future | WeChat MP AppID |
🔗 MCP Integration
Start the MCP server to query pipeline state from any MCP client:
python -m mcp.server
Then in Claude Code: "What articles were published today?"
🗺 Roadmap
- Phase 1 ✅ Core pipeline (Capture → Analyze → Write → Judge)
- Phase 2 ✅ Adapt + Publish agents, platform connectors
- Phase 3 🚧 MCP Server, full test coverage, CI/CD
- Phase 4 🔮 Human-in-the-loop review, A/B testing, analytics dashboard