Pipeline Dual-Agent para Desenvolvimento Assistido por IA
BazooAlchemist/dual-agent-dev-pipeline
Metodologia + pipeline de dupla auditoria para coding com IA
Ver no GitHubMetodologia estruturada de 12 passos para desenvolvimento assistido por IA, com dupla auditoria cruzada entre agentes (Claude + Codex). Reduz erros típicos de LLMs através de gates de validação independentes e esquemas de risco em três níveis.
Instalação
npx convez-ai BazooAlchemist/dual-agent-dev-pipeline1 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
- Estruturar workflows de desenvolvimento com assistentes de IA
- Implementar auditoria cruzada entre modelos de linguagem
- Reduzir erros de diagnóstico e over-engineering em código gerado
Veredito

Projeto maduro e bem documentado que formaliza boas práticas para desenvolvimento com assistentes de IA. Excelente para equipes que buscam rigor e reprodutibilidade em pipelines de coding assistido, mas requer investimento em compreensão da metodologia antes do deployment.
README
A tradução para o português deste README ainda está na fila.
Dual-Agent Dev Pipeline
A field-tested AI-assisted development pipeline methodology + Codex bridge reference deployment for Claude Code CLI.
Bilingual Summary / 中英双语摘要
EN: This repository provides a structured development pipeline designed specifically for AI-assisted coding workflows. It contains methodology, adapter guides, annotated configuration examples, canonical audit schemas, and a Codex bridge reference deployment for cross-model audit. The pipeline reduces the class of errors LLMs are most prone to — skipping diagnosis, over-engineering, ignoring the root cause — by enforcing 12 steps guarded by two independent audit gates. Each step is designed to catch one category of error; skipping one means accepting that category.
CN: 本仓库提供一套专门为 AI 辅助编码设计的分阶段开发流水线:核心方法论(与工具无关的通用规则)、适配器指南(具体工具的能力映射)、带注释的配置示例、canonical audit schema、以及基于 Codex 的双模型互审参考部署。该流水线通过 12 个步骤 + 两道独立审计闸门,系统性降低 LLM 最常见的错误类型——跳过诊断、过度设计、忽视根因。每步防一类错误,跳一步 = 接受该类错误。
Repository Structure
dual-agent-dev-pipeline/
|
|-- README.md Entry point (this file)
|-- AGENTS.md Repo-local agent workflow rules
|-- POLISHING.md Repository polishing plan and next steps
|-- PHILOSOPHY.md Design philosophy & why this pipeline exists
|-- QUICKSTART.md 5-step minimal onboarding (~30 min)
|
|-- core/ Methodology — tool-agnostic, portable to any AI coding tool
| |-- 01-pipeline-overview.md 12-step pipeline panorama & decision flow
| |-- 02-pipeline-steps.md Detailed step walkthroughs, rationale, common mistakes
| |-- 03-risk-tiers.md Low / Mid / High risk tier definitions & escalation rules
| |-- 04-dual-agent-audit.md Dual-agent cross-audit protocol (capability-trigger based)
| |-- 05-three-layer-architecture.md 3-layer architecture: Method / Behavior / Process Gates
|
|-- adapters/ Tool mappings — how capabilities map to concrete tools
| |-- tool-mapping-table.md Capability -> reference tool -> alternative mapping table
| |-- claude-code.md Onboarding guide for Claude Code CLI
| |-- codex-auditor.md Interface definition for Codex as external audit agent
|
|-- examples/ Annotated config examples — copy, edit, deploy
| |-- CLAUDE.md.example
| |-- dev-pipeline-skill.md
| |-- governance-dual-audit.md
| |-- settings.json.example
|
|-- deploy/ Runtime reference deployment helpers ([reference] — tested on 1 env)
| |-- codex-bridge/
| |-- README.md Manual deployment steps
| |-- setup-codex-bridge.sh [reference] Setup script
| |-- health-check.sh [reference] Health check script
| |-- codex-bridge-launchd.plist launchd plist template
|
|-- appendix/ Non-normative appendices
| |-- real-world-cases.md Anonymized real-world case studies (3 sub-cases)
| |-- file-trigger-examples.md File-based trigger examples (illustrative only)
|-- schema/ Canonical JSON schemas for audit bridge messages
| |-- audit-request.schema.json
| |-- audit-response.schema.json
|
|-- scripts/
| |-- validate-repo.sh Repository validation checks
Three Entry Paths / 三条入口
Choose the path that matches your goal:
Entry A — "I want to understand the methodology"
README -> PHILOSOPHY.md -> core/01-pipeline-overview.md -> core/04-dual-agent-audit.md
Ideal for architects and team leads evaluating whether this approach fits their workflow.
Entry B — "I want the minimal working setup in ~30 minutes"
README -> QUICKSTART.md -> examples/CLAUDE.md.example -> examples/governance-dual-audit.md
Ideal for individual developers ready to adopt the pipeline immediately.
Entry C — "I want to deploy the Codex bridge reference deployment"
Entry B first -> deploy/codex-bridge/README.md -> (optional) deploy/codex-bridge/*.sh
Adds the cross-model audit agent for automated Gate audits.
Prerequisites
- macOS (launchd-based service management)
- Claude Code CLI (for pipeline execution)
- Node.js >= 18 (for Codex bridge)
- git (for version control)
- (Optional) Codex CLI — for dual-agent audit setup via Entry C
Quick Start
See QUICKSTART.md for a 5-step minimal onboarding path (~30 minutes target time, actual time may vary).
License
Boundary Statement / 边界声明
In scope:
- AI-assisted development pipeline methodology for Claude Code CLI + Codex
- Risk tier classification and dual-agent cross-audit protocol
- Annotated configuration templates (CLAUDE.md, settings.json, governance rules, dev-pipeline skill)
- Codex bridge reference deployment for macOS
- Canonical request/response schema for external audit messages
Out of scope / 不解决什么:
- This is not a general-purpose CI/CD pipeline. It is specifically designed for AI-assisted code generation workflows.
- The deploy scripts are reference deployment helpers tested on a single environment. They assume you provide or adapt the actual bridge worker.
- The pipeline does not replace security audits, unit test coverage requirements, or manual code reviews.
- Non-macOS platform deployment is not covered. Windows and Linux users will need to adapt the launchd plist and shell scripts to their platform equivalents.
- The methodology does not guarantee zero bugs. It reduces the probability of LLM-pattern errors through systematic gating.
Return Path
Questions, issues, or contributions? Open an issue or pull request in this repository.