GStack for OpenCode 🚀

The ultimate multi-agent AI programming team, powered by OpenCode.
Porting Garry Tan's GStack methodology with "Best-of-Breed" model selection and advanced tool-use capabilities.

---

English | 中文

🌐 Overview

GStack for OpenCode transforms your AI from a single assistant into a Dynamic Expert Team. By leveraging OpenCode's agentic framework, it assigns specialized roles to the most suitable LLMs (Claude, Gemini, DeepSeek, Kimi), creating a workflow that outperforms single-model tools like Claude Code in complex, long-term engineering tasks.

✨ Key Features

• 🛡️ Multi-Model Mastery: Use Claude 3.5 Sonnet for logic, Gemini 1.5 Pro for architecture (2M context), and DeepSeek/Kimi for cost-effective reasoning.
• ⚔️ Advanced Tool-Use: Agents are unlocked with bash (shell execution), webfetch (real-time research), and edit (file modification) capabilities.
• 🧩 Zero Context Pollution: Each role runs in an isolated agent session, ensuring critical architectural decisions aren't diluted by implementation details.
• 🇨🇳 Domestic Ready: First-class support for DeepSeek, Kimi, and Zhipu GLM via OpenAI-compatible endpoints.

🏗️ Workflow Architecture

graph TD
    A[CEO /plan-ceo] -->|Verify Value| B[Eng Manager /plan-eng]
    B -->|Design Architecture| C[Developer Default]
    C -->|Implementation| D[Reviewer /review]
    D -->|Security Audit| E[Security /security]
    E -->|UI/UX Review| F[Design /design]
    F -->|Validation| G[QA /qa]
    G -->|Documentation| H[Docs /docs]
    H -->|Preparation| I[Ship /ship]
    I -->|Release| J(Success)

Loading

⚖️ Why GStack? (Pros & Cons)

🚀 What Problems Does It Solve?

1. Context Pollution: Single-model sessions often get "confused" by previous implementation details. GStack isolates high-level architecture from low-level coding.
2. High Costs: Stop using expensive Claude 3.5 for simple unit tests. Offload tasks to DeepSeek or GPT-4o mini.
3. Vendor Lock-in: Claude Code is locked to Anthropic. GStack gives you the freedom to swap models mid-flow.
4. Domestic Barrier: Native support for Chinese developers using DeepSeek, Kimi, and GLM without VPN hurdles.

✅ Pros

• Specialized Intelligence: Each role uses the model that excels at that specific task (e.g., Gemini for long-context architecture).
• Autonomous Tool-Use: Agents can now run bash to explore your codebase and webfetch to research external documentation.
• Structured Workflow: Forces a "Think-Design-Code-Verify" cycle, significantly reducing technical debt.
• Open Source & Extensible: Easily add your own roles (e.g., DBA, DevOps) via Markdown templates.

⚠️ Cons (Current Limitations)

• Setup Overhead: Requires manual environment variable configuration for domestic models (due to OpenCode CLI bugs).
• Dependency: Performance is tied to the stability of the OpenCode CLI and chosen model's tool-calling capability.
• Learning Curve: Users need to understand the "Role-playing" concept to get the best results.

🎭 The Expert Team

Role	Command	Tool-set	Core Mission
CEO	/plan-ceo	🌐 Web	Product vision, MVP scope, and business value verification.
Eng Manager	/plan-eng	💻 Bash, 🌐 Web	Technical architecture, system boundaries, and tech-stack selection.
Reviewer	/review	💻 Bash, 📝 Edit	Critical bug hunting, performance optimization, and logic review.
Security	/security	💻 Bash, 🌐 Web	Vulnerability scanning, privacy audit, and security hardening.
Designer	/design	🌐 Web	UI/UX consistency, accessibility (A11y), and user flow optimization.
QA	/qa	💻 Bash, 📝 Edit	Edge-case testing, test generation, and automated bug fixing.
Docs	/docs	📝 Edit, 🌐 Web	README auditing, API documentation, and technical writing.
Ship	/ship	💻 Bash	Release checklists, CHANGELOG generation, and rollback planning.

🚀 Quick Start

1. Installation

# Clone the repository
git clone https://github.com/yandong2023/gstack-opencode.git && cd gstack-opencode

# Install agents, commands and skills
make install

# Validate your environment
make validate

2. Domestic Model Configuration (DeepSeek/Kimi)

If you are using domestic models, we recommend the Environment Variable Workaround to bypass CLI configuration bugs:

# Add to your .zshrc or .bashrc
export OPENAI_API_KEY="your-deepseek-api-key"
export OPENAI_BASE_URL="https://api.deepseek.com/v1"

3. Usage

Enter your project directory and start the magic:

opencode
/plan-ceo "I want to add a real-time notification system."

---

🌐 概述

GStack for OpenCode 将你的 AI 从单一助手转变为一支 动态专家团队。通过利用 OpenCode 的 Agent 框架，它将专门的角色分配给最适合的大模型（Claude, Gemini, DeepSeek, Kimi），创造出在处理复杂工程任务时超越 Claude Code 等单一模型工具的工作流。

✨ 核心特性

• 🛡️ 多模型协作：让 Claude 3.5 Sonnet 负责逻辑，Gemini 1.5 Pro 负责架构（200万上下文），DeepSeek/Kimi 负责高性价比推理。
• ⚔️ 深度工具集成：专家角色已解锁 bash（命令行）、webfetch（联网研究）和 edit（文件修改）权限。
• 🧩 零上下文污染：每个角色运行在独立的 Agent 会话中，确保架构决策不会被琐碎的实现细节干扰。
• 🇨🇳 国产模型优化：原生支持 DeepSeek、Kimi、智谱 GLM 等国内顶尖模型。

🎭 专家团队介绍

• CEO (/plan-ceo): 负责“重构问题本质”。不再直接讨论怎么写，而是讨论“值不值得做”。
• 工程经理 (/plan-eng): 负责“定义边界”。通过 bash 探索代码库，设计不带偏见的技术方案。
• QA 工程师 (/qa): 负责“寻找失败”。通过 bash 运行测试，并使用 edit 权限自动修复 Bug。
• 安全专家 (/security): 负责“红队演练”。审计敏感数据泄露和注入风险。

💡 最佳实践

1. 先沟通需求：永远从 /plan-ceo 开始，AI 会帮你砍掉 30% 不必要的复杂性。
2. 给 QA 权限：当 QA 要求运行 npm test 时，点击允许，它能自己找到并修复报错。
3. 大上下文优势：如果你有 Gemini 1.5 Pro，建议将 /plan-eng 的模型设为 Gemini，它能“瞬间”理解你的整个代码库。

🤝 贡献

我们欢迎任何形式的贡献！请参阅 CONTRIBUTING.md。

📜 许可证

本项目采用 MIT 许可证。

---

🙏 致谢

• Garry Tan - GStack 方法论创始人
• OpenCode - 开源 AI 编程利器