长期用 Claude Code 或 Codex 写项目的人，真该把 Trellis 装上试试。

说它是现在最接近“让 AI 记住你项目“的方案，不算夸张。

很多人写着写着觉得 AI 越来越不靠谱，第一反应是去换更强的模型、或者把提示词堆得更长。但问题常常不在模型——而在它每次都是“空着脑子“进场：项目结构、命名规范、技术选型、上次做到哪，全得从头再喂一遍。

Trellis 直接把这事根治了。

它在你项目里建一个 .trellis/ 目录，把规范、任务、进度、踩过的坑都沉淀进去，跟代码一起进版本库。下次 AI 一进来先读这些，自己就知道该守什么规矩、做到哪一步、接下来干嘛。

相当于给项目配了一份“常驻记忆“，换谁来接手都不断片。

而且它不只是帮你塞上下文，是一整套开发工作流：先把需求问清楚，再动手写，写完对着你的规范和测试自查，最后把这次的经验写回项目——下次更聪明。

还有两点很多人没用上：复杂任务它能自己拆开、开几个分身并行干，互不打架；团队里一个人定的规范，全队的 AI 都跟着守，新人接手直接继承。它也不挑工具——Claude Code、Codex、Cursor 等 14 个 AI 编程平台都能装。

裸用 AI，像请了个聪明但每天失忆的临时工。 配上 Trellis，才开始有一支记得住、守规矩的开发团队的样子。

开源，目前 9000+ star、每周 5000+ 下载。 GitHub：http://github.com/mindfold-ai/Trellis…

mindfold-ai/Trellis

Source: https://github.com/mindfold-ai/Trellis

An out-of-the-box engineering framework for AI coding.
_{AI writes code fast, but every session it starts from scratch — no memory of your project, your conventions, or your team's requirements. Trellis persists specs, tasks, and memory into your repo, so any coding agent works to your engineering standards.}

简体中文 • Docs • Quick Start • Supported Platforms • Use Cases

Why Trellis?

Prerequisites:

• Node.js >= 18
• Python >= 3.9

Quick Start

# 1. Install Trellis
npm install -g @mindfoldhq/trellis@latest

# 2. Initialize in your repo
trellis init -u your-name

# 3. Or initialize with the platforms you actually use
trellis init --cursor --opencode --codex -u your-name

See the Quick Start and Supported Platforms guides for setup details.

How to Use

The workflow is simple:

1. Describe what you want in natural language.
2. Brainstorm with the AI one question at a time until the PRD is clear, then implementation begins.
3. Let it run — the AI calls Trellis Implement and auto-checks the result against specs, lint, type-check, and tests.
4. Type /trellis:finish-work when the work is done or the session context fills up. Trellis archives the task and updates journals.

How It Works

Trellis runs a 4-phase loop with auto-invoked skills and sub-agents:

1. Plan — trellis-brainstorm walks through requirements one question at a time and writes prd.md. Research-heavy items go to a trellis-research sub-agent. The result is curated specs + research files referenced from implement.jsonl / check.jsonl.
2. Implement — a trellis-implement sub-agent writes code from the PRD with the curated context auto-injected, no git commit.
3. Verify — a trellis-check sub-agent reviews the diff against specs and runs lint, type-check, and tests, self-fixing where it can.
4. Finish — a final check runs, then trellis-update-spec promotes new learnings back into .trellis/spec/ so the next session starts smarter.

Resources

FAQ

Star History

Community & Resources

• Official Docs
• GitHub Issues
• Discord
• Tech Blog

Official Repository • AGPL-3.0 License • Built by Mindfold