Trellis

AI has strong coding capabilities, but if your repository's codebase keeps growing and you keep doing pure vibe coding, your project will almost certainly become increasingly messy and hard to maintain and manage.

So why don't human engineers run into this problem? I thought about my own development workflow, and it goes roughly like this:

1. Receive a requirement, and clarify exactly what needs to be done
2. My brain automatically does a RAG lookup, recalling relevant reusable code kits, development standards, where they're stored, and how to use them
3. Start coding
4. Review my code against the standards I remember for how this type of requirement should be implemented in the project
5. Commit and push

So why can't AI do this? On one hand, people who do vibe coding might be complete beginners with no coding foundation, so they don't even know what standardized coding looks like. On the other hand, engineers swamped with requirements and tight schedules don't have time to run a full standardized workflow for every single requirement (if they did, they might as well write the code themselves).

All in all, to solve this problem, we built an AI workflow internally that mimics human engineers' development patterns. We codify all code structure standards and development guidelines into documentation, and also log commit-related information into these docs. Before every conversation with AI, we only inject the relevant context it needs via the corresponding workflow. This way, before actual coding starts, you don't get a pre-trained agent with strong coding skills that randomizes its output, but a 'code engineer' that's familiar with your project's recent commits and knows exactly how to implement the current requirement's related code.

After coding is done, there's also a forced context injection process for code review, to prevent forgetting coding standards due to excessive context during coding. After work is completed, there's a dedicated process to check if the coding produced or modified any existing development standards, and update the standards accordingly, so the framework gets better the more you use it.

In short, we've open-sourced this project, named Trellis. AI capabilities are like ivy: strong and tenacious, but it grows wildly. We hope our framework acts like a trellis, guiding it to follow standards as it grows.

Original repo: https://github.com/mindfold-ai/Trellis

Short README: https://github.com/mindfold-ai/Trellis/blob/main/README_CN.md

Detailed concept introduction: https://github.com/mindfold-ai/Trellis/blob/main/docs/guide-zh.md

(If you know k8s, check this doc to better understand our concept: https://github.com/mindfold-ai/Trellis/blob/main/docs/use-k8s-to-know-trellis-zh.md)

If you find this useful, please give us a star ~