mindfold-ai/Trellis

What it solves

AI coding agents often start every session from scratch, lacking memory of project-specific conventions, team requirements, and previous progress. Trellis provides a persistent engineering framework that stores specifications, tasks, and memory directly in the repository, ensuring AI agents follow established engineering standards across different sessions and platforms.

How it works

Trellis implements a four-phase loop using specialized sub-agents:

1. Plan: A brainstorming agent (trellis-brainstorm) clarifies requirements and creates a PRD (Product Requirements Document), while a research agent (trellis-research) handles complex items.
2. Implement: An implementation agent (trellis-implement) writes the code based on the PRD and injected context.
3. Verify: A verification agent (trellis-check) reviews the diff against specifications and runs linting, type-checking, and tests, fixing errors automatically.
4. Finish: The system updates specifications (trellis-update-spec) to promote new learnings back into the project's shared rules, archiving the task.

Who it’s for

It is designed for solo developers and engineering teams who use AI coding agents and want to maintain consistent standards, shared project memory, and a structured workflow across multiple AI platforms.

Highlights

• Auto-injected specs: Automatically provides relevant project conventions to AI agents.
• Task-centered workflow: Keeps PRDs and task status organized within the repo.
• Project memory: Uses workspace journals to preserve context between sessions.
• Multi-platform support: Works across 22 different AI coding platforms.
• Ablation testing: Allows users to temporarily remove Trellis to compare project state with and without the framework.