explaingit

lonelyherbivore/trellis-herbivore

23TypeScriptAudience · developerComplexity · 3/5ActiveLicenseSetup · moderate

TLDR

A fork of the Trellis AI coding workflow tuned for Claude Code, with Chinese docs and explicit gates for planning, strategy decisions, and multi-layer reviews.

Mindmap

mindmap
  root((Trellis-Herbivore))
    Inputs
      Natural language request
      Project spec files
      Task PRD and design
    Outputs
      Task folder artefacts
      Implementation plan
      Review reports
    Use Cases
      Plan before coding with Claude
      Enforce review gates
      Capture decisions in repo
    Tech Stack
      TypeScript
      Node 18
      Python 3.9
      npm

Things people build with this

USE CASE 1

Set up a planning-first Claude Code workflow for a new project

USE CASE 2

Force explicit subagent and worktree choices before coding starts

USE CASE 3

Add staged code, spec, and architecture reviews to a team repo

USE CASE 4

Keep AI coding context in repo files instead of long chat threads

Tech stack

TypeScriptNode.jsPythonnpm

Getting it running

Difficulty · moderate Time to first run · 30min

Needs Node 18 plus Python 3.9 and a global npm install, then a per-project trellis init before the workflow is usable.

AGPL-3.0, you can use and modify the code but derivative works and networked services must also be open-sourced under AGPL.

In plain English

Trellis-Herbivore is a customised fork of an existing AI coding tool called Trellis, tuned specifically for use with Claude Code, the command-line coding assistant. The original Trellis is described as a team AI coding harness with its own ideas about specs, tasks, and workflow. This branch keeps all of that intact and adds extra structure on top, with documentation written in Chinese so a team can read and review the artefacts directly. The core idea is that important context should live in repository files, not in a chat window that grows longer and longer. Project rules go into a spec folder, and each task gets its own folder with a product requirements document, a design file, an implementation plan, and small JSON files that load context for the implementation and check steps. When a user asks for something in plain language, the workflow first turns the request into a task, then runs a brainstorm step and a so-called grill step to challenge the requirements, then forces explicit decisions before any code is written. Before implementation starts, the developer has to pick three things: whether the current chat session writes the code or a subagent does, whether changes happen on the current branch or in a separate worktree, and whether the work follows the default Trellis flow or test-driven development. The aim is to make these choices visible in the task documents instead of being implicit. Quality control is split into layers. After implementation there is a general check, then separate spec, code, and architecture reviews, an optional deeper architectural review for sensitive tasks, a post-merge review, and a final build and test. Installation is one npm command for Node 18 or newer with Python 3.9 or newer. The project is released under the AGPL-3.0 license.

Copy-paste prompts

Prompt 1
Walk me through running trellis init in my Node 18 project and what files it creates
Prompt 2
I want to add a new feature, take me through trellis-brainstorm and trellis-grill-me step by step
Prompt 3
Show me how to configure the implementation strategy as subagent plus worktree for my next task
Prompt 4
Run trellis-check, trellis-spec-review, trellis-code-review on the task in .trellis/tasks/latest and summarise findings
Prompt 5
Explain the difference between trellis-code-architecture-review and trellis-improve-codebase-architecture deep-review
Open on GitHub → Explain another repo

Generated 2026-05-22 · Model: sonnet-4-6 · Verify against the repo before relying on details.