---
title: "Four layers"
description: "「Four layers」：yokoten 使用指南的一章。"
canonical: https://yokoten.app/zh-tw/docs/guide/four-layers
lang: en
source: https://github.com/yokoten-labs/yokoten/blob/v0.1.2/docs/guide.md
ref: "v0.1.2"
---

> For AI agents: the documentation index is at https://yokoten.app/llms.txt. To set up yokoten in a project, follow https://yokoten.app/agents.md.

> 這一頁還沒有繁體中文版，下面是英文原文。

# Four layers

| Layer | What | How |
|-------|------|-----|
| **Machine** | Skills, agents, and two user-level hooks, the same on every computer | `yokoten install`, or the Claude Code plugin |
| **Project** | Rules, the commit check, the tool-call hook, CI, issue and PR templates, Dependabot (monthly, one pull request), a security policy | `yokoten init`, then the `yokoten` skill for an old project |
| **Server** | Branch and tag rulesets, squash merge, code owners, a bot with limited rights, the security switches | Set once by the owner; `doctor` lists the steps |
| **Rules** | `AGENTS.md`: twelve core rules, plus the project's own | Every rule names its check |

An old project isn't a blank slate. Some of its habits are better than ours. So the
`yokoten` skill goes through the practices one at a time with you, and each one gets a
decision: **as-is**, **lighter** (and what's left out), or **not now** (with the reason
and when to look again). "Not now" goes into `yokoten.json`, and `doctor` stops asking.

## A first draft of the table

`yokoten adopt` drafts that table before the talk starts. It reads the project's files
and says what the project does today for each step: **done**, **missing**, **partial**
or **unknown**. Partial means the project has its own way, such as husky hooks, or a CI
that doesn't run yokoten. Unknown means yokoten can't check the step from here, such as
a server setting; `yokoten doctor` reads it. A model then proposes a decision and a one-sentence reason
for each row.

- The model gets only those short facts. It never gets file contents.
- yokoten runs its privacy check on the request first. On a hit, it sends nothing.
- With no key, a failed call or a bad answer, rules draft the table instead: done or
  missing is as-is, partial is lighter, unknown is as-is once `doctor` has checked it.
  The output says which one drafted it.
- `adopt` never writes a file. You still decide every row.
- A partial row is not lost when you run `init`: it imports `AGENTS.md` from your
  `CLAUDE.md`, and it adds the commit check to husky's `pre-commit`.

The model runs on Meta's model API, through its OpenAI-compatible endpoint
(`https://api.meta.ai/v1`, model `muse-spark-1.3`). Set `LLAMA_API_KEY` or `MODEL_API_KEY`
to your key. `LLAMA_API_BASE` and `LLAMA_MODEL` point it at any other OpenAI-compatible
endpoint and model.

---

上一章: [What a session sees](https://yokoten.app/zh-tw/docs/guide/what-a-session-sees.md)
下一章: [Works with](https://yokoten.app/zh-tw/docs/guide/works-with.md)