junhanliu-dev/espalier-engineering

espalier 训练果树沿着墙壁平整生长——修剪、布线、高效。Espalier 对你的 AI 编码代理也做同样的事： 它读取你的代码库，发现其中已有的模式，并将其编码为机器强制的护栏。生成的代码第一次尝试就沿着你的约定生长，而不是第五次。

快速开始

就这些。每个仓库只需1~10到15分钟的设置，然后每次未来的更改都会重复使用。

问题所在

AI程序员写的代码很合理，但不适合你的代码库。他们发明你已有的助手，挑选你不用的库，throw，你的仓库多年前就在Result 上标准化。这不是模型智能的问题——而是不成文规则的问题。模型无法读取你团队的Slack历史记录。

它是如何运作的

1. /espalier-init 发现 — 平行侦察特工读取你的代码，提取没人写下的约定。
2. 它会写下它们——作为每个项目espalier/目录：始终加载规则、阶段加载技能、子代理定义、维基和程序钩子。

An espalier trains a fruit tree to grow flat along a wall — pruned, wired, productive. Espalier does the same for your AI coding agents: it reads your codebase, discovers the patterns already in it, and encodes them as machine-enforced guardrails. Generated code grows along your conventions on the first try, not the fifth.

Quick start

/plugin marketplace add Junhanliu-dev/espalier-engineering
/plugin install espalier-engineering@espalier-engineering
/espalier-init

That’s it. One ~10–15 minute setup per repo, then every future change reuses it.

The problem

AI coders write plausible code that doesn’t fit your codebase. They invent helpers you already have, pick libraries you don’t use, and throw where your repo standardised on Result years ago. It’s not a model-intelligence problem — it’s an unwritten rules problem. The model can’t read your team’s Slack history.

How it works

1. /espalier-init discovers — parallel scout agents read your code and extract the conventions nobody wrote down.
2. It writes them down — as a per-project espalier/ directory: always-loaded rules, phase-loaded skills, sub-agent definitions, a wiki, and programmatic hooks.
3. Every change is gated — a coder agent writes, a separate reviewer agent (different tool set) checks against the same rules, and hooks block pushes that skip the gates.
4. It stays honest — drift detection flags docs the code has outgrown; refreshes are gated, never silent.

espalier/
├── rules/        # always loaded: coding standards, security, production standards
├── skills/       # the pipelines: /espalier, /espalier-fix, /espalier-ask, …
├── agents/       # harness-coder / harness-reviewer / harness-security
├── wiki/         # architecture, data models, critical paths
├── hooks/        # programmatic gates: layer checks, pre-push, drift detect