插件构建工具

一个以评估为先的Claude代码插件构建工具,通过原始决策记录设计插件,在编写代码前生成可运行的评估套件,并在构建过程中持续验证直到套件通过。

  • by Aznatkoiny
  • Python
  • 本地部署
  • 更新于
在 GitHub 查看 Aznatkoiny/claude-dev-toolkit

使用说明

claude-dev-toolkit

Development home of plugin-forge — the eval-first plugin builder for Claude Code.

The eval suite is the contract; green is the definition of done.

plugin-forge replaced this repo's earlier toolkit (reference skills + builder agents + scaffold commands). Where the old toolkit scaffolded files and hoped, plugin-forge designs plugins through Primitive Decision Records, generates a runnable evaluation suite before any code exists, freezes it, builds under a headless goal loop until the suite is green, and verifies against a holdout tranche the builder never saw. The legacy components live on in git history.

Repository layout

Path What it is
plugins/plugin-forge/ The plugin — see its README for the full pipeline
docs/plans/ The design record: research provenance, synthesis decisions D1–D10, binding contracts
claude-code-ref/ Official Claude Code documentation snapshots used as research ground truth
.claude-plugin/marketplace.json Marketplace manifest (this repo installs as a marketplace)

Install

Released standalone at Aznatkoiny/plugin-forge (recommended):

/plugin marketplace add Aznatkoiny/plugin-forge
/plugin install plugin-forge@plugin-forge

Also available via zAI-Skills: /plugin install plugin-forge@zAI-Skills.

From this repo (dev):

claude --plugin-dir ./plugins/plugin-forge

Develop

# hot-reload loop
claude --plugin-dir ./plugins/plugin-forge     # then /reload-plugins after edits

# validation gate
claude plugin validate ./plugins/plugin-forge --strict

# dogfood regression suite — deterministic, zero tokens, seconds to run
./plugins/plugin-forge/bin/forge-eval run --suite regression \
  --plugin-dir ./plugins/plugin-forge --evals-dir ./plugins/plugin-forge/evals

Releases are cut by syncing plugins/plugin-forge/ to the standalone repo and tagging (see the v0.1.0 release).

License

MIT