VASMC

🌍 English | 🇨🇳 中文

---

🌍 English

Decentralized Markdown prompt compiler for LLM skills: URL imports, deterministic builds, cross-language outputs, and AI build work orders.

In the LLM era, Markdown has become source code: system prompts, skill packs, and instruction sets are all written in Markdown. VASMC treats them as first-class compilation targets by resolving @import dependencies, cross-compiling language blocks, and generating AI build work orders for semantic tasks that belong to the current AI rather than a deterministic tool.

In AI-native systems, the context window is both execution space and data space. The model cannot reliably separate instructions from data after they enter the prompt. VASMC establishes a deterministic control point before that execution surface exists: every token entering a system prompt is explicitly declared, assembled, traceable to a human-authorized source, and classified by output format.

👉 Read the full design spec

📦 Core Features

• Input surface sovereignty: Compile-time control over content sources. Every token entering the execution surface has traceable human authorization, and prompt / doc format classification acts as a compile-time safety primitive.
• Decentralized package management: Fetch Markdown modules directly from URLs. No registry, no intermediary.
• Deterministic builds: SHA-256 lock files (vasmc-lock.yaml) make builds reproducible.
• Cross-compilation: AST-level language block filtering (<!-- lang:xx -->) generates language-specific outputs. Missing language coverage is handed to the current AI through vasmc build work orders.
• Two import modes:
  • @import:link rewrites aliases to local relative paths while preserving hyperlink structure.
  • @import:inline expands remote content inline for assembling large prompt contexts.

🛠️ Packages And Commands

Package	Command	Description
@vasm/core	No bin	Shared deterministic compiler core and VASM protocol implementation
@vasm/cli	vasmc	AI build, dependency management, and follow-up work orders
@vasm/console	vasm-console	Human-facing console with optional external-model lint / diff tools

📌 Versioning And Release

VASMC uses Changesets to manage npm workspace versions. @vasm/core, @vasm/cli, and @vasm/console are currently released as a fixed version group, so they always share the same version. This matches the publish model: vasmc and vasm-console bundle core into their command binaries, so core behavior changes usually affect both command artifacts.

Before publishing, run npm run changeset to declare the affected packages and SemVer bump, then run npm run release:version to write package versions and changelogs. Run npm run release:check before publishing to verify tests, builds, self-compilation, and npm pack dry-run. Use npm run release:publish for npm publishing. After publishing succeeds, run npm run release:github to verify package tags, create the vX.Y.Z aggregate tag, push the current branch and tags, and create a GitHub Release with GitHub CLI.

🚀 Quick Start

npm install -g @vasm/cli
vasmc init                    # create vasmc-build.yaml
vasmc add https://example.com/skill.md --alias my-skill
vasmc build                   # build the workspace and generate follow-up work orders for the current AI

For full CLI usage, see Help And Usage.

---

🇨🇳 中文

去中心化的 LLM Prompt 编译器 — 模块化导入、交叉编译、AI 原生的编译编排，以及输入面主权保障。

在大语言模型时代，Markdown 已经演变为源代码 — System Prompt、技能包、指令集全部用 Markdown 编写。VASMC 将它们视为一等编译目标：解析 @import 依赖、跨语种交叉编译，并编排 AI 编辑器处理那些属于智能而非工具的语义任务。

AI-Native 系统中，整个上下文窗口既是执行空间也是数据空间——LLM 无法区分指令与数据。VASMC 在执行面形成之前建立确定性的控制点：所有进入 system prompt 的内容，都经过编译链的显式声明与组装，来源可追溯，格式有分类。

👉 阅读完整的设计规范

📦 核心特性

• 输入面主权: 编译期内容来源管控——所有进入执行面的 token 都有可追溯的人类授权来源，prompt / doc 格式分类是编译时的安全分类原语。
• 去中心化包管理: 直接通过 URL 拉取 Markdown 模块 — 无注册表，无中间人。
• 确定性构建: SHA-256 锁文件 (vasmc-lock.yaml) 确保构建可复现。
• 交叉编译: AST 级语言块过滤 (<!-- lang:xx -->) 生成各语种产物；未覆盖语种由 vasmc build 工作单交给当前 AI 处理。
• 双模式引入:
  • @import:link — 别名重写为本地相对路径（保留超链接结构）。
  • @import:inline — 内联展开远程内容（组装大型 Prompt 上下文）。

🛠️ 包与命令边界

包	命令	说明
@vasm/core	无 bin	共享确定性编译核心与 VASM 协议实现
@vasm/cli	vasmc	AI build、依赖管理、后续工作单
@vasm/console	vasm-console	人用控制台，包含可选外部模型 lint/diff

📌 版本与发布

VASMC 使用 Changesets 管理 npm workspace 版本。@vasm/core、@vasm/cli、@vasm/console 当前作为 fixed version group 发布：三者始终保持同一版本号。原因是 vasmc 与 vasm-console 的发布产物都会内置 core，core 行为变化通常也意味着两个命令产物的语义变化。

发布前先用 npm run changeset 声明变更影响和 SemVer 级别，再用 npm run release:version 写入版本号与 changelog。正式发布前执行 npm run release:check，确认测试、构建、自举编译和 npm pack dry-run 均通过。npm 发布使用 npm run release:publish。发布成功后执行 npm run release:github：它会检查 package tags、创建 vX.Y.Z 聚合 tag、push 当前分支与 tags，并通过 GitHub CLI 创建 GitHub Release。

🚀 快速上手

npm install -g @vasm/cli
vasmc init                    # 生成 vasmc-build.yaml
vasmc add https://example.com/skill.md --alias my-skill
vasmc build                   # 编译工作区，并为当前 AI 生成语义工作单

完整 CLI 用法请参阅 帮助与用法文档。