llm-code-format

From source code to Markdown to LLMs and back again

A TypeScript library for parsing and serializing multiple code files in Markdown format, specifically designed for Large Language Model (LLM) interactions. The idea of this comes from the fact that when you ask an LLM to write code for you with multiple files, different models have different "preferences" when it comes to representing code. This library aims to provide a consistent way to parse and serialize code files in Markdown format, regardless of the format used by the LLM. It may also be useful for other applications that need to work with multiple code files in Markdown format, such as being able to extract code files from a Markdown document or generate a Markdown document from code files.

See also editcodewithai, a higher level code editing tool built on llm-code-format.

Features

• Parse code blocks from various Markdown formats including:

  • Bold Format (**filename.js**)
  • Backtick-Heading Format (### filename.js``)
  • Standard Heading Format (### filename.js)
  • Colon Format (filename.js:)
  • Hash Format (# filename.js)
  • File Bold Format (### File: **filename.js**)
  • Heading Bold Format (### **filename.js**)
  • Numbered Bold Format (1. **filename.js**)
  • Numbered Backtick Format (### 1. filename.js``)

• Streaming parser for real-time code editing views

• Serialize code files back to Markdown in a consistent format

• TypeScript support with full type definitions

• Zero dependencies

• Comprehensive test coverage

Installation

npm install llm-code-format

Usage Examples

Basic Example Usage

import { parseMarkdownFiles, formatMarkdownFiles } from "llm-code-format";

// Parse Markdown content containing code blocks
const markdownString = `
**index.html**

\`\`\`html
<h1>Hello World</h1>
\`\`\`

**styles.css**

\`\`\`css
body { color: blue; }
\`\`\`
`;

const { files, format } = parseMarkdownFiles(markdownString);
console.log(format); // "Bold Format"
console.log(files); // Object with filenames as keys and contents as values

// Serialize files back to Markdown
const files = {
  "index.html": "<h1>Hello World</h1>",
  "styles.css": "body { color: blue; }",
};

const markdown = formatMarkdownFiles(files);

Recipe: Extract Files from Blog Posts

This "recipe" shows how to extract example code from blog posts such that they can be run (works with Astro):

import { parseMarkdownFiles } from "llm-code-format";
import fs from "fs";
import path from "path";

const blogDir = "src/pages/blog";
const publicDir = "public/examples";

// Read all markdown files from blog directory
fs.readdirSync(blogDir)
  .filter((fileName) => fileName.endsWith(".md"))
  .forEach((fileName) => {
    const postName = fileName.replace(".md", "");
    const filePath = path.join(blogDir, fileName);
    const markdownString = fs.readFileSync(filePath, "utf8");

    // Use llm-code-format to extract code blocks from markdown!
    const { files } = parseMarkdownFiles(markdownString);

    if (Object.keys(files).length === 0) return;
    const outputDirectory = path.join(publicDir, postName);
    if (fs.existsSync(outputDirectory)) {
      fs.rmSync(outputDirectory, { recursive: true });
    }
    fs.mkdirSync(outputDirectory);
    Object.entries(files).forEach(([name, content]) => {
      fs.writeFileSync(path.join(outputDirectory, name), content);
    });
  });

API

parseMarkdownFiles(markdownString: string, format?: string)

Parses a Markdown string containing code blocks and returns an object with:

• files: Object with filenames as keys and file contents as values
• format: String indicating the detected format

Optional format parameter to specify a particular format to parse:

• 'backtick-heading'
• 'file-bold'
• 'numbered-backtick'
• 'heading-bold'
• 'standard-heading'
• 'colon'
• 'bold'
• 'hash'
• 'numbered-bold'

formatMarkdownFiles(files: FileCollection)

Converts a FileCollection object into a Markdown string using the Bold Format.

StreamingMarkdownParser

A streaming parser that processes Markdown content chunk by chunk, emitting callbacks when file names or code blocks are detected. This is particularly useful for real-time code editing views where users can see LLMs edit specific files as they generate content.

import {
  StreamingMarkdownParser,
  StreamingParserCallbacks,
} from "llm-code-format";

// Define callbacks for file name changes, code lines, and non-code lines
const callbacks: StreamingParserCallbacks = {
  onFileNameChange: async (fileName, format) => {
    console.log(`File changed to: ${fileName} (${format})`);
    // Can now perform async operations like database saves
    await saveFileMetadata(fileName, format);
  },
  onCodeLine: async (line) => {
    console.log(`Code line: ${line}`);
    // Can format code asynchronously
    await formatAndSaveCodeLine(line);
  },
  onNonCodeLine: async (line) => {
    console.log(`Comment/text: ${line}`);
    // Can process documentation asynchronously
    await processDocumentationLine(line);
  },
  onFileDelete: async (fileName) => {
    console.log(`File marked for deletion: ${fileName}`);
    // Handle file deletion when empty code blocks are detected
    await deleteFile(fileName);
  },
};

// Create a new parser instance with the callbacks
const parser = new StreamingMarkdownParser(callbacks);

// Process chunks as they arrive (e.g., from an LLM streaming response)
await parser.processChunk("**index.html**\n");
await parser.processChunk("```html\n");
await parser.processChunk("<h1>Hello World</h1>\n");
await parser.processChunk("</html>\n");
await parser.processChunk("```\n");

// Flush any remaining content when the stream ends
await parser.flushRemaining();

Handling File Deletion

The streaming parser can detect when files should be deleted by recognizing empty code blocks. When a file header is followed by an empty code block (or one containing only whitespace), the onFileDelete callback is triggered:

**fileToDelete.js**

```
```

**anotherFileToDelete.css**

```

```

This pattern signals that these files should be removed from your project.

Methods

• constructor(callbacks: StreamingParserCallbacks): Creates a new parser instance with the specified callbacks.
• processChunk(chunk: string): Processes a chunk of text from the stream. Returns a Promise that resolves when all callbacks have completed.
• flushRemaining(): Processes any remaining content in the buffer after the stream has ended. Returns a Promise that resolves when all callbacks have completed.

Callback Interface

type StreamingParserCallbacks = {
  onFileNameChange: (fileName: string, format: string) => Promise<void>;
  onCodeLine: (line: string) => Promise<void>;
  onNonCodeLine?: (line: string) => Promise<void>;
};

• onFileNameChange: Called when a file header is detected outside a code fence. Returns a Promise that should resolve when processing is complete.
• onCodeLine: Called for each line emitted from inside a code fence. Returns a Promise that should resolve when processing is complete.
• onNonCodeLine: Called for each line outside code fences that is not a file header. Returns a Promise that should resolve when processing is complete.

Currently, the streaming parser only supports the "Bold Format" (**filename.js**), but is designed to be extensible for supporting additional formats in the future.

Contributing

Contributions welcome! Please read the CONTRIBUTING.md guidelines before submitting a PR. All changes should start with creating an issue first.