PatBat - Constrained Code Generation

A tool for constrained code generation with integrated validation, powered by AI. Use it as a library in your TypeScript/JavaScript projects or directly from the command line.

Prerequisites

• Bun Runtime
• Environment Variables:
  • For Anthropic: ANTHROPIC_API_KEY
  • For OpenAI: OPENAI_API_KEY

Installation

bunx patbat-codegen

Usage

SDK Mode

import { AITransformData } from "patbat-codegen";

const result = await AITransformData(
  {
    // Core inputs
    problem: "transform data to get average age and oldest person",
    inputSchema: "z.object({ name: z.string(), age: z.number() })",
    inputData: [
      { name: "Alice", age: 25 },
      { name: "Bob", age: 30 }
    ],
    outputSchema: "z.object({ averageAge: z.string(), oldestPerson: z.string() })"
  },
  {
    // Optional configuration
    failOnInvalidData: false,
    returnGeneratedCode: false,
    minSamplesCount: 1,
    maxSampleTokenSize: 1000,
    validateOutput: false,
    aiAdapterConfig: {
      provider: 'anthropic',
      model: 'claude-3-5-haiku-20241022',
      temperature: 0
    }
  }
);

CLI Mode

The CLI supports multiple ways to provide input and configuration:

1. Single Input File

bunx patbat-codegen --input-path input.json

Where input.json contains all required fields (inputSchema, inputData, outputSchema).

2. Component Files

You can provide each component separately:

bunx patbat-codegen \
  --input-schema-path schema.ts \
  --input-data-path data.json \
  --output-schema-path output-schema.ts

3. Unix-Style Pipeline

Use stdin for any component while specifying which one via the --stdin flag:

# Pipe in complete input
cat input.json | bunx patbat-codegen --stdin input

# Pipe in just the data (or any other component via --stdin)
cat data.json | bunx patbat-codegen \
  --stdin inputData \
  --input-schema-path schema.ts \
  --output-schema-path output-schema.ts

CLI Options

Core Options

• --input-path <path> - Path to complete input JSON file
• --config-path <path> - Path to configuration JSON file
• --problem <text> - Override or set the problem description

Component File Options

• --input-schema-path <path> - Path to input schema file
• --input-data-path <path> - Path to input data JSON file
• --output-schema-path <path> - Path to output schema file

Input Method

• --stdin <type> - Read from stdin: "input", "inputData", "inputSchema", or "outputSchema"

Debug Options

• --debug - Write generated code to code.tmp.ts
• --stats - Write full response to stats.tmp.json

Output Handling

• stdout contains only the transformation result
• Warnings and errors are written to stderr
• Debug files (if enabled) are written to:
  • code.tmp.ts - Generated TypeScript code
  • stats.tmp.json - Complete execution statistics

Configuration Options

Input Schema

type CodegenInput = {
  problem: string;        // Transformation description
  inputSchema: string;    // Zod schema for input validation
  inputData: any[];      // Array of data to transform
  outputSchema: string;   // Zod schema for output validation
}

Configuration Schema

type CodegenConfig = {
  failOnInvalidData: boolean;     // Stop on validation errors
  returnGeneratedCode: boolean;   // Include code in response
  minSamplesCount: number;        // Minimum samples for context
  maxSampleTokenSize: number;     // Max tokens for sampling
  validateOutput: boolean;        // Validate transformation result
  aiAdapterConfig: {
    provider: 'anthropic' | 'openai-standard' | 'openai-custom';
    model: string;
    temperature?: number;
    maxTokens?: number;
  }
}

Testing

Run the CLI tests with:

bash ./samples/cli-sample1/run-test.sh

Future Improvements

1. Smart Sample Selection
   Planning to integrate the "pickasso" npm package for more intelligent data sampling.

2. Modular Architecture
   The following modules are being prepared for separate npm packages:

   • aidap - AI adapter module for different providers
   • zodinator - Zod convenience utilities

   These modules are already well-separated in the codebase and will be published independently.

3. maxRetries, timeoutMs, lifecycle hooks are not implemented yet on this branch

4. Correct shebangs for the right runtimes during build