Dynamic Pre-Processor for GitHub Copilot

A "Middleware" Agent for VS Code Chat. This extension intercepts your chat prompts, uses a fast, lightweight AI model to pre-process them (formatting, sanitization, or planning), and then forwards the optimized result to your main reasoning model (like Gemini 2.5 Pro, GPT-4, etc.) for the final answer.

🚀 Why use this?

Standard Copilot is great, but sometimes you need to transform data before asking a question.

• Example: "Convert this JSON to TOON format before analyzing."
• Example: "Anonymize sensitive IDs in this log file before fixing the bug."
• Example: "Summarize this massive file into a plan, then ask the main model to implement it."

Instead of doing this manually in two steps, this extension automates the chain: User Input -> Mini Model (Pre-Processor) -> Main/Selected Model -> Final Answer

✨ Key Features

• ⚡ Dual-Engine Architecture: Uses a cheap/fast model (e.g., gpt-5-mini) for text manipulation and a heavy model (e.g., Claude Sonnet 4.5) for deep reasoning.
• 🧠 Smart Context Logic:
  • Auto-Magic Context: If you don't select any code, it automatically reads your active editor tab.
  • Dual-Budgeting: Sends a concise "summary" context to the Mini model (to prevent crashes) but the Full-Fidelity file content to the Main model.
• 🛡️ Safety First: Dynamic limit detection ensures you never hit "Context Window Exceeded" errors, regardless of which model you select.
• ⚙️ Fully Configurable: You define exactly what the Pre-processor should do via settings.

⚙️ Configuration

Go to Settings (Ctrl+,) and search for Pre Processor.

📖 Usage

1. Open GitHub Copilot Chat.
2. Select your desired Main/Selected Model from the dropdown (e.g., Claude Sonnet 4.5).
3. Type @preProcessor followed by your query.

Example Scenario: The "TOON" Converter

1. Set Instructions: In settings, set preProcessor.instructions to:

   "Check if there is any JSON in the user prompt. Convert that JSON into a TOON format. Do not answer the user question yet, just output the transformed prompt."

2. The Prompt:

   @preProcessor fix the logic error in this request: { "id": 1, "active": false }
   

3. What Happens:
   • Step 1 (Mini Model): Sees the JSON. Converts it to TOON. Rewrites the prompt.
   • Step 2 (Main Model): Receives: "Fix the logic error in this request: [TOON Data]" and the full context of your open file.
   • Result: The Main Model answers your question using the cleaned data.

🏗️ Architecture

graph LR
    A[User Input] --> B{Instructions Set?};
    B -- No --> C[Direct Pass-through];
    B -- Yes --> D[Step 1: Mini Model];
    D -- "Context (Limited)" --> E[Refine/Format Prompt];
    E --> F[Step 2: Main Model];
    C --> F;
    F -- "Context (Full Fidelity)" --> G[Final Answer];