Skip to content

GitHub Copilot SDK Advanced Tutorial

Author: Fu-Jie | Version: 1.0.0 | Project: OpenWebUI Extensions

This tutorial guides you through unleashing the full potential of the GitHub Copilot SDK plugin, specifically focusing on advanced file generation, BYOK customization, and complex task orchestration.

1. The 3-Step File Delivery Protocol

This is one of the most powerful features of this plugin. Agents can generate real physical files (e.g., .xlsx, .pdf, .csv) in their isolated workspace and publish them for you to download.

Automated Execution Logic:

  1. Write (Local): The Agent creates a file using code within its isolated directory (the current directory . during Python execution).
  2. Publish: The Agent automatically calls publish_file_from_workspace(filename='report.xlsx').
  3. Link: The plugin handles S3 or local storage mapping, bypasses RAG interference, and returns a secure link like /api/v1/files/.../content.

[!TIP] User Command Tip: You can simply tell the Agent: "Analyze the previous data and export an Excel sheet for me." It will automatically trigger this entire sequence.

2. Advanced BYOK (Bring Your Own Key) Mode

If you don't have a GitHub Copilot subscription or want to use high-end models from OpenAI/Anthropic directly, you can use the BYOK mode.

How to Configure:

  1. Set Base URL: e.g., https://api.openai.com/v1.
  2. Set API Key: Enter your key in your personal settings (Valves).
  3. Real-time Model Refresh: The plugin features a Config-Aware Refresh mechanism. When you modify the API Key or Base URL, simply refresh the model selector in the UIโ€”the plugin will automatically fetch the latest available models from the backend.

3. Workspace Isolation & Debugging

Every chat session has a physically isolated folder, ensuring that files from different tasks do not interfere with each other.

Physical Path Rules:

  • In-Container Path: /app/backend/data/copilot_workspace/{user_id}/{chat_id}/
  • Agent's Perspective: It sees the . directory as the path mentioned above.

Debugging Pro Tips:

  1. Enable DEBUG Valve: Set DEBUG to True in the configuration.
  2. Check the Console: Open browser developer tools (F12) -> Console.
  3. Capture Paths: You will see logs like ๐Ÿ“‚ Workspace Resolved: /.../, which helps you confirm exactly where the Agent is writing its files.

4. Deep Analysis Bypassing RAG

Traditional file uploads in OpenWebUI trigger vectorization (RAG), which might not be precise enough for large-scale data analysis.

Advantages of this Plugin: - When used with the Files Filter plugin, the Agent can directly read every byte of the raw file. - It can analyze every row of a CSV as if it were running a script locally, preventing information loss caused by retrieval-based slicing.

5. Common Interaction Examples

  • Data Conversion: "Convert this JSON content into a beautifully formatted Word document and provide a download link."
  • Code Review: "Read all .py files in the workspace, find potential bugs, and publish the suggestions as a Markdown report."
  • Chart Generation: "Generate an Excel report based on the financial data and use Python to draw a trend chart for me."

๐Ÿš€ Real-world Example: Automated Financial Analysis

Scenario

A user uploads a raw sales log named sales_data.csv and asks the AI to generate a summarized Excel report.

1. User Command

"Analyze sales_data.csv in the current directory, calculate total revenue per product category, and export an Excel file named category_summary.xlsx for me."

2. Agent Execution Flow

The Agent performs the following steps autonomously:

  • Step 1: Write and Run Python Code 
    import pandas as pd
# Direct file access in the isolated workspace (Bypassing RAG for 100% accuracy)
df = pd.read_csv('sales_data.csv')
summary = df.groupby('Category')['Revenue'].sum().reset_index()
# Save the result locally
summary.to_excel('category_summary.xlsx', index=False)
  • Step 2: Call the Publishing Tool The Agent calls: publish_file_from_workspace(filename="category_summary.xlsx")
  • Step 3: Deliver the Link The tool returns a download_url, which the Agent presents to the user.

3. Final Result

The Agent responds:

"Analysis complete! I have summarized the revenue by category. You can download your report here:

๐Ÿ“Š Download: Category_Summary.xlsx"

Execution Screenshot

GitHub Copilot SDK Test Result

โญ Continuous Improvement

If you encounter any issues or have suggestions for new features, feel free to submit an Issue or participate in discussions on OpenWebUI Extensions.