Module IV — Exporting Code and Debugging with GitHub Copilot

Note

Duration: ~15 minutes This is the module where GitHub Copilot becomes your primary tool. You will export the agent from Foundry Toolkit for VS Code to Python code, encounter a real-world runtime error, and use Copilot to diagnose and fix it. This teaches not just how to use Copilot, but how to think with Copilot.

Objectives

Export the Compliance Agent from Foundry Toolkit for VS Code to Python code via View Code
Set up the project directory, install dependencies, and verify the .env configuration
Run the agent and encounter a common Azure runtime error
Use GitHub Copilot Agent mode to diagnose, trace, and fix the error
Successfully run the agent from the command line

Step 1: Export the Agent Code

In the Foundry Toolkit for VS Code Agent Builder (where you saved the ComplianceAgent in Module III), locate the View Code button at the top-right corner of the screen.
Click View Code.
You will be prompted to choose a location to save the agent code. A dialog will appear:
```
Choose the folder where your agent root folder will be located
```
Select the default location (e.g., C:\Users\<username>\AIToolkitAgents) or click Browse to choose a different folder (e.g., inside your workshop repository under agent/).
Press Enter to confirm. A new VS Code window will open with the generated agent project.

Step 2: Explore the Generated Project Structure

The generated project will contain the following key files:

ComplianceAgent/
├── run_agent.py           ← Main agent script (entry point)
├── requirements.txt       ← Python dependencies
├── .env                   ← Environment variables (endpoint, model name)
└── README.md              ← Auto-generated documentation

Understanding `run_agent.py`

The main script typically contains:

Imports — Azure AI Foundry SDK, agent classes
Configuration — Endpoint and model name from environment variables
Agent creation — Creates the agent with instructions and Azure AI Search tool
Conversation — Sends a hardcoded user message and prints the response

Tip

To understand the generated code, you can ask GitHub Copilot in Ask mode:

Open Copilot Chat (Ctrl+Alt+I), ensure run_agent.py is referenced as context, and send:

Explain what this script does, step by step. Focus on how the agent 
connects to Azure AI Foundry and uses Azure AI Search for retrieval.

Step 3: Verify the .env Configuration

Open the .env file and verify the following values:

PROJECT_ENDPOINT=https://<your-foundry-hub>.services.ai.azure.com/api/projects/<your-project-name>
MODEL_DEPLOYMENT_NAME=gpt-4o

Find Your Project Endpoint

If you need to find your endpoint:

Option A — From AI Foundry Portal:
- Open AI Foundry
- Navigate to your project
- Click Overview in the left sidebar
- Copy the Project endpoint URL

Option B — Using Azure CLI (ask Copilot):

Open Copilot Chat in Agent mode and send:

Using Azure CLI, find the AI Foundry project endpoint for the 
"Compliance-Sentinel" project in the "compliance-agent-rg" resource group. 
Run the command and show me the endpoint URL.

GitHub Copilot will generate and run the appropriate az command.

Important

Ensure the PROJECT_ENDPOINT URL is correct and the MODEL_DEPLOYMENT_NAME matches the name of your GPT-4o deployment from Module II. If these values are incorrect, the agent will fail to connect.

Step 4: Install Dependencies

Open a terminal in the project directory.
Install the required Python packages:
```
pip install -r requirements.txt
```
Verify the installation:
```
pip list | findstr azure
```
You should see packages like:
- azure-ai-projects
- azure-ai-agents
- azure-identity

Note

If you see dependency conflicts, try creating a virtual environment first:

python -m venv .venv
.venv\Scripts\activate
pip install -r requirements.txt

Step 5: Run the Agent (and Encounter the Error)

Ensure you are authenticated with Azure:
```
az login
```
Run the agent script:
```
python run_agent.py
```

The Expected Error

There is a high probability that you will encounter the following error:

UnicodeDecodeError: 'utf-8' codec can't decode byte 0x8b in position 1: invalid start byte

This is a known issue with the Azure AI Foundry SDK when using the aiohttp transport layer. Do not panic — this is an opportunity to learn how to debug with GitHub Copilot.

Note

If you do NOT encounter this error and the agent runs successfully — congratulations! Skip ahead to Step 7 and proceed to Module V. The error depends on your specific Python package versions and Azure SDK combination.

Step 6: Debug the Error with GitHub Copilot

This is the most powerful part of the workshop. You will use GitHub Copilot to trace through the error, understand the root cause, and generate a fix — all without leaving VS Code.

6.1 Share the Error with Copilot

Open GitHub Copilot Chat (Ctrl+Alt+I).
Switch to Agent mode using the mode selector at the top of the chat panel.

Copy the full error traceback from the terminal and paste it into Copilot Chat with the following prompt:

I am running an Azure AI Foundry agent and getting this error:

UnicodeDecodeError: 'utf-8' codec can't decode byte 0x8b in position 1: 
invalid start byte

The error occurs when the agent tries to process the response from 
Azure AI Search. 

The file run_agent.py is the entry point.

Diagnose this error and explain the root cause.

6.2 Understanding the Root Cause

GitHub Copilot will trace through the error and identify the root cause. Here is what it will find:

Error Chain:

aiohttp → azure-core → ContentDecodePolicy → auto_decompress=False

Root Cause Explanation:

The Azure REST API returns responses with gzip compression (Content-Encoding: gzip).
The Azure SDK's aiohttp transport layer has auto_decompress set to False on the aiohttp.ClientSession.
However, the SDK's ContentDecodePolicy expects the response body to be already decompressed UTF-8 text.
Since the body is still gzip-compressed (raw bytes starting with 0x1f 0x8b — the gzip magic number), the UTF-8 decoder fails at byte 0x8b.

In simpler terms: Azure's SDK explicitly tells aiohttp "don't decompress automatically," but then the response pipeline tries to decode the still-compressed bytes as UTF-8 text — which fails.

6.3 Ask Copilot to Fix the Issue

Now, send the following detailed prompt to Copilot in Agent mode:

You are debugging an Azure AI Foundry agent runtime.

Error:
UnicodeDecodeError: 'utf-8' codec can't decode byte 0x8b in position 1: 
invalid start byte

Trace through:
aiohttp → azure-core → ContentDecodePolicy → auto_decompress=False

Root cause:
Azure SDK explicitly sets auto_decompress=False on the aiohttp ClientSession,
but the response is gzip-encoded. The ContentDecodePolicy then tries to 
decode raw gzip bytes as UTF-8, which fails.

Goal:
Fix this issue in run_agent.py without breaking the Azure SDK request pipeline.

Constraints:
- Azure sets auto_decompress=False intentionally
- The response body IS gzip compressed
- We need to decompress it before the SDK tries to decode it
- The fix should be a safe monkey-patch or middleware, not a fork of azure-core

Fix the issue safely and explain the change. Apply the fix to run_agent.py.

6.4 The Fix

GitHub Copilot will generate a patch that intercepts the response and decompresses gzip content before the SDK's ContentDecodePolicy processes it. The fix typically looks like this:

import gzip
from azure.core.pipeline.policies import SansIOHTTPPolicy

class DecompressPolicy(SansIOHTTPPolicy):
    """Middleware to decompress gzip responses before ContentDecodePolicy sees them."""
    
    def on_response(self, request, response):
        encoding = response.http_response.headers.get("Content-Encoding", "")
        if "gzip" in encoding.lower():
            raw = response.http_response.body()
            if raw[:2] == b'\x1f\x8b':  # gzip magic number
                decompressed = gzip.decompress(raw)
                response.http_response._body = decompressed
                # Remove the Content-Encoding header to prevent double decompression
                if "Content-Encoding" in response.http_response.headers:
                    del response.http_response.headers["Content-Encoding"]

Note

The exact implementation may vary depending on what Copilot generates. The key principle is:

Intercept the response before ContentDecodePolicy processes it
Check if the body starts with gzip magic bytes (0x1f 0x8b)
Decompress the body
Replace the raw body with the decompressed content

6.5 Apply the Fix

GitHub Copilot in Agent mode will propose changes to run_agent.py.
Review the proposed changes carefully.
Click Accept to apply the fix.
Copilot may also need to add the gzip import and register the policy in the client pipeline.

Step 7: Run the Agent Successfully

After applying the fix, run the agent again:

python run_agent.py

Expected output: The agent should now successfully:

Connect to Azure AI Foundry
Create the agent with Azure AI Search tool
Process the hardcoded user query
Retrieve relevant compliance documents
Generate a structured compliance report

Example successful output:

## Compliance Compass Report

**Scenario:** Vendor risk assessment for AI analytics vendor in Singapore
**Generated:** March 18, 2026
**Confidence:** High

### Risk Overview
- **Risk Score:** 7 / 10 (High)
- **Risk Level:** High
- The scenario involves cross-border processing of customer transaction data, 
  triggering RBI data localization requirements and potential GDPR implications.

### Key Regulatory Findings
- **RBI Data Localization:** Payment data must be stored in India. Processing 
  abroad is permitted during the transaction phase only.
- **DPDP Act 2023:** Transfer of personal data outside India requires consent 
  and compliance with Section 16 provisions.

### Retrieved Knowledge Base References
- RBI Data Localization Guidelines: "All data relating to payment systems must 
  be stored in a system only in India."
- DPDP Act 2023: "The Central Government may restrict transfer of personal data 
  to such country or territory outside India as notified."

### Recommended Actions
**Immediate Actions:**
- Conduct a data residency assessment for the Singapore vendor
- Add contractual clauses requiring Indian data storage

**Medium-Term Actions:**
- Perform a vendor cybersecurity audit under RBI vendor onboarding guidelines

**Escalation Option:**
- If the vendor cannot comply with data localization, consider India-based 
  infrastructure alternatives

### Next Steps
For further analysis, provide additional scenario details as needed.

What You Learned About Copilot Debugging

This module demonstrated a real-world debugging workflow with GitHub Copilot:

Step	What You Did	Copilot Feature
1	Pasted the error traceback	Agent mode context understanding
2	Asked for root cause analysis	Deep code tracing across libraries
3	Provided constraints for the fix	Structured prompt engineering
4	Applied the generated fix	Agent mode file editing
5	Verified the fix works	Iterative debugging cycle

Key Insight: The power of Copilot debugging comes from giving it context (the error), constraints (what cannot change), and a goal (what success looks like). This is how you think with Copilot, not just use it.

Prerequisites Checklist

Before moving to Module V, confirm:

Agent code exported from Foundry Toolkit for VS Code via View Code
.env file configured with correct endpoint and model name
Dependencies installed from requirements.txt
Error diagnosed and fixed using GitHub Copilot (if encountered)
Agent runs successfully from the command line
Structured compliance report generated in terminal output

Key Takeaways

View Code in Foundry Toolkit for VS Code generates a complete, runnable agent project from your visual prototype.
Real-world Azure SDK errors (like the gzip/UTF-8 decode issue) are common and can be systematically debugged with Copilot.
Structured prompts for Copilot (error + trace + constraints + goal) produce better fixes than vague "fix this" requests.
GitHub Copilot in Agent mode can read files, trace code paths, generate patches, and apply them — all in a single conversation.

Click Next to proceed to Module V: Adding Interactive Input.

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Module IV — Exporting Code and Debugging with GitHub Copilot

Objectives

Step 1: Export the Agent Code

Step 2: Explore the Generated Project Structure

Understanding `run_agent.py`

Step 3: Verify the .env Configuration

Find Your Project Endpoint

Step 4: Install Dependencies

Step 5: Run the Agent (and Encounter the Error)

The Expected Error

Step 6: Debug the Error with GitHub Copilot

6.1 Share the Error with Copilot

6.2 Understanding the Root Cause

6.3 Ask Copilot to Fix the Issue

6.4 The Fix

6.5 Apply the Fix

Step 7: Run the Agent Successfully

What You Learned About Copilot Debugging

Prerequisites Checklist

Key Takeaways

FilesExpand file tree

04_Code_Export_Debug.md

Latest commit

History

04_Code_Export_Debug.md

File metadata and controls

Module IV — Exporting Code and Debugging with GitHub Copilot

Objectives

Step 1: Export the Agent Code

Step 2: Explore the Generated Project Structure

Understanding run_agent.py

Step 3: Verify the .env Configuration

Find Your Project Endpoint

Step 4: Install Dependencies

Step 5: Run the Agent (and Encounter the Error)

The Expected Error

Step 6: Debug the Error with GitHub Copilot

6.1 Share the Error with Copilot

6.2 Understanding the Root Cause

6.3 Ask Copilot to Fix the Issue

6.4 The Fix

6.5 Apply the Fix

Step 7: Run the Agent Successfully

What You Learned About Copilot Debugging

Prerequisites Checklist

Key Takeaways

Understanding `run_agent.py`