OpenSpace/gdpval_bench/skills/execute-code-fallback/SKILL.md

name	description
execute-code-fallback	Fallback from execute_code_sandbox to run_shell when e2b sandbox is unavailable

Execute Code Fallback Pattern

When execute_code_sandbox fails due to e2b unavailability, immediately fall back to run_shell with embedded Python code. This ensures task continuity when the sandbox service is down.

When to Apply

Use this pattern when you observe:

• execute_code_sandbox returns errors mentioning "e2b", "sandbox unavailable", or connection timeouts
• Repeated sandbox execution failures (2+ attempts)
• Error messages indicating the code execution environment is unreachable

Fallback Procedure

Step 1: Detect the Failure

Identify that execute_code_sandbox has failed. Common error indicators:

• "e2b service unavailable"
• "Sandbox connection failed"
• "Execution environment not reachable"
• Timeout errors during code execution

Step 2: Switch to run_shell with Embedded Python

Instead of:

execute_code_sandbox(code="...")

Use:

run_shell(command="python3 -c '...your Python code...'")

Step 3: Install Dependencies First (If Needed)

If your Python code requires external packages, install them first:

run_shell(command="pip install pandas requests matplotlib")

Then execute your main code:

run_shell(command="python3 << 'EOF'
import pandas as pd
import requests

# Your code here
print("Success")
EOF
")

Step 4: Use Heredoc for Multi-line Code

For complex Python scripts, use heredoc syntax for cleaner code:

run_shell(command="python3 << 'PYTHON_SCRIPT'
import json
import os

# Complex logic here
data = {'key': 'value'}
with open('output.json', 'w') as f:
    json.dump(data, f)

print('File created successfully')
PYTHON_SCRIPT
")

Complete Example

Scenario: You need to process a CSV file and generate a report.

Original approach (sandbox):

execute_code_sandbox(code="""
import pandas as pd
df = pd.read_csv('data.csv')
summary = df.describe()
print(summary)
""")

Fallback approach (run_shell):

# First install dependencies if needed
run_shell(command="pip install pandas --quiet")

# Then execute the code
run_shell(command="python3 << 'EOF'
import pandas as pd
df = pd.read_csv('data.csv')
summary = df.describe()
print(summary)
EOF
")

Important Considerations

1. State Persistence: Unlike execute_code_sandbox, run_shell executions may not share state between calls. Save intermediate results to files if needed.

2. Working Directory: Ensure you're operating in the correct directory. Use pwd to verify or include cd /path/to/workdir in your commands.

3. Python Version: Use python3 explicitly to avoid ambiguity. Verify with python3 --version if needed.

4. Error Handling: Check the stdout/stderr from run_shell to confirm success. Failed Python scripts will return non-zero exit codes.

5. Security: Be cautious when embedding user-provided data into shell commands. Escape appropriately or use file-based input.

6. Performance: For large computations, run_shell may be slower than sandbox. Consider breaking into smaller steps if timeouts occur.

Quick Reference

Task	Sandbox Approach	Fallback Approach
Simple calculation	execute_code_sandbox(code="print(2+2)")	run_shell(command="python3 -c 'print(2+2)'")
Install + run	execute_code_sandbox(code="import pkg; ...")	run_shell(command="pip install pkg && python3 -c '...'")
Multi-line script	execute_code_sandbox(code="...")	run_shell(command="python3 << 'EOF'...EOF")
File I/O	execute_code_sandbox(code="...")	run_shell(command="python3 << 'EOF'...EOF")

Recovery Checklist

• Confirm execute_code_sandbox failure (not a code bug)
• Switch to run_shell immediately
• Install required packages with pip install
• Use heredoc for multi-line Python
• Verify output and handle errors
• Save intermediate results to files if multi-step