Code Execution Fallback

Code Execution Fallback & Workspace Anchoring

This skill provides a robust pattern for executing code when the primary method fails, combined with proper workspace path management to prevent file location errors.

Core Techniques

1. Workspace Path Anchoring

Always establish and verify your working directory at the start of any task:

# At the beginning of any code execution
import os
workspace_path = os.getcwd()
print(f"Working directory: {workspace_path}")

# In shell scripts
pwd
echo "Current directory: $(pwd)"

Why: Prevents files from being written to unexpected locations when agents switch between tools.

2. Execution Fallback Ladder

When execute_code_sandbox fails, follow this escalation pattern:

Level 1: Retry with Simpler Code

• Simplify the code structure
• Remove complex dependencies
• Add explicit error handling

Level 2: Use run_shell with Heredoc

When sandbox execution repeatedly fails, switch to shell execution:

python3 << 'EOF'
import os
import pandas as pd

# Your code here
data = {"col1": [1, 2, 3], "col2": ["a", "b", "c"]}
df = pd.DataFrame(data)
df.to_csv("output.csv", index=False)
print("File written successfully")
EOF

Key points:

• Use << 'EOF' (quoted) to prevent variable expansion
• Include all imports and dependencies inline
• Add explicit success/failure messages

Level 3: Delegate to shell_agent

For complex multi-step tasks with error recovery needs:

Task: Create a data processing pipeline that reads CSV, transforms data, and outputs Excel
Requirements:
- Handle missing values
- Apply transformations
- Write to ./output/ directory
- Retry on transient errors

3. Explicit Path Management

Always use absolute or explicitly relative paths:

# BAD - relies on implicit working directory
df.to_csv("output/data.csv")

# GOOD - explicit path anchoring
import os
base_path = os.getcwd()
output_dir = os.path.join(base_path, "output")
os.makedirs(output_dir, exist_ok=True)
df.to_csv(os.path.join(output_dir, "data.csv"))

# BAD
cd some_dir && python script.py

# GOOD
SCRIPT_DIR="$(cd "$(dirname "${BASH_SOURCE[0]}")" && pwd)"
cd "$SCRIPT_DIR"
python script.py

Decision Tree

execute_code_sandbox fails?
├── Yes, with syntax/import errors → Fix code, retry Level 1
├── Yes, with timeout/resource errors → Use Level 2 (run_shell heredoc)
├── Yes, with unknown/unclear errors → Use Level 3 (shell_agent)
└── No, success → Verify output file exists at expected path

Common Failure Scenarios & Solutions

| Error Type | Likely Cause | Recommended Fallback | |------------|--------------|---------------------| | ModuleNotFoundError | Missing packages | runshell with pip install first | | Timeout | Long-running operation | shellagent with progress tracking | | PermissionError | Wrong directory | Verify workspace path, use explicit paths | | Unknown error | Sandbox limitations | runshell or shellagent |

Example: Robust File Generation

# Step 1: Anchor workspace
import os
workspace = os.getcwd()
print(f"Workspace: {workspace}")

# Step 2: Create output directory explicitly
output_path = os.path.join(workspace, "deliverables")
os.makedirs(output_path, exist_ok=True)

# Step 3: Generate content with error handling
try:
    # Your generation logic here
    with open(os.path.join(output_path, "report.txt"), "w") as f:
        f.write("Content here")
    print(f"Success: File written to {output_path}")
except Exception as e:
    print(f"Error: {e}")
    # Signal to escalate to run_shell or shell_agent
    raise

Anti-Patterns to Avoid

• ❌ Assuming current directory without verification
• ❌ Using relative paths like ../output/file.txt without context
• ❌ Repeatedly retrying failed execute_code_sandbox without changing approach
• ❌ Not checking if output files exist after generation
• ❌ Mixing implicit and explicit path styles in same task

Verification Checklist

After any code execution:

• [ ] Confirm working directory was verified at start
• [ ] Confirm output files exist at expected paths
• [ ] Confirm file contents are non-empty and valid
• [ ] If execution failed, escalate to next fallback level within 2 retries

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: HKUDS
• Source: HKUDS/OpenSpace
• License: MIT

Install and usage instructions live in the source repository linked above.