code-interpreter（代码解释器）

Code Interpreter

A general-purpose code execution environment powered by AWS Bedrock AgentCore Code Interpreter. Run code, execute shell commands, and manage files in a secure sandbox.

Available Tools

• execute_code(code, language, output_filename): Execute Python, JavaScript, or TypeScript code.
• execute_command(command): Execute shell commands.
• file_operations(operation, paths, content): Read, write, list, or remove files in the sandbox.
• ci_push_to_workspace(paths): Save sandbox files to the shared workspace (S3). Omit paths to save all files in the sandbox root.

Tool Parameters

execute_code

execute_command

file_operations

tool_input Examples

execute_code — text output

{
  "code": "import pandas as pd\ndf = pd.DataFrame({'A': [1,2,3], 'B': [4,5,6]})\nprint(df.describe())",
  "language": "python"
}

execute_code — generate chart

{
  "code": "import matplotlib\nmatplotlib.use('Agg')\nimport matplotlib.pyplot as plt\nimport numpy as np\nx = np.linspace(0, 10, 100)\nplt.figure(figsize=(10,6))\nplt.plot(x, np.sin(x))\nplt.title('Sine Wave')\nplt.savefig('sine.png', dpi=300, bbox_inches='tight')\nprint('Done')",
  "language": "python",
  "output_filename": "sine.png"
}

execute_command — install a package

{
  "command": "pip install yfinance"
}

execute_command — check environment

{
  "command": "python --version && pip list | head -20"
}

file_operations — write a file

{
  "operation": "write",
  "content": [{"path": "config.json", "text": "{\"key\": \"value\"}"}]
}

file_operations — list files

{
  "operation": "list",
  "paths": ["."]
}

file_operations — read a file

{
  "operation": "read",
  "paths": ["output.csv"]
}

When to Use This Skill

Use code-interpreter as a sandbox for testing and prototyping code. For production tasks (creating documents, charts, presentations), prefer specialized skills.

Do NOT use for:

• Formatting or displaying code examples (respond directly with markdown code blocks)
• Explaining code or algorithms (respond directly with text)
• Simple calculations you can do mentally (just provide the answer)
• Any task that doesn't require actual code execution

Code Interpreter vs Code Agent

Workspace Integration

All files go to the code-interpreter/ namespace — a flat, session-isolated space separate from office documents.

Sandbox → Workspace (save outputs):

// Save a specific file after execution
{ "tool": "ci_push_to_workspace", "paths": ["chart.png", "results.json"] }

// Save everything in the sandbox root
{ "tool": "ci_push_to_workspace" }

// Alternative: save a single file inline during execute_code
{ "tool": "execute_code", "output_filename": "chart.png", "code": "..." }

Uploaded files (auto-preloaded):

Files uploaded by the user (e.g. ZIP archives) are automatically available in the sandbox — no manual loading needed. Just use them directly in execute_code.

Read saved files via workspace skill:

workspace_read("code-interpreter/chart.png")
workspace_read("code-interpreter/results.json")
workspace_list("code-interpreter/")

Text files (.py, .csv, .json, .txt, etc.) are transferred as-is. Binary files (.png, .pdf, .xlsx, etc.) are handled via base64 encoding automatically.

Environment

• Languages: Python (recommended, 200+ libraries), JavaScript, TypeScript
• Shell: Full shell access via execute_command
• File system: Persistent within session; use file_operations to manage files
• Session state: Variables and files persist across multiple calls within the same session
• Network: Internet access available (can use requests, urllib, curl)

Supported Languages

• Python (recommended) — 200+ pre-installed libraries covering data science, ML, visualization, file processing
• JavaScript — Node.js runtime, useful for JSON manipulation, async operations
• TypeScript — TypeScript runtime with type checking

Pre-installed Python Libraries

Data Analysis & Visualization

Machine Learning & AI

Mathematical & Optimization

File Processing & Documents

Image & Media

Data Storage & Formats

Web & API

Utilities

For the full list of 200+ libraries with versions, run: execute_command(command="pip list")

Usage Patterns

Pattern 1: Data Analysis

import pandas as pd
import numpy as np

df = pd.DataFrame({
    'date': pd.date_range('2024-01-01', periods=100),
    'revenue': np.random.normal(1000, 200, 100),
    'costs': np.random.normal(700, 150, 100),
})
df['profit'] = df['revenue'] - df['costs']

print("=== Summary Statistics ===")
print(df.describe())
print(f"\nTotal Profit: ${df['profit'].sum():,.2f}")
print(f"Profit Margin: {df['profit'].mean() / df['revenue'].mean() * 100:.1f}%")

Pattern 2: Visualization (with output_filename)

import matplotlib
matplotlib.use('Agg')
import matplotlib.pyplot as plt
import numpy as np

fig, axes = plt.subplots(2, 2, figsize=(14, 10))

categories = ['Q1', 'Q2', 'Q3', 'Q4']
values = [120, 150, 180, 210]
axes[0,0].bar(categories, values, color='#2196F3')
axes[0,0].set_title('Quarterly Revenue')

x = np.linspace(0, 10, 50)
axes[0,1].plot(x, np.sin(x), 'b-', linewidth=2)
axes[0,1].set_title('Trend')

sizes = [35, 30, 20, 15]
axes[1,0].pie(sizes, labels=['A','B','C','D'], autopct='%1.1f%%')
axes[1,0].set_title('Market Share')

x = np.random.normal(50, 10, 200)
y = x * 1.5 + np.random.normal(0, 15, 200)
axes[1,1].scatter(x, y, alpha=0.5, c='#FF5722')
axes[1,1].set_title('Correlation')

plt.tight_layout()
plt.savefig('dashboard.png', dpi=300, bbox_inches='tight')
print('Dashboard saved')

Pattern 3: Machine Learning

from sklearn.model_selection import train_test_split
from sklearn.ensemble import RandomForestClassifier
from sklearn.metrics import classification_report
from sklearn.datasets import load_iris

iris = load_iris()
X_train, X_test, y_train, y_test = train_test_split(
    iris.data, iris.target, test_size=0.3, random_state=42
)

model = RandomForestClassifier(n_estimators=100, random_state=42)
model.fit(X_train, y_train)
y_pred = model.predict(X_test)

print(classification_report(y_test, y_pred, target_names=iris.target_names))

Pattern 4: SQL with DuckDB

import duckdb
import pandas as pd

orders = pd.DataFrame({
    'order_id': range(1, 101),
    'customer': [f'Customer_{i%20}' for i in range(100)],
    'amount': [round(50 + i * 3.5, 2) for i in range(100)],
})

result = duckdb.sql("""
    SELECT customer, COUNT(*) as cnt, ROUND(SUM(amount), 2) as total
    FROM orders GROUP BY customer
    HAVING COUNT(*) >= 3 ORDER BY total DESC LIMIT 10
""").df()
print(result.to_string(index=False))

Pattern 5: Fetch Data from Web

import requests
import pandas as pd

response = requests.get("https://api.example.com/data")
data = response.json()
df = pd.DataFrame(data)
print(df.head())

Pattern 6: Multi-step Workflow (session state persists)

Call 1: execute_code → load and clean data, store in variable `df`
Call 2: execute_code → analyze `df`, generate chart, save as PNG
Call 3: execute_code → export results to CSV
Call 4: file_operations(operation="read") → download the CSV

Variables (df) and files persist across calls in the same session.

Important Rules

1. matplotlib.use('Agg') before import matplotlib.pyplot — sandbox has no display.
2. Use print() for text output — stdout is how results are returned.
3. output_filename must match exactly — the filename in plt.savefig() or wb.save() must match the output_filename parameter.
4. Use execute_command for shell tasks — ls, pip install, curl, etc.
5. Use file_operations for file management — read/write/list/remove files explicitly.
6. Session state persists — variables and files remain across calls. Use this for multi-step workflows.

Common Mistakes to Avoid

• Forgetting matplotlib.use('Agg') before import matplotlib.pyplot as plt
• Using plt.show() instead of plt.savefig() — there is no display
• Typo in output_filename — must match the file saved by the code exactly
• Using execute_code for shell tasks — use execute_command instead
• Writing binary files via file_operations — use execute_code to generate binary files, then download with output_filename