PraisonAIUI/docs/features/code-execution.md

Code Execution

Sandboxed code execution with language allowlists and SDK integration. Wraps praisonaiagents.tools.code.execute_code when available, falls back to safe simulation mode.

Quick Start

# List supported languages
curl http://localhost:8083/api/code/languages

# Execute code
curl -X POST http://localhost:8083/api/code/execute \
  -H "Content-Type: application/json" \
  -d '{"code":"print(\"Hello World\")","language":"python"}'

How It Works

graph TD
    A["Code Submission"] --> B{"SandboxExecutionManager"}
    B --> C{"SDK Available?"}
    C -->|Yes| D["praisonaiagents\nexecute_code()"]
    C -->|No| E["Simulated Mode\n(safe, no execution)"]
    B --> F{"Language Allowed?"}
    F -->|No| G["400 Error\n+ allowed list"]

1. Code is submitted with a target language
2. The manager checks if the language is in the allowlist (default: python, javascript, bash)
3. If praisonaiagents is installed, delegates to execute_code for real execution
4. Otherwise returns a simulation response with the code size

Approval & Auth

Every POST /api/code/execute records an approval decision (tool_name: "execute_code", risk_level: "medium") through the approvals manager:

• Default policy (risk_threshold: "high") auto-approves and records the decision; the response includes approval_id.
• always_deny_tools: ["execute_code"] → 403 {"status": "denied"}.
• A stricter policy (e.g. risk_threshold: "low") → 202 {"status": "pending_approval", "approval_id": "..."}. After approval via POST /api/approvals/{id}/approve, resubmit the same code with "approval_id". Each approval works once and only for that exact code.

When auth is enabled (/api/auth/config mode other than none), code execution, approve/deny, and PUT /api/approvals/policies return 401 to unauthenticated callers. The recorded approver is the authenticated identity, not a value from the request body.

Configuration

from praisonaiui.features.code_execution import SandboxExecutionManager, set_execution_manager

# Custom configuration
mgr = SandboxExecutionManager(
    sandbox=True,                              # Enable sandbox mode
    timeout=60,                                # Execution timeout
    allowed_languages=["python", "javascript"] # Only allow these languages
)
set_execution_manager(mgr)

REST API

Endpoint	Method	Description
/api/code/languages	GET	List supported languages
/api/code/execute	POST	Execute code in sandbox

POST /api/code/execute

// Request
{"code": "print('hello')", "language": "python", "timeout": 30}

// Response (simulated mode)
{"language": "python", "status": "simulated",
 "output": "[Sandbox] Code received (15 chars, python)",
 "sandbox": true, "note": "Install praisonaiagents for real execution"}

// Response (disallowed language → 400)
{"status": "error", "error": "Language 'ruby' not allowed",
 "allowed": ["python", "bash", "javascript"]}

Related

• Skills — Code execution tools in the skill catalog