# Minimal Computer Use Agent loop using Anthropic + pyautogui
import anthropic, pyautogui, base64, io
from PIL import ImageGrab

client = anthropic.Anthropic()
MAX_STEPS = 20

def capture_screen() -> str:
    """Take screenshot, return as base64 PNG"""
    img = ImageGrab.grab()
    buf = io.BytesIO()
    img.save(buf, format="PNG")
    return base64.b64encode(buf.getvalue()).decode()

def execute_action(action: dict):
    """Map model output → real OS action"""
    t = action["type"]
    if   t == "click":  pyautogui.click(action["x"], action["y"])
    elif t == "type":   pyautogui.typewrite(action["text"], interval=0.05)
    elif t == "scroll": pyautogui.scroll(action.get("delta", -3))
    elif t == "key":    pyautogui.hotkey(*action["keys"])

def run_agent(task: str):
    messages = []
    for step in range(MAX_STEPS):
        # 1. Capture current screen state
        screenshot = capture_screen()

        # 2. Ask VLM: what action to take next?
        messages.append({"role": "user", "content": [{
            "type": "image",
            "source": {"type": "base64", "media_type": "image/png", "data": screenshot}
        }, {"type": "text", "text": f"Task: {task}\nWhat is the next action? Reply with JSON: {{type, x?, y?, text?, keys?}}\nOr reply DONE if complete."}]})

        response = client.messages.create(
            model="claude-opus-4-6",
            max_tokens=512,
            messages=messages
        )

        reply = response.content[0].text.strip()
        messages.append({"role": "assistant", "content": reply})

        # 3. Check if done
        if "DONE" in reply:
            print("✅ Task complete")
            break

        # 4. Execute the action
        import json
        action = json.loads(reply)
        execute_action(action)

run_agent("Open Chrome, go to wikipedia.org, search for 'Turing machine'")