#!/usr/bin/env python3 """ REAL Gemini Deep Think Workflow - Windows Native Runs on Windows Python with real Chrome profile for authenticated access. This script is designed to run via Windows PowerShell: python E:\genesis-system\scripts\deep_think_windows.py Key differences from WSL version: - Uses Windows paths (backslash) - Can launch real Chrome executable - Has access to Windows DPAPI for cookie decryption - Chrome + Playwright can communicate natively """ import os import sys import time import json from datetime import datetime # Windows paths CHROME_EXE = r"C:\Program Files (x86)\Google\Chrome\Application\chrome.exe" CHROME_USER_DATA = r"C:\Users\P3\AppData\Local\Google\Chrome\User Data" SCREENSHOTS_DIR = r"E:\genesis-system\deep_think_results\screenshots" RESULTS_DIR = r"E:\genesis-system\deep_think_results" os.makedirs(SCREENSHOTS_DIR, exist_ok=True) step_log = [] workflow_steps = [] def log(msg, level="INFO"): ts = datetime.now().strftime("%H:%M:%S") line = f"[{ts}] [{level}] {msg}" print(line, flush=True) step_log.append(line) def screenshot(page, name): path = os.path.join(SCREENSHOTS_DIR, f"{name}.png") try: page.screenshot(path=path, full_page=False) log(f"Screenshot: {name}") except Exception as e: log(f"Screenshot failed '{name}': {e}", "WARN") return path def run(): from playwright.sync_api import sync_playwright, TimeoutError as PlaywrightTimeoutError log("=" * 60) log("GEMINI DEEP THINK - WINDOWS NATIVE") log(f"Started: {datetime.now().isoformat()}") log(f"Chrome: {CHROME_EXE}") log(f"Profile: {CHROME_USER_DATA}") log("=" * 60) TEST_PROMPT = "What is the single most important strategic decision for a new AI voice widget product targeting Australian digital agencies in 2026?" result = { "status": "unknown", "logged_in": False, "mode_selected": None, "deep_think_found": False, "response_text": "", "workflow_steps": [], "screenshots": [], "errors": [], } with sync_playwright() as p: # Launch Chrome with real user profile log("\n[STEP 1] Launching Chrome with real Windows profile...") try: browser = p.chromium.launch_persistent_context( user_data_dir=CHROME_USER_DATA, executable_path=CHROME_EXE, headless=False, slow_mo=300, args=[ "--profile-directory=Default", "--no-first-run", "--no-default-browser-check", "--disable-component-update", "--disable-extensions-except=", # disable extensions to avoid conflicts ], viewport={"width": 1440, "height": 900}, ignore_default_args=["--enable-automation"], # avoid detection ) log("[OK] Chrome launched with real profile!") workflow_steps.append({ "step": 1, "action": "Launch Chrome with real profile", "code": f'p.chromium.launch_persistent_context(user_data_dir=r"{CHROME_USER_DATA}", executable_path=r"{CHROME_EXE}", headless=False)', "note": "Must run on Windows Python, not WSL" }) except Exception as e: log(f"[ERROR] Chrome launch failed: {e}", "ERROR") result["errors"].append(f"Chrome launch: {e}") log("[FALLBACK] Trying plain Chromium...") browser_obj = p.chromium.launch(headless=False, slow_mo=200) browser = browser_obj workflow_steps.append({ "step": 1, "action": "Launch Chromium (fallback - no profile)", }) # Get page if hasattr(browser, 'pages') and browser.pages: page = browser.pages[0] else: page = browser.new_page() try: # ---- STEP 2: Navigate to Gemini ---- log("\n[STEP 2] Navigating to gemini.google.com...") page.goto("https://gemini.google.com", timeout=45000, wait_until="domcontentloaded") time.sleep(6) screenshot(page, "WIN_01_loaded") log(f"URL: {page.url}") log(f"Title: {page.title()}") # ---- STEP 3: Check login state ---- log("\n[STEP 3] Checking login state...") elements = page.evaluate(""" () => { const els = document.querySelectorAll('button, [role="button"], a, [aria-label]'); return Array.from(els) .filter(el => { const rect = el.getBoundingClientRect(); return rect.width > 0 && rect.height > 0; }) .map(el => ({ tag: el.tagName, text: el.textContent?.trim()?.substring(0, 60), aria: el.getAttribute('aria-label') || '', })); } """) log(f"Visible elements: {len(elements)}") for el in elements[:25]: if el.get('text') or el.get('aria'): log(f" [{el['tag']}] '{el.get('text','')}' / '{el.get('aria','')}'") # Detect login has_signin = any('sign in' in (el.get('text') or '').lower() for el in elements) has_mode_picker = any('mode picker' in (el.get('aria') or '').lower() for el in elements) has_account = any( any(x in (el.get('aria') or '').lower() for x in ['account', 'profile', 'google account']) for el in elements ) log(f"Sign in visible: {has_signin}") log(f"Mode picker visible: {has_mode_picker}") log(f"Account element visible: {has_account}") is_logged_in = has_account or (not has_signin and has_mode_picker) result["logged_in"] = is_logged_in log(f"LOGIN STATE: {'LOGGED IN' if is_logged_in else 'NOT LOGGED IN'}") screenshot(page, "WIN_02_login_state") if not is_logged_in: log("[WARN] Not logged in! Taking screenshot for documentation.", "WARN") result["status"] = "needs_login" result["message"] = "Chrome profile cookies not accessible - need manual login" # ---- STEP 4: Open mode picker ---- log("\n[STEP 4] Opening mode picker...") mode_picker = page.locator('[aria-label="Open mode picker"]') if mode_picker.count() > 0: log(f"Mode picker found: '{mode_picker.text_content()}'") mode_picker.click() time.sleep(2) screenshot(page, "WIN_03_mode_dropdown") # Get all elements in the dropdown dropdown_items = page.evaluate(""" () => { const els = document.querySelectorAll( '[role="menuitem"], [role="option"], [role="menuitemradio"], li' ); return Array.from(els) .filter(el => { const rect = el.getBoundingClientRect(); return rect.width > 0 && rect.height > 0; }) .map(el => ({ tag: el.tagName, text: el.textContent?.trim()?.substring(0, 100), aria: el.getAttribute('aria-label'), role: el.getAttribute('role'), disabled: el.getAttribute('aria-disabled'), 'data-mode-id': el.getAttribute('data-mode-id'), 'data-test-id': el.getAttribute('data-test-id'), x: Math.round(el.getBoundingClientRect().x), y: Math.round(el.getBoundingClientRect().y), })); } """) log(f"Dropdown items: {len(dropdown_items)}") for item in dropdown_items: if item.get('text'): log(f" [{item['tag']}] '{item.get('text','')}' disabled={item.get('disabled')} mode-id={item.get('data-mode-id','')}") # Save dropdown analysis with open(os.path.join(RESULTS_DIR, "dropdown_analysis.json"), 'w') as f: json.dump(dropdown_items, f, indent=2) # Filter available (non-disabled) modes available_modes = [item for item in dropdown_items if item.get('text') and item.get('disabled') != 'true'] log(f"\nAvailable (non-disabled) modes:") for m in available_modes: log(f" - '{m.get('text','')}'") # Check for Deep Think modes deep_think_available = [m for m in available_modes if any(x in (m.get('text') or '').lower() for x in ['deep think', 'thinking', 'pro'])] if deep_think_available: target = deep_think_available[0] log(f"[OK] Selecting: '{target.get('text')}'") # Click by position page.mouse.click(target['x'] + 20, target['y'] + 10) time.sleep(2) screenshot(page, "WIN_04_mode_selected") result["mode_selected"] = target.get('text', 'Unknown') result["deep_think_found"] = True workflow_steps.append({ "step": 4, "action": "Select thinking mode", "mode": target.get('text'), "code": f'page.mouse.click({target["x"] + 20}, {target["y"] + 10})', "alternative": 'page.locator(\'[data-test-id="bard-mode-option-thinking"]\').click()', }) else: log("[INFO] No deep think modes available (may need subscription)") page.keyboard.press("Escape") result["mode_selected"] = "Fast (default)" # ---- STEP 5: Enter prompt ---- log(f"\n[STEP 5] Entering prompt...") input_el = page.locator('[aria-label="Enter a prompt for Gemini"]').first if input_el.count() == 0: input_el = page.locator('div[contenteditable="true"]').first if input_el.count() > 0: input_el.click() time.sleep(0.5) page.keyboard.type(TEST_PROMPT, delay=15) time.sleep(1) screenshot(page, "WIN_05_prompt_entered") log(f"Prompt entered: '{TEST_PROMPT[:50]}...'") workflow_steps.append({ "step": 5, "action": "Enter prompt", "selector": '[aria-label="Enter a prompt for Gemini"]', "code": 'page.locator(\'[aria-label="Enter a prompt for Gemini"]\').click(); page.keyboard.type(prompt)', }) else: log("[ERROR] Could not find input!", "ERROR") result["status"] = "no_input" return result # Submit send_btn = page.locator('[aria-label="Send message"]') if send_btn.count() > 0: send_btn.click() log("Submitted via Send button") else: page.keyboard.press("Enter") log("Submitted via Enter") workflow_steps.append({ "step": 6, "action": "Submit prompt", "selector": '[aria-label="Send message"]', "code": 'page.locator(\'[aria-label="Send message"]\').click()', }) time.sleep(4) screenshot(page, "WIN_06_submitted") # ---- STEP 6: Wait for response ---- log(f"\n[STEP 6] Waiting for response (mode: {result.get('mode_selected', 'unknown')})...") log("Deep Think can take 60-120 seconds...") # Check for thinking block (appears before response in Deep Think) time.sleep(5) thinking_block = page.locator('[class*="thinking"], [aria-label*="thinking"], .thought-block').count() log(f"Thinking block visible: {thinking_block > 0}") screenshot(page, "WIN_07_thinking_start") # Wait loop prev_len = 0 stable_count = 0 full_response = "" for i in range(60): time.sleep(2) elapsed = (i + 1) * 2 current_response = page.evaluate(""" () => { const responses = document.querySelectorAll('model-response'); if (responses.length > 0) { return responses[responses.length - 1].innerText || ''; } return ''; } """) or "" is_generating = page.evaluate(""" () => !!document.querySelector('[aria-busy="true"], .generating, [class*="loading"]') """) if len(current_response) == prev_len: stable_count += 1 else: stable_count = 0 prev_len = len(current_response) full_response = current_response if elapsed % 10 == 0: log(f" {elapsed}s: response_len={len(current_response)}, generating={is_generating}, stable={stable_count}") if stable_count >= 4 and not is_generating and elapsed > 20 and len(current_response) > 100: log(f"[OK] Response complete after {elapsed}s!") full_response = current_response break if i == 59: log("[WARN] Timeout reached, using what we have") full_response = current_response screenshot(page, "WIN_08_response_complete") result["response_text"] = full_response result["status"] = "success" if len(full_response) > 100 else "partial" log(f"\nResponse length: {len(full_response)} chars") log(f"Preview: {full_response[:400]}") except Exception as e: log(f"[ERROR] {e}", "ERROR") import traceback traceback.print_exc() result["errors"].append(str(e)) result["status"] = "error" try: screenshot(page, "WIN_ERROR") except: pass finally: log("Closing browser...") browser.close() # Save results result["workflow_steps"] = workflow_steps result["step_log"] = step_log result["screenshots"] = sorted(os.listdir(SCREENSHOTS_DIR)) results_path = os.path.join(RESULTS_DIR, "windows_results.json") with open(results_path, 'w') as f: json.dump(result, f, indent=2, default=str) log(f"Results saved: {results_path}") return result if __name__ == "__main__": result = run() print(f"\n{'='*60}") print(f"STATUS: {result.get('status')}") print(f"LOGGED IN: {result.get('logged_in')}") print(f"MODE: {result.get('mode_selected')}") print(f"DEEP THINK: {result.get('deep_think_found')}") if result.get('response_text'): print(f"\nFULL RESPONSE:\n{result['response_text']}")