#!/usr/bin/env python3 """ REAL Gemini Deep Think Workflow Script Uses existing Chrome profile to access gemini.google.com """ import os import sys import time from datetime import datetime from playwright.sync_api import sync_playwright, TimeoutError as PlaywrightTimeoutError SCREENSHOTS_DIR = "/mnt/e/genesis-system/deep_think_results/screenshots" CHROME_USER_DATA = "/mnt/c/Users/P3/AppData/Local/Google/Chrome/User Data" CHROME_PROFILE = "Default" os.makedirs(SCREENSHOTS_DIR, exist_ok=True) def screenshot(page, name): path = f"{SCREENSHOTS_DIR}/{name}.png" page.screenshot(path=path, full_page=False) print(f"[SCREENSHOT] Saved: {path}") return path def wait_and_screenshot(page, name, wait_ms=2000): time.sleep(wait_ms / 1000) return screenshot(page, name) def run(): print("=" * 60) print("GEMINI DEEP THINK BROWSER AUTOMATION") print(f"Started: {datetime.now().isoformat()}") print("=" * 60) with sync_playwright() as p: print("\n[STEP 1] Launching Chrome with existing user profile...") try: # Try launching with persistent context (uses real Chrome profile) browser = p.chromium.launch_persistent_context( user_data_dir=CHROME_USER_DATA, channel="chrome", # Use real Chrome not Chromium headless=False, # Must be visible for real Chrome profile slow_mo=500, args=[ f"--profile-directory={CHROME_PROFILE}", "--no-first-run", "--disable-blink-features=AutomationControlled", ] ) print("[OK] Chrome launched with persistent context") page = browser.pages[0] if browser.pages else browser.new_page() except Exception as e: print(f"[WARN] Chrome launch with profile failed: {e}") print("[FALLBACK] Trying Chromium without profile...") browser_obj = p.chromium.launch(headless=False, slow_mo=300) browser = browser_obj page = browser.new_page() try: print("\n[STEP 1] Navigating to gemini.google.com...") page.goto("https://gemini.google.com", timeout=30000) time.sleep(4) screenshot(page, "01_initial_load") current_url = page.url print(f"[URL] Current URL: {current_url}") page_title = page.title() print(f"[TITLE] Page title: {page_title}") # Check login state if "accounts.google.com" in current_url or "signin" in current_url.lower(): print("\n[STEP 2] NOT LOGGED IN - redirected to Google login") screenshot(page, "02_login_page") print("[INFO] Login required. Cannot auto-login per protocol.") print("[INFO] Please log in manually and re-run this script.") return {"status": "needs_login", "url": current_url} # Check if we're on Gemini if "gemini.google.com" in current_url: print("[OK] On Gemini - checking login state...") screenshot(page, "02_gemini_page") # Look for login indicators page_content = page.content() # Check for common logged-in indicators logged_in_indicators = [ page.locator('[data-test-id="profile-icon"]').count() > 0, page.locator('[aria-label="Google Account"]').count() > 0, page.locator('img[alt*="profile"]').count() > 0, ] logged_out_indicators = [ page.locator('text=Sign in').count() > 0, "Sign in" in page.content(), ] print(f"[CHECK] Logged-in indicators: {logged_in_indicators}") print(f"[CHECK] Logged-out indicators: {logged_out_indicators}") print("\n[STEP 3] Analyzing Gemini UI structure...") time.sleep(3) screenshot(page, "03_gemini_ui_full") # Get page structure print("[INFO] Looking for sidebar elements...") # Try to find sidebar / left panel sidebar_selectors = [ '[data-test-id="sidebar"]', '.sidebar', 'nav', '[role="navigation"]', '[aria-label="Main menu"]', '.side-nav', 'mat-sidenav', ] for sel in sidebar_selectors: count = page.locator(sel).count() if count > 0: print(f"[FOUND] Sidebar selector '{sel}': {count} elements") # Look for model selector print("\n[INFO] Looking for model selector...") model_selectors = [ '[data-test-id="model-selector"]', '[aria-label*="model"]', '[aria-label*="Model"]', 'button:has-text("Gemini")', 'button:has-text("2.5")', 'button:has-text("Pro")', '[class*="model"]', 'mat-select', ] for sel in model_selectors: try: count = page.locator(sel).count() if count > 0: print(f"[FOUND] Model selector '{sel}': {count} elements") elem = page.locator(sel).first print(f" Text: {elem.text_content()[:100] if elem.text_content() else 'N/A'}") except Exception as e: pass # Look for Deep Think / thinking mode print("\n[INFO] Looking for Deep Think option...") deep_think_selectors = [ 'button:has-text("Deep Think")', '[aria-label*="Deep Think"]', '[aria-label*="Think"]', 'text=Deep Think', '[data-test-id*="think"]', '[class*="think"]', 'button:has-text("Think")', ] for sel in deep_think_selectors: try: count = page.locator(sel).count() if count > 0: print(f"[FOUND] Deep Think selector '{sel}': {count} elements") elem = page.locator(sel).first print(f" Text: {elem.text_content()[:100] if elem.text_content() else 'N/A'}") except Exception as e: pass # Screenshot the full page with scroll print("\n[INFO] Taking full page screenshots...") page.evaluate("window.scrollTo(0, 0)") screenshot(page, "04_top_of_page") # Get all visible text to understand the page structure visible_text = page.evaluate(""" () => { const elements = document.querySelectorAll('button, [role="button"], a, [role="menuitem"]'); return Array.from(elements).slice(0, 50).map(el => ({ tag: el.tagName, text: el.textContent?.trim()?.substring(0, 100), aria: el.getAttribute('aria-label'), class: el.className?.substring(0, 100), id: el.id })); } """) print("\n[INFO] Visible interactive elements:") for elem in visible_text[:30]: if elem.get('text') or elem.get('aria'): print(f" [{elem['tag']}] text='{elem.get('text', '')}' aria='{elem.get('aria', '')}' id='{elem.get('id', '')}'") # Try to find and click model selector to see options print("\n[STEP 3b] Trying to access model selector...") try: # Common Gemini model button location model_btn = page.locator('button').filter(has_text="2.5").first if model_btn.count() > 0: print("[OK] Found model button with '2.5'") screenshot(page, "05_before_model_click") model_btn.click() time.sleep(2) screenshot(page, "06_model_dropdown_open") else: # Try generic model area selectors_to_try = [ 'button[jsname]', '[data-test-id="bard-mode-selector"]', '.model-selector', ] for sel in selectors_to_try: elems = page.locator(sel).all() for elem in elems[:5]: txt = elem.text_content() or "" if any(x in txt for x in ['Gemini', 'Pro', 'Flash', 'Ultra', 'model']): print(f"[FOUND] Model-related element: {txt[:50]}") except Exception as e: print(f"[WARN] Model selector interaction failed: {e}") screenshot(page, "07_current_state") print("\n[STEP 4] Attempting to find chat input and Deep Think toggle...") # Find the chat input area input_selectors = [ 'textarea[placeholder*="message"]', 'textarea[placeholder*="Enter"]', 'textarea[placeholder*="Ask"]', 'div[contenteditable="true"]', 'rich-textarea', '[data-test-id="user-input"]', 'textarea', ] input_found = False for sel in input_selectors: count = page.locator(sel).count() if count > 0: print(f"[FOUND] Chat input '{sel}': {count} elements") input_found = True break # Look for deep think button near input area print("\n[INFO] Looking for Deep Think toggle near input area...") bottom_area = page.evaluate(""" () => { // Get all elements in bottom 30% of page const height = window.innerHeight; const elements = document.querySelectorAll('button, [role="button"], [role="switch"], [role="checkbox"]'); return Array.from(elements) .filter(el => { const rect = el.getBoundingClientRect(); return rect.top > height * 0.6; }) .slice(0, 30) .map(el => ({ tag: el.tagName, text: el.textContent?.trim()?.substring(0, 100), aria: el.getAttribute('aria-label'), class: el.className?.substring(0, 100), })); } """) print("[INFO] Bottom area elements:") for elem in bottom_area: if elem.get('text') or elem.get('aria'): print(f" [{elem['tag']}] text='{elem.get('text', '')}' aria='{elem.get('aria', '')}'") screenshot(page, "08_bottom_area") # Now attempt the actual workflow print("\n[STEP 5] Attempting Deep Think workflow...") # Step 5a: Find model selector and switch to 2.5 Pro print("[5a] Looking for model selector button...") # Gemini typically has a model selector in the top bar or near input page_html_snippet = page.evaluate(""" () => document.querySelector('header')?.innerHTML?.substring(0, 2000) || document.querySelector('.app-header')?.innerHTML?.substring(0, 2000) || 'no header found' """) print(f"[HTML] Header snippet: {page_html_snippet[:500]}") screenshot(page, "09_pre_workflow") print("\n[STEP 5b] Attempting to find and click Deep Think mode...") # Try clicking on the thinking/model area # Gemini UI often has a button at bottom left of input area deep_think_attempt_selectors = [ '[aria-label*="Deep Think"]', '[aria-label*="think"]', 'button:has-text("Deep Think")', '[data-test-id*="deep"]', 'button:has-text("1.5")', # old model buttons 'button:has-text("Advanced")', 'mat-slide-toggle', '[role="switch"]', ] deep_think_found = False for sel in deep_think_attempt_selectors: try: count = page.locator(sel).count() if count > 0: print(f"[FOUND] Potential Deep Think: '{sel}' ({count} elements)") elem = page.locator(sel).first print(f" Text: '{elem.text_content()[:100] if elem.text_content() else ''}'") deep_think_found = True except Exception as e: pass if not deep_think_found: print("[INFO] Deep Think not found via direct selectors - doing full DOM scan...") all_interactive = page.evaluate(""" () => { const els = document.querySelectorAll('[role="button"], button, [role="switch"], [role="checkbox"], [role="menuitem"]'); return Array.from(els).map(el => ({ text: el.textContent?.trim()?.substring(0, 80), aria: el.getAttribute('aria-label'), role: el.getAttribute('role'), class: el.className?.substring(0, 80), })).filter(el => el.text || el.aria); } """) print(f"[INFO] Total interactive elements: {len(all_interactive)}") for elem in all_interactive: txt = (elem.get('text') or '').lower() aria = (elem.get('aria') or '').lower() if any(x in txt + aria for x in ['deep', 'think', 'pro', '2.5', 'model', 'ultra', 'reason']): print(f" RELEVANT: text='{elem.get('text', '')}' aria='{elem.get('aria', '')}'") # Final state screenshot screenshot(page, "10_final_state") # Try to interact with the chat - send test message if input found if input_found: print("\n[STEP 6] Attempting to send test prompt...") test_prompt = "What is the single most important strategic decision for a new AI voice widget product targeting Australian digital agencies in 2026?" try: # Try textarea first textarea = page.locator('textarea').first if textarea.count() > 0: textarea.click() time.sleep(0.5) textarea.fill(test_prompt) time.sleep(1) screenshot(page, "11_prompt_entered") # Try to submit with Enter textarea.press("Enter") print("[OK] Prompt submitted via Enter key") time.sleep(5) screenshot(page, "12_after_submit") # Wait for response (up to 120 seconds) print("[WAIT] Waiting for Deep Think response (up to 120s)...") for i in range(24): # 24 x 5s = 120s time.sleep(5) screenshot(page, f"13_response_progress_{i:02d}") # Check if still loading loading = page.locator('[aria-label*="loading"], .loading, [class*="loading"]').count() if loading == 0 and i > 3: print(f"[OK] Response appears complete after {(i+1)*5}s") break # Get response text response_selectors = [ '[data-test-id="response-container"]', '.response-content', 'model-response', '[class*="response"]', 'message-content', ] response_text = "" for sel in response_selectors: elems = page.locator(sel).all() if elems: response_text = "\n".join([e.text_content() or "" for e in elems[-3:]]) if response_text.strip(): print(f"[OK] Got response via '{sel}'") break if not response_text: # Fallback - get all text from main content area response_text = page.evaluate(""" () => { const main = document.querySelector('main') || document.body; return main.innerText?.substring(0, 5000); } """) print(f"\n[RESPONSE PREVIEW]\n{response_text[:500]}...") screenshot(page, "14_final_response") return { "status": "success", "response_text": response_text, } except Exception as e: print(f"[ERROR] Failed to send prompt: {e}") screenshot(page, "11_input_error") print("\n[INFO] Generating workflow documentation...") # Collect all screenshots taken screenshots = sorted([f for f in os.listdir(SCREENSHOTS_DIR) if f.endswith('.png')]) print(f"[INFO] Screenshots taken: {len(screenshots)}") for s in screenshots: print(f" - {s}") return { "status": "partial", "screenshots": screenshots, "message": "UI analyzed but workflow may need adjustment" } except Exception as e: print(f"[ERROR] Main execution failed: {e}") import traceback traceback.print_exc() try: screenshot(page, "ERROR_state") except: pass return {"status": "error", "error": str(e)} finally: print("\n[INFO] Closing browser...") browser.close() if __name__ == "__main__": result = run() print(f"\n[RESULT] {result}")