#!/usr/bin/env python3 """ REAL Gemini Deep Think Workflow - v3 Correct Chrome paths discovered: - Executable: /mnt/c/Program Files (x86)/Google/Chrome/Application/chrome.exe - User Data: /mnt/c/Users/P3/AppData/Local/Google/Chrome/User Data """ import os import sys import time import json 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_EXE = "/mnt/c/Program Files (x86)/Google/Chrome/Application/chrome.exe" CHROME_USER_DATA = "/mnt/c/Users/P3/AppData/Local/Google/Chrome/User Data" 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) step_log.append(line) return line def screenshot(page, name): path = f"{SCREENSHOTS_DIR}/{name}.png" try: page.screenshot(path=path, full_page=False) log(f"Screenshot: {path}") except Exception as e: log(f"Screenshot failed '{name}': {e}", "WARN") return path def get_interactive_elements(page): """Get all visible interactive elements from the page""" return page.evaluate(""" () => { const els = document.querySelectorAll( 'button, [role="button"], a[href], [role="switch"], ' + '[role="checkbox"], [role="combobox"], [aria-haspopup], ' + 'textarea, [contenteditable="true"], input' ); return Array.from(els).map(el => { const rect = el.getBoundingClientRect(); return { tag: el.tagName, text: el.textContent?.trim()?.substring(0, 80) || '', aria: el.getAttribute('aria-label') || '', role: el.getAttribute('role') || '', class: el.className?.toString()?.substring(0, 80) || '', id: el.id || '', 'data-test-id': el.getAttribute('data-test-id') || '', visible: rect.width > 0 && rect.height > 0, y: Math.round(rect.y), x: Math.round(rect.x), }; }).filter(el => el.visible && (el.text || el.aria)); } """) def run(): log("=" * 60) log("GEMINI DEEP THINK BROWSER AUTOMATION v3") log(f"Started: {datetime.now().isoformat()}") log(f"Chrome: {CHROME_EXE}") log(f"Profile: {CHROME_USER_DATA}") log("=" * 60) result = { "status": "unknown", "logged_in": False, "deep_think_found": False, "model_25_pro_found": False, "prompt_sent": False, "response_text": "", "ui_elements": {}, "workflow_steps": [], "screenshots": [], "errors": [], } with sync_playwright() as p: # ---- STEP 1: Launch Chrome with real user profile ---- log("\n[STEP 1] Launching Chrome with Windows user 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", ], viewport={"width": 1440, "height": 900}, ignore_default_args=["--enable-automation"], # avoid bot detection ) log("[OK] Chrome launched with real Windows profile") workflow_steps.append({ "step": 1, "action": "Launch Chrome", "code": f'p.chromium.launch_persistent_context(user_data_dir="{CHROME_USER_DATA}", executable_path="{CHROME_EXE}", headless=False)', "note": "Uses real Chrome profile with existing session cookies" }) except Exception as e: log(f"[ERROR] Chrome launch with profile failed: {e}", "ERROR") result["errors"].append(f"Chrome launch: {e}") log("[FALLBACK] Launching plain Chromium (no profile - no session)") browser = p.chromium.launch( headless=False, slow_mo=200, args=["--disable-blink-features=AutomationControlled"], ) workflow_steps.append({ "step": 1, "action": "Launch Chromium (fallback)", "note": "Real Chrome profile failed - using bare Chromium (will need login)" }) # Get or create page if hasattr(browser, 'pages') and browser.pages: page = browser.pages[0] elif hasattr(browser, 'new_page'): page = browser.new_page() 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) # Allow JS to fully load screenshot(page, "01_gemini_loaded") current_url = page.url log(f"URL: {current_url}") log(f"Title: {page.title()}") workflow_steps.append({ "step": 2, "action": "Navigate to Gemini", "url": "https://gemini.google.com", "code": 'page.goto("https://gemini.google.com", timeout=45000)', "note": "Wait 6s after domcontentloaded for full JS initialization" }) # ---- STEP 3: Detect login state ---- log("\n[STEP 3] Detecting login state...") # Multiple strategies to detect login elements = get_interactive_elements(page) log(f"Total interactive elements: {len(elements)}") # Log all elements for debugging for el in elements: log(f" ELEM: tag={el['tag']} text='{el['text'][:50]}' aria='{el['aria'][:50]}'") # Check login indicators has_signin = any( 'sign in' in (el.get('text') or '').lower() or 'sign in' in (el.get('aria') or '').lower() for el in elements ) has_profile = any( any(x in (el.get('aria') or '').lower() for x in ['google account', 'account', 'profile', 'menu']) for el in elements ) has_chat_input = any( el['tag'] in ['TEXTAREA'] or 'contenteditable' in el.get('class', '').lower() or any(x in (el.get('aria') or '').lower() for x in ['message', 'chat', 'input', 'type']) for el in elements ) log(f"Sign in button present: {has_signin}") log(f"Profile element present: {has_profile}") log(f"Chat input present: {has_chat_input}") # Look specifically for mode picker (appears in logged-in state) mode_picker_elements = [el for el in elements if 'mode picker' in (el.get('aria') or '').lower() or 'fast' in (el.get('text') or '').lower() or '2.5' in (el.get('text') or '').lower() ] is_logged_in = len(mode_picker_elements) > 0 or (has_chat_input and not has_signin) result["logged_in"] = is_logged_in log(f"Login status: {'LOGGED IN' if is_logged_in else 'NOT LOGGED IN'}") if not is_logged_in: screenshot(page, "02_NOT_LOGGED_IN") log("[CRITICAL] Not logged in. Chrome profile cookies didn't transfer.", "WARN") log("Manual intervention needed: Kinan needs to authenticate via browser", "WARN") # Document the state we found result["status"] = "needs_login" result["message"] = "Chrome session not active in Playwright context" result["workflow_steps"] = workflow_steps # Still continue to map UI structure for documentation log("Continuing to document UI structure for workflow reference...") else: log("[OK] LOGGED IN - session active!") screenshot(page, "02_LOGGED_IN") # ---- STEP 4: Map the full Gemini UI ---- log("\n[STEP 4] Mapping Gemini UI structure...") screenshot(page, "03_full_ui") # Find key UI components ui_components = { "mode_picker": None, "model_selector": None, "deep_think_toggle": None, "chat_input": None, "send_button": None, "write_tool": None, "research_tool": None, "tools_button": None, } for el in elements: text_lower = (el.get('text') or '').lower() aria_lower = (el.get('aria') or '').lower() combined = text_lower + ' ' + aria_lower if 'mode picker' in aria_lower or 'fast' == text_lower: ui_components["mode_picker"] = el log(f"[FOUND] Mode picker: text='{el['text']}' aria='{el['aria']}'") if '2.5' in text_lower or 'pro' in text_lower or 'model' in aria_lower: ui_components["model_selector"] = el if 'deep think' in combined or ('think' in combined and 'deep' in combined): ui_components["deep_think_toggle"] = el result["deep_think_found"] = True if el['tag'] in ['TEXTAREA', 'INPUT'] or 'contenteditable' in text_lower: ui_components["chat_input"] = el if 'send message' in aria_lower: ui_components["send_button"] = el if 'write' == text_lower or 'write, button' in aria_lower: ui_components["write_tool"] = el if 'research' == text_lower or 'research, button' in aria_lower: ui_components["research_tool"] = el if 'tools' == text_lower: ui_components["tools_button"] = el result["ui_elements"] = {k: v for k, v in ui_components.items() if v} log(f"\nUI Components found: {[k for k,v in ui_components.items() if v]}") # Save UI map with open("/mnt/e/genesis-system/deep_think_results/ui_map_v3.json", 'w') as f: json.dump({"elements": elements, "components": ui_components}, f, indent=2, default=str) # ---- STEP 5: Click mode picker to see model options ---- log("\n[STEP 5] Opening mode picker to see model/mode options...") if ui_components["mode_picker"]: mode_picker_el = ui_components["mode_picker"] log(f"Clicking mode picker: text='{mode_picker_el['text']}'") try: # Click by position page.mouse.click( mode_picker_el['x'] + 10, mode_picker_el['y'] + 10 ) time.sleep(2) screenshot(page, "04_mode_picker_opened") # Get dropdown items dropdown = page.evaluate(""" () => { // Look for visible dropdown/menu const selectors = [ '[role="menu"] [role="menuitem"]', '[role="listbox"] [role="option"]', '[role="dialog"] button', '.dropdown-item', '[data-option-index]', ]; for (const sel of selectors) { const els = document.querySelectorAll(sel); if (els.length > 0) { return { selector: sel, items: Array.from(els).map(el => ({ text: el.textContent?.trim()?.substring(0, 100), aria: el.getAttribute('aria-label'), role: el.getAttribute('role'), })) }; } } // Fallback: all newly visible elements const allEls = document.querySelectorAll('button, [role="button"]'); return { selector: 'fallback', items: Array.from(allEls) .filter(el => { const rect = el.getBoundingClientRect(); return rect.width > 0 && rect.height > 0; }) .map(el => ({ text: el.textContent?.trim()?.substring(0, 80), aria: el.getAttribute('aria-label'), role: el.getAttribute('role'), })) .slice(0, 20) }; } """) log(f"Dropdown selector: {dropdown.get('selector')}") log(f"Dropdown items ({len(dropdown.get('items', []))}):") for item in dropdown.get('items', []): if item.get('text') or item.get('aria'): log(f" - text='{item.get('text', '')}' aria='{item.get('aria', '')}'") # Check for Deep Think / 2.5 Pro in dropdown items = dropdown.get('items', []) deep_think_items = [i for i in items if 'deep think' in (i.get('text') or '').lower() or 'deep think' in (i.get('aria') or '').lower() ] pro_25_items = [i for i in items if '2.5 pro' in (i.get('text') or '').lower() or '2.5 pro' in (i.get('aria') or '').lower() ] if deep_think_items: log(f"[FOUND] Deep Think option: {deep_think_items[0]}") result["deep_think_found"] = True # Click it if dropdown.get('selector') != 'fallback': deep_think_btn = page.locator( f"{dropdown['selector']}" ).filter(has_text="Deep Think").first if deep_think_btn.count() > 0: deep_think_btn.click() time.sleep(2) screenshot(page, "05_deep_think_selected") log("[OK] Deep Think selected!") workflow_steps.append({ "step": 5, "action": "Select Deep Think", "selector": f'{dropdown["selector"]}:has-text("Deep Think")', "code": 'page.locator(\'[role="menuitem"]:has-text("Deep Think")\').click()', }) elif pro_25_items: log(f"[FOUND] 2.5 Pro option: {pro_25_items[0]}") result["model_25_pro_found"] = True else: log("[INFO] Deep Think / 2.5 Pro not visible in dropdown") # Close dropdown page.keyboard.press("Escape") time.sleep(0.5) workflow_steps.append({ "step": 5, "action": "Open mode picker", "selector": '[aria-label="Open mode picker"]', "code": 'page.locator(\'[aria-label="Open mode picker"]\').click()', "found_items": [i.get('text', '') for i in items if i.get('text')], }) except Exception as e: log(f"Mode picker interaction failed: {e}", "WARN") result["errors"].append(f"Mode picker: {e}") page.keyboard.press("Escape") # ---- STEP 6: Find chat input and send prompt ---- log("\n[STEP 6] Finding chat input and sending test prompt...") screenshot(page, "06_before_input") TEST_PROMPT = "What is the single most important strategic decision for a new AI voice widget product targeting Australian digital agencies in 2026?" # Try multiple input selectors input_selectors = [ 'div[contenteditable="true"]', 'textarea', 'rich-textarea textarea', '[role="textbox"]', '[aria-label*="message"]', '[aria-label*="Enter"]', '[placeholder*="message"]', 'p[contenteditable="true"]', ] input_found = False for sel in input_selectors: try: el = page.locator(sel).first if el.count() > 0 and el.is_visible(): log(f"[OK] Found input via: {sel}") el.click() time.sleep(0.5) # Use type for contenteditable, fill for textarea if 'contenteditable' in sel or el.evaluate('el => el.contentEditable') == 'true': page.keyboard.type(TEST_PROMPT, delay=20) else: el.fill(TEST_PROMPT) time.sleep(1) screenshot(page, "07_prompt_typed") log(f"Prompt typed: '{TEST_PROMPT[:50]}...'") input_found = True result["prompt_sent"] = True workflow_steps.append({ "step": 6, "action": "Enter prompt", "selector": sel, "prompt": TEST_PROMPT, "code": f'page.locator("{sel}").first.click(); page.keyboard.type(prompt)', }) break except Exception as e: log(f"Input selector {sel} failed: {e}", "WARN") if not input_found: log("[ERROR] Could not find chat input!", "ERROR") result["status"] = "no_input" result["errors"].append("No chat input found") else: # Submit the prompt log("Submitting prompt...") send_btn = page.locator('[aria-label="Send message"]') if send_btn.count() > 0 and send_btn.is_visible(): send_btn.click() log("[OK] Clicked Send button") workflow_steps.append({ "step": 7, "action": "Submit prompt", "selector": '[aria-label="Send message"]', "code": 'page.locator(\'[aria-label="Send message"]\').click()', }) else: page.keyboard.press("Enter") log("[OK] Pressed Enter to submit") workflow_steps.append({ "step": 7, "action": "Submit prompt", "method": "Enter key", "code": 'page.keyboard.press("Enter")', }) time.sleep(4) screenshot(page, "08_after_submit") # ---- STEP 7: Wait for response ---- log("\n[STEP 7] Waiting for response (up to 120 seconds)...") log("Note: Deep Think responses can take 60-120 seconds") max_wait = 120 poll_interval = 3 last_content = "" stable_count = 0 for i in range(max_wait // poll_interval): time.sleep(poll_interval) elapsed = (i + 1) * poll_interval # Get current page text to detect changes current_content = page.evaluate(""" () => { const response = document.querySelector( 'model-response, [data-response-index], .response-content, ' + '[class*="response-container"]' ); if (response) return response.innerText?.substring(0, 200); return document.body.innerText?.substring(0, 500); } """) or "" # Check for loading indicators is_loading = page.evaluate(""" () => { const loaders = document.querySelectorAll( '[aria-busy="true"], [class*="loading"], [class*="spinner"], ' + '.generating, [class*="generating"]' ); return loaders.length; } """) if current_content == last_content: stable_count += 1 else: stable_count = 0 last_content = current_content if i % 5 == 0: # Every 15 seconds log(f" {elapsed}s: loading_elements={is_loading}, stable_count={stable_count}") screenshot(page, f"09_waiting_{i:02d}") # Consider done if stable for 3 polls and not loading if stable_count >= 3 and is_loading == 0 and elapsed > 15: log(f"[OK] Response stable after {elapsed}s") break screenshot(page, "10_response_final") # ---- STEP 8: Extract response ---- log("\n[STEP 8] Extracting response text...") response_text = page.evaluate(""" () => { // Try specific response containers first const responseSelectors = [ 'model-response', '[data-response-index]', '.response-content', '[class*="response-container"]', '[class*="model-response"]', '.markdown-container', ]; for (const sel of responseSelectors) { const els = document.querySelectorAll(sel); if (els.length > 0) { const texts = Array.from(els) .map(el => el.innerText?.trim()) .filter(t => t && t.length > 100); if (texts.length > 0) { return {selector: sel, text: texts[texts.length - 1]}; } } } // Fallback: get main content area const main = document.querySelector('main, [role="main"]'); return { selector: 'main fallback', text: (main || document.body).innerText?.substring(0, 10000) }; } """) if response_text and response_text.get('text'): result["response_text"] = response_text['text'] result["response_selector"] = response_text.get('selector') log(f"[OK] Response extracted via: {response_text.get('selector')}") log(f"Response length: {len(response_text['text'])} chars") log(f"\nPREVIEW:\n{response_text['text'][:500]}\n...") sel_used = response_text.get('selector', 'model-response') workflow_steps.append({ "step": 8, "action": "Extract response", "selector": sel_used, "code": f'response = page.locator("{sel_used}").last.text_content()', }) result["status"] = "success" else: log("[WARN] Could not extract response text", "WARN") result["status"] = "partial" except Exception as e: log(f"[CRITICAL ERROR] {e}", "ERROR") import traceback traceback.print_exc() result["errors"].append(str(e)) result["status"] = "error" try: screenshot(page, "ERROR_state") except: pass finally: log("\nClosing browser...") try: browser.close() except: pass # Save results result["workflow_steps"] = workflow_steps result["step_log"] = step_log result["screenshots"] = sorted(os.listdir(SCREENSHOTS_DIR)) results_path = "/mnt/e/genesis-system/deep_think_results/v3_results.json" with open(results_path, 'w') as f: json.dump(result, f, indent=2, default=str) log(f"\nResults 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"Deep Think found: {result.get('deep_think_found')}") print(f"Prompt sent: {result.get('prompt_sent')}") print(f"Screenshots: {len(result.get('screenshots', []))}") if result.get('response_text'): print(f"\nRESPONSE (first 500 chars):\n{result['response_text'][:500]}") if result.get('errors'): print(f"\nErrors: {result['errors']}")