#!/usr/bin/env python3 """ REAL Gemini Deep Think Workflow - v4 Key discoveries from v3: - Gemini accessible without Google auth (shows Fast/Thinking/Pro modes) - Mode picker selector: button[aria-label="Open mode picker"] - Modes available: Fast, Thinking, Pro (NOT "Deep Think" literally) - "Thinking" = standard thinking, "Pro" = deep think (long form) - Input selector: div[aria-label="Enter a prompt for Gemini"] - Response selector: model-response This version: selects "Pro" mode (maximum thinking) then sends prompt """ 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" 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) def screenshot(page, name): path = f"{SCREENSHOTS_DIR}/{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(): log("=" * 60) log("GEMINI DEEP THINK v4 - SELECT THINKING/PRO MODE") log(f"Started: {datetime.now().isoformat()}") 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", "mode_selected": None, "response_text": "", "workflow_steps": [], "screenshots": [], "errors": [], } with sync_playwright() as p: # Launch Chromium (headless=False required for full JS) log("\n[STEP 1] Launching browser...") browser = p.chromium.launch( headless=False, slow_mo=200, args=[ "--disable-blink-features=AutomationControlled", "--start-maximized", ] ) page = browser.new_page(viewport={"width": 1440, "height": 900}) try: # ---- 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(5) screenshot(page, "01_loaded") log(f"URL: {page.url}") log(f"Title: {page.title()}") workflow_steps.append({ "step": 2, "action": "Navigate", "url": "https://gemini.google.com", "wait": "5 seconds after domcontentloaded", }) # ---- Click mode picker (currently shows "Fast") ---- log("\n[STEP 3] Opening mode picker...") # The mode picker button has aria-label="Open mode picker" mode_picker = page.locator('[aria-label="Open mode picker"]') if mode_picker.count() == 0: log("[WARN] Mode picker not found, trying text-based selector") mode_picker = page.locator('button:has-text("Fast")') if mode_picker.count() > 0: log(f"[OK] Mode picker found, current text: '{mode_picker.text_content()}'") screenshot(page, "02_before_mode_click") mode_picker.click() time.sleep(2) screenshot(page, "03_mode_dropdown_open") log("Mode picker dropdown opened") workflow_steps.append({ "step": 3, "action": "Open mode picker", "selector": '[aria-label="Open mode picker"]', "code": 'page.locator(\'[aria-label="Open mode picker"]\').click()', "note": "This opens a dropdown with Fast / Thinking / Pro modes" }) # Get all visible elements to find mode options all_elements = page.evaluate(""" () => { const els = document.querySelectorAll('li, [role="menuitem"], [role="option"], [role="listitem"]'); const result = []; for (const el of els) { const rect = el.getBoundingClientRect(); if (rect.width > 0 && rect.height > 0) { result.push({ tag: el.tagName, text: el.textContent?.trim()?.substring(0, 100), aria: el.getAttribute('aria-label'), role: el.getAttribute('role'), class: el.className?.toString()?.substring(0, 60), x: Math.round(rect.x), y: Math.round(rect.y), }); } } return result; } """) log(f"Dropdown elements found: {len(all_elements)}") for el in all_elements: log(f" [{el['tag']}] text='{el.get('text','')}' role='{el.get('role','')}' y={el.get('y')}") # Target selectors for mode options (discovered from v3) # v3 found: "Fast Answers quickly", "Thinking Solves complex problems", "Pro Thinks longer for advanced math & code" # Try to select "Pro" mode (deepest thinking) pro_selectors = [ 'li:has-text("Pro")', '[role="menuitem"]:has-text("Pro")', '[role="option"]:has-text("Pro")', 'button:has-text("Pro")', 'div:has-text("Pro Thinks longer")', ] thinking_selectors = [ 'li:has-text("Thinking")', '[role="menuitem"]:has-text("Thinking")', '[role="option"]:has-text("Thinking")', ] mode_selected = False # Try Pro first (deepest thinking = "Deep Think") for sel in pro_selectors: try: el = page.locator(sel).first if el.count() > 0 and el.is_visible(): log(f"[OK] Found Pro mode via: {sel}") log(f" Text: '{el.text_content()[:60]}'") el.click() time.sleep(2) screenshot(page, "04_pro_mode_selected") result["mode_selected"] = "Pro (Deep Think)" mode_selected = True workflow_steps.append({ "step": 4, "action": "Select Pro mode (Deep Think)", "selector": sel, "code": f'page.locator("{sel}").first.click()', "note": "Pro = deepest thinking mode (equivalent to Deep Think)", }) break except Exception as e: log(f" Selector {sel} failed: {e}", "WARN") # Try clicking by position based on discovered y-coordinates if not mode_selected and all_elements: pro_elements = [el for el in all_elements if 'pro' in (el.get('text') or '').lower()] if pro_elements: pro_el = pro_elements[0] log(f"[OK] Clicking Pro by position: ({pro_el['x']}, {pro_el['y']})") page.mouse.click(pro_el['x'] + 20, pro_el['y'] + 10) time.sleep(2) screenshot(page, "04_pro_clicked_by_pos") result["mode_selected"] = "Pro (by position)" mode_selected = True # Fallback to Thinking mode if not mode_selected: for sel in thinking_selectors: try: el = page.locator(sel).first if el.count() > 0 and el.is_visible(): log(f"[OK] Selecting Thinking mode via: {sel}") el.click() time.sleep(2) screenshot(page, "04_thinking_mode_selected") result["mode_selected"] = "Thinking" mode_selected = True break except: pass if not mode_selected: log("[WARN] Could not select thinking mode - using Fast mode", "WARN") page.keyboard.press("Escape") result["mode_selected"] = "Fast (default)" else: log("[WARN] Mode picker button not found!", "WARN") result["mode_selected"] = "Unknown (mode picker not found)" # ---- Check mode is selected ---- time.sleep(1) mode_btn_text = page.locator('[aria-label="Open mode picker"]').text_content() if \ page.locator('[aria-label="Open mode picker"]').count() > 0 else "unknown" log(f"Current mode button: '{mode_btn_text}'") screenshot(page, "05_mode_confirmed") # ---- Enter prompt ---- log(f"\n[STEP 5] Entering test prompt...") log(f"Prompt: {TEST_PROMPT}") # Use the aria-label selector discovered in v3 input_el = page.locator('[aria-label="Enter a prompt for Gemini"]') if input_el.count() == 0: input_el = page.locator('div[contenteditable="true"]').first if input_el.count() > 0 and input_el.is_visible(): log("[OK] Found chat input") input_el.click() time.sleep(0.5) # Type the prompt character by character page.keyboard.type(TEST_PROMPT, delay=15) time.sleep(1) screenshot(page, "06_prompt_entered") 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] Chat input not found!", "ERROR") result["status"] = "no_input" result["errors"].append("Chat input not found") return result # ---- Submit prompt ---- log("\n[STEP 6] 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": 6, "action": "Submit prompt", "selector": '[aria-label="Send message"]', "code": 'page.locator(\'[aria-label="Send message"]\').click()', }) else: page.keyboard.press("Enter") log("[OK] Enter key to submit") time.sleep(4) screenshot(page, "07_submitted") # ---- Wait for response ---- log(f"\n[STEP 7] Waiting for response (mode: {result['mode_selected']})...") log("Deep Think / Pro mode can take 60-120 seconds. Being patient...") # Check if thinking indicator appears time.sleep(5) screenshot(page, "08_thinking_start") # Check for "Thinking..." or similar indicator page_text_early = page.evaluate("() => document.body.innerText?.substring(0, 1000)") has_thinking_indicator = any(x in page_text_early.lower() for x in ['thinking', 'processing', 'generating']) log(f"Thinking indicator present: {has_thinking_indicator}") # Wait loop - 120 seconds total prev_content_hash = "" stable_rounds = 0 response_text = "" for i in range(60): # 60 x 2s = 120s time.sleep(2) elapsed = (i + 1) * 2 # Extract current response text current = page.evaluate(""" () => { // Try model-response first (confirmed working in v3) const responses = document.querySelectorAll('model-response'); if (responses.length > 0) { const last = responses[responses.length - 1]; return last.innerText?.substring(0, 10000) || ''; } // Fallback return document.body.innerText?.substring(0, 2000) || ''; } """) or "" # Check loading state is_generating = page.evaluate(""" () => { // Check for various generating indicators const indicators = [ document.querySelector('[aria-label*="Generating"]'), document.querySelector('[aria-label*="Loading"]'), document.querySelector('.generating'), document.querySelector('[aria-busy="true"]'), document.querySelector('[class*="loading"]'), document.querySelector('[class*="spinner"]'), ]; return indicators.some(el => el !== null); } """) content_hash = hash(current[:200]) if content_hash == prev_content_hash: stable_rounds += 1 else: stable_rounds = 0 prev_content_hash = content_hash response_text = current if elapsed % 15 == 0: # Log every 15 seconds log(f" {elapsed}s: generating={is_generating}, stable={stable_rounds}, response_len={len(current)}") screenshot(page, f"09_progress_{elapsed:03d}s") # Done when stable for 3 rounds (6 seconds) and not generating if stable_rounds >= 3 and not is_generating and elapsed > 20: log(f"[OK] Response complete after ~{elapsed}s!") response_text = current break # Also check for response length > threshold (real response) if len(current) > 200 and not is_generating and elapsed > 30: log(f"[OK] Response ready ({len(current)} chars) after {elapsed}s") response_text = current break screenshot(page, "10_response_complete") # ---- Extract full response ---- log("\n[STEP 8] Extracting full response...") full_response = page.evaluate(""" () => { const responses = document.querySelectorAll('model-response'); if (responses.length > 0) { const last = responses[responses.length - 1]; return last.innerText || last.textContent || ''; } // Fallback to main content const main = document.querySelector('main, [role="main"]'); return (main || document.body).innerText?.substring(0, 15000) || ''; } """) or response_text result["response_text"] = full_response result["status"] = "success" if len(full_response) > 100 else "partial" log(f"Response length: {len(full_response)} chars") log(f"\nRESPONSE PREVIEW:\n{full_response[:600]}\n...") workflow_steps.append({ "step": 8, "action": "Extract response", "selector": "model-response", "code": 'response = page.locator("model-response").last.inner_text()', "result_length": len(full_response), }) 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, "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)) with open("/mnt/e/genesis-system/deep_think_results/v4_results.json", 'w') as f: json.dump(result, f, indent=2, default=str) log(f"Results saved: v4_results.json") return result if __name__ == "__main__": result = run() print(f"\n{'='*60}") print(f"STATUS: {result.get('status')}") print(f"MODE: {result.get('mode_selected')}") print(f"Screenshots: {len(result.get('screenshots', []))}") if result.get('response_text'): print(f"\nFULL RESPONSE:\n{result['response_text']}")