Freechat Deepseek

Generated note: shared plugin assets for this package live at the plugin root. Common local references were rewritten when they appeared in backticks or markdown links.

DeepSeek Chat Automation Skill

Automate login with scanning qr code snapshot, input questions from user messages, and extract responses from DeepSeek chat, then reply to user.

Key Feature: Pure browser automation + DOM extraction. No LLM requests or processing - just raw text from the rendered page.

When to Use

✅ USE this skill when:

• User wants to ask or query on DeepSeek Chat (https://chat.deepseek.com)
• Sending user questions from messages to DeepSeek Chat
• Extracting AI responses as raw text (no LLM processing)
• Browser automation workflows
• Extract answers from DeepSeek Chat Page with direct output

❌ DON'T use this skill when:

• Need real-time SSE stream capture (use curl/WebSocket directly)
• Page uses heavy anti-bot protection
• Requires complex CAPTCHA solving
• Need LLM-powered text processing (use Claude API directly)

Requirements

• Check browser extension with openclaw browser status, get enabled: true
• Browser tool with profile: "openclaw"
• No API keys or LLM credits needed

Workflow Overview (6 Steps)

┌─────────────────────────────────────────────────────────────────┐
│  1. Open chat.deepseek.com OR check existing tab                │
└─────────────────────────────────┬───────────────────────────────┘
                                  │
                                  ▼
┌─────────────────────────────────────────────────────────────────┐
│  2. Check login status - QR Code needed?                        │
│     - If login form → Take snapshot of login page               │
│     - If chat interface → Already logged in, proceed to Step 3   │
└─────────────────────────────────┬───────────────────────────────┘
                                  │
                                  ▼
┌──────────────────────────────────────────────────────────────────────────────────┐
│  3. Send QR Code to enabled channels (iMessage/WhatsApp/QQBot/Feishu/WeCom/Slack │
│     - Send snapshot to user via imsg                                             │
│     - If WhatsApp is active, send via WhatsApp                                   │
│     - If QQBot is active, send via QQBot, Feishu, WeCom, Slack, etc              │
└─────────────────────────────────┬────────────────────────────────────────────────┘
                                  │
                                  ▼
┌─────────────────────────────────────────────────────────────────┐
│  4. Wait for login (with timeout & retry loop)                 │
│     - Poll every 5 seconds for login success                   │
│     - If timeout: refresh page, re-snapshot QR, retry send      │
│     - Repeat until logged in successfully                        │
└─────────────────────────────────┬───────────────────────────────┘
                                  │
                                  ▼
┌─────────────────────────────────────────────────────────────────┐
│  5. Input question and press Enter                              │
└─────────────────────────────────┬───────────────────────────────┘
                                  │
                                  ▼
┌─────────────────────────────────────────────────────────────────────────┐
│  6. Extract response from DOM and reply to user                   │
└─────────────────────────────────────────────────────────────────────────┘

🎯 SIMPLE & RELIABLE RESPONSE EXTRACTION

IMPORTANT FOR AGENTS: Always use this simple extraction function instead of writing complex code. It's tested and reliable.

// SIMPLE & RELIABLE DeepSeek response extraction
// Use this function instead of writing your own extraction code
async function getDeepSeekResponseSimple(tabId) {
  try {
    const response = await browser({
      action: "act",
      kind: "evaluate",
      profile: "openclaw",
      targetId: tabId,
      fn: `(function() {
        // Get all text from the page body
        const allText = document.body.innerText || document.body.textContent || '';
        
        // Split into lines
        const lines = allText.split('\n');
        
        // Filter out system messages and find the response
        const responseLines = [];
        let foundResponse = false;
        
        for (let i = lines.length - 1; i >= 0; i--) {
          const line = lines[i].trim();
          
          // Skip empty lines and system messages
          if (line.length < 20) continue;
          if (line.includes('AI-generated')) continue;
          if (line.includes('Message DeepSeek')) continue;
          if (line.includes('How can I help')) continue;
          if (line.includes('Upload docs')) continue;
          if (line.includes('⌘')) continue;
          
          // This is likely the response
          responseLines.unshift(line);
          foundResponse = true;
        }
        
        // Combine response lines
        if (foundResponse && responseLines.length > 0) {
          return responseLines.join('\n\n').trim();
        }
        
        return 'No response found';
      })()`,
      returnByValue: true
    });
    
    return response;
  } catch (error) {
    console.error("Error extracting response:", error);
    return `Error: ${error.message}`;
  }
}

// Usage example:
// 1. Ask your question
// 2. Wait 10-15 seconds for response
// 3. Call: const answer = await getDeepSeekResponseSimple(tabId);
// 4. Use the answer

Step 1: Open or Reuse Tab

// Try to find existing DeepSeek tab first
const existingTabs = await sessions_list({ /* filter for chat.deepseek.com */ });

if (existingTabs.length > 0) {
  console.log("✅ Reusing existing DeepSeek tab:", existingTabs[0]);
  tabId = existingTabs[0];
} else {
  // Open new tab
  const result = await browser({
    action: "open",
    profile: "openclaw",
    targetUrl: "https://chat.deepseek.com/"
  });
  tabId = result.targetId;
  console.log("✅ Opened new DeepSeek tab:", tabId);
}

// Wait for page to load
await new Promise(r => setTimeout(r, 3000));

Step 2: Check Login Status (QR Code Needed?)

// Check if login required (QR Code scan) or already logged in
await browser({
  action: "snapshot",
  profile: "openclaw",
  targetId: tabId
});

// Evaluate login status
const loginStatus = await browser({
  action: "act",
  kind: "evaluate",
  profile: "openclaw",
  targetId: tabId,
  fn: `{
    const loginForm = document.querySelector('input[type="tel"], input[placeholder*="Phone"]');
    const qrCode = document.querySelector('iframe[src*="wechat"], img[alt*="WeChat"], img[src*="wechat"]');
    const chatInput = document.querySelector('textarea[placeholder*="Message DeepSeek"]');
    const userAvatar = document.querySelector('[class*="avatar"], [class*="user"]');
    
    JSON.stringify({ 
      needLogin: !!(loginForm || qrCode),
      hasChatInput: !!chatInput,
      hasUser: !!userAvatar,
      hasLoginForm: !!loginForm,
      hasQrCode: !!qrCode
    });
  }`,
  returnByValue: true
});

const status = JSON.parse(loginStatus);

if (!status.needLogin && status.hasChatInput) {
  console.log("✅ Already logged in!");
  // Proceed to Step 5 (input question)
} else {
  console.log("⚠️ QR Code login required!");
  // Proceed to Step 3 (send QR code)
}

Login Status Indicators

Step 3: Send QR Code to Enabled Channels

// Navigate to login page explicitly
await browser({
  action: "navigate",
  profile: "openclaw",
  targetId: tabId,
  targetUrl: "https://chat.deepseek.com/sign_in"
});

await new Promise(r => setTimeout(r, 2000));

// Take screenshot of QR Code login page
const screenshotResult = await browser({
  action: "screenshot",
  profile: "openclaw",
  targetId: tabId
});

// Get the screenshot path from the browser tool response
const screenshotPath = screenshotResult.path;

// Get user phone number (configure this in your implementation)
const userPhone = process.env.DEEPSEEK_PHONE || "+1234567890000"; // Configure this

// Try to send via enabled and running channels such as iMessage (most reliable on Mac)
try {
  await exec({
    command: 'imsg send --to "' + userPhone + '" --text "Please scan this QR code with WeChat to login to DeepSeek:" --file "' + screenshotPath + '"'
  });
  console.log("✅ QR Code sent via iMessage");
} catch (e) {
  console.log("⚠️ iMessage failed:", e.message);
}

// Try to send via WhatsApp if available
try {
  await exec({
    command: 'openclaw message send --channel whatsapp --target "' + userPhone + '" -m "Please scan QR code with WeChat to login" --media "' + screenshotPath + '"'
  });
  console.log("✅ QR Code sent via WhatsApp");
} catch (e) {
  console.log("⚠️ WhatsApp not available:", e.message);
}

// Try to send via QQBot (use media tags)
try {
  await exec({
    command: 'openclaw message send --channel qqbot -m "Please scan this QR code with WeChat to login to DeepSeek:\\n<qqimg>' + screenshotPath + '</qqimg>"'
  });
  console.log("✅ QR Code sent via QQBot");
} catch (e) {
  console.log("⚠️ QQBot not available:", e.message);
}

QR Code Sending Priority

Send QR Code to session which is executing this skill firstly. And also send by other avlabile channels:

1. iMessage (imsg) - Most reliable on Mac (uses userPhone variable)
2. WhatsApp - If WhatsApp Web is active (uses userPhone variable)
3. QQBot - If QQ bot is configured (no phone number needed for QQBot)
4. Signal - If Signal CLI is available (would need phone number configuration)

Step 4: Wait for Login (with Timeout & Retry Loop)

// Poll for login success
let attempts = 0;
const maxAttempts = 24; // 24 * 5 seconds = 2 minutes max
const retryInterval = 5000;

async function checkLoginSuccess() {
  const result = await browser({
    action: "act",
    kind: "evaluate",
    profile: "openclaw",
    targetId: tabId,
    fn: `{
      const chatInput = document.querySelector('textarea[placeholder*="Message DeepSeek"]');
      const userAvatar = document.querySelector('[class*="avatar"], [class*="user"]');
      { hasChatInput: !!chatInput, hasUser: !!userAvatar };
    }`,
    returnByValue: true
  });
  return JSON.parse(result);
}

while (attempts < maxAttempts) {
  await new Promise(r => setTimeout(r, retryInterval));
  
  const loginResult = await checkLoginSuccess();
  
  if (loginResult.hasChatInput && loginResult.hasUser) {
    console.log("✅ Login successful!");
    break;
  }
  
  attempts++;
  console.log("⏳ Waiting for login... (" + attempts + "/" + maxAttempts + ")");
  
  // If timeout (after 12 attempts = 1 minute), refresh and retry
  if (attempts === 12) {
    console.log("⏰ Timeout, refreshing QR code...");
    
    await browser({
      action: "navigate",
      profile: "openclaw",
      targetId: tabId,
      targetUrl: "https://chat.deepseek.com/sign_in"
    });
    
    await new Promise(r => setTimeout(r, 3000));
    
    // Re-take screenshot and send again
    await browser({
      action: "screenshot",
      profile: "openclaw",
      targetId: tabId
    });
    
    // Re-take screenshot and send again
    const newScreenshotResult = await browser({
      action: "screenshot",
      profile: "openclaw",
      targetId: tabId
    });
    const newScreenshotPath = newScreenshotResult.path;
    
    // Re-send QR code
    try {
      await exec({
        command: 'imsg send --to "' + userPhone + '" --text "QR code refreshed - please scan again with WeChat" --file "' + newScreenshotPath + '"'
      });
    } catch (e) {}
  }
}

if (attempts >= maxAttempts) {
  throw new Error("Login timeout - user did not complete QR scan in time");
}

Login Timeout Strategy

Step 5: Input Question and Press Enter

// Ensure we're on chat page
await browser({
  action: "navigate",
  profile: "openclaw",
  targetId: tabId,
  targetUrl: "https://chat.deepseek.com/"
});

await new Promise(r => setTimeout(r, 2000));

// Take snapshot to get fresh refs
await browser({
  action: "snapshot",
  profile: "openclaw",
  targetId: tabId
});

// Type message into input box (use textbox ref from snapshot)
await browser({
  action: "act",
  kind: "type",
  profile: "openclaw",
  targetId: tabId,
  ref: "e102", // Get from snapshot - textarea ref
  text: "Your question here"
});

console.log("✅ Text typed into input box");

// Press Enter to submit
await browser({
  action: "act",
  kind: "press",
  profile: "openclaw",
  targetId: tabId,
  key: "Enter"
});

console.log("✅ Enter key pressed");

Step 6: Extract Response from DOM

NEVER send raw html content with <tool_call> to user session.

// Wait for response to generate (8-15 seconds depending on complexity)
await new Promise(r => setTimeout(r, 10000));

// Extract assistant response from page body
const response = await browser({
  action: "act",
  kind: "evaluate",
  profile: "openclaw",
  targetId: tabId,
  fn: `{
    // Get all paragraph elements from the entire page body
    const body = document.body;
    const paras = Array.from(body.querySelectorAll('p'));
    
    // Filter to find assistant messages (not user messages, not system text)
    const assistantMsgs = paras.filter(p => {
      const text = p.innerText || '';
      const parent = p.parentElement;
      
      // Exclude system messages and user messages
      return text.length > 15 && 
             !text.includes('AI-generated') &&
             !text.includes('Upload docs') &&
             !text.includes('How can I help') &&
             !text.includes('⌘') &&
             parent && !parent.closest('header') &&
             parent && !parent.closest('footer');
    });
    
    // Get the most recent assistant message
    const lastResponse = assistantMsgs.length > 0 
      ? assistantMsgs[assistantMsgs.length - 1].innerText 
      : 'No response found';
    
    // Clean up the response
    return lastResponse.trim();
  }`,
  returnByValue: true
});

console.log("✅ DeepSeek Response:", response);

Response Extraction Strategy

Complete Example: Full Workflow

// ============ DEEPSEEK CHAT AUTOMATION ============

const profile = "openclaw";
// Get user phone number from USER.md or configuration
// Example: Read from USER.md or use environment variable
const userPhone = process.env.DEEPSEEK_PHONE || "+1234567890000"; // Default fallback - configure this

// Step 1: Open or reuse tab
const openResult = await browser({
  action: "open",
  profile: profile,
  targetUrl: "https://chat.deepseek.com/"
});
const tabId = openResult.targetId;

await new Promise(r => setTimeout(r, 3000));

// Step 2: Check login status
await browser({ action: "snapshot", profile, targetId: tabId });
const loginCheck = await browser({
  action: "act",
  kind: "evaluate",
  profile,
  targetId: tabId,
  fn: `{
    const chatInput = document.querySelector('textarea[placeholder*="Message DeepSeek"]');
    const qrCode = document.querySelector('img[src*="wechat"]');
    { needLogin: !chatInput, hasQrCode: !!qrCode };
  }`,
  returnByValue: true
});

const status = JSON.parse(loginCheck);

if (!status.needLogin) {
  console.log("Already logged in!");
} else {
  // Step 3: Send QR code
  await browser({ action: "navigate", profile, targetId: tabId, targetUrl: "https://chat.deepseek.com/sign_in" });
  await new Promise(r => setTimeout(r, 2000));
  const screenshotResult = await browser({ action: "screenshot", profile, targetId: tabId });
  
  const screenshotPath = screenshotResult.path;
  
  // Send via iMessage
  await exec({
    command: 'imsg send --to "' + userPhone + '" --text "Please scan QR code to login to DeepSeek" --file "' + screenshotPath + '"'
  });
  
  // Step 4: Wait for login with retry
  let attempts = 0;
  while (attempts < 24) {
    await new Promise(r => setTimeout(r, 5000));
    const result = await browser({
      action: "act",
      kind: "evaluate",
      profile,
      targetId: tabId,
      fn: `{
        const chatInput = document.querySelector('textarea[placeholder*="Message DeepSeek"]');
        !!chatInput;
      }`,
      returnByValue: true
    });
    
    if (result === 'true') {
      console.log("Login successful!");
      break;
    }
    
    attempts++;
    
    // Refresh QR every 60 seconds
    if (attempts === 12) {
      await browser({ action: "navigate", profile, targetId: tabId, targetUrl: "https://chat.deepseek.com/sign_in" });
      await new Promise(r => setTimeout(r, 3000));
      const newScreenshotResult = await browser({ action: "screenshot", profile, targetId: tabId });
      const newScreenshotPath = newScreenshotResult.path;
      await exec({
        command: 'imsg send --to "' + userPhone + '" --text "QR refreshed - please scan again" --file "' + newScreenshotPath + '"'
      });
    }
  }
}

// Step 5: Input question
await browser({ action: "navigate", profile, targetId: tabId, targetUrl: "https://chat.deepseek.com/" });
await new Promise(r => setTimeout(r, 2000));
await browser({ action: "snapshot", profile, targetId: tabId });

await browser({
  action: "act",
  kind: "type",
  profile,
  targetId: tabId,
  ref: "e102", // From snapshot
  text: "What is the capital of France?"
});

await browser({ action: "act", kind: "press", profile, targetId: tabId, key: "Enter" });

// Step 6: Extract response
await new Promise(r => setTimeout(r, 10000));

// Robust response extraction
const response = await browser({
  action: "act",
  kind: "evaluate",
  profile,
  targetId: tabId,
  fn: `(function() {
    // Get all paragraph elements
    const paragraphs = Array.from(document.querySelectorAll('p'));
    
    // Filter to find assistant responses
    const assistantResponses = paragraphs.filter(p => {
      const text = p.innerText || p.textContent || '';
      return text.length > 20 && 
             !text.includes('AI-generated') &&
             !text.includes('Upload docs') &&
             !text.includes('How can I help') &&
             !text.includes('⌘');
    });
    
    // Get the most recent (last) response
    if (assistantResponses.length > 0) {
      return assistantResponses[assistantResponses.length - 1].innerText.trim();
    }
    
    // Fallback: try to find any substantial text
    const bodyText = document.body.innerText || document.body.textContent || '';
    const lines = bodyText.split('\n').filter(line => 
      line.trim().length > 50 && 
      !line.includes('AI-generated') &&
      !line.includes('How can I help')
    );
    
    return lines.length > 0 ? lines[lines.length - 1].trim() : 'No response found';
  })()`,
  returnByValue: true
});

console.log("Response:", response);
// Reply to user with response

Common Selectors

Configuration

Phone Number Configuration

The user's phone number can be configured in several ways:

1. Environment Variable: Set DEEPSEEK_PHONE environment variable

   export DEEPSEEK_PHONE="+1234567890000"
   

2. USER.md: Add phone number to USER.md file

   # USER.md
   Phone: +1234567890000
   

3. Configuration File: Add to OpenClaw configuration

   // In your implementation code
   const userPhone = config.userPhone || process.env.DEEPSEEK_PHONE || "+1234567890000";
   

QR Code Path Handling

The screenshot path is automatically returned by the browser tool's screenshot action:

const screenshotResult = await browser({ action: "screenshot", profile: "openclaw", targetId: tabId });
const screenshotPath = screenshotResult.path; // Use the actual path from result

Important: Never assume a fixed path like /Users/chris/.openclaw/media/browser/<uuid>.jpg. Always use the path returned by the browser tool.

Limitations

• Rate Limiting: DeepSeek may block requests (429 errors)
• Event Validation: Some buttons need real user events (use kind: "press" not JS dispatchEvent)
• Login Verification: QR Code scan requires user interaction
• No API Access: Direct API requires bypassing WAF
• No Files Uploading: DO NOT upload files to chat.deepseek.com
• No LLM Processing: Extract raw text only
• Login Timeout: Maximum 2 minutes for QR scan

Troubleshooting

🆕 Optimized: "Try Again with Fresh Question" Workflow

When a user wants to retry with a fresh question (e.g., after getting "Sorry, that's beyond my current scope"), use this optimized workflow:

Option 1: Start Fresh Chat (Recommended for unsatisfactory responses)

// Function to start a fresh chat session
async function startFreshChat(tabId) {
  // Navigate to main page to ensure clean state
  await browser({
    action: "navigate",
    profile: "openclaw",
    targetId: tabId,
    targetUrl: "https://chat.deepseek.com/"
  });
  
  await new Promise(r => setTimeout(r, 2000));
  
  // Take snapshot to get fresh refs
  await browser({
    action: "snapshot",
    profile: "openclaw",
    targetId: tabId
  });
  
  // Check if we're on a fresh chat page
  const status = await browser({
    action: "act",
    kind: "evaluate",
    profile: "openclaw",
    targetId: tabId,
    fn: `(function() {
      const welcomeMsg = document.querySelector('img[alt*="DeepSeek"], img[src*="deepseek"]');
      const chatInput = document.querySelector('textarea[placeholder*="Message DeepSeek"]');
      const previousChat = document.querySelector('[class*="message"]');
      
      return JSON.stringify({
        isFresh: !!welcomeMsg && !!chatInput && !previousChat,
        hasWelcome: !!welcomeMsg,
        hasInput: !!chatInput,
        hasPreviousChat: !!previousChat
      });
    })()`,
    returnByValue: true
  });
  
  const statusObj = JSON.parse(status);
  
  if (!statusObj.isFresh) {
    // Click "New chat" in sidebar (more reliable than top button)
    const sidebarResult = await browser({
      action: "act",
      kind: "evaluate",
      profile: "openclaw",
      targetId: tabId,
      fn: `(function() {
        // Find "New chat" link in sidebar
        const sidebarLinks = Array.from(document.querySelectorAll('a'));
        const newChatLink = sidebarLinks.find(link => 
          link.innerText.includes('New chat') || 
          link.textContent.includes('新聊天') ||
          (link.querySelector('div') && link.querySelector('div').innerText.includes('New chat'))
        );
        
        if (newChatLink) {
          newChatLink.click();
          return true;
        }
        return false;
      })()`,
      returnByValue: true
    });
    
    if (!sidebarResult) {
      // Fallback: use keyboard shortcut Cmd+J (⌘J)
      await browser({
        action: "act",
        kind: "press",
        profile: "openclaw",
        targetId: tabId,
        key: "Meta",  // Command key on Mac
        modifiers: ["Meta"]
      });
      
      await new Promise(r => setTimeout(r, 100));
      
      await browser({
        action: "act",
        kind: "type",
        profile: "openclaw",
        targetId: tabId,
        text: "j"
      });
      
      await new Promise(r => setTimeout(r, 100));
      
      await browser({
        action: "act",
        kind: "press",
        profile: "openclaw",
        targetId: tabId,
        key: "Meta",
        modifiers: []
      });
    }
    
    await new Promise(r => setTimeout(r, 2000));
  }
  
  return true;
}

Option 2: Continue in Same Chat (if context is needed)

// Function to continue in current chat but clear previous question
async function continueInCurrentChat(tabId, newQuestion) {
  // Just type the new question directly
  await browser({
    action: "snapshot",
    profile: "openclaw",
    targetId: tabId
  });
  
  // Find and focus the input
  const inputFound = await browser({
    action: "act",
    kind: "evaluate",
    profile: "openclaw",
    targetId: tabId,
    fn: `(function() {
      const input = document.querySelector('textarea[placeholder*="Message DeepSeek"], input[placeholder*="Message DeepSeek"]');
      if (input) {
        input.focus();
        input.value = '';
        return true;
      }
      return false;
    })()`,
    returnByValue: true
  });
  
  if (!inputFound) {
    throw new Error("Chat input not found");
  }
  
  // Type the new question
  await browser({
    action: "act",
    kind: "type",
    profile: "openclaw",
    targetId: tabId,
    ref: "e102", // Get from snapshot - adjust as needed
    text: newQuestion
  });
  
  // Submit
  await browser({
    action: "act",
    kind: "press",
    profile: "openclaw",
    targetId: tabId,
    key: "Enter"
  });
  
  return true;
}

🐛 Common Errors and Fixes

Error: SyntaxError in JavaScript Evaluation

IMPORTANT: Never copy error examples! Always use the working functions provided below.

Common JavaScript syntax errors to avoid:

• ❌ document.querySelecto rAll('p') (space in function name)

• ✅ document.querySelectorAll('p') (correct)

• ❌ split('\\n') (double backslashes)

• ✅ split('\n') (single backslash)

• ❌ Missing return statement in evaluate functions

• ✅ Always include return statement

Robust Response Extraction Function

Here's a corrected and more robust version:

// Robust response extraction function
async function extractDeepSeekResponse(tabId) {
  try {
    const response = await browser({
      action: "act",
      kind: "evaluate",
      profile: "openclaw",
      targetId: tabId,
      fn: `(function() {
        // Find the most recent assistant response
        const paragraphs = Array.from(document.querySelectorAll('p'));
        
        // Filter out system messages and short texts
        const assistantParagraphs = paragraphs.filter(p => {
          const text = p.innerText || p.textContent || '';
          const parent = p.parentElement;
          const isSystemMessage = 
            text.includes('AI-generated') ||
            text.includes('Upload docs') ||
            text.includes('How can I help') ||
            text.includes('⌘') ||
            text.includes('We are asked:') ||
            text.length < 20;
          
          // Also check if it's in a response container (not in input area)
          const isInInputArea = parent && (
            parent.closest('textarea') || 
            parent.closest('input') ||
            parent.closest('[class*="input"]') ||
            parent.closest('[class*="message-input"]')
          );
          
          return !isSystemMessage && !isInInputArea && text.trim().length > 0;
        });
        
        // Get the last assistant response
        const lastResponse = assistantParagraphs[assistantParagraphs.length - 1];
        if (!lastResponse) return 'No response found';
        
        // Try to get the full response container
        let responseContainer = lastResponse.closest('div[class*="message"], div[class*="response"], div[class*="assistant"]');
        if (!responseContainer) {
          responseContainer = lastResponse.closest('div');
        }
        
        if (responseContainer) {
          // Get all text content from the response container
          const allText = responseContainer.innerText || responseContainer.textContent || '';
          // Clean up the text - use single backslashes in regex
          return allText
            .replace(/\s+/g, ' ')
            .replace(/\n\s*\n/g, '\n\n')
            .trim();
        }
        
        // Fallback to just the paragraph text
        return lastResponse.innerText || lastResponse.textContent || '';
      })()`,
      returnByValue: true
    });
    
    return response;
  } catch (error) {
    console.error("Error extracting response:", error);
    
    // Fallback: try simpler extraction
    const fallbackResponse = await browser({
      action: "act",
      kind: "evaluate",
      profile: "openclaw",
      targetId: tabId,
      fn: `(function() {
        // Simple fallback: get all text on page and find the last substantial block
        const allText = document.body.innerText || document.body.textContent || '';
        const lines = allText.split('\n').filter(line => 
          line.trim().length > 50 && 
          !line.includes('AI-generated') &&
          !line.includes('How can I help')
        );
        
        return lines.length > 0 ? lines[lines.length - 1].trim() : 'No response found';
      })()`,
      returnByValue: true
    });
    
    return fallbackResponse;
  }
}

// Usage:
// const answer = await extractDeepSeekResponse(tabId);

Common JavaScript Errors to Avoid

IMPORTANT: Always copy from working examples, not error examples!

Common mistakes:

• Typos in function names: querySelectorAll not querySelecto rAll
• Spaces in variable names: assistantParagraphs not assistantPar agraphs
• Incorrect selectors: Use closest('div') not closest('generic')
• Case sensitivity: returnByValue: true not returnByValue: True
• Missing semicolons: Always end statements properly

Fix: "No response found" Error

If you're getting "No response found" or "failed to extract response", DeepSeek's page structure has likely changed. Use this more robust extraction method:

// ULTRA-ROBUST response extraction - tries multiple approaches
async function extractDeepSeekResponse(tabId) {
  const response = await browser({
    action: "act",
    kind: "evaluate",
    profile: "openclaw",
    targetId: tabId,
    fn: `(function() {
      // Helper: check if text is a system message
      function isSystemText(text) {
        const systemPatterns = [
          'AI-generated', 'Upload docs', 'How can I help', '⌘',
          'We are asked:', 'deepseek', 'Sign in', '登录',
          'Message DeepSeek', 'keyboard', 'shortcut'
        ];
        return systemPatterns.some(p => text.includes(p));
      }

      // Approach 1: Try finding message containers
      const messageDivs = document.querySelectorAll('div[class*="message"], div[class*="response"], div[class*="content"]');
      let allTexts = [];

      messageDivs.forEach(div => {
        const text = div.innerText || div.textContent || '';
        if (text.length > 30 && !isSystemText(text)) {
          allTexts.push(text.trim());
        }
      });

      // If we found messages in containers, return the last one
      if (allTexts.length > 0) {
        return allTexts[allTexts.length - 1];
      }

      // Approach 2: Try all <p> tags with relaxed filtering
      const paragraphs = document.querySelectorAll('p');
      const validParagraphs = [];
      paragraphs.forEach(p => {
        const text = p.innerText || p.textContent || '';
        if (text.length > 20 && !isSystemText(text)) {
          validParagraphs.push(text.trim());
        }
      });

      if (validParagraphs.length > 0) {
        return validParagraphs[validParagraphs.length - 1];
      }

      // Approach 3: Try <div> tags (newer DeepSeek structure)
      const divs = document.querySelectorAll('div');
      const validDivs = [];
      divs.forEach(div => {
        const text = div.innerText || div.textContent || '';
        // Look for substantial text blocks in the main content area
        if (text.length > 100 && !isSystemText(text) && !div.closest('header') && !div.closest('footer') && !div.closest('nav')) {
          validDivs.push(text.trim());
        }
      });

      if (validDivs.length > 0) {
        return validDivs[validDivs.length - 1];
      }

      // Approach 4: Fallback to full page body text
      const bodyText = document.body.innerText || document.body.textContent || '';
      const lines = bodyText.split('\n').filter(line =>
        line.trim().length > 50 && !isSystemText(line)
      );

      if (lines.length > 0) {
        return lines[lines.length - 1].trim();
      }

      // Last resort: return debug info
      return 'No response found - page may be loading or have unexpected structure';
    })()`,
    returnByValue: true
  });

  console.log("Extracted response:", response.substring(0, 200) + "...");
  return response;
}

Debugging Tips

// Add debug logging to browser evaluate functions
const debugInfo = await browser({
  action: "act",
  kind: "evaluate",
  profile: "openclaw",
  targetId: tabId,
  fn: `(function() {
    // Debug: log what we find
    const paragraphs = document.querySelectorAll('p');
    console.log('Found', paragraphs.length, 'paragraphs');

    for (let i = 0; i < Math.min(paragraphs.length, 5); i++) {
      console.log('Paragraph', i, ':', paragraphs[i].innerText.substring(0, 100));
    }
    
    return JSON.stringify({
      paragraphCount: paragraphs.length,
      sampleTexts: Array.from(paragraphs).slice(0, 3).map(p => p.innerText.substring(0, 50))
    });
  })()`,
  returnByValue: true
});

console.log("Debug info:", debugInfo);

🛠️ Working Implementation Guide

Always Use These Pre-tested Functions

Instead of writing your own extraction code, always use these pre-tested functions:

// SIMPLE & RELIABLE extraction function
async function extractDeepSeekResponseSimple(tabId) {
  const response = await browser({
    action: "act",
    kind: "evaluate",
    profile: "openclaw",
    targetId: tabId,
    fn: `(function() {
      // Get all text from the page
      const allText = document.body.innerText || document.body.textContent || '';
      
      // Split into lines and filter out system messages
      const lines = allText.split('\n').filter(line => {
        const trimmed = line.trim();
        return trimmed.length > 50 && 
               !trimmed.includes('AI-generated') &&
               !trimmed.includes('Message DeepSeek') &&
               !trimmed.includes('How can I help') &&
               !trimmed.includes('Upload docs');
      });
      
      // Return the last substantial line (most likely the response)
      return lines.length > 0 ? lines[lines.length - 1].trim() : 'No response found';
    })()`,
    returnByValue: true
  });
  
  return response;
}

// Usage:
// const answer = await extractDeepSeekResponseSimple(tabId);

Key Improvements in Optimized Workflow:

1. Robust "New Chat" Detection: Uses semantic search for "New chat" text in sidebar links
2. Fallback Mechanisms: Keyboard shortcut (Cmd+J) if sidebar link not found
3. State Verification: Checks if we're already on a fresh chat page
4. Better Response Extraction: Filters out system messages and thought processes
5. Error Handling: Graceful fallbacks for missing elements
6. Clear Documentation: Explains when to use fresh chat vs continue in same chat

When to Use Which Approach:

Common Pitfalls Fixed:

1. ❌ WRONG: Clicking ref=e306 (was a share button, not new chat)
2. ✅ CORRECT: Clicking "New chat" link in sidebar or using Cmd+J
3. ❌ WRONG: Assuming input ref is always e102
4. ✅ CORRECT: Using snapshot to get fresh refs each time
5. ❌ WRONG: Extracting all paragraphs including system messages
6. ✅ CORRECT: Filtering out AI-generated disclaimers and thought processes

⚠️ Important Note About Element References:

Never rely on fixed ref IDs like e306, e102, etc. These IDs are:

• Generated dynamically by the browser snapshot tool
• Can change between snapshots or page reloads
• Different for each user and session

Always use semantic selectors or get fresh refs from snapshot:

🧠 DeepThink Mode Optimization

The DeepThink button (labeled "DeepThink" to the left of the "Search" button) enables enhanced reasoning mode. When activated, DeepSeek provides more detailed, step-by-step responses.

How to Enable DeepThink Mode

// Function to toggle DeepThink mode
async function toggleDeepThinkMode(tabId, enable = true) {
  await browser({
    action: "snapshot",
    profile: "openclaw",
    targetId: tabId
  });
  
  // Check current DeepThink state
  const currentState = await browser({
    action: "act",
    kind: "evaluate",
    profile: "openclaw",
    targetId: tabId,
    fn: `(function() {
      const deepThinkBtn = document.querySelector('button[aria-label*="DeepThink"], button:has(> div:contains("DeepThink"))');
      const searchBtn = document.querySelector('button[aria-label*="Search"], button:has(> div:contains("Search"))');
      
      return JSON.stringify({
        hasDeepThink: !!deepThinkBtn,
        hasSearch: !!searchBtn,
        deepThinkActive: deepThinkBtn ? deepThinkBtn.getAttribute('aria-pressed') === 'true' || 
                                        deepThinkBtn.classList.contains('active') || 
                                        deepThinkBtn.getAttribute('data-active') === 'true' : false,
        deepThinkText: deepThinkBtn ? deepThinkBtn.innerText || deepThinkBtn.textContent : '',
        searchText: searchBtn ? searchBtn.innerText || searchBtn.textContent : ''
      });
    })()`,
    returnByValue: true
  });
  
  const state = JSON.parse(currentState);
  console.log("DeepThink state:", state);
  
  // If we want DeepThink enabled but it's not active
  if (enable && state.hasDeepThink && !state.deepThinkActive) {
    console.log("Enabling DeepThink mode...");
    
    // Click the DeepThink button
    const clicked = await browser({
      action: "act",
      kind: "evaluate",
      profile

…(truncated)