Skip to content
Open
Show file tree
Hide file tree
Changes from all commits
Commits
File filter

Filter by extension

Filter by extension

Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
45 changes: 43 additions & 2 deletions src/lib/agent/BrowserAgent.prompt.ts
Original file line number Diff line number Diff line change
Expand Up @@ -8,6 +8,16 @@ You are now operating in EXECUTION MODE. You will be provided with:

Your primary responsibility is to interpret each action and translate it into the correct tool calls, executing them within the browser environment.

# ⚠️ CRITICAL: YOUTUBE VIDEO TASKS
**If the current page is YouTube (youtube.com/watch or youtube.com/shorts) AND the task involves video content:**
- Summarize video → MUST use youtube_tool(action='summarize')
- Question about video → MUST use youtube_tool(action='ask', question='...')
- Get transcript → MUST use youtube_tool(action='transcript')

**DO NOT use click/scroll/extract for YouTube VIDEO CONTENT - these tools CANNOT access video transcripts or content.**
Browser automation can only see the page UI (buttons, titles), not the actual video content or transcript.
Only youtube_tool can analyze what's IN the video.

# STEP BY STEP EXECUTION PROCESS

1. **Analyze the context:** Review the user task, current state, execution history, challenges, and reasoning done so far to understand the user's goal. This will give you enough context to understand what has been carried out so far and what should be done next.
Expand Down Expand Up @@ -37,6 +47,8 @@ Your primary responsibility is to interpret each action and translate it into th
- "Scroll to [element]" → LOOK at screenshot, find element's nodeId label, use scroll(nodeId)
- "Press [key]" → use key(key)
- "Extract [data]" → use extract(format, task)
- "Summarize YouTube video" → use youtube_tool(action='summarize')
- "Ask about YouTube video" → use youtube_tool(action='ask', question='...')
- "Submit form" → LOOK at screenshot, find submit button's nodeId label, click(nodeId)
↳ If click fails → use visual_click("submit button description")

Expand Down Expand Up @@ -112,6 +124,7 @@ Tab Control:

Data Operations:
- extract(format, task): Extract structured data matching JSON schema
- youtube_tool(action, question?, videoUrl?): Analyze YouTube videos (requires Gemini API)
- pdf_extract(format, task?, page?, pages?): Extract data from PDF documents with selective PDF.js capabilities
↳ Raw metadata: format={metadata: true} (no LLM cost)
↳ Raw text: format={text: true} (no LLM cost)
Expand Down Expand Up @@ -181,6 +194,18 @@ You do NOT perform actions yourself. Your role is to propose clear, actionable n
- After each round of execution, review the history and updated state, and refine your plan and suggest next steps as needed.
- When the task is fully complete, provide a final answer and set \`taskComplete=true\`. Answer must be grounded based on latest browser state and screenshot.

# ⚠️ CRITICAL: YOUTUBE VIDEO TASK DETECTION
**Before planning ANY actions, check if this is a YouTube video task:**
- Is the URL youtube.com/watch or youtube.com/shorts?
- Does the task ask to: summarize, analyze, get transcript, or ask questions about the VIDEO CONTENT?

**If YES to both → Your plan MUST include: "Use youtube_tool to [action]"**
- Example: "Use youtube_tool to summarize the video"
- Example: "Use youtube_tool to answer the question about [topic]"

**DO NOT propose:** "Click Show transcript", "Extract page content", "Scroll down", "Click More actions"
**WHY:** Browser automation CANNOT access video content - only the page UI (buttons, titles). The youtube_tool uses Gemini's native video understanding to analyze the actual video content and transcript.

# STEP BY STEP REASONING

1. **Analysis of User Query, Execution History and Current/Updated Browser State:**
Expand Down Expand Up @@ -210,9 +235,13 @@ ${toolDescriptions}
- Google Docs: document reading, writing, and formatting
- Notion: note and database management

**Always prefer MCP for these services over browser automation when possible.**
**Always prefer MCP for these services over browser automation when possible.**
Example: Use "Use MCP to search Gmail for unread emails" instead of "Navigate to gmail.com".

**For YouTube video analysis, always use youtube_tool instead of extract or browser automation.**
The extract tool cannot access video content - only youtube_tool can analyze videos.
Example: Use "Use youtube_tool to summarize" instead of "Extract page content".

# PDF HANDLING GUIDANCE

**Detecting PDFs:** If the current page is a PDF (URL ends with .pdf, title contains "PDF", or browser state shows PDF viewer controls like page navigation, zoom, download), use specialized PDF tools instead of general browser automation.
Expand All @@ -236,7 +265,7 @@ Example: Use "Use MCP to search Gmail for unread emails" instead of "Navigate to

**Page Selection:**
- \`page: [3, 5]\` for specific pages
- \`pages: {start: 1, end: 5}\` for ranges
- \`pages: {start: 1, end: 5}\` for ranges
- \`pages: "all"\` for entire document (default)
- **50-Page Limit:** All PDF.js operations are limited to the first 50 pages to prevent resource exhaustion. If a document has more than 50 pages, request only the needed 50 pages as a max of 50 pages will be processed.

Expand Down Expand Up @@ -308,6 +337,18 @@ You do NOT perform actions yourself. Your role is to manage the TODO list, analy
- After each round of execution, review the history and updated state, update the TODO list progress, and refine your plan and suggest next steps as needed.
- When all TODO items are complete, provide a final answer and set \`allTodosComplete=true\`. Answer must be grounded based on latest browser state and screenshot.

# ⚠️ CRITICAL: YOUTUBE VIDEO TASK DETECTION
**Before planning ANY actions, check if this is a YouTube video task:**
- Is the URL youtube.com/watch or youtube.com/shorts?
- Does the task ask to: summarize, analyze, get transcript, or ask questions about the VIDEO CONTENT?

**If YES to both → Your plan MUST include: "Use youtube_tool to [action]"**
- Example: "Use youtube_tool to summarize the video"
- Example: "Use youtube_tool to answer the question about [topic]"

**DO NOT propose:** "Click Show transcript", "Extract page content", "Scroll down", "Click More actions"
**WHY:** Browser automation CANNOT access video content - only the page UI (buttons, titles). The youtube_tool uses Gemini's native video understanding to analyze the actual video content and transcript.

# STEP BY STEP REASONING

1. **Analysis of User Query, Execution History, TODO Progress and Current/Updated Browser State:**
Expand Down
2 changes: 2 additions & 0 deletions src/lib/agent/BrowserAgent.ts
Original file line number Diff line number Diff line change
Expand Up @@ -53,6 +53,7 @@ import {
GetSelectedTabsTool,
DateTool,
MCPTool,
YouTubeTool,
PdfExtractTool,
} from "@/lib/tools";
import { GlowAnimationService } from '@/lib/services/GlowAnimationService';
Expand Down Expand Up @@ -288,6 +289,7 @@ export class BrowserAgent {
this.toolManager.register(CelebrationTool(this.executionContext)); // Celebration/confetti tool
this.toolManager.register(DateTool(this.executionContext)); // Date/time utilities
this.toolManager.register(BrowserOSInfoTool(this.executionContext)); // BrowserOS info tool
this.toolManager.register(YouTubeTool(this.executionContext)); // YouTube video analysis

// External integration tools
this.toolManager.register(MCPTool(this.executionContext)); // MCP server integration
Expand Down
2 changes: 1 addition & 1 deletion src/lib/tools/PdfExtract.ts
Original file line number Diff line number Diff line change
Expand Up @@ -75,7 +75,7 @@ AI-powered extraction (uses LLM):

pages: z
.union([
z.literal("all"),
z.enum(["all"]), // Using enum instead of literal for Gemini API compatibility (Gemini doesn't support JSON Schema "const" keyword)
z.object({
start: z.number(),
end: z.number()
Expand Down
241 changes: 241 additions & 0 deletions src/lib/tools/YouTubeTool.ts
Original file line number Diff line number Diff line change
@@ -0,0 +1,241 @@
import { DynamicStructuredTool } from '@langchain/core/tools'
import { z } from 'zod'
import { ExecutionContext } from '@/lib/runtime/ExecutionContext'
import { PubSubChannel } from '@/lib/pubsub/PubSubChannel'
import { Logging } from '@/lib/utils/Logging'
import { LLMSettingsReader } from '@/lib/llm/settings/LLMSettingsReader'
import { BrowserOSProvider } from '@/lib/llm/settings/browserOSTypes'

// Input schema
const YouTubeToolInputSchema = z.object({
action: z.enum(['summarize', 'ask', 'transcript'])
.describe('Action to perform: summarize (get video summary), ask (answer question about video), transcript (get full transcript)'),
question: z.string().optional()
.describe('Question to ask about the video (required for "ask" action)'),
videoUrl: z.string().url().optional()
.describe('YouTube video URL (auto-detects from current tab if not provided)')
})

type YouTubeToolInput = z.infer<typeof YouTubeToolInputSchema>

export function YouTubeTool(context: ExecutionContext): DynamicStructuredTool {
return new DynamicStructuredTool({
name: 'youtube_tool',
description: `Analyze YouTube videos using Gemini's native video understanding. Capabilities:

• summarize: Get key takeaways and main points from any video
• ask: Ask specific questions about video content (requires "question" parameter)
• transcript: Get full video transcript with timestamps

AUTO-DETECTION: If no videoUrl provided, analyzes the YouTube video in the current tab.

NOTE: Requires Gemini API key configured in settings.

EXAMPLES:
- "Summarize this YouTube video"
- "What does the video say about artificial intelligence?"
- "Get the full transcript of this video"`,

schema: YouTubeToolInputSchema,

func: async (args: YouTubeToolInput) => {
try {
context.incrementMetric('toolCalls')

// Check if Gemini is available
const geminiAvailable = await _isGeminiAvailable()
if (!geminiAvailable) {
return JSON.stringify({
ok: false,
error: 'YouTube video analysis requires Gemini API. Please configure a Gemini API key in settings to use this feature.'
})
}

// Get video URL
context.getPubSub().publishMessage(
PubSubChannel.createMessage('Analyzing YouTube video...', 'thinking')
)

let videoUrl = args.videoUrl
if (!videoUrl) {
const page = await context.browserContext.getCurrentPage()
videoUrl = page.url()
}

// Validate YouTube URL
const videoId = _extractVideoId(videoUrl)
if (!videoId) {
return JSON.stringify({
ok: false,
error: 'Not a valid YouTube video URL. Please navigate to a YouTube video or provide a video URL.'
})
}

// Process with Gemini
return await _processWithGemini(videoUrl, args, context)

} catch (error) {
context.incrementMetric('errors')
Logging.log('YouTubeTool', `Error: ${error instanceof Error ? error.message : String(error)}`, 'error')
return JSON.stringify({
ok: false,
error: `YouTube tool failed: ${error instanceof Error ? error.message : String(error)}`
})
}
}
})
}

// ============= Helper Functions =============

/**
* Check if Gemini is available and configured
*/
async function _isGeminiAvailable(): Promise<boolean> {
try {
const config = await LLMSettingsReader.readAllProviders()
const geminiProvider = config.providers.find((p: BrowserOSProvider) => p.type === 'google_gemini' && p.apiKey)
return !!geminiProvider
} catch (error) {
return false
}
}

/**
* Process YouTube video using Gemini's native support via REST API
*/
async function _processWithGemini(
videoUrl: string,
args: YouTubeToolInput,
context: ExecutionContext
): Promise<string> {
try {
// Get Gemini configuration
const config = await LLMSettingsReader.readAllProviders()
const geminiProvider = config.providers.find((p: BrowserOSProvider) => p.type === 'google_gemini' && p.apiKey)

if (!geminiProvider || !geminiProvider.apiKey) {
throw new Error('Gemini API key not found')
}

// Prepare the prompt based on action
let promptText: string
switch (args.action) {
case 'summarize':
promptText = `Please summarize this YouTube video in 3 sections:
1. **Main Topic**: What is this video about? (1-2 sentences)
2. **Key Takeaways**: The most important points (3-5 bullet points)
3. **Notable Insights**: Any particularly interesting or valuable information`
break

case 'ask':
if (!args.question) {
return JSON.stringify({
ok: false,
error: 'Question is required for "ask" action. Please provide a question parameter.'
})
}
promptText = `Answer this question about the YouTube video: "${args.question}"

Provide a clear, direct answer based on the video content. If the answer includes specific quotes or references, include them.`
break

case 'transcript':
promptText = `Extract and provide the full transcript of this YouTube video with timestamps if available.`
break

default:
throw new Error('Invalid action')
}

context.getPubSub().publishMessage(
PubSubChannel.createMessage('Processing with Gemini...', 'thinking')
)

// Use Gemini REST API directly with proper fileData format
const modelId = geminiProvider.modelId || 'gemini-2.5-flash'
const apiUrl = `https://generativelanguage.googleapis.com/v1beta/models/${modelId}:generateContent`

const requestBody = {
contents: [{
parts: [
{ text: promptText },
{ file_data: { file_uri: videoUrl } }
Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

logic: incorrect file_data format - Gemini API expects YouTube URLs to be passed differently

The current code passes the YouTube URL directly as file_uri, but Gemini's API doesn't accept YouTube URLs in the file_data.file_uri field. That field is for files uploaded via the File API. For YouTube videos, you should pass the URL directly in a text part or use a different approach.

Check Gemini documentation for correct YouTube video analysis format.

Prompt To Fix With AI
This is a comment left during a code review.
Path: src/lib/tools/YouTubeTool.ts
Line: 163:163

Comment:
**logic:** incorrect `file_data` format - Gemini API expects YouTube URLs to be passed differently

The current code passes the YouTube URL directly as `file_uri`, but Gemini's API doesn't accept YouTube URLs in the `file_data.file_uri` field. That field is for files uploaded via the File API. For YouTube videos, you should pass the URL directly in a text part or use a different approach.

Check Gemini documentation for correct YouTube video analysis format.

How can I resolve this? If you propose a fix, please make it concise.

]
}],
generationConfig: {
temperature: 0.3,
maxOutputTokens: 2000
}
}

const response = await fetch(apiUrl, {
method: 'POST',
headers: {
'Content-Type': 'application/json',
'x-goog-api-key': geminiProvider.apiKey
},
body: JSON.stringify(requestBody)
})

if (!response.ok) {
const errorText = await response.text()
Logging.log('YouTubeTool', `Gemini API error: ${response.status}`, 'error')
throw new Error(`Gemini API error: ${response.status} - ${errorText}`)
}

const data = await response.json()

// Extract the text from Gemini's response
const responseText = data.candidates?.[0]?.content?.parts?.[0]?.text

if (!responseText) {
throw new Error('No text content in Gemini response')
}

return JSON.stringify({
ok: true,
output: args.action === 'transcript'
? { transcript: responseText }
: args.action === 'ask'
? { answer: responseText }
: { summary: responseText }
})

} catch (error) {
Logging.log('YouTubeTool', `Gemini processing error: ${error}`, 'error')
throw error
}
}

/**
* Extract video ID from various YouTube URL formats
*/
function _extractVideoId(url: string): string | null {
try {
// youtube.com/watch?v=VIDEO_ID
const watchMatch = url.match(/[?&]v=([^&]+)/)
if (watchMatch) return watchMatch[1]

// youtu.be/VIDEO_ID
const shortMatch = url.match(/youtu\.be\/([^?&]+)/)
if (shortMatch) return shortMatch[1]

// youtube.com/embed/VIDEO_ID
const embedMatch = url.match(/youtube\.com\/embed\/([^?&]+)/)
if (embedMatch) return embedMatch[1]

// youtube.com/v/VIDEO_ID
const vMatch = url.match(/youtube\.com\/v\/([^?&]+)/)
if (vMatch) return vMatch[1]

// youtube.com/shorts/VIDEO_ID
const shortsMatch = url.match(/youtube\.com\/shorts\/([^?&]+)/)
if (shortsMatch) return shortsMatch[1]

return null
} catch (error) {
Logging.log('YouTubeTool', `Error extracting video ID: ${error}`, 'error')
return null
}
}
1 change: 1 addition & 0 deletions src/lib/tools/index.ts
Original file line number Diff line number Diff line change
Expand Up @@ -35,3 +35,4 @@ export * from './GetSelectedTabsTool'
export * from './GroupTabsTool'
export * from './BrowserOSInfoTool'
export * from './DateTool'
export * from './YouTubeTool'