Gemini Search Agent

This agent provides advanced web search capabilities using the Gemini CLI in headless mode with tool restriction. All search functionality is handled through the Gemini CLI's google_web_search tool, NOT Claude's internal web search.

The agent uses gemini -p (headless mode) with --yolo (auto-approval) and restricts the Gemini CLI to only use the google_web_search tool via .gemini/settings.json.

Features

• Context isolation for better token savings
• Smart caching with 1-hour TTL
• Auto-retry logic with exponential backoff
• Web search functionality using Gemini CLI with grounded web server
• Usage analytics tracking
• Dynamic content extraction from websites
• False positive validation of results

Search Patterns

• Uses Gemini CLI headless mode: gemini -p "/tool:googleSearch query:\"search query\" raw:true" --yolo --output-format json -m "gemini-2.5-flash"
• Restricts Gemini to only google_web_search tool via settings.json
• Not using Claude's internal web search functionality
• Handles different query types (factual, research, news)
• Processes search results with relevance scoring
• Extracts clean text content from web pages via Gemini
• Validates results against original query
• No direct web scraping or HTTP requests

Usage

/search [query] - Perform a web search using Gemini CLI
/search-stats - View usage statistics
/clear-cache - Clear the search cache

Caching System

• Results cached for 1 hour
• MD5 keying for efficient storage
• Automatic cache cleanup
• Cache hit analytics tracking

Error Handling

• Exponential backoff on failures
• Fallback search engine selection
• Graceful degradation
• Comprehensive logging

Content Extraction

• Extracts clean text from web pages using grounded web server
• Removes HTML tags and formatting
• Validates content relevance
• Handles multiple content sources

Result Validation

• Calculates relevance scores
• Filters false positives
• Provides quality assessment
• Warns for low-relevance content

Important Note

This plugin uses the Gemini CLI in headless mode (gemini -p) with the --yolo flag for automated tool approval. The .gemini/settings.json file restricts the Gemini CLI to exclusively use the google_web_search tool.

What this means:

• All web searches go through Google's web search via Gemini's API
• No Claude internal search functionality is triggered
• No direct web scraping or HTTP requests are made
• Results are grounded in real web content retrieved by Gemini
• The Gemini agent intelligently decides when to use the search tool based on the query