tool-search（工具搜索）

References (archive): SCAFFOLD_SKILLS_ARCHIVE_MAP.md — embedding/semantic tool discovery from everything-claude-code backend-patterns, tdd-workflow.

Tool Search Skill

Identity

Tool Search - Provides semantic tool discovery using embeddings to scale from dozens to thousands of tools with 90%+ context reduction.

Capabilities

• Semantic Tool Search: Find relevant tools based on task context
• Embedding-Based Matching: Use embeddings for accurate tool discovery
• On-Demand Loading: Load tools only when needed
• Context Efficiency: 90%+ reduction in tool definition tokens

The Problem

Traditional tool loading:

• All tools loaded upfront
• 58 tools = ~55K tokens
• Context fills quickly
• Hard to scale beyond ~100 tools

The Solution

Tool Search with Embeddings:

• Only Tool Search Tool loaded initially (~500 tokens)
• Tools discovered on-demand via semantic search
• 3-5 relevant tools loaded per search (~3K tokens)
• Total: ~8.7K tokens vs. ~77K traditional (85% reduction)

How It Works

1. Initial State: Only Tool Search Tool + critical tools loaded
2. Tool Discovery: Agent searches for tools based on task
3. Semantic Matching: Embeddings match tools to task context
4. Tool Expansion: Matching tools expanded into full definitions
5. Tool Use: Agent uses discovered tools

Configuration

MCP Configuration (.claude/.mcp.json)

{
  "betaFeatures": ["advanced-tool-use-2025-11-20"],
  "toolSearch": {
    "enabled": true,
    "autoEnableThreshold": 20,
    "defaultDeferLoading": true
  },
  "mcpServers": {
    "repo": {
      "deferLoading": true,
      "alwaysLoadTools": ["search_code", "read_file"]
    },
    "github": {
      "deferLoading": true,
      "alwaysLoadTools": ["create_pull_request", "get_issue"]
    }
  }
}

Always Load Critical Tools

Keep 3-5 most-used tools always loaded:

• Core file operations: read_file, write_file, search_code
• Essential integrations: create_pull_request, get_issue
• Frequently used: take_screenshot, navigate_page

Usage Patterns

When to Use Tool Search

Most Beneficial When:

• Tool definitions consuming >10K tokens
• Tool library has 10+ tools
• Experiencing tool selection accuracy issues
• Building MCP-powered systems with multiple servers

Less Beneficial When:

• Small tool library (<10 tools)
• All tools used frequently in every session
• Tool definitions are compact

Tool Discovery

Agent Workflow:

1. Agent needs capability (e.g., "create a pull request")
2. Agent searches: "github pull request creation"
3. Tool Search returns: create_pull_request tool
4. Tool expanded into full definition
5. Agent uses tool

Example:

User: "Create a pull request for my changes"

Agent searches: "github pull request creation"
Tool Search finds: create_pull_request tool
Tool loaded and used

Best Practices

1. Clear Tool Names and Descriptions

Good:

{
  "name": "search_customer_orders",
  "description": "Search for customer orders by date range, status, or total amount. Returns order details including items, shipping, and payment info."
}

Bad:

{
  "name": "query_db_orders",
  "description": "Execute order query"
}

2. System Prompt Guidance

Add guidance in agent prompts:

You have access to tools for Slack messaging, Google Drive file management,
Jira ticket tracking, and GitHub repository operations. Use the tool search
to find specific capabilities when needed.

3. Keep Critical Tools Always Loaded

Don't defer loading for:

• Core file operations
• Essential integrations
• Frequently used tools

4. Monitor Tool Usage

Track which tools are discovered:

• Most searched tools
• Tool discovery patterns
• Context savings achieved

Implementation

Embedding-Based Tool Search

The tool search uses embeddings to match tools to queries:

1. Tool Indexing: Create embeddings for all tool definitions
2. Query Embedding: Create embedding for user query
3. Similarity Search: Find tools with similar embeddings
4. Tool Expansion: Load matching tools into context

Tool Search Tool

The Tool Search Tool itself:

• Searches tool library semantically
• Returns relevant tools based on query
• Expands tools into full definitions
• Maintains tool index

Benefits

Context Efficiency

• 85% reduction in tool definition tokens
• 52.5% total context (down from 87%)
• Within optimal range (60-70% target)

Improved Accuracy

• 11% improvement in tool selection accuracy
• 79.5% → 88.1% (Opus 4.5)
• Better tool matching for complex queries

Scalability

• Scales to thousands of tools
• No context limit concerns
• Dynamic tool discovery

Examples

Example 1: GitHub Operations

User: "Create a pull request"

Agent workflow:
1. Searches: "github pull request creation"
2. Tool Search finds: create_pull_request tool
3. Tool loaded (3K tokens)
4. Agent uses tool
5. Total context: ~8.7K tokens (vs. 55K traditional)

Example 2: File Operations

User: "Search for authentication code"

Agent workflow:
1. Searches: "code search file operations"
2. Tool Search finds: search_code, read_file tools
3. Tools loaded (5K tokens)
4. Agent uses tools

Example 3: Multiple Integrations

User: "Check Slack messages and create Jira ticket"

Agent workflow:
1. Searches: "slack message reading"
2. Tool Search finds: read_slack_message tool
3. Searches: "jira ticket creation"
4. Tool Search finds: create_jira_ticket tool
5. Both tools loaded (6K tokens total)

Integration

With MCP Servers

Tool search works with MCP servers:

• GitHub MCP: 35 tools → 3-5 loaded on-demand
• Slack MCP: 11 tools → 2-3 loaded on-demand
• Custom MCPs: Any number of tools → Loaded as needed

With Agent System

All agents benefit from tool search:

• Reduced context usage
• Better tool selection
• Scalable tool libraries

Troubleshooting

Tools Not Found

• Check tool names and descriptions are clear
• Verify tool search is enabled
• Review search queries
• Check tool index is up to date

Context Still High

• Verify deferLoading is enabled
• Check alwaysLoadTools list (should be minimal)
• Review tool definitions (may be too verbose)
• Monitor actual tool usage

Tool Selection Issues

• Improve tool descriptions
• Add more context to search queries
• Review tool naming conventions
• Check embedding quality

Integration with Programmatic Tool Calling (PTC)

Tool Search works excellently with Programmatic Tool Calling:

1. Tool Search finds relevant tools (on-demand loading)
2. PTC orchestrates tools efficiently (reduced context)
3. Result: Optimal tool usage with minimal token consumption

Example Workflow:

# Tool Search finds tools
tools = search_tools("github issue management")

# PTC orchestrates multiple tool calls
team = await get_team_members("engineering")
issues = await asyncio.gather(*[
    get_issue(member["github_username"]) for member in team
])
# Only final results in context, not all intermediate data

See PTC Patterns Guide for comprehensive PTC documentation.

Related Documentation

• Advanced Tool Use - Comprehensive tool use guide
• PTC Patterns - Programmatic Tool Calling patterns
• Context Optimization - Context management

References

• Tool Search with Embeddings Cookbook
• Programmatic Tool Calling Cookbook
• Advanced Tool Use Documentation

# Search for git-related tools
node .claude/tools/tool_search.mjs --query "git"

# Search for database tools
node .claude/tools/tool_search.mjs --query "database" --limit 3

# Search for testing tools
node .claude/tools/tool_search.mjs --query "testing"

</usage_example>

Memory Protocol (MANDATORY)

Before starting: Read .claude/context/memory/learnings.md

After completing:

• New pattern -> .claude/context/memory/learnings.md
• Issue found -> .claude/context/memory/issues.md
• Decision made -> .claude/context/memory/decisions.md

ASSUME INTERRUPTION: If it's not in memory, it didn't happen.