Feature Request: Dynamic Tool/Skill Loading (Tool Search Tool pattern)

Summary

Request for dynamic skill loading to reduce context token usage.

Problem

Currently, all configured skills are loaded into the system prompt on every request (~3-5K tokens for 20+ skills). This is wasteful when most requests only need 1-2 skills.

Proposed Solution

Implement a "Tool Search Tool" pattern:

1. Index skills semantically - Create embeddings of skill descriptions
2. Query before loading - On each request, search for relevant skills based on user message
3. Load dynamically - Only inject 2-3 most relevant skills into context

Benefits

• 5-15% token reduction per request (depending on skill count)
• Better scaling as skill library grows
• Aligns with patterns used by Denoland and other agent frameworks

References

• Tool Search Tool Pattern - Deno
• Similar to RAG but for tools instead of documents

Use Case

Power users with 20+ skills configured who want to optimize token consumption without sacrificing capability.

---
Submitted via OpenClaw agent