[AI] Tool selection at scale — pre-filter tools before LLM #20

Problem

The ReAct agent currently sends ALL registered tool descriptions to the LLM on every request. At 14 tools (~2K tokens) this works fine. At 30+ tools, the tool list alone eats most of the 4096-token context window, leaving no room for reasoning.

Goal

Implement a pre-filtering layer that narrows 30+ tools to the 3-5 most relevant before passing them to the LLM.

Proposed Approach (RLM-inspired)

1. Phase 1: Use existing keyword matching + ability descriptions to score and rank candidate tools (no LLM cost, instant)
2. Phase 2: Inject only selected tool schemas into the ReAct system prompt
3. Phase 3: Add a search-tools meta-ability so the LLM can request additional tools mid-loop if needed

Key Files

• src/extensions/services/message-router.js — existing keyword matching
• src/extensions/services/react-agent.js — system prompt construction
• src/extensions/services/chat-orchestrator.js — tool list building

Context

• See CLOUDFEST_HACKATHON.md "Architecture & Research > Tool Selection at Scale" for full design
• See docs/ai-fundamentals/11-tool-selection-at-scale.md
• The keywords array and description field on every ability are the existing index

Skills needed

AI/ML, JS