MaTriXy/Agent-Registry

• **Archetype**: agent-optimization-tool
• **Stars**: 8
• **Last pushed**: 2026 (v2.0.1)
• **License**: MIT
• **Discovered**: 2026-04-13
• **Skills found**: 1 (lazy-loading agent registry)

Summary

A lazy-loading system for Claude Code agents that reduces context window usage by 70-90% through on-demand loading rather than eager loading. Features smart BM25 + keyword matching search engine for intent-based agent discovery, interactive migration UI with multi-level selection, and lightweight index architecture. Real-world test with 140 agents showed context reduction from 16.4k to 2.7k tokens (83% decrease). Includes UserPromptSubmit hook for automatic agent discovery.

Assessment

HIGH VALUE for babysitter process discovery optimization. The lazy-loading pattern and context reduction techniques directly apply to babysitter's process library management. The BM25 search algorithm, automatic agent discovery via prompt analysis, and token efficiency optimization are highly relevant. The shift from eager to lazy loading with metadata-only indexes provides a proven pattern for scaling large process libraries without context window exhaustion.

Extraction Priority

• High
• Rationale: The lazy-loading pattern, BM25 search implementation, and context optimization techniques are directly transferable to babysitter's process discovery. The automatic intent analysis and confidence-based matching (score ≥ 0.5) provide sophisticated process recommendation capabilities. Real-world validation with 83% token reduction proves effectiveness.

Processes

• **Lazy-Loading Process Discovery**: Replace eager loading with metadata-only indexes for large process libraries
• **Intent-Based Process Recommendation**: Analyze user prompts to automatically suggest relevant processes
• **BM25 Search for Process Discovery**: Implement BM25 + keyword matching for intelligent process search
• **Context Window Optimization**: Reduce token usage through selective loading and metadata caching
• **Confidence-Based Process Matching**: Score process relevance and auto-inject high-confidence matches

Plugin Ideas

None - the functionality is best implemented as core babysitter enhancement rather than external plugin.

Patterns

• Lazy loading with metadata-only indexes
• BM25 + keyword search for relevance scoring
• Automatic prompt analysis for intent detection
• Multi-level selection UI with keyboard navigation
• Content hash-based change detection
• Confidence thresholding (≥ 0.5) for auto-injection
• Graceful fallback to text mode

Library Mapping

Plugin Marketplace Mapping