@skillbolt/search

Intelligent skill search for Skillbolt. Finds task-relevant skills using LLM-guided multi-level tree traversal.

Features

• Multi-level tree descent — Navigates capability tree with LLM selection at each level
• Parallel branch search — Explores multiple branches concurrently
• Skill pruning — Deduplication and workflow-stage ranking
• Early stop optimization — Stops when enough relevant skills are found
• Event system — Real-time progress callbacks

Quick Start

import { Searcher, createSearchConfig } from '@skillbolt/search';
import { loadTree } from '@skillbolt/tree';
import { createLLMAdapter } from '@skillbolt/core';

const llm = createLLMAdapter();
const tree = await loadTree('tree.yaml');
const config = createSearchConfig({ maxSkills: 10 });
const searcher = new Searcher({ llm, tree, config });

const result = await searcher.search('create a landing page');
for (const skill of result.selectedSkills) {
  console.log(`${skill.name}: ${skill.description}`);
}
console.log(`LLM calls: ${result.llmCalls}`);

CLI

skill skill-search "create a promotional video for my research paper"
skill skill-search "fix login bug" --max-skills 5 --json

API

| Export | Description | | -------------------- | --------------------------------------------- | | Searcher | Main search class with recursive tree descent | | createSearchConfig | Create search configuration | | SearchResult | Query results with selected skills and stats | | SearchEvent | Progress event types for real-time tracking | | createEventEmitter | Simple callback-based event system |

Configuration

| Option | Default | Description | | -------------- | ------- | -------------------------- | | maxParallel | 4 | Concurrent branch searches | | pruneEnabled | true | Enable dedup and ranking | | temperature | 0.3 | LLM sampling temperature | | maxSkills | 10 | Maximum skills to return |

License

MIT