claude-knowledge-sdk

@jadecli/claude-knowledge-sdk

Distilled knowledge from Anthropic's Claude ecosystem — as code, not docs.

A TypeScript SDK + Claude Code plugin that captures the patterns, types, and tooling from Claude Code v2.1.83, the Claude Agent SDK, claude-code-actions, and 8 Anthropic engineering blog posts. Includes 4 Claude Code skills, TypeScript LSP support, CI/CD workflows, and an llms.txt documentation crawler.

What This Is

Instead of reading docs, you import them:

import {
  runLoop,              // Agent loop wrapping SDK query()
  orchestrateResearch,  // Lead + parallel subagent fan-out/fan-in
  recursiveResearch,    // Research → evaluate → improve loop
  parseLlmsTxt,         // Parse llms.txt documentation indexes
  fetchAllKnowledge,    // Fetch docs from 3 Anthropic surfaces
  calculateBudget,      // Context window management
  generateOtelEnvVars,  // OTel configuration generator
  MODEL_PRICING,        // Current model pricing table
} from '@jadecli/claude-knowledge-sdk';

Quick Start

# Install
npm install @jadecli/claude-knowledge-sdk

# Fetch documentation index
npx ck fetch-docs --priority critical

# Run a recursive research query
npx ck research "how does Claude Code handle context compaction"

# Generate OTel monitoring config
npx ck otel-setup --backend prometheus

Claude Code Plugin

This repo is also a Claude Code plugin with 4 skills and TypeScript LSP:

# Load as plugin (dev)
claude --plugin-dir ~/repos/claude-knowledge-sdk-typescript

# Reload after changes
/reload-plugins

# Use skills
/claude-knowledge-sdk:doc-fetcher
/claude-knowledge-sdk:research-loop
/claude-knowledge-sdk:otel-tracker
/claude-knowledge-sdk:llms-txt-crawler

Skills

Skill	Description
doc-fetcher	Local documentation oracle — fetches and caches Anthropic docs
research-loop	Multi-agent recursive research with lead + subagent pattern
otel-tracker	OTel monitoring setup generator for Prometheus/SigNoz/Grafana
llms-txt-crawler	Parses llms.txt files, crawls doc URLs, generates Scrapy spiders

llms.txt Crawler

The llms-txt-crawler skill can:

1. Parse any site's llms.txt file (the LLM-friendly doc index standard)
2. Crawl discovered documentation URLs via WebFetch
3. Generate a full Scrapy spider project for bulk crawling with ClaudeBot user-agent
4. Build a local knowledge index at ~/.claude/knowledge/

Known endpoints: code.claude.com/docs/llms.txt, platform.claude.com/llms.txt

Architecture

Built from 5 distilled sources:

Source	What It Provides
anthropics/claude-code v2.1.83	CLI surface, hooks, skills, plugins, monitoring
@anthropic-ai/claude-agent-sdk v0.2.33	query(), AgentDefinition, subagents, V2 sessions
anthropics/claude-code-action	CI/CD integration, GitHub Actions patterns
anthropics/claude-code-monitoring-guide	OTel + Prometheus stack
anthropic.com/engineering (8 posts)	Multi-agent patterns, context engineering, tool design

CI/CD

Three GitHub Actions workflows are included:

1. CI Gate (ci.yml)

Runs on push to main and PRs: typecheck → build → test → lint. All must pass.

2. Claude PR Review (claude-review.yml)

Uses anthropics/claude-code-action@v1 for automated PR review.

3. Security Review (security.yml)

Uses anthropics/claude-code-security-review@main for AI security scanning.

Required GitHub Secrets

Secret	Source	Used By
CLAUDE_CODE_OAUTH_TOKEN	Run claude setup-token locally	claude-review.yml
CLAUDE_API_KEY	Anthropic API key with Claude Code usage enabled	security.yml

Prerequisites:

• Install the Claude GitHub App on your repo (run /install-github-app in Claude Code)
• Security review is not hardened against prompt injection — only review trusted PRs

Code Standards

Follows Boris Cherny's "Programming TypeScript" discipline:

• Branded types prevent mixing AgentId with SessionId
• Result<T, E> replaces try/catch — no exceptions cross boundaries
• Discriminated unions model every state transition with assertNever()
• noUncheckedIndexedAccess catches undefined array/object access
• Readonly by default — all types use readonly properties

Key Patterns Implemented