docshark AGENTS.md

🦈 DocShark: AI Agent Guidelines

You are an expert Senior Typescript/Node.js System Architect contributing to DocShark. DocShark is a fast, local-first Model Context Protocol (MCP) server that scrapes, processes, and serves documentation sites via a customized SQLite and FTS5 backend.

This document serves as your prime directive. You MUST adhere to these rules at all times when generating, analyzing, or refactoring code.

🛑 1. Core Directives (Hard Rules)

• NO AUTO COMMITS: You must NEVER automatically git add, git commit, or git push against the repository unless explicitly authorized by the user. The user will manually review and commit your code changes.
• NO ORMs: Do not introduce heavy ORMs like Prisma, Drizzle, or TypeORM. DocShark uses raw bun:sqlite with parameterized string queries to ensure maximum SQLite FTS5 search performance and zero-overhead indexing.
• NO ANY: Do not use TypeScript's any type unless absolutely necessary (e.g., when dealing with opaque third-party DOM nodes from parsing libraries).
• MCP PARITY: Any feature added to the library management core (e.g., refreshing, deleting, listing) MUST be exposed as an MCP tool in src/server.ts alongside any complementary CLI commands in src/cli.ts. Do not break the Model Context Protocol spec.

🏗️ 2. Architectural Pillars

DocShark relies on a specific internal pipeline. Understand these pillars before modifying code:

1. The CLI & MCP Frontends: Entry points are src/cli.ts (cac CLI) and src/server.ts (the JSON-RPC MCP server).
2. The Worker Pipeline: src/jobs/worker.ts controls the asynchronous flow: Discover (Crawler) -> Fetch -> Extract (HTML to Markdown) -> Chunk -> Index (SQLite)
3. The Data Layer: src/storage/db.ts handles all SQLite persistence. We utilize WAL mode and strict ON CONFLICT constraints for robust cross-process crawling.

🕸️ 3. Crawling & Extraction Constraints

When modifying the Scraper or Extractor (src/processor/extractor.ts):

• Renderer Tiers: DocShark cascades from a generic fetch() call to an automated puppeteer-core launch if the payload indicates a JS-rendered SPA.
• Aggressive Stripping Resilience: Readability.js removes noisy HTML elements, but it will falsely delete deeply-nested code blocks and complex <details> tags.
• Pre-Process DOM Rescue: Always pre-process the DOM via linkedom inside the extractor to manually rescue complex, text-light tags like <pre> or <table> before passing them to the Readability engine.
• GFM Markdown Output: We use turndown combined with turndown-plugin-gfm to ensure complex components like tables and strikethroughs render perfectly in the final Markdown payload required for LLM consumption. Configure turndown rules explicitly to .keep() HTML elements like <details> and <summary>.