Research Agent Session Notes (2026-08-01)

Concepts

• Skill ≈ reusable capability module; under the hood usually extra instructions in context + optional tools/schemas.
• Tooling for terminal: use child_process.spawn / execFile, not raw exec for untrusted input.
• Sandboxing: for AI-generated or user-influenced commands, run inside Docker/VM, not directly on host.

Images

• node:20-bookworm-slim: Debian 12, glibc, apt, ~200 MB with Node.
• node:alpine: Alpine Linux, musl, apk, smaller (~130–150 MB with Node), more compatibility caveats.
• Bun: oven/bun (e.g. oven/bun:1-debian, oven/bun:canary).

Project direction

• Build a Node/TS service:
  • Core: queue, browser pool (Playwright), extraction, SQLite vault.
  • API: internal methods/HTTP routes.
  • MCP: expose search_web, fetch_url, browse_page, save_source, query_vault, etc.
  • Auth: API key between agent/client and service.
• Existing work: browser MCPs + deep-research MCPs exist; room for a polished Node-first, research+browser-focused product.

Next steps

• Decide: core service only, or core + MCP wrapper.
• Draft folder layout:
  • src/server.ts, src/orchestrator.ts, src/queue.ts
  • src/workers/{browser,fetch,analysis}-worker.ts
  • src/browser/playwright.ts
  • src/db/{schema,index}.ts