The chat was hallucinating install commands, losing job context across follow-ups, and sometimes returning cards without useful conversational text. This update gives the AI stronger product/data rails while keeping GPT-4o/OpenClaw as the conversational layer.

The site already had feed UI, Firecrawl configuration, and DB-aware article detail code, but there was no crawler/writer pipeline that could discover sources, generate original Agent Mag coverage, record ingestion metadata, and publish validated articles. /read and the public machine-readable surfaces were still mostly tied to MDX, so generated articles would not consistently show up where readers, search engines, or AI crawlers expect them.

The workflow also needed a clear operator notification loop. Without a status email, you have to manually inspect GitHub Actions or database documents to know whether scraping, writing, and publishing actually worked.