Deep Research Systematic, evidence-based research that produces cited, source-backed reports. Uses the filesystem as working memory -- scraped content, extracted notes, and source metadata are saved to disk rather than held in context. This keeps the context window lean and makes research resumable, searchable, and reusable. Architecture Core principle: Write to disk aggressively, read back selectively. Never hold raw scraped content in context longer than it takes to extract findings. The LLM context should contain only the current working set -- not the entire research corpus. Research Para…