docs: JARVIS spec simplification research

[bdkaat]

Summary

• Research analysis evaluating the JARVIS build spec against AO's existing architecture
• Identifies MVP features vs unnecessary complexity across 6 areas: real-time (Convex vs SSE), supervisor (LangGraph vs TypeScript), data layer (PostgreSQL+pgvector+Redis vs flat files), subprocess orchestration, auth, and self-improvement
• Proposes a simplified architecture delivering 80% of the value at 20% of the complexity — ~7 days of incremental AO improvements instead of a ground-up rewrite with 10+ new dependencies

Key Findings

• Skip Convex: SSE already works, Convex creates architectural mismatch with flat-file storage
• Skip LangGraph: ~120 lines of TypeScript replaces the entire Python supervisor graph
• Skip PostgreSQL + pgvector + Redis: Flat files handle tens of sessions; SQLite is the natural upgrade path
• Skip FastAPI: No reason to add Python to a TypeScript codebase
• Defer auth: Localhost-only tool doesn't need Auth.js v5 + RBAC
• Defer self-modification: Use prompt refinement (lessons-learned.md) instead

Test plan

• Review research analysis for accuracy and completeness
• Validate that proposed simplified architecture covers JARVIS spec requirements
• Check that recommended phase ordering is sound

🤖 Generated with Claude Code