Chat engine CLI prototype — persona + RAG + history → Claude streaming, save messages

[Punkte]

What to build

A working chat engine that assembles a dynamic system prompt (persona card + RAG excerpts + recent history + anti-hallucination guardrails), calls Claude with streaming, and persists the exchange. Delivered as a CLI script so quality can be validated before any UI is built.

Acceptance criteria

• POST /chat accepts { conversationId, message } and streams SSE response
• System prompt includes: persona card, top-5 RAG excerpts, last 20 chat messages, anti-hallucination rules
• Response streams token-by-token via Server-Sent Events
• User and assistant messages saved to chat_messages after each turn
• CLI script (scripts/chat.ts) lets you converse via terminal without a UI
• Manual test: responses match the target person's writing style (length, punctuation, expressions)
• max_tokens: 300 enforced to prevent essay-length replies

Blocked by

• Persona extraction — sample → LLM extraction → merge → store PersonaCard JSON #5
• RAG retriever — vector search endpoint returning top-K chunks for a query #6