# Next Session: Phone Conversation Loop — Continue Testing + Voice Personality

**Use this prompt to continue from Session 394 (2026-04-02)**

---

```
Continue from Session 394. Phone conversation loop is WORKING (5 turns, 3 commits deployed on Panda). Two research tasks:

## 1. VOICE PERSONALITY RESEARCH (PRIORITY)
Rajesh wants Annie to have a warm, caring, kind, helpful personality — NOT cold/generic.
The Nemotron Nano responses were correct but impersonal ("I'm a custom-built personal assistant").

Research needed:
- How production voice assistants (Replika, Character.AI, Pi, Hume) achieve warm personality
- Open-source system prompts / persona engineering best practices
- Research papers on conversational AI personality (warmth, empathy, engagement)
- How to tune via system prompt vs fine-tuning vs voice selection
- Current state: prompt_builder uses hardcoded fallback (SOUL.md/USER.md/RULES.md missing on Panda)
- Annie's workspace files need to be created on Panda: ~/.her-os/annie/{SOUL,USER,RULES,TOOLS,MEMORY}.md

## 2. CONTINUE LIVE TESTING
Still untested:
- Barge-in (interrupt Annie mid-speech)
- Tool use (web_search, search_memory)
- Auto-answer daemon mode
- Model singleton fix for daemon (currently reloads models per call)

## What's Working (DO NOT re-implement)
- 5-turn conversation: Whisper base STT → Nemotron Nano LLM → Kokoro TTS
- FrameBroadcaster (independent queues), queue drain, 1.5s echo guard
- Panda infra: .env, pixel-pin, ADB, BT, GPU all verified
- Commits: f62374b, 7654b99, db2ba08 — all deployed on Panda
```