# Session 275 Prompt

## Context
Session 273 created `docs/learn-annie-architecture.html` — a 1,279-line learning page mapping Claude Code's architecture onto Annie. It confirmed they share the same while-loop architecture and identified three growth vectors. Session 274 was a no-op (context carried forward).

## What to do

1. **Review the learning page** — Open `docs/learn-annie-architecture.html` in the browser and verify it renders correctly (dark/light themes, all 9 chapters, code blocks, comparison tables, Think About It reveals). Take a screenshot if possible.

2. **Pending from previous sessions:**
   - Deploy to Titan — `ssh titan "cd ~/workplace/her/her-os && git pull"` then restart context-engine
   - Factory reset — `scripts/reset-data.sh --force` on Titan (still pending from session 261)

3. **Growth vectors identified from Claude Code architecture analysis** (pick one to start):
   - **Sub-agent delegation** (biggest gap) — Annie has zero sub-agents. Claude Code has 6 types. Design how Annie could spawn research/draft/memory-dive agents with fresh context windows.
   - **Multi-layer configuration** — Move from 1-2 context layers to 5 (household → user → relationship → personal → session). Enables multi-user households.
   - **Explicit hooks system** — Move from implicit hooks (session start/end) to configurable event-driven hooks like Claude Code's 20+ event types.

## Key files
- `docs/learn-annie-architecture.html` — The new learning page (1,279 lines, 9 chapters)
- `docs/learn-contextual-memory.html` — Reference design system (same aesthetics)
- `services/annie-voice/bot.py` — Annie's Pipecat pipeline (the while loop)
- `services/annie-voice/context_loader.py` — Context assembly (6 parallel sources)
- `services/annie-voice/tools.py` — Tool definitions and handlers
- `services/annie-voice/context_management.py` — Anthropic compaction betas
- `services/annie-voice/text_llm.py` — Text chat with tool-call loop (max 5 rounds)