Agent-Native Smart Office

Policy-governed smart office demo: Sense → Reason → Decide → Act → Verify → Explain (HVAC, lighting, plug loads).

Planning docs

Doc	Role
docs/plans/Task-Plan-Agent-Native-Smart-Building-Demo.pdf	Original Task-Plan backbone
docs/roadmap.md	Living product/demo plan
docs/structure.md	Why this layout (apps vs packages, hardware, scenarios)

Repository layout (GitHub-facing)

apps/
  demo-web/       # Software demo UI (Vercel / Hugging Face)  ← build first
  demo-edge/      # Local edge runtime for physical path     ← when boards online
packages/         # Shared domain, contracts, policy, agent-core
hardware/         # Inventory, roles, wiring, per-device tests, curated evidence
firmware/         # MCU nodes (later)
scenarios/        # Simulation / demo scenario fixtures
demo/
  physical/       # How to run tabletop / live physical show
  assets/         # Fresh pitch media after retests
scripts/          # Local helpers
docs/             # Plans + structure (this team’s short docs only)

Why not only apps/web? See docs/structure.md. Short version: Turborepo’s apps/web is fine for a single product site; this project has web demo + physical demo + lab evidence, so names are intent-based (demo-web, demo-edge) and non-deployables live outside apps/.

Local-only (gitignored): .venv/, secrets, large models.

What goes where (common asks)

Need	Folder
Web pitch UI + agent interaction view	apps/demo-web
MQTT / board bridge run locally	apps/demo-edge
Board roles (Pi gateway, Jetson vision…)	hardware/roles
Jetson / Pi / camera / MCU test results	hardware/tests/<device>
Scenario JSON for sim agents	scenarios/samples
Physical demo script for presenters	demo/physical
Screenshots after UI works	demo/assets

Next focus

1. Read Task-Plan + docs/roadmap.md + docs/structure.md.
2. Implement apps/demo-web with fixtures from scenarios/.
3. Parallel: fill hardware/inventory + roles, retest boards into hardware/tests/*.
4. Keep Streamlit out — web path only for software demo.

Invariants

• Policy/safety before any physical act.
• LLMs do not dispatch device commands.
• Sent/ACK ≠ verified; label simulated vs physical.
• No measured 10–20% claims without evidence.
• Hardware optional; software demo must stand alone.