VTech NightLight Integration

Why this exists

The VTech V-Hush Mini is a combined nightlight and sound machine. Out of the box, it is controllable only through the vendor’s mobile app — one Bluetooth connection at a time, no automation hooks, no way to integrate it into a broader smart home. If you want the soother to turn on when a door sensor trips, or dim in stages through the evening on a schedule, you are out of luck.

This integration solves that. It is an unofficial Home Assistant custom component that controls the device locally over Bluetooth LE, bypassing VTech’s cloud and phone app entirely. It exposes power, RGB color, brightness, a toddler lock, a sleep timer, and the device’s on-board weekly routines — all as native Home Assistant entities that participate in automations, scenes, and dashboards. The goal is straightforward: interoperability with hardware you own, built by observing BLE traffic. It ships no vendor software and is not affiliated with VTech.

Engineering approach

The integration spans several HA entity platforms — light, switch, sensors, number, select, and button — built on bleak and bleak-retry-connector for resilient Bluetooth communication that works through ESPHome Bluetooth proxies.

The BLE protocol required non-trivial reverse engineering. The device expects an access-right handshake before it will accept commands. It publishes state changes through BLE notifications, so physical button presses on the device sync back into Home Assistant as local push updates — no polling needed. The implementation handles MTU-aware packet fragmentation and reassembly, bounded write timeouts to prevent hung connections, and background auto-reconnect when the device goes out of range and returns.

The most interesting protocol detail is the routine payload. The device stores a weekly schedule of power/color/sound configurations, and the integration fully encodes and decodes that payload. Critically, it preserves unknown bytes — fields the integration does not yet model are round-tripped intact so that writing a routine back to the device never clobbers data the integration cannot interpret. The protocol was reverse-engineered by capturing the official app’s BLE traffic using an Android HCI snoop log, analyzing it in Wireshark, building a bleak prototype, and diffing to discover the missing handshake. All documentation is written from that primary research, with no vendor source code or binaries reproduced.

Agentic development

This is where the project gets most interesting from a software-engineering perspective. The entire integration was a one-person, AI-orchestrated build. I acted as architect and reviewer; the implementation was driven by AI coding agents operating under a shared operating contract.

A single AGENTS.md file defines the autonomous delivery loop: build before committing, always work on a feature branch, open a draft PR unprompted, monitor CI, push fixes until all checks pass, mark the PR ready for review unprompted, and leave the final merge to me. CLAUDE.md and GEMINI.md are thin pointers to that one source of truth, keeping the contract consistent regardless of which agent is working. The git history shows work genuinely split across multiple agents — Claude Opus, Gemini, and Antigravity — all following conventional-commit discipline with co-author trailers, with me steering scope and reviewing each PR.

The most compelling piece is an agent-driven hardware test loop. An agent deploys integration and dashboard-card changes to a live Home Assistant host, restarts it, then visually verifies the rendered card using Playwright — screenshot comparison plus console-error assertions — before considering the change done. This closes the loop from code change to real device behavior, something that is normally a manual step in IoT development.

Quality is gated by CI on every PR: pytest for unit tests, ruff for linting, bandit for security analysis, plus Home Assistant’s own hassfest and HACS validation workflows to ensure the integration meets community standards. A secret-scanning regime prevents credentials from entering the repository.