Lofi

Foundation for an ESP32-S3 lo-fi box that keeps multiple devices in time over ESP-NOW and can be prototyped in a browser first.

Current Shape

• crates/lofi-core: no_std timing, sync, packet, transport, scheduled events, deterministic role-based sample scheduling, fixed-pack lookup, stateless sample playback, and bounded mix DSP.
• crates/lofi-app: no_std device runtime — the code that runs identically on hardware and in the simulator. A Device owns one box's clock, transport, groove, event queue, play state, and LCD model; it renders mono audio and a 128x64 SSD1306 framebuffer from shared mesh state. Neither the firmware nor the simulator re-implements the musical behavior.
• crates/lofi-sim: host simulation kernel (library + WAV bin) that drives real Devices through a simulated ESP-NOW network (drift, loss, latency), a scheduled drop, and a stereo monitor mix. Pan/solo/volume are listener-side controls; a real box just drives one mono speaker.
• crates/lofi-web: allocation-free no_std WebAssembly ABI around one real lofi-app::Device. Every virtual module is a separate WASM instance with independent memory, clock, peer table, and synth state.
• apps/mesh-lab: the sole interactive app, built with Vite, React, and TypeScript. Its AudioWorklet hosts the WASM devices and a bounded ESP-NOW substrate that transports their real encoded beacon/probe frames with simulated disconnects, loss, latency, jitter, offsets, and drift.
• proto/lofi/v1/lofi.proto: protobuf source of truth for mesh, transport, scheduled event, and groove-state messages.
• docs/: product, mesh sync, hardware portability, music engine, and simulator UI notes.

Browser mesh lab

npm install
npm run dev

Open http://localhost:5173/ in the host browser and press Start. The Vite predev hook builds the Rust WASM module and stages it for the app. Three WASM devices start by default. Add/remove modules, disconnect individual modules, change loss/latency/jitter, alter device clock drift, and monitor root election, multi-role assignment, peer count, and sync quality in real time. The lab supports ten modules; every box remains locally musical while the full group keeps distinct rhythm and tonal responsibilities. Sequencing, sample playback, arrangement, filtering, wire encoding, peer tracking, and clock discipline remain in the same no_std runtime intended for firmware. JavaScript only models the radio medium, mixes mono speaker outputs for monitoring, and renders controls.

The music evolves automatically at shared eight-bar phrase boundaries. The settings panel counts down to the next synchronized change; there is no user-facing seed or fixed tape selector.

Open http://localhost:5173/judge for the blinded eight-bar listening study. Its tick/cross judgements, structured tags, and optional notes persist locally and on this development box for reproducible taste-model analysis.

Exact browser-path WAV render

npm run build:web
node tools/listen-qa/render.mjs --seed 2 --nodes 5 --duration 96 \
  --output target/listen-qa/seed-2.wav

This executes the production AudioWorklet and one WASM instance per module, including the simulated mesh substrate and browser listener mix. See Listen QA for the automated and human acceptance gates.

The current groove is sample-only. A fixed 5.95 MiB pack supplies 192 harvested elements. The audible path selects one source-coherent scene per session and deals source-matched, transient-bounded drum one-shots plus grid-conformed bass, harmony, melody, and texture loops across modules. A bounded deterministic arrangement changes rhythm density and stem prominence every eight bars. No oscillator, unrelated sample shuffle, allocator, or mutable playback cursor runs in the audio path.

Hardware Direction

The firmware target should use esp-hal plus esp-radio/ESP-NOW on ESP32-S3 in no_std. Current Espressif Rust radio support exposes ESP-NOW through esp-radio with the esp-now and unstable features. The radio runtime may require esp-alloc/esp-rtos, but the audio engine and musical core should remain fixed-capacity and allocation-free.

ESP-NOW v2 supports 1470-byte payloads, but the current sync frame is 40 bytes so it keeps airtime low and remains comfortably below v1's 250-byte payload size.

Next hardware milestones:

1. Add an lofi-firmware-esp32s3 crate with esp-hal, Embassy timers, ESP-NOW receive/send tasks, LCD, button input, and I2S DAC output.
2. Feed every received ESP-NOW packet into lofi_core::protocol::Frame::decode.
3. Use a monotonic microsecond hardware timer as the local clock source for ClockModel.
4. Schedule audio edges from root time via ClockModel::local_from_root.

Design Docs

• Product Notes
• Mesh Sync
• Hardware Portability
• Music Engine
• Listen QA
• AI Content Pipeline
• Simulator UI
• Commercialization Roadmap