ruvnet/RuView

This system is a comprehensive, Rust-based platform for contactless human sensing and spatial intelligence using WiFi Channel State Information (CSI), mmWave radar, and other RF modalities. It provides a full-stack solution that includes low-level signal processing, neural network inference for pose estimation and vital sign monitoring, and a privacy-preserving architecture with cryptographic verification of sensor data and model accuracy. The system integrates with smart home ecosystems like Home Assistant and HomeKit, offering both edge-computing capabilities via WebAssembly modules and a robust cloud/backend infrastructure for data aggregation, visualization, and AI-assisted development workflows.

The \references\ directory now contains a complete, production-ready reference implementation of the InvisPose system, which enables real-time full-body human pose estimation through walls using commodity WiFi hardware. This addition includes the core neural network architecture (CSI phase sanitization, modality translation network, and DensePose-RCNN integration), a FastAPI-based REST and WebSocket API for real-time data streaming, and an interactive web dashboard with live signal visualization and hardware configuration panels. The package also provides Python scripts for model training and performance benchmarking, alongside comprehensive documentation detailing hardware requirements, deployment steps, and application use cases in healthcare, retail, and security.

references · high confidence

Added a Python-based evaluation script (\run\_zeroshot\_eval.py\) and its supporting data/metric utilities (\bidmc\_windows.py\) to compare zero-shot inference from pretrained models (Amazon Chronos-Bolt and TimesFM) against RuForecast's native naive baselines (last-value, seasonal-naive) on the BIDMC dataset. The script mirrors the evaluation logic, windowing, and metrics of the existing Rust \evaluate\ CLI to ensure apples-to-apples comparison, allowing users to run and compare performance metrics (MAE, WQL) for these external models without fine-tuning.

scripts/ruforecast-zeroshot · high confidence

A new Jinja2 template and its Weaver configuration have been added to the Rust registry templates. This template generates a Rust module containing constants for OpenTelemetry semantic conventions, including schema URLs, attribute keys, and event names, derived from the central telemetry registry. It also includes a test to ensure that instrumentation code only uses literals registered in the semconv registry, ensuring consistency between the generated constants and the actual telemetry calls.

templates · high confidence

New Python scripts in the \scripts/fal-visual-teacher\ directory enable the creation of a synthetic visual teacher LoRA for RuView's room visualization aesthetic. \extract\_frames.py\ pulls frames from existing fal.ai-generated room visualization videos to build a training dataset with style-specific captions. \train\_lora.py\ packages this data, uploads it to fal.ai, and submits a job to the \fal-ai/flux-lora-fast-training\ endpoint to generate the LoRA weights. \validate.py\ then performs a test inference using the trained model to confirm it produces the expected visual style (e.g., glowing pose overlays, dark tech aesthetic) without influencing any biometric or sensor data.

scripts/fal-visual-teacher · high confidence

The AetherArena benchmark now uses an append-only, hash-chained ledger (ledger.jsonl) to record results, ensuring that any tampering with past entries is detectable via cryptographic verification. This change introduces the underlying infrastructure (ledger\_tools.py) for maintaining this integrity, alongside the initial population of the ledger with RuView's MM-Fi benchmark scores across in-domain, cross-subject, and cross-environment protocols.

aether-arena/ledger · high confidence

The archive/v1 location now contains the full WiFi-DensePose v1 API package (version 1.1.0), including the FastAPI application entry point, dependency injection for pose/stream/hardware services, and middleware for JWT authentication and sliding-window rate limiting. It exposes REST endpoints for health/readiness, pose estimation (current, analyze, historical, zone occupancy), and WebSocket streams for real-time pose and event data, with first-message JWT authentication on WebSocket connections to prevent token leakage in URLs.

archive/v1 · high confidence

This update introduces the ESP32-C6-specific capability modules (ADR-110) and the adaptive sensing controller (ADR-081) to the firmware node. The C6 modules include low-power LP-core wake-on-motion hibernation, ESP-NOW-based cross-node time synchronization, antenna selection for the Seeed XIAO C6 board, and a soft-AP with Wi-Fi 6 (HE) support for bench testing. The adaptive controller implements a closed-loop system that dynamically adjusts the CSI capture profile, channel, and mesh role based on radio health and edge-derived motion/presence scores, running on fast, medium, and slow control loops. The build system (CMakeLists.txt) is updated to compile these new sources and their dependencies, with conditional compilation for C6 targets and optional features like AMOLED displays and WASM.

firmware/esp32-csi-node/main · high confidence

The wifi-densepose-wasm-edge crate now ships a unified EdgePipeline that wires approximately 64 edge skills, including new AI Security modules (PromptShield for replay/injection/jamming detection, BehavioralProfiler for anomaly scoring, and AdversarialDetector for signal integrity) and autonomous reasoning modules (PsychoSymbolic inference and SelfHealingMesh topology analysis). A new example (run\_all\_skills) demonstrates the pipeline, and Criterion benchmarks measure host latency for key hot paths. The WASM build is now configured with a strict 128 KB memory limit and reduced target features, and non-finite host floats are sanitized at the WASM↔host boundary to prevent NaN propagation.

v2/crates/wifi-densepose-wasm-edge · high confidence

This change introduces an experimental integration that allows a LiDAR-capable iPhone to act as a geometry sensor for RuView. It includes a native iOS SwiftUI app (requiring iOS 17+) that captures depth data via ARKit's scene depth, and a Node.js WebSocket relay that streams this data to a browser-based point cloud viewer. The native component sends downsampled, compact depth frames (u16 millimeters + confidence) over WebSocket, while the web component decodes and renders the point cloud. This is an experimental feature; physical device validation is not yet complete, and production use requires additional security measures like TLS and authentication.

integrations · high confidence

The new \homecore-automation\ crate introduces a YAML-based automation engine that parses trigger, condition, and action rules and executes them against the HOMECORE event bus. It supports state-changed, time-based, template, and service-call triggers, along with state, numeric, template, and logical (and/or/not) conditions. Actions include service calls, state setting, delays, scene activation, and conditional branching (\choose\). The engine enforces four run modes—Single, Restart, Queued, and Parallel—with an optional \max\ concurrency cap to bound resource usage. To prevent denial-of-service via malicious templates, template rendering is bounded by a 1,000,000-instruction fuel limit and a 64 KiB source length cap, while delay values are safely clamped to prevent panics on invalid inputs.

v2/crates/homecore-automation · high confidence

The homecore-server binary now exposes a backend-for-frontend (BFF) gateway that aggregates Home Assistant-compatible endpoints (rooms, COGs, appliance metrics) and reverse-proxies the calibration service, while also providing a bounded startup restore for entity/device registries and recorder states. It adds an optional HomeKit Accessory Protocol (HAP) server for HomeKit controller integration, a plugin subsystem supporting both native Rust and WebAssembly (via Wasmtime) plugins with signed/unsigned policy controls, and an intent-handling pipeline for voice/automation commands. These changes expand the server's capabilities from a basic state machine to a full integration hub with dashboard aggregation, smart-home bridging, and extensible plugin architecture.

v2/crates/homecore-server/src · high confidence

Added the initial SQL schema for the Claude Flow V3 Memory Database, defining core tables for memory entries, pattern learning with confidence scoring and temporal decay, and learning trajectories. Also added a state configuration file initializing the swarm with a hierarchical topology and specialized strategy.

.swarm · high confidence

This change introduces the first version of the Pose Estimation Cog, a component that subscribes to WiFi CSI data and emits 17-keypoint COCO pose estimation events. The release includes build scaffolding and manifests for both x86\_64 and ARM architectures, along with a pre-trained model artifact (pose\_v1.safetensors). The cog is configured to emit frames by default using a minimum confidence threshold of 0.185, which corresponds to the model's validation PCK@50 score, ensuring that inference output is visible out-of-the-box despite the model currently being below the target accuracy for fine-grained joint detection.

v2/crates/cog-pose-estimation/cog · high confidence

This change introduces the initial project structure for the RuView Desktop application, a cross-platform tool for managing ESP32 WiFi sensing networks built with Tauri v2 and React. The commit adds the core configuration files, including the Tauri capability manifest (default.json) which defines permissions for core operations and file dialogs, the build script, and generated ACL schemas. It also includes a README detailing the application's architecture, UI pages (Dashboard, Nodes, Flash, OTA, etc.), and design system, alongside a daemon-state configuration file for internal tooling.

v2/crates/wifi-densepose-desktop · high confidence

This change introduces the complete source code for the WiFi-DensePose mobile application, built on React Native and Expo. The app provides a cross-platform interface (iOS, Android, and Web) for real-time WiFi sensing, featuring live 3D Gaussian splat visualizations, vital sign monitoring (breathing and heart rate), and a Mass Casualty Assessment (MAT) dashboard. It connects to a sensing server via WebSocket and includes a robust offline fallback that generates synthetic data when the server is unreachable. The release also includes a full suite of unit tests for the app's components, hooks, and services to ensure stability.

ui/mobile · high confidence

Adds the Beamforming Feedback Layer for Detection (BFLD) crate, implementing a privacy-first WiFi sensing pipeline. The system enforces structural invariants to keep raw beamforming data and identity embeddings strictly in-RAM, with automatic zeroization on drop. It features a coherence gate with hysteresis and debouncing to stabilize risk-based decisions (Accept, PredictOnly, Reject, Recalibrate), and applies strict privacy gating to output events, stripping identity-derived fields like risk scores and signature hashes at higher privacy classes. The pipeline includes Home Assistant MQTT auto-discovery, availability topic management, and a Soul Signature matcher for identity correlation, all wired through a configurable emitter pipeline.

v2/crates/wifi-densepose-bfld/src · high confidence

This change adds the foundational \homecore\ Rust crate, implementing the core state machine (using a concurrent DashMap-backed store), a typed event bus (split into system and domain channels via Tokio broadcast), and in-memory entity and device registries. It also introduces a \homecore-plugin-example\ crate demonstrating the ADR-128 WASM plugin host ABI, showing how guest plugins can subscribe to state changes and update entity states via the host. The implementation includes benchmarks for the state machine hot paths and strict entity ID validation to prevent memory DoS.

v2/crates/homecore · high confidence

Adds the HOMECORE-UI frontend for the Cognitum Appliance, providing a single-page application with a top navigation bar, a left-side sub-navigation for nine sections (Dashboard, SEED Fleet, Entities, Rooms, COGs, Calibration, Events, Audit, and Settings), and a hash-based router. The UI includes a design system with CSS tokens and app styles, a JavaScript API client that communicates with the homecore-server BFF gateway (with a dev-only mock layer for demo mode), and a shared WebSocket event bus for real-time updates. The dashboard features panels for appliance health, SEED fleet status, entity states, room sensing, COG runtime, calibration wizard, event history, audit logs, and settings.

v2/crates/homecore-server/ui · high confidence

This change adds the \cog-person-count\ component, a learned multi-person counter that replaces the previous heuristic (\subcarrier\_diversity / dedup\_factor\) with a Candle neural network. The cog polls the sensing server's \/api/v1/sensing/latest\ endpoint, runs the \count\_v1.safetensors\ model on CSI windows, and emits calibrated \person.count\ events (including confidence and probability distributions) to stdout. It supports multi-node fusion via confidence-weighted log-sum and includes signed binaries for ARM and x86\_64, along with a configuration schema for runtime parameters like polling interval and model path. The current v0.0.2 model is trained on limited data and should be treated as a working pipeline scaffold rather than a production-ready counter.

v2/crates/cog-person-count/cog · high confidence

This change introduces the \@ruv/ruview-cli\ package, providing a command-line interface for streaming live CSI frames, running single-shot pose and person-count inference, listing edge module cogs, and managing background training jobs. It also refactors the \@ruvnet/rvagent\ MCP server to include a hardened Streamable HTTP transport that enforces origin validation, bearer token authentication, and request body size limits to prevent DoS, while adding intelligent binary detection for Cog executables that respects CPU architecture to ensure the correct binary is used on different appliances.

tools/ruview-mcp · high confidence

Adds the \ruview\ plugin (v0.3.0) to the marketplace, providing Claude Code skills, commands, and agents for end-to-end WiFi-DensePose workflows including ESP32 hardware setup, model training, and sensing applications. The plugin registers the \@ruvnet/rvagent\ MCP server to enable agentic flows and includes a mirrored set of Codex prompts to ensure operator parity across both AI coding assistants.

plugins/ruview · high confidence

The homecore-recorder crate now persists Home Assistant state changes and domain events to a local SQLite database, mirroring the HA recorder schema v48. To prevent memory-DoS, history queries are capped at 1,000,000 rows and startup restores at 100,000 rows. State attributes are deduplicated using an FNV-1a 64-bit hash to reduce storage I/O. The recorder includes a background listener that subscribes to state machine broadcasts and handles lag gracefully. A semantic search index is also provided (via the \ruvector\ feature) using deterministic hash-based embeddings for k-NN similarity search.

v2/crates/homecore-recorder/src · high confidence

This change adds the \wifi-densepose-privshield-harness\ (WiFi Veil), a development aid for the WiFi Veil privacy-shield crate. It provides a dependency-free CLI (\wifi-densepose-privshield-harness\) that includes a read-only \guidance\ surface for capability navigation, an \init\/\doctor\ workflow to verify the \@metaharness/kernel\ and host adapter installation, a cost-optimal \route\ command for model tier selection, and a \flywheel\ command that demonstrates a synthetic self-improvement loop (propose → evaluate → gate → promote) with signed, replayable lineage. The harness is explicitly scoped to development: it does not run a WiFi radio, does not emit RF, and does not replace the crate's own validation gates (\cargo test -p wifi-densepose-privshield\).

harness/wifi-densepose-privshield · high confidence

This change adds the \wifi-densepose-privshield\ crate, which implements a compliant-waveform defense called WiFi Veil (VEIL). The shield shapes a device's outgoing beamforming feedback using keyed Givens rotations to prevent unauthorized passive sniffers from re-identifying individuals, while preserving link throughput for legitimate receivers. The crate provides a deterministic, dependency-free synthetic model to demonstrate the defense, including a compliance module that verifies the shield is energy-preserving and non-jamming. It also ships a custom terminal harness (\veil\) for interactive visualization and optimization of the shield's parameters, alongside a web-based console dashboard.

v2/crates/wifi-densepose-privshield · high confidence

The \wifi-densepose-core\ crate now provides the foundational data structures, traits, and utilities for the WiFi-DensePose system. It defines core types such as \CsiFrame\, \PoseEstimate\, and \Keypoint\, along with a comprehensive error hierarchy (\CoreError\, \SignalError\, etc.). The crate introduces a physics-constrained pose refinement system (\PoseRefinementV1\) that applies additive corrections based on physical constraints (e.g., bone lengths, floor penetration) and tracks provenance via \PhysicsProvenance\. It also establishes a canonical frame contract (\CanonicalFrame\) for deterministic serialization and includes trait abstractions for signal processing, neural inference, and data storage.

v2/crates/wifi-densepose-core · high confidence

This change introduces three new crates that establish the foundational data models for the RuView perception substrate. The \ruview-ontology\ crate defines the canonical \Site ▸ Building ▸ Floor ▸ Space ▸ Zone\ containment hierarchy and leaf entities (Sensor, Person, Object, Observation, Track, Event), enforcing single-parent invariants and carrying mandatory evidence levels and provenance. The \ruview-attest\ crate implements authenticated sensor identity (ADR-305) by modeling the chain of custody from device to signed measurement, including replay and freshness checks. The \ruview-evidence\ crate provides an append-only accuracy ledger (ADR-304) that records per-context performance metrics and enforces a strict evidence ladder (L0–L5) to ensure accuracy claims are traceable and cannot be silently upgraded.

v2/crates/ruview-attest, v2/crates/ruview-evidence, v2/crates/ruview-ontology · high confidence

The \ruforecast-autogenous-bridge\ crate adds a local-development verification layer that cryptographically signs and verifies RuForecast hyperparameter candidates before promotion. By requiring signed receipts from independent judges on synthetic corpora and enforcing a non-inferiority margin on weighted quantile loss, this bridge ensures that only models proven to beat their parent baselines are promoted, adding a defense-in-depth check to the existing Darwin gate.

v2/crates · high confidence

The new \homecore-migrate\ tool allows users to import Home Assistant entity and device registries, configuration entries, secrets, and automations into HOMECORE storage. The CLI provides \import-entities\, \import-devices\, and \import-config-entries\ commands to convert HA \.storage\ JSON files into HOMECORE-compatible formats, and \inspect-\*\ commands to preview data without writing. The tool ensures data safety by using atomic writes that refuse to overwrite existing destination files unless the \--force\ flag is used, and it includes security hardening by redacting secret values from error messages during parsing.

v2/crates/homecore-migrate · high confidence

The \cog-person-count\ component replaces the previous slot-heuristic approach with a learned inference engine that outputs a categorical distribution over person counts (0–7) alongside a calibrated confidence score. It introduces multi-node fusion logic that combines per-node predictions using confidence-weighted log-sums, ensuring that higher-confidence nodes have greater influence on the final result. To prevent untrustworthy multi-occupant counts, predictions exceeding the trained class range (0–1) are flagged as \low\_confidence\ and clamped in reported values. The CLI now emits a real, signed manifest with valid \binary\_sha256\ and Ed25519 signatures, and the runtime publishes structured \person.count\ events including 95% confidence intervals and raw probability distributions for downstream consumers.

v2/crates/cog-person-count/src · high confidence

The \wifi-densepose-wifiscan\ crate now provides a unified BSSID acquisition layer with platform-specific adapters: a native \wlanapi.dll\ FFI scanner for Windows that reads the driver's cached BSS list directly (replacing the slower \netsh\ subprocess and enabling higher polling rates), a \netsh\ fallback for compatibility, and new Linux (\iw\) and macOS (\CoreWLAN\ via Swift helper) scanners. It introduces a \BssidRegistry\ for tracking access points with running RSSI statistics, a \MultiApFrame\ for aggregating multi-AP signal data into a single timestamped snapshot, and a pluggable \WlanScanPort\ trait to decouple the scanning mechanism from the sensing pipeline.

v2/crates/wifi-densepose-wifiscan · high confidence

The new \nvsim\ crate provides a deterministic, forward-only simulation pipeline for NV-diamond ensemble magnetometers, allowing users to synthesize magnetic-field traces without physical hardware. It models the full sensing path—from scene primitives (dipoles, current loops, ferrous objects) and material attenuation to NV-ensemble physics and ADC digitization—emitting fixed-layout \MagFrame\ binary records. The simulator guarantees byte-identical outputs for the same inputs and seed, supported by a SHA-256 proof-bundle system that detects any drift in physics constants, PRNG streams, or frame formats.

v2/crates/nvsim · high confidence

Added a reference calibration service in the \aether-arena/calibration\ directory that fits tiny (\~11 KB) LoRA adapters to adapt a shared WiFi-CSI pose model to specific rooms or subjects. The service includes Python scripts (\calibrate.py\, \cog\_calibrate.py\, \infer.py\) and a model definition (\model.py\) supporting two distinct model architectures: a transformer-based MM-Fi model (producing \.npz\ adapters) and a conv+MLP 'cog' model (producing \.safetensors\ adapters for the Rust runtime). The addition is accompanied by self-contained end-to-end regression tests (\test\_calibration.py\, \test\_cog\_calibration.py\) that verify the calibration pipeline, adapter size, and inference correctness on synthetic data.

aether-arena/calibration · high confidence

The new \cog-pose-estimation\ Cog now provides 17-keypoint pose estimation from WiFi CSI data, introducing a runtime configuration system that allows users to set a \min\_confidence\ threshold to control frame emission (defaulting to the model's typical confidence of 0.185 to prevent silent suppression of events). Additionally, the inference engine supports an optional \--adapter\ CLI flag to load per-room LoRA calibration adapters, enabling users to recover high-accuracy pose tracking in unseen rooms or for specific individuals by applying low-rank deltas to the model's pose head.

v2/crates/cog-pose-estimation/src · high confidence

Adds a new \ruview-pointcloud\ binary that fuses real-time WiFi CSI data (from ESP32 nodes) with camera depth estimation to generate a unified 3D point cloud. The service exposes an HTTP viewer (defaulting to loopback \127.0.0.1:9880\) and supports CLI commands for capturing PLY files, running demos, and recording CSI fingerprints. It includes a cross-platform camera capture module, a CSI pipeline for motion and vital sign detection, and a brain-bridge module to periodically sync spatial observations to an external agent.

v2/crates/wifi-densepose-pointcloud/src · high confidence

The \harness/homecore\ directory now contains the \homecore\ developer metaharness, a bounded CLI and MCP server that maps Homecore capabilities to reviewed source, tests, and ADRs. It exposes read-only MCP tools for guidance, diagnostics, and memory search, while reserving verification and workspace writes for explicit local CLI execution. The harness is WASM-first, loading the \@metaharness/kernel\ WebAssembly backend before falling back to native or JavaScript, and enforces a default-deny tool policy with strict bounds on execution. It also includes a local 'brain' corpus of reviewed knowledge, provenance manifest generation and verification scripts, and host-specific configuration templates for Claude Code and Codex.

harness/homecore · high confidence

The new \wifi-densepose-nn\ crate provides the neural network inference engine for WiFi-based DensePose estimation, supporting ONNX Runtime (default), PyTorch via \tch-rs\, and Candle backends with optional CUDA and TensorRT acceleration. It includes a DensePose head for 24-part body segmentation and UV regression, a modality translator for CSI-to-visual feature mapping, and an RF encoder with seven multi-task heads (pose, presence, count, activity, vitals, gait, identity) featuring per-head predictive uncertainty. The ONNX backend implements zero-copy input handling for contiguous tensors and guards against OOM from unresolved dynamic dimensions, while benchmarks are provided for native convolution, ONNX inference, and concurrency.

v2/crates/wifi-densepose-nn · high confidence

The new \wifi-densepose-vitals\ crate provides a four-stage pipeline to extract heart rate and respiratory rate from ESP32 Channel State Information (CSI). It includes a preprocessor for static component suppression, extractors for breathing (0.1–0.5 Hz) and heart rate (0.8–2.0 Hz) using bandpass filtering and autocorrelation, and an anomaly detector for clinical alerts like apnea or tachycardia. The implementation uses \VecDeque\ for O(1) sliding window performance and includes a ground-truth module for time-aligning and evaluating vitals against reference devices.

v2/crates/wifi-densepose-vitals · high confidence

The \ruview-auth\ crate now provides the core authentication infrastructure for RuView, enabling it to act as an OAuth resource server that verifies Cognitum access tokens offline. This includes a JWKS cache that tolerates network outages on Pi-class hardware by serving cached keys when WAN is lost, and a strict verification policy that rejects long-lived setup/workload tokens, enforces ES256 signatures, and uses \client\_id\ as the audience claim. For users, this enables a new interactive sign-in flow (\wifi-densepose login\) that supports both browser-based PKCE login and an out-of-band paste fallback for headless environments, with credentials securely stored and automatically refreshed.

v2/crates/ruview-auth · high confidence

This change adds three new crates to the RuView perception substrate. The \ruview-ood\ crate implements the ADR-300 staleness guard, classifying inferences as KNOWN, DEGRADED, or UNKNOWN based on calibration certificate validity, signal quality, and domain drift; it suppresses confident classifications when the domain is UNKNOWN. The \ruview-witness\ crate provides an append-only, hash-linked audit chain that records the reasoning stages of an inference (from RF observation to policy decision) to ensure tamper-evidence and provenance. The \ruview-scorecard\ crate introduces a multi-domain benchmarking model that tracks performance per operating domain (e.g., room-known, unseen) with confidence intervals, enforcing a promotion gate that prevents models from advancing if any specific domain's performance regresses.

v2/crates/ruview-ood, v2/crates/ruview-scorecard, v2/crates/ruview-witness · high confidence

This change adds a new alerting subsystem and a complete REST API for the WiFi-DensePose MAT module. The alerting system automatically generates, dispatches, and manages alerts based on survivor triage status, vital signs, and location data, including features like escalation timeouts, priority calculation, and mass-casualty assessment. The new REST API exposes endpoints for managing disaster events, scan zones, survivors, and alerts, along with a WebSocket stream for real-time updates on survivor detections and alert status changes. This provides rescue teams with a structured interface to monitor ongoing operations and receive critical notifications.

v2/crates/wifi-densepose-mat/src · high confidence

This change adds the \ruview-certify\ crate, which implements signed, expiring capability certificates (ADR-318) that bind specific capabilities (Presence, Pose) to a room, hardware, and calibration context, ensuring claims are bounded and verifiable rather than unconditional. It also adds the \ruview-policy\ crate, which serves as the authorization gate for governed actions (ADR-321, ADR-327), enforcing fail-closed decisions based on certificate validity, domain state (Known/Degraded/Unknown), and assurance requirements, while introducing a governed intent and approval workflow for execute-mode actions.

v2/crates/ruview-policy · high confidence

This change adds the \wifi-densepose-engine\ crate, which serves as the composition root for the sensing pipeline. It implements a \StreamingEngine\ that fuses multistatic signal data, enforces privacy controls (demoting privacy classes when contradictions are detected), and records a cryptographic witness for every output to ensure traceability of signal evidence, model versions, and calibration IDs. Additionally, it introduces a \MeshGuard\ component that dynamically monitors the coupling graph of sensing nodes using a min-cut algorithm to detect partition risks and trigger recalibration recommendations when the mesh structure degrades.

v2/crates/wifi-densepose-engine/src · high confidence

The system now combines multiple sensor observations (e.g., WiFi, BLE, mmWave) into a single, probabilistic world state using inverse-variance weighting, which sharpens confidence when sources agree and explicitly flags irreconcilable conflicts or insufficient coverage. Additionally, a new ground-truth validation plane allows comparing these RF estimates against independent reference sensors (such as wearables or cameras) to generate graded evidence reports, ensuring that the system's accuracy claims are backed by measurable agreement rather than inference alone.

(repo-wide) · high confidence

The pose-fusion UI has been rebranded as RuView and rebuilt from the ground up to support a new dual-modal (video + WiFi CSI) pose estimation pipeline. The demo now features a high-fidelity CanvasRenderer that draws detailed skeleton overlays with distinct styling for fingers, toes, and the neck, alongside real-time CSI amplitude heatmaps and 2D embedding space visualizations. Under the hood, the system uses a CnnEmbedder that leverages RuVector WASM attention mechanisms (with a JS fallback) to extract features from both video frames and CSI pseudo-images. A FusionEngine dynamically weights the video and CSI embeddings based on signal quality (brightness, motion, SNR) to produce a fused pose estimate, which is then decoded into 26 keypoints by the PoseDecoder. The CsiSimulator handles both live WebSocket connections (with ticket-based auth and verified-frame gating to prevent false 'LIVE' states) and synthetic demo data that correlates with video motion, including through-wall persistence. The main orchestration (main.js) ties these components together, auto-connecting to local sensing servers and providing a comprehensive UI with confidence bars, latency tracking, and cross-modal similarity metrics.

ui/pose-fusion/js · high confidence

The Observatory UI is now available, providing a 3D visualization of WiFi sensing data using Three.js. This interface features a dark, cinematic theme with a HUD overlay that displays vital signs, signal metrics, and scenario status. It supports 12 distinct sensing scenarios (such as fall detection, vital signs monitoring, and occupancy tracking) with auto-cycling and manual selection. The visualization includes animated wireframe human figures, RSSI waveform graphs, phase constellation plots, and post-processing effects like bloom and vignette, all configurable via an in-app settings dialog.

ui/observatory · high confidence

The dashboard now provides a complete, browser-based interface for the nvsim simulator, featuring a dark/light theme, a responsive 4-zone layout (rail, topbar, sidebar, main, inspector, console), and a home view for new users. Key capabilities include a live console with a REPL for running simulations and verifying SHA-256 witnesses, an App Store for managing 65+ hot-loadable WASM edge modules, and a 'Ghost Murmur' research view that audits NV-diamond heartbeat detection claims against physics literature. The dashboard also includes a help center, a debug HUD, and supports both local WASM execution and optional WebSocket transport to a host server.

dashboard · high confidence

The \harness/ruview\ directory now provides a complete operator harness for the RuView WiFi-sensing system, enabling AI agents (Claude Code and Codex) to onboard hardware, provision ESP32 nodes, calibrate rooms, and train pose models. This release introduces a strict honesty guardrail that prevents overstating accuracy by requiring all performance claims to be tagged as MEASURED, CLAIMED, or SYNTHETIC and compared against a mean-pose baseline. The harness includes a CLI and MCP server with a default-deny tool policy, ensuring that sensitive operations like hardware flashing or workspace modifications require explicit confirmation and grants. It also adds support for Cognitum Spaces via OAuth for reading spatial data, a source-cited 'brain' for repository guidance, and a 'flywheel' system for automated, human-gated evolution of the harness's own policies.

harness/ruview · high confidence

The AetherArena spatial-intelligence benchmark is now available as a public Hugging Face Space, providing a vendor-neutral, tamper-evident leaderboard for RF/WiFi sensing models. The interface displays a hash-chained witness ledger of scored results (starting with RuView's CSI-Transformer entries) and allows users to filter by category (pose, presence). The Space is configured to run on Python 3.12 with Gradio 5.9.1, ensuring compatibility and stability for the public-facing benchmark display.

aether-arena/space · high confidence

A new JSONL file, data/clone-data.rvf, has been added to store long-term records of repository clone and view counts, preserving data beyond GitHub's native 14-day retention window. The file contains metadata and daily snapshots for the period of May 5–18, 2026, providing a historical baseline for traffic analysis.

data · high confidence

The \homecore-hap\ crate introduces a native Rust implementation of the HomeKit Accessory Protocol (HAP) R2, replacing the previous \hap\ 0.1 pre-release dependency. This change provides a fail-closed network foundation with authenticated cryptographic pairing (SRP-6a and X25519/Ed25519) and encrypted transport. It exposes 11 specific accessory types (including Lightbulb, Switch, Sensors, Lock, and SecuritySystem) by mapping HOMECORE entities to HAP characteristics, enabling secure discovery and control via mDNS and a bounded TCP server.

v2/crates/homecore-hap · high confidence

This change introduces a suite of new configuration files for the \.claude\ agent system, defining specialized AI agents for code quality analysis, system architecture design, web browser automation, and distributed consensus protocols. The new agents include a Code Quality Analyzer for reviewing code smells and technical debt, a System Architecture Designer for high-level design decisions, a Browser Agent for web automation tasks, and a set of consensus coordinators (Byzantine, Gossip, Raft, Quorum, CRDT, Security, and Performance) for managing distributed system integrity and synchronization. These files establish the triggers, capabilities, tool permissions, and behavioral constraints for these autonomous agents within the development workflow.

.claude · high confidence

Added two Rust examples for the Beamforming Feedback Layer for Detection (BFLD) to help users integrate the feature. The \bfld\_minimal\ example demonstrates the core operator flow: constructing a pipeline with a signature hasher, feeding sensing inputs and identity embeddings, and receiving privacy-gated events as JSON. The \bfld\_handle\ example illustrates the recommended production pattern using a worker thread, showing the full lifecycle from publishing availability and discovery payloads to driving frames via a handle and performing graceful shutdown.

v2/crates/wifi-densepose-bfld/examples · high confidence

This location introduces the build infrastructure and configuration for the new ESP32 CSI Node firmware (ADR-018), supporting both ESP32-S3 and ESP32-C6 targets. The CMakeLists.txt now reads the version from version.txt to ensure the binary description matches the release tag. The firmware is configured to require ESP-IDF v5.4+ and includes specific SDK configurations for 4MB, 8MB, and 16MB flash layouts, as well as overlays for display-less boards (DevKitC-1) and XIAO C6 antenna selection. A new provision.py script allows users to write WiFi credentials and node configuration to the device's NVS partition without recompiling, using an additive-by-default merge strategy to prevent accidental overwrites of existing settings.

firmware/esp32-csi-node · high confidence

Added a new firmware project in firmware/esp32-hello-world that performs a comprehensive hardware capability discovery on boot. The application probes and prints details for chip info, memory (internal DRAM, PSRAM, DMA), flash storage, partition tables, and WiFi capabilities (including CSI support). It is configured to support both ESP32-S3 and ESP32-C6 targets, with the default configuration set for ESP32-C6 (4MB flash, WiFi CSI enabled). The build system uses ESP-IDF v6+ compatible CMake files and includes a detailed sdkconfig with SOC\\ definitions for the C6 target.

firmware/esp32-hello-world · high confidence

Added a suite of shell scripts in scripts/gcp to provision, run, and teardown GCP instances for three distinct workloads: OccWorld retraining on an 8× A100 cluster, Cosmos-Transfer2.5-2B evaluation on a single A100 80GB instance, and ruview-swarm MARL training on an L4 instance. The scripts handle instance creation with zone fallbacks, environment setup (Conda/Rust), model weight downloads, remote execution of training/evaluation jobs via SSH, and checkpoint/result synchronization back to the local machine.

scripts/gcp · high confidence

A new frontend example has been added at examples/frontend, providing a dark-only web UI built with Lit 3, TypeScript, and Vite. This example includes a REST and WebSocket API client (HomecoreClient) that mirrors the homecore-api JSON shapes, along with Lit components for the application shell, entity state cards, entity forms, and modals. It features a simple client-side router for Dashboard, States, Services, and Settings pages, and includes unit tests for the API client, state card rendering, and design token validation.

examples/frontend · high confidence

Introduces the \cog-ha-matter\ cog, a Seed-installable wrapper around the ADR-115 MQTT publisher that enables automatic Home Assistant discovery via mDNS (advertising \\_ruview-ha.\_tcp\) and provides an optional embedded MQTT broker. The cog includes a tamper-evident audit log built on an Ed25519-signed, SHA-256 hash-chained witness system to ensure state transitions are verifiable, and supports privacy mode to strip biometric data from published semantic primitives.

v2/crates/cog-ha-matter/src · high confidence

This crate introduces the RuVector v2.0.4 integration layer for the WiFi-DensePose signal-processing pipeline and Multi-AP Triage (MAT) module, implementing seven integration points including subcarrier partitioning, spectrogram gating, and triangulation. It adds deterministic benchmarks for Approximate Nearest Neighbor (ANN) search performance (linear scan vs. HNSW) and Coordinate Remote Viewing (CRV) pipeline stages, alongside coverage measurement harnesses for the RaBitQ sketch estimator.

v2/crates/wifi-densepose-ruvector · high confidence

The examples directory now includes a unified RuView Live dashboard (ruview\_live.py) that fuses WiFi CSI and mmWave sensor data to display real-time vitals, stress, and environment metrics, alongside dedicated scripts for sleep apnea screening, stress monitoring, and blood pressure estimation. This release also introduces a new web UI frontend built with Lit 3 and TypeScript, providing a Home Assistant-style interface for the sensing stack.

examples · high confidence

The \v2/crates/wifi-densepose-signal/src/ruvsense\ directory now contains a comprehensive set of new modules that implement the core signal-processing pipeline. This includes \calibration.rs\ for empty-room baseline calibration with support for HT20/HT40/HE20/HE40 PHY tiers, \cir.rs\ for Channel Impulse Response estimation via ISTA sparse recovery, and \coherence.rs\/\coherence\_gate.rs\ for computing per-link coherence scores and applying threshold-based gating to Kalman filter updates. Additional modules provide \array\_coordinator.rs\ for node admission and directional evidence, \adversarial.rs\ for detecting spoofed or injected signals, \attractor\_drift.rs\ for longitudinal biomechanical drift analysis, and \cross\_room.rs\ for maintaining identity continuity across rooms. These files collectively establish the foundational signal ingestion, validation, and feature-extraction logic for the RuvSense subsystem.

v2/crates/wifi-densepose-signal/src/ruvsense · high confidence

The ui/components directory now includes a suite of new JavaScript modules that implement the core user interface for the WiFi-DensePose application. DashboardTab provides a system health overview with real-time data source indicators (distinguishing between live hardware, simulated, and invented data) and feature status. LiveDemoTab manages the primary pose detection experience, including auto-start logic, WebSocket connection handling, and integration with the PoseDetectionCanvas for rendering skeletons, heatmaps, and trails. SensingTab introduces a 3D Gaussian-splat visualization for WiFi CSI data, powered by Three.js, alongside an HUD for RSSI, motion, and classification metrics. Supporting components include HardwareTab for antenna/CSI simulation, ModelPanel and TrainingPanel for managing ML models and training jobs, SettingsPanel for configuration, and TabManager for accessible navigation. These changes introduce new interactive capabilities and visualizations for monitoring and controlling the sensing system.

ui/components · high confidence

The \ui/services\ directory now contains a comprehensive set of JavaScript services that centralize communication with the backend. \api.service.js\ provides a unified HTTP client with request/response interceptors and automatic bearer token handling. \websocket.service.js\ and \websocket-client.js\ manage WebSocket connections, including ADR-272 ticket-based authentication and reconnection logic. \sensing.service.js\ handles the sensing data stream with explicit state management for live, simulated, and reconnecting modes. Supporting services include \pose.service.js\ for pose estimation streams, \data-processor.js\ for transforming API data into visualization formats, \health.service.js\ for system monitoring, \model.service.js\ for model lifecycle management, \stream.service.js\ for streaming controls, and \training.service.js\ for training and recording operations. A test file \websocket.service.test.mjs\ validates the ticket exchange and token stripping logic.

ui/services · high confidence

This change introduces a suite of new client-side utility modules in ui/utils to enhance the user interface. It adds an Activity Log for real-time system event monitoring, a Command Palette for keyboard-driven navigation and actions, and a Data Source Banner to clearly distinguish between live, simulated, and synthetic data. Additional utilities include a Backend Detector for automatic mock server fallback, a Connection Status widget, Data Export functionality, and support for internationalization (English/Polish), mobile navigation, and idle management.

ui/utils · high confidence

This change introduces a comprehensive benchmarking suite in the \benchmarks\ directory. It adds \edge-latency/RESULTS.md\ and \edge-skills/RESULTS.md\ to document host-measured steady-state inference latencies for wasm-edge skills and cogs, alongside synthetic-ground-truth validation results for edge skills, explicitly distinguishing these from unmeasured ESP32 hardware targets and cold-start metrics. It also adds the \wiflow-std/\ directory containing Python scripts (\eval\_repro.py\, \quantize\_bench.py\, \onnx\_bench.py\, etc.) and a detailed \RESULTS.md\ that reproduce, refute, and optimize the WiFlow-STD (DY2434) WiFi-pose SOTA model, including dataset corruption detection, retraining verification, and ONNX/INT8 edge optimization benchmarks.

benchmarks · high confidence

Added three new Python example scripts demonstrating contactless sensing capabilities. The room environment monitor (examples/environment/room\_monitor.py) fuses WiFi CSI, mmWave radar, and light sensor data to track occupancy, vital signs, and ambient conditions. The sleep apnea screener (examples/sleep/apnea\_screener.py) uses mmWave respiratory rate data to detect apnea and hypopnea events, estimating the Apnea-Hypopnea Index (AHI). The stress monitor (examples/stress/hrv\_stress\_monitor.py) calculates Heart Rate Variability (HRV) metrics like SDNN and RMSSD from mmWave heart rate data to estimate real-time stress levels.

examples/environment, examples/sleep, examples/stress · high confidence

Added a new example demonstrating contactless guest sentiment analysis using WiFi Channel State Information (CSI) from ESP32-S3 nodes coordinated by a Cognitum Seed appliance. The package includes a JSON schema defining the 8-dimensional happiness vector (covering metrics like gait, fluidity, and social energy), a Python query tool for kNN search and drift monitoring, and a shell script for provisioning multi-node swarms. This enables users to deploy privacy-preserving, radio-wave-based sentiment tracking without cameras or microphones.

examples/happiness-vector · high confidence

The \examples/through-wall\ area now includes a complete, self-contained in-browser WiFlow trainer (\wiflow\_browser.html\) that performs a 4-stage gated flow (Calibrate, Capture, Train, Infer) using TensorFlow.js and MediaPipe Pose entirely in the browser, ensuring the training and inference camera frames align. It also adds a live sensing demo (\index.html\) that visualizes real-time WiFi CSI motion, presence, and coarse localization from an ESP32-S3, and a live pose inference viewer (\pose.html\) that overlays WiFi-inferred skeletons on a camera feed. Supporting tools include a Python capture script (\wiflow\_capture.py\) for generating paired CSI-pose datasets, an A/B validation script (\wiflow\_ab.py\) to test for temporal leakage, and a threaded static server (\serve.py\) to host the pages.

examples/through-wall · high confidence

Added two new Python scripts in the examples/medical directory that enable contactless vital sign monitoring using a 60 GHz mmWave radar (Seeed MR60BHA2) connected to an ESP32-C6. The bp\_estimator.py script estimates blood pressure trends in real-time by analyzing heart rate variability (HRV) metrics like SDNN and LF/HF ratio, supporting both uncalibrated trend tracking and calibrated absolute estimates. The vitals\_suite.py script provides a broader 10-capability suite including continuous heart rate and breathing rate monitoring, HRV stress analysis, sleep stage classification, apnea event detection, cough and snoring detection, activity state recognition, and a meditation quality scorer. A comprehensive README documents the hardware requirements, usage instructions, accuracy expectations, and medical disclaimers for these research and wellness-tracking tools.

examples/medical · high confidence

A new server component has been introduced to expose the nvsim simulation pipeline via a REST API and a binary WebSocket stream. Users can now manage simulation state (scenes, configuration, seeds) and control execution (run, pause, step, reset) through HTTP endpoints, while receiving real-time MagFrame data via WebSocket. The server also provides endpoints for generating and verifying deterministic witnesses and exporting proof bundles, ensuring byte-identical results across WASM and server transports. A Dockerfile is included to containerize the service, exposing port 7878 with health checks.

v2/crates/nvsim-server · high confidence

The \wifi-densepose-calibration\ crate introduces a complete per-room calibration pipeline. It defines a guided enrollment protocol (\enrollment.rs\, \anchor.rs\) that captures specific human postures (standing, sitting, lying, breathing, moving, sleeping) against an empty-room baseline, applying a quality gate to reject poor captures. The system extracts statistical features (\extract.rs\) and trains a versioned \SpecialistBank\ (\bank.rs\) containing small models for presence, posture, breathing, heartbeat, restlessness, and anomaly detection. It also records transceiver geometry (\geometry.rs\) and generates a deterministic geometry embedding (\geometry\_embedding.rs\) to condition future models, and produces signed, versioned room-fingerprint certificates (\certificate.rs\) to track calibration provenance and detect baseline drift.

v2/crates/wifi-densepose-calibration · high confidence

The examples/research-sota directory has been reorganized into nine thematic folders, each containing Python scripts and READMEs that document specific research findings. The 01-physics-floor folder introduces scripts for Time-of-Arrival Cramer-Rao Lower Bound calculations and multi-scatterer Fresnel zone modeling, quantifying the performance gap between idealized and realistic human body models. The 02-placement folder adds scripts for 3D antenna placement strategies, demonstrating that ceiling-only mounting fails to cover floor-level targets and identifying optimal anchor heights for vital-signs and pose-estimation use cases.

examples/research-sota · high confidence

Added the \wifi-densepose-sar\ crate, a standalone research module that simulates synthetic stepped-frequency multi-position RF measurements and reconstructs a 3D reflectivity field using delay-and-sum backprojection. This crate provides the core reconstruction primitive for a handheld through-wall RF imaging device, including a forward measurement simulator, a voxel-grid reconstruction kernel, and a point-cloud extraction step. It is explicitly a synthetic, evidence-level L0 research tool and does not include hardware drivers or real-world calibration. The implementation features an optimized backprojection kernel that uses incremental phasor rotation to achieve a measured \~4.4-4.5x speedup over the initial version, and includes comprehensive physics validation tests that verify range/cross-range resolution and pose-error sensitivity against closed-form radar formulas.

v2/crates/wifi-densepose-sar · high confidence

This change introduces three new Rust crates that form the foundation for RuView's closed-loop, information-driven sensing capabilities. The \ruview-active\ crate provides a synthetic, L0 planning model for a closed-loop RF experiment controller, allowing the system to propose controllable measurement configurations (channel, bandwidth, cadence, antenna aperture) based on per-zone uncertainty and last response, with a first-class handling of unknown states. The \ruview-counterfactual\ crate implements a generative spatial model that scores scene hypotheses (e.g., occupant counts and positions) against observed RF link measurements using an RF twin, returning a best-explanation or a first-class UNKNOWN verdict. Finally, the \ruview-infogain\ crate introduces a value-of-information scheduler that ranks candidate sensor actions by their predicted uncertainty reduction relative to their modeled resource costs (compute, energy, bandwidth), enabling efficient allocation of scarce edge resources.

(repo-wide) · high confidence

The repository has been reorganized with the root directory flattened to v2. This change introduces a new Rust library, ruvector-crv, which integrates the Coordinate Remote Viewing (CRV) protocol into the ruvector ecosystem by mapping its six-stage signal line methodology to vector database subsystems (using hyperbolic embeddings, multi-head attention, GNNs, and spiking neural networks). Additionally, a new HTML example, mat-dashboard.html, has been added to v2/examples to demonstrate a WiFi-Mat Disaster Response Dashboard interface.

v2/examples, v2/patches, v2/patches/ruvector-crv · high confidence

The new \ruview-offaxis\ crate implements a clean-room, head-coupled off-axis perspective projection (ADR-324) in Rust, exposing a WASM surface for browser demos. It provides asymmetric frustum and screen-aligned view matrices based on physical screen corners and eye position, alongside a One-Euro filter for tracking-noise smoothing and a Tier B coarse-parallax stage that maps RF signal-field peaks to bounded eye offsets. The crate is dependency-free on native targets and includes benchmarks for the per-frame hot path.

v2/crates/ruview-offaxis · high confidence

Added a suite of new scripts to support the WiFi-DensePose and WiFlow sensing pipelines. This includes \align-ground-truth.js\ to time-align camera keypoints with CSI recordings for supervised training, and \apnea-detector.js\ for pre-screening breathing disorders from RF data. Several benchmarking tools were introduced: \benchmark-model.py\ for ONNX model performance, \benchmark-rf-scan.js\ for RF scan metrics, \benchmark-rtl8721dx-csi.py\ for RTL8721Dx hardware throughput, and \benchmark-ruvllm.js\ and \benchmark-wiflow.js\ for specific model architectures. Additionally, \c6-presence-watcher.py\ bridges ESP32-C6 presence data to HomeKit, and \calibrate-camera-room.py\ assists in 3D room calibration for pose training.

scripts · high confidence

The wifi-densepose-signal crate now includes a suite of new modules for CSI preprocessing and feature extraction, including hardware normalization (ADR-027), conjugate multiplication for CFO/SFO cancellation, Hampel filtering for outlier removal, and feature extraction for amplitude, phase, and Doppler profiles. Additionally, new binaries \calibration\_proof\_runner\ and \cir\_proof\_runner\ have been added to verify the deterministic, cross-platform reproducibility of the empty-room baseline calibration (ADR-135) and the CIR estimator (ADR-134) by comparing SHA-256 hashes of their outputs against expected values.

v2/crates/wifi-densepose-signal/src · high confidence

A new set of five browser-based demos has been added to examples/three.js, showcasing the ADR-097 sensing-helpers scene with increasing complexity. The suite includes basic helper visualizations, a cinematic camera with pseudo-CSI overlays, GLTF skinned mesh animation blending, and an FBX-based Mixamo X Bot loader. The final demo introduces a real-time pipeline that uses MediaPipe Pose to drive a live IK retargeting of the Mixamo character, with an optional live overlay from an ESP32 CSI feed. A local Python server is provided to serve these demos with appropriate no-cache headers, and the repository now includes a .gitignore to manage large Mixamo assets and diagnostic screenshots.

examples/three.js · high confidence

The \wifi-densepose-train\ crate introduces a complete training pipeline for WiFi-DensePose, including dataset loading (MM-Fi and deterministic synthetic), subcarrier interpolation, and a training orchestrator gated by the \tch-backend\ feature. It adds a benchmark harness for bfee parsing, split assignment, and leakage auditing, alongside a Python script for int8 quantization of the WiFlow-STD model. The crate implements an ablation evaluation harness that measures privacy leakage (membership inference) and latency, and introduces a metric-locked pose-accuracy system that explicitly handles PCK normalization ambiguities (torso-diameter vs. bounding-box vs. absolute). Additionally, it provides an AetherArena score runner with witness chains for reproducible benchmarking and a training verification binary that uses deterministic proofs to ensure model learning.

v2/crates/wifi-densepose-train · high confidence

The \wifi-densepose-physics\ crate now provides a kinematic physics engine that refines raw pose observations by enforcing anatomical and environmental constraints. The engine applies bone-length consistency, joint-angle limits, floor-plane contact hypotheses, and temporal velocity/acceleration bounds to correct joint positions. It supports multiple operational modes (Audit, ShadowCorrect, Off) and includes a configurable validation layer to ensure runtime limits are safe. Deterministic golden-result verification and split-manifest leakage checks are included to ensure reproducible behavior and data integrity.

scripts/pose-physics, v2/crates/wifi-densepose-physics · high confidence

The repository now includes foundational configuration files to standardize the development environment and support AI coding assistants. A \.dockerignore\ file excludes build artifacts, caches, and environment files from Docker contexts. \.gitattributes\ enforces LF line endings for the contributor harness to ensure deterministic hashing across operating systems. \.gitmodules\ registers multiple external dependencies (midstream, ruvector, rvcsi, ruv-neural, rufield, ruview-swarm, worldgraph, metaharness, ruforecast) as git submodules. Additionally, \.mcp.json\ configures the 'claude-flow' Model Context Protocol server for use with AI agents, and \AGENTS.md\ provides a comprehensive operating contract, repository map, and validation instructions for Codex and other AI assistants.

(repo-wide) · high confidence

The wifi-densepose-sar-harness is introduced as a new agent harness for the wifi-densepose-sar RF tomography crate, featuring a four-agent workflow (architect, implementer, reviewer, test-writer) and a CLI with init, doctor, route, and flywheel commands. It integrates three self-improvement and cost-optimization features: Darwin Mode for sandboxed, measurable harness self-mutation; a cost-optimal model router that directs queries to the cheapest tier meeting a 0.8 quality bar; and a synthetic flywheel demo that runs a propose-evaluate-gate-promote loop with signed, replayable lineage. The harness is built on the @metaharness/kernel (WASM/native) and @metaharness/host-claude-code, with a smoke test to verify the install and a vitest configuration that handles shebang stripping for test imports.

harness/wifi-densepose-sar · high confidence

The sensing-server crate now includes a comprehensive README detailing its architecture, features (including UDP CSI ingestion, vital sign detection, and WebSocket broadcasting), and usage examples. A SECURITY.md document has been added to outline the UDP data plane threat model, specifying the new \--udp-bind\, \--udp-allow\, and \--udp-insecure-lan\ controls for restricting sensor frame sources. Additionally, an MQTT throughput micro-benchmark (\benches/mqtt\_throughput.rs\) and a runnable publisher example (\examples/mqtt\_publisher.rs\) have been introduced to validate the performance and wiring of the Home Assistant auto-discovery and state publishing pipeline.

v2/crates/wifi-densepose-sensing-server · high confidence

The \ruview-unified\ crate introduces a comprehensive RF spatial world model and active sensing control plane. It provides hardware adapters that normalize diverse inputs (Wi-Fi CSI, FMCW radar, UWB, Bluetooth Channel Sounding, 5G NR) into a canonical \RfTensor\, including a live hardware test example for ESP32 nodes. The model features a universal RF foundation encoder, a Gaussian map for spatial representation with physics-informed inverse updates, and a programmable perception control plane that manages sensing tasks, active actions, and spatial freshness. It also includes a strict anti-leakage evaluation protocol to ensure model generalization across different environments and devices.

v2/crates/wifi-densepose-hardware · high confidence

The \homecore-assist\ crate introduces a voice-activated intent recognition and execution pipeline. It provides a \RegexIntentRecognizer\ for classic pattern matching and a \SemanticIntentRecognizer\ that uses deterministic feature-hashing embeddings with an in-memory cosine k-NN index for semantic similarity matching. The pipeline includes five built-in intent handlers mirroring Home Assistant's classic intents (Turn On, Turn Off, Set Brightness, Nevermind, Cancel All) and a \RufloRunner\ trait that bridges to external LLM agents via subprocess or resolves intents locally. It also defines a satellite voice session protocol for handling audio streams and transcripts.

v2/crates/homecore-assist · high confidence

The pose-fusion demo now includes pre-built WebAssembly packages for the ruvector-attention library and a CNN embedder. The ruvector-attention package exposes six attention mechanisms (Scaled Dot-Product, Multi-Head, Hyperbolic, Linear, Flash, and Local-Global) along with Mixture of Experts and CGT Sheaf Attention, plus training utilities like Adam/AdamW optimizers, InfoNCE loss, and learning rate scheduling. A build script (build.sh) is provided to regenerate these WASM artifacts using wasm-pack, and browser-compatible JavaScript wrappers are included to load the WASM modules via fetch() for use in the demo.

ui/pose-fusion, ui/pose-fusion/pkg · high confidence

The UI has been completely rebuilt as a modular, modern web application featuring a new tabbed interface (Dashboard, Hardware, Live Demo, Sensing, Training) and a comprehensive CSS design system with dark/light mode support. Key additions include a Sensing tab with 3D Gaussian-splat signal field visualization, a dual-modal pose estimation page (pose-fusion.html) combining video and WiFi CSI data, and an Observatory view for scenario-based monitoring. The application now supports PWA installation, includes a command palette, keyboard shortcuts, toast notifications, and an activity log, and introduces a data source indicator banner to clearly distinguish between live ESP32, reconnecting, and simulated data streams.

ui · high confidence

This change introduces the initial firmware scaffolding for the WiFi Veil privacy shield, rebranded from 'Privshield'. It adds a portable C core (\veil\_shield\) that implements a keyed Givens-rotation obfuscation algorithm, verified via host-side tests for energy conservation, reversibility, and cross-language determinism with the Rust reference crate. Additionally, it provides ESP32-specific ESP-IDF components (\veil\_ris\_controller\ and \veil\_sensing\_detector\) designed as supporting roles: the former drives an external Reconfigurable Intelligent Surface (RIS) to perturb sensing directions, and the latter detects sensing solicitations to trigger the shield. These ESP32 components are build-only skeletons (SYNTHETIC/L0) with hardware paths marked as TODO, as the ESP32 cannot directly shape its own beamforming feedback due to closed PHY constraints.

firmware/privshield · high confidence

The \wifi-veil\ area is now a self-contained repository extracted from the RuView monorepo, providing a complete development and deployment environment for the privacy shield. It includes a GitHub Actions CI pipeline that enforces honesty invariants, runs Rust tests/lints/WASM builds, validates the portable C firmware core, and executes the npm MetaHarness smoke tests. A dedicated GitHub Pages workflow automatically deploys the self-contained WiFi Veil Console (\ui/veil-console.html\) as a static site. The repository also introduces a portable C shield core (\firmware/core/\) with host-validated tests, per-provider firmware scaffolds (OpenWiFi, OpenWRT, Nexmon, ESP32) for future hardware integration, and a dependency-free npm MetaHarness (\harness/\) for development guidance and workflow automation.

wifi-veil · high confidence

The \wifi-densepose\ command-line tool now supports Cognitum sign-in via \login\, \logout\, and \whoami\ commands, replacing static API token sharing with scoped OAuth sessions (read-only by default, with optional \--admin\ and \--spaces\ flags). Users can now read tenant-scoped spatial hierarchy data (sites, buildings, floors, spaces, zones, entities, events, alerts) using the \spaces\ command. The CLI also introduces a guided per-room calibration workflow (\enroll\, \train-room\, \room-status\, \room-watch\) that walks anchor sequences against a baseline to build specialist banks, and provides a new \calibrate-serve\ subcommand that exposes an HTTP API for remote UI-driven baseline capture.

v2/crates/wifi-densepose-cli · high confidence

The Python package has been modernized to version 2.0.0, introducing a new PyO3-based binding layer that exposes the Rust sensing stack directly. Users can now install the core DSP via \wifi-densepose\ or use the \ruview\ meta-package for the same API under a brand-facing namespace. This release adds three optional SOTA subsystems available as extras: \\[aether\]\ for contrastive CSI embeddings and re-identification, \\[meridian\]\ for cross-environment domain generalization, and \\[mat\]\ for disaster-survivor detection and triage. The package also includes new bindings for 802.11 beamforming feedback (BFLD) and skeletal pose keypoints, along with comprehensive benchmarks and examples for these new capabilities.

python · high confidence

The Docker image now includes the native Rust port of Home Assistant (homecore-server) alongside the sensing server and the Home Assistant discovery cog (cog-ha-matter). To address a security vulnerability where the server could start unauthenticated and bound to all interfaces, the entrypoint now refuses to start the sensing server with default settings (bind 0.0.0.0, no API token); operators must explicitly opt-in to unauthenticated access or configure a bearer token. Additionally, the image now bundles Python sensing capabilities, supports optional OTLP log export via OpenTelemetry, and fixes port binding to match documentation (3000/3001).

docker · high confidence

The HomeCore API now enforces bearer-token authentication on all REST endpoints, including the previously unauthenticated \/api/\ status route, and validates tokens via a configurable whitelist (\HOMECORE\_TOKENS\) rather than accepting any non-empty value. WebSocket connections are similarly gated, rejecting invalid tokens and fixing a previous issue where responses were discarded. Additionally, CORS is restricted to an explicit allowlist (defaulting to local development ports) instead of permitting all origins, and the server binary defaults to loopback-only binding to prevent accidental network exposure.

v2/crates/homecore-api/src · high confidence

Third-party dependencies (midstream, ruvector, sublinear-time-solver, metaharness, rufield, rvcsi) are now managed as git submodules instead of static copies or other mechanisms. A new vendor/README.md provides instructions for initializing submodules via \git submodule update --init --recursive\ or \git clone --recurse-submodules\, and notes that an automated workflow checks for upstream updates every 6 hours.

vendor · high confidence

The AETHER contrastive-embedding and CSI-to-pose transformer logic has been extracted into a new \wifi-densepose-aether\ crate that relies solely on \std\ (no external ML or async dependencies). This allows the Python \wifi\_densepose\[aether\]\ wheel to bind the embedding surface without pulling in the server's Axum, Tokio, or worldgraph dependencies, keeping the wheel size within the 5 MB budget. The crate exposes the full embedding pipeline—including the ProjectionHead with optional LoRA adapters, the Graph Transformer, SONA online adaptation (LoRA + EWC++), and sparse inference/quantization helpers—while the parent \wifi-densepose-sensing-server\ re-exports these modules so existing public APIs remain unchanged. A new native parity test ensures the Rust implementation matches committed golden vectors, closing a gap where only the Python binding was previously validated in CI.

v2/crates/wifi-densepose-aether · high confidence

The original Python implementation (\archive/v1/\) is now explicitly deprecated and frozen, with clear documentation directing users to the maintained \v2/\ Rust workspace and official pip wheels for new work. To ensure scientific reproducibility and prevent accidental regression in the signal-processing pipeline, the archive retains a load-bearing deterministic proof bundle (\archive/v1/data/proof/\). This bundle includes a synthetic CSI reference signal generator, verification scripts, and expected SHA-256 hashes that CI uses to validate pipeline determinism across platforms. The archive also contains legacy API documentation and test results for historical reference, but no new features or refactoring are permitted in this location.

archive · high confidence

The pose-fusion UI now features a new dark theme with a deep blue background, green accent colors, and specific typography (Inter and JetBrains Mono). This style sheet defines the layout for a dual-modal demo, including a header with status indicators, a main video panel with mirrored video/canvas overlays, and side panels. The visual design matches the 'Observatory' aesthetic, providing a polished interface for the RuView brand.

ui/pose-fusion/css · high confidence

The \wifi-densepose-occworld-candle\ crate provides a native Rust implementation of the OccWorld TransVQVAE world model using the Candle framework, replacing the previous Python bridge to eliminate IPC overhead. The engine features a real, deterministic convolutional encoder and decoder that produce input-dependent outputs without randomness in the forward path, and it includes an \honesty\ flag (\weights\_trained\) to clearly distinguish predictions from trained checkpoints versus deterministic but untrained dummy weights. To ensure robustness, the crate enforces strict input validation—rejecting degenerate shapes like zero-batch or over-capacity frame counts with clear \ShapeMismatch\ errors—and fixes a critical checkpoint-loading bug where int32 tensors were incorrectly mapped to int64, which previously caused crashes on malformed or attacker-supplied SafeTensors files.

v2/crates/wifi-densepose-occworld-candle · high confidence

The \homecore-plugins\ crate now enforces security controls that were previously deferred: Ed25519 signature and SHA-256 integrity verification (P4) are active, rejecting plugins with mismatched hashes, invalid signatures, or untrusted publishers, and capability isolation (P5) is active, ensuring plugins can only write to entity domains explicitly declared in their manifest permissions.

v2/crates/homecore-plugins · high confidence

The \wifi-densepose-rufield\ crate now supports a second sensing modality on the field surface: ultrasonic range profiles from BatVu recordings, in addition to the existing WiFi CSI data. This new module configures the upstream adapter to output coarse, P1-classified profiles suitable for network egress, while strictly rejecting raw P0 data and simulated events in production contexts. The bridge also implements a critical privacy mapping that prevents RuView's 'Derived' class (which carries identity embeddings) from being incorrectly exposed as low-privacy P1 data; instead, it is mapped to the biometric/identity tier (P4 or P5) and blocked from network egress by a fail-closed gate that only allows P1/P2 data.

v2/crates/wifi-densepose-rufield · high confidence

The UI now automatically determines the backend address from the current page origin, allowing it to function correctly across different deployment environments (such as Docker on port 3000 or local development on port 8080) without manual configuration. Additionally, WebSocket connections now dynamically match the page's protocol (using wss:// for HTTPS and ws:// for HTTP), fixing connectivity issues that occurred when the UI was served over plain HTTP but forced a secure WebSocket connection.

ui/config · high confidence

The legacy wifi-densepose v1.x package is now a tombstone that raises an ImportError directing users to upgrade to v2.0.0, which wraps the new Rust-based stack. To support the rebranding, a new ruview package has been introduced as a thin alias that re-exports all public symbols from wifi-densepose, allowing users to import from either name interchangeably while accessing the same underlying compiled Rust core and Python facade.

python/ruview-meta, python/tombstone · high confidence

Introduced the \wifi\_densepose\ package to resolve import errors when using the library. This new package provides a high-level \WiFiDensePose\ facade that wraps the underlying \ServiceOrchestrator\, exposing a simplified API with \start()\, \stop()\, and \get\_latest\_poses()\ methods, along with context-manager support for easier resource management.

_wifi\densepose · high confidence

This update refreshes the release binaries for the ESP32-C6 (c6-adr110, c6-onboarding) and ESP32-S3 (s3-adr110, s3-fair-adr110) nodes to version 0.6.7. The primary fix addresses a display-less board issue where CSI data frames were not being captured (yielding 0 pps); this is resolved by ensuring the C6 console RX remains alive and USB onboarding is shipped, resulting in verified data capture (0 to 27 pps) on ESP32-C6 hardware. The release includes updated SHA256 checksums and initial OTA data for all supported board variants.

_firmware/esp32-csi-node/release\bins · high confidence

Added UI test suite and runner; Added benchmarks for streaming engine cycle and mesh guard performance; Added comprehensive benchmarks for signal processing and calibration modules; Added comprehensive test coverage for the BFLD crate; Added host-based fuzzing stubs for ESP-IDF and FreeRTOS; Added host-side unit tests for adaptive controller, feature state, and mesh logic; Added integration tests for API authentication, history, and WebSocket behavior; Added integration tests for the disaster response pipeline and triage logic; Added performance benchmarks for detection algorithms and CSI parsing; Added regression tests for Docker entrypoint and Cargo.lock security invariants; Added smoke test fixtures for spatial-intelligence benchmark; Added smoke tests for person-count inference and fusion logic; Added smoke tests for pose estimation inference engine; Added steady-state CPU inference benchmarks for person-count and pose-estimation cogs; Added test suites for the SENSE-BRIDGE MCP server and CLI scaffolding; Added tests for calibration drift detection, serialization, and CIR estimation; Added tests for camera-room calibration pipeline; Benchmark harness for pointcloud splatting optimization; Host-side test and fuzzing harness for ESP32 CSI firmware.

This change adds a locked dependency file for the v1 pipeline verification to ensure deterministic floating-point results, pinning numpy to 2.4.2 and scipy to 1.17.1. It also introduces the build manifests for the new PyO3-based Python wheel (wifi-densepose v2.0.0), the ruview meta-package, and several new JavaScript/TypeScript harnesses and dashboards (nvsim-dashboard, homecore-frontend, homecore, ruview, privshield, sar, iphone-lidar-web), establishing the dependency baselines for these components.

(dependencies) · high confidence

Written by watchdog.canine.dev from the codebase's own history, inside the signed delivery this page is composed from.

• CAI 41 → 57 (+16.2)
• Rubric changed (rubric-2026.08.15 → rubric-2026.09.15) — scores are not directly comparable.

• Code Health 38 → 61 (+23.5)
• Architecture 97 → 73 (-24.7)
• Maturity 89 → 93 (+4.9)
• Readiness 30 → 67 (+36.9)
• Security 37 → 68 (+31.7)
• Event Sourcing 100 (new)
• Accessibility 45 (new)

• (anonymous) (cognitive 18) (harness/ruview/src/mcp-server.js)
• (anonymous) (cognitive 21) (v2/crates/homecore-server/ui/js/panels/entities.js)
• (anonymous) (cognitive 22) (v2/crates/wifi-densepose-desktop/ui/.vite/deps/chunk-JCH2SJW3.js)
• (anonymous) (cyclomatic 16) (v2/crates/homecore-server/ui/js/panels/entities.js)
• (anonymous) (cyclomatic 22) (v2/crates/wifi-densepose-desktop/ui/.vite/deps/chunk-JCH2SJW3.js)
• ADR not followed: .issue-177-body (docs/adr/.issue-177-body.md)
• Coverage not measured — test suite did not build
• Critical CVE: [GHSA redacted] (requirements.txt)
• Critical CVE: [GHSA redacted] (requirements.txt)
• Critical CVE: [GHSA redacted] (dashboard/package-lock.json)
• Critical CVE: [GHSA redacted] (requirements.txt)
• Critical CVE: [GHSA redacted] (aether-arena/space/requirements.txt)
• Critical CVE: [GHSA redacted] (ui/mobile/package-lock.json)
• Decision and consequences are absent despite strong findings: the body names no explicit decision (e.g., add production docs URL, implement calibration progress stream) and no trade-offs for those changes (docs/qe-reports/05-quality-experience.md)
• Dimension evaluation failed
• Duplicated block (10 lines × 2) (archive/v1/tests/fixtures/api_client.py)
• Duplicated block (10 lines × 2) (archive/v1/tests/fixtures/csi_data.py)
• Duplicated block (11 lines × 2) (archive/v1/scripts/test_api_endpoints.py)
• Duplicated block (11 lines × 2) (archive/v1/tests/fixtures/csi_data.py)
• Duplicated block (11 lines × 3) (archive/v1/scripts/test_api_endpoints.py)
• …and 231 more

• (anonymous) (cognitive 49) (ui/pose-fusion/pkg/ruvector-attention/ruvector_attention_browser.js)
• (anonymous) (cyclomatic 50) (ui/pose-fusion/pkg/ruvector-attention/ruvector_attention_browser.js)
• API Redundancy/Confusion: with_weights_and_adapter is a convenience method that duplicates the functionality of calling with_weights and with_adapter sequentially. While not strictly an error, it creates a branching API surface where the combined method might have different side effects or ordering guarantees than the sequential calls.
• ActivityHead::train (cognitive 34) (v2/crates/ruview-unified/src/heads.rs)
• AdaptivePoolingAttacker::enroll (cognitive 17) (v2/crates/wifi-densepose-privshield/src/attacker.rs)
• AdaptivePoolingAttacker::enroll (cognitive 17) (wifi-veil/src/attacker.rs)
• AnomalyDetector::process_frame (cognitive 39) (v2/crates/wifi-densepose-wasm-edge/src/adversarial.rs)
• AnomalyDetector::process_frame (cyclomatic 24) (v2/crates/wifi-densepose-wasm-edge/src/adversarial.rs)
• ApneaDetector.ingest (cognitive 24) (scripts/apnea-detector.js)
• ApneaDetector.ingest (cyclomatic 17) (scripts/apnea-detector.js)
• App.App (cognitive 24) (v2/crates/wifi-densepose-desktop/ui/src/App.tsx)
• App.App (cyclomatic 30) (v2/crates/wifi-densepose-desktop/ui/src/App.tsx)
• AppStateInner::effective_source (cognitive 30) (v2/crates/wifi-densepose-sensing-server/src/main.rs)
• AppStateInner::effective_source (cyclomatic 16) (v2/crates/wifi-densepose-sensing-server/src/main.rs)
• ArrayCoordinator::coordinate (cognitive 18) (v2/crates/wifi-densepose-signal/src/ruvsense/array_coordinator.rs)
• AsymmetricConvBlock.forward (cognitive 51) (scripts/wiflow-model.js)
• AsymmetricConvBlock.forward (cyclomatic 17) (scripts/wiflow-model.js)
• AttractorDetector::process_frame (cognitive 27) (v2/crates/wifi-densepose-wasm-edge/src/lrn_anomaly_attractor.rs)
• AttractorDetector::process_frame (cyclomatic 20) (v2/crates/wifi-densepose-wasm-edge/src/lrn_anomaly_attractor.rs)
• AutomationEngine::start_event_loop (cognitive 18) (v2/crates/homecore-automation/src/engine.rs)
• …and 1452 more

• 169 commits — 120 feature/other, 49 fixes

• v2/crates — 52 commits
• (repo) — 41 commits
• firmware/esp32-csi-node — 23 commits
• docs/adr — 14 commits
• (root) — 6 commits
• .github/workflows — 4 commits
• harness/ruview — 4 commits
• docs/research — 3 commits
• firmware/privshield — 3 commits
• v2/Cargo.lock — 3 commits
• docs/csi-frame-selection-measured.md — 2 commits
• docs/validation — 2 commits
• harness/wifi-densepose-privshield — 2 commits
• wifi-veil/.github — 2 commits
• .claude/scheduled_tasks.lock — 1 commit
• docs/huggingface — 1 commit
• docs/user-guide.md — 1 commit
• integrations/iphone-lidar — 1 commit
• scripts/fal-visual-teacher — 1 commit
• scripts/ruforecast-zeroshot — 1 commit

• fix: Merge branch 'fix-1936-work' into codex-calib-work
• fix: Merge pull request #1588 from ruvnet/fix/issue-triage-batch-1
• fix: Merge pull request #1800 from TsinbeiLabs/fix/calibration-wall-clock-gate
• fix: Merge pull request #1879 from ruvnet/fix/hf-csi-embedding-metrics
• fix: Merge pull request #1945 from ruvnet/codex/adr-348-collision-fix
• fix: chore(forecast): bump ruforecast submodule to governance/spend enforcement fix
• fix: chore(forecast): bump ruforecast submodule to the governance/spend fix
• fix: fix(calibration): bind field models to stable CSI grids
• fix: fix(calibration): bind server receipts to room identity
• fix: fix(calibration): bind the grid the radio keeps emitting, not the sparse minority
• fix: fix(calibration): expire live field models safely
• fix: fix(calibration): gate finalize on wall-clock duration, expose elapsed_s and frames/s
• fix: fix(calibration): learn held out empty residuals
• fix: fix(calibration): only the grid-bound radio may write the single-link baseline
• fix: fix(calibration): preserve expired status instead of collapsing to none
• fix: fix(calibration): register calibrated-presence-evidence schema via field_bridge
• fix: fix(calibration): require unique raw CSI samples
• fix: fix(calibration): separate receipt schema from telemetry literals
• fix: fix(calibration): tolerate measured UDP reordering
• fix: fix(engine): wire node positions into StreamingEngine's governed fuser (#1880)
• …and 149 more

Written by watchdog.canine.dev from the codebase's own history, inside the signed delivery this page is composed from.

ruvnet/RuView was measured the same way every project in this corpus was: the same rubric, at a pinned commit, with the result published in full. Point a surveyor at a repository you know and see whether you agree with it.