64× memory compression. 14 000 QPS. Zero external dependencies. Pure Rust.
ruvector's new ruvector-residual-vq crate ships Residual Vector Quantization (RVQ) —
the quantization scheme behind Meta's EnCodec, Google SoundStream, and LanceDB's default
compressor since 2024 — as a standalone, dependency-light Rust library.
midstreamer@0.3.1 + ruflo Step 2 + Step 3 + alpha.38Shipped 2026-05-14. Real QUIC end-to-end, no stubs.
| Layer | Where | Version |
|---|---|---|
Upstream — midstreamer npm package |
ruvnet/midstream#81 merged |
0.3.1 on npm |
| Ruflo — federation transport loader (Step 2) | ruflo#2007 merged earlier today |
in alpha.37 |
| Ruflo — Rust federation peer crate (Step 3) | ruflo#2009 merged |
alpha.38 |
| Release | ruflo#2010 merged |
published |
Date: 2026-05-14 · Versions shipped: @claude-flow/cli@3.7.0-alpha.37, claude-flow@3.7.0-alpha.37, ruflo@3.7.0-alpha.37
ruvnet ships two related-but-distinct Rust workspaces: ruvnet/midstream (AIMDS safety gates + math/scheduling primitives) and ruvnet/agentic-flow (transport adapters, including QUIC). The names suggest overlap, and the question — can midstreamer give ruflo real QUIC + Tailscale-style networking + Rust in-flight agentics? — has come up enough times to deserve a written answer.
The short answer: midstreamer alone, no. midstreamer-quic (the underlying Rust crate, cross-pollinated with agentic-flow's bridge layer), yes — with a concrete three-step plan that lands in ruflo behind opt-in env flags.
Published 2026-05-14. Same MCP-tool surface, strictly better detection, accurate audit counters, one CVE cleared. If you're using aidefence_* MCP tools (the 3-gate pattern) through any ruflo plugin, you get the improvements automatically when the workspace package bumps.
aidefence (also published as aidefense — same code, two spellings) is the npm side of the AIMDS (AI Manipulation Defense System) project — a production-ready security middleware that protects AI applications from prompt injection, manipulation, PII leakage, and adversarial inputs. The library splits the defense into four latency tiers: detection (<10ms pattern matching + PII sanitization), behavioral analysis (<100ms anomaly + LTL policy checking), formal verification (<500ms theorem proving), and adaptive re
Audience: anyone who installed neural-trader between v2.5.0 (Nov 2025) and v2.7.5 (today) and wondered why nothing worked.
neural-trader@2.7.6 is on npm. It fixes a four-bug cascade that made the package nonfunctional from v2.5.0 onward:
.gitignore mistake that silently dropped three load-bearing JS modules from the published tarball, so require('neural-trader') always threw Cannot find module. Fixed in 2.7.5.
Detailed reference companion to the RuView tutorial "Pi 5 + Hailo Cluster: Building a Cognitive RF Observer with rvcsi". This is the unabridged recipe — every command, every config, every unit file. The tutorial is the narrative; this is the manual.
Target cluster: 4× Raspberry Pi 5 (8GB), 2× Hailo-8 M.2 HAT, 1× workstation (Linux, ≥16GB RAM), all on Tailscale.
ruvector-ivfpq brings production-grade IVF-PQ (Inverted File Index + Product Quantization) to the Rust vector search ecosystem with 21× memory compression and 94.7% recall@10 at 9,481 QPS.
Vector search is the backbone of modern AI applications — RAG pipelines, semantic search, recommendation systems, and multimodal retrieval all depend on finding the nearest embedding vectors efficiently at scale. The challenge: at 10M+ vectors, brute-force exact search is too slow and graph-based indexes (HNSW, DiskANN) use too much memory.
TL;DR —
ruvectorships its first IVF (Inverted File) index — the search structure behind FAISS, Qdrant, and Milvus — in pure safe Rust (#![forbid(unsafe_code)], no C/BLAS, WASM-friendly). A redundant secondary-list assignment recovers the recall classic IVF loses near cluster boundaries: +31.8 percentage-point recall@10 atnprobe=1over IVFFlat, at identical memory, thanks to a deduplicating shared-block layout. Auto-vectorised distance kernels push ~130k–270k single-thread QPS on a desktop x86-64 core.
Repository: https://github.com/ruvnet/ruvector
Crate: crates/ruvector-rairs · PR: ruvnet/RuVector#459
ADR: docs/adr/ADR-193-rairs-ivf.md
⚠️ Provenance. Early drafts of this work cited a "RAIRS / SEIL, Yang & Chen, SIGMOD 2026, arXiv:2601.07183" paper. That reference is unverified — treat the names as our own and the crate as an orig
Status: Research note — non-binding survey. Nothing here is an accepted decision. Date: 2026-05-11 · Author: research pass (Claude Code) · Scope owner: ruv Companion docs:
docs/research/sota/2026-Q2-rf-sensing-and-edge-rust.md,docs/research/sota-surveys/wifi-sensing-ruvector-sota-2026.md,docs/research/sota-surveys/ruview-multistatic-fidelity-sota-2026.mdMaps onto ADRs: 015, 016, 017, 024, 027, 028, 029–032, 039, 040, 069, 081, 084–086, 095, 096
Date: 2026-05-11
Subjects: NousResearch/hermes-agent ("the agent that grows with you") vs. ruvnet/ruflo (formerly claude-flow, v3.6.x — "multi-agent AI orchestration for Claude Code")
Type: Bidirectional capability gap analysis + strategic recommendations
Evidence quality: Hermes facts sourced from its public README, docs site (hermes-agent.nousresearch.com/docs), and DeepWiki — the repo was not cloned, so tool/skill counts and module names are approximate. Ruflo facts sourced from its own README / CLAUDE.md / docs, which carry vendor-stated benchmarks (treat "89% routing accuracy", "150×–12,500× search", etc. as claims, not independent measurements). Confidence: medium-high on architecture/positioning, medium on exact counts.