The archive.

Project releases, commits and company announcements from Reuven Cohen (rUv) and Cognitum One.

Ruflo 3.50 to 3.52: Claude Code mods, the /ruflo console and hardened guards

— Project releases

Over four days Ruflo moved onto Claude Code's new mods API. Version 3.50 made it an opt-in mod with in-process routing and a live swarm pane, 3.51 added a one-command /ruflo console and turned the mods on by default in new projects, and the 3.52 plugin release on October 5 closed the guard bypasses found in an overnight test run.

RuView sensing fixes: no vital-sign numbers from an empty room or stale nodes

— Project releases

Three merged RuView fixes between September 30 and October 4 tighten when the sensing server reports anything. Stale nodes no longer inflate room counts, a breathing signal can no longer be mistaken for a confident 120 BPM heart rate, and vital-sign numbers are now published only with fresh, calibrated evidence that exactly one person is present.

RuView 0.9.1: watch WiFi sensing nodes live inside Claude Code

— Project releases

The RuView npm toolkit 0.9.1, merged and published on October 2, adds a Claude Code mod that opens a /ruview pane beside the conversation with node health, a CSI waterfall and a radar view. It also adds a spectrum capture option and a token-authenticated HTTP mode for its MCP server.

RuView ESP32 firmware 0.8.13 pre-release: stalled sensors recover themselves

— Experimental

PRE-RELEASE. RuView firmware 0.8.13, published September 29 from an unmerged pull request, adds a watchdog that restarts sensing when data stops flowing and fixes a timing mismatch that dropped about a quarter of collected CSI. The measurements come from one ESP32-C6 board with deliberately triggered faults.

RuVector KGE 0.2.0: local knowledge-graph link prediction with HolE and RotatE

— Project releases

@ruvector/kge, published to npm on September 28 from RuVector PR #1057, trains knowledge-graph embeddings locally and answers which entity most plausibly completes a fact. It ships a CLI, a TypeScript API and a localhost server, with a native addon and a WASM fallback. We installed it and ran its first commands.

RuVector WASM 0.2.0 makes browser HNSW search real instead of silently flat

— Project releases

The browser build of RuVector accepted a request for an HNSW index but quietly used exact flat search, because the core library disables its native HNSW in WebAssembly. Pull request #1035, merged September 27, adds a portable in-browser graph index and reports which backend is actually running. It shipped as @ruvector/wasm 0.2.0.

RuView firmware fixes: ESP32-C6 onboarding restored and a leaner 4 MB build

— Project releases

Two merged RuView firmware fixes landed on September 26 and 27. The first gives 8 MB boards the storage partition their logger needed and removes the logger from 4 MB builds, which the release notes say made a physical ESP32-C6 image 43,056 bytes smaller and ended a boot error. The second restores native USB onboarding on the ESP32-C6.

Ruflo 3.46.1 security fix: Windows browser tools no longer pass input through a shell

— Project releases

On Windows machines without a global agent-browser install, Ruflo's browser MCP tools fell back to npx through the command shell, so text supplied to the tools could be read as shell syntax. Release 3.46.1 on September 26 starts the process directly with an argument list and fails closed. It closes a week that also shipped four feature and fix releases, summarised here.

ruPet is a desktop robot pet for ChatGPT and Codex, with an optional Ruflo-aware skill

— Project releases

ruPet is a new public repository released on September 22. Version 1.0.0 is an animated desktop pet in the v2 pet format; version 1.1.0, the same afternoon, adds an optional $rupet skill and prompts that can use Ruflo memory and routing when those tools are available. The pet itself is artwork only and runs no code.

ruQu ships an offline laboratory for observer consistency experiments

— Experimental

A pull request merged on September 20 adds a reproducible command-line laboratory that runs Bell and CHSH protocols against the bundled quantum simulator, binds every run into a hash-linked evidence chain, and rejects synthetic evidence when a run claims to be a replay of captured data.

MCP Studio is a starter kit for building a ChatGPT plugin on a real MCP server

— Project releases

A new public repository appeared on September 17 and reached its current shape on September 18. MCP Studio is a web playground and a working Model Context Protocol server in one project, with twelve read-only tools, eight resources, five dashboard templates and a protocol test suite covering thirty assertions.

Ruflo 3.42.4 stops smart memory search from reporting the wrong similarity number

— Project releases

A memory search in smart mode was putting an internal fusion score into the field named similarity, so a result whose real relevance was 0.8 could be displayed as 0.03. The September 17 release puts the raw relevance back in similarity, moves the fusion score to its own field and adds fourteen regression tests.

RuVector adds an optional native graph engine without replacing the one it trusts

— Project releases

A September 16 commit makes a compiled minimum-cut package an optional dependency. The existing implementation stays the default and the fallback, because fuzzing it against brute force over 3,844 random graphs found no mismatches. What the native package adds is capability, not a correction.

RuView ties a calibration receipt to the room it was actually measured in

— Project releases

A sensing system is only trustworthy if a calibration cannot be quietly reused somewhere it does not apply. A merged pull request on September 16 binds calibration receipts to process, session, room and model identity, and refuses to finish a calibration until every requested node has contributed.

rultra puts its sensor box behind an MCP server

— Project releases

rultra is a Raspberry Pi 5 sensor box in Rust that refuses to call a device working until someone has actually observed it working. A September 13 commit adds an MCP server exposing ruv:// resources and an end-to-end suite, so an agent can read the box's honest self-assessment.

Open Claude Code tracks upstream nightly by nightly

— Project releases

A clean room open source implementation of the Claude Code CLI: async generator agent loop, 25 tools, four MCP transports, six permission modes, hooks and sessions. Nightly releases track the upstream version it mirrors.

Ruflo gives ChatGPT its own federation key so it can finally speak

— Project releases

The federation relay refuses any event signed by someone other than the connection that authenticated. ChatGPT Federation had no key of its own, so it could read the federation but never publish into it. A merged pull request on September 12 gives it a key, its own authenticated connection and three narrow tools.

The RuV Stack gets a change feed you can check the provenance of

— Project releases

Keeping up with two hundred public repositories by reading push notifications does not work. A September 12 commit on the ruvnet entrypoint adds a provenance-aware change feed, and follow-up commits on September 14 expose how fresh the underlying snapshot actually is.

Agentic DevOps v2 plans deployments instead of running them

— Project releases

The v2 rewrite replaces model driven shell execution with a deterministic local planner. It turns a small deployment specification into Docker, Kubernetes and CI files, shows exactly what changes, and never touches your cluster.

Agentic Reports rebuilds around evidence you can check

— Project releases

Agentic Reports turns a pile of documents into an extractive report where every passage carries its source, capture time and content hash. A September 11 commit revived the project around verified citations, a six tool MCP server and a local CLI.

Dynamo MCP v2 retires executable templates

— Project releases

Project scaffolding is a quiet security hole: templates that run code during generation. Dynamo v2 shows the exact files first, verifies a template fingerprint, then writes the project without running template scripts or downloading code.

ONNX Agent v2 checks whether a quantized model is still the same model

— Project releases

Quantization is usually sold as free speed. A September 11 commit turns ONNX Agent into a qualification pipeline that measures both halves of the claim: numerical parity against an independent reference, and an actual measured latency improvement gate.

Reflective Engineer v2 turns promotion decisions into a replayable gate

— Project releases

A September 11 commit rebuilds Reflective Engineer as an evidence review console. Import an experiment report, replay its quality, safety, cost, latency and regression checks in the browser, a CLI or an MCP host, and get the same verdict from all three.

Agent Name Service v2 makes an agent prove it owns its name

— Project releases

A name is only useful if someone can prove they control it. The September 10 revival gives Agent Name Service real Ed25519 bindings with a pinned issuer and expiry, a SQLite registry with durable revocation, and audience-bound single-use challenges.

Federated MCP reads the federation and labels every result untrusted

— Project releases

A September 10 revival turns Federated MCP into a bounded reader for public RuFlo federation activity. It reads x.ruv.io through the official SDK, marks everything it returns as untrusted observation, and cannot publish, mint invites or hold an admin token.

Guardrail checks the data before the agent acts on it

— Project releases

Guardrail is a small, bounded policy checker for structured data, revived as a v2 alpha with a local CLI, MCP tools and a policy resource. Missing fields deny; present but falsy values stay valid.

Ruflo 3.41.2 stops a memory curate from shrinking your index

— Project releases

Ruflo gives Claude Code a coordination layer: swarms, routing and memory that survives between sessions. Release 3.41.2 fixes a bug where the automatic memory curator could overwrite a hand written index, plus daemon consent and sibling disclosure fixes.

LatentMesh keeps agents talking where the internet does not reach

— Experimental

LatentMesh coordinates agents over LoRa, radio, Bluetooth, WiFi or plain audio. The September work adds a native runtime slice and fail closed calibration qualification — and the 0.2.0 release is an honest record of an idea that did not work.

rGi keeps an agent's work alive between prompts

— Experimental

rGi is a perpetual agentic runtime built on the RuV stack in Rust, WebAssembly and TypeScript. Two September 8 commits laid the foundation: governed execution, grounded adaptation and durable evaluation gates.

Dream Machine re-checks its evidence before anything gets promoted

— Experimental

Dream Machine runs one honest repository evolution cycle per night and never merges by itself. A September change makes the witness re-verify the evidence read set immediately before promotion, closing the gap between measuring and deciding.

ruClip closes a nonce replay race in its credential path

— Experimental

ruClip is a control plane for running a company where the employees are agents. A September fix closes a time of check to time of use race that allowed nonce replay against actor credentials and human identity attestations.

WorldGraph adds bounded capture playback to its digital twin

— Project releases

WorldGraph models a physical space as a typed, provenance tracked graph and ships an AI agent that helps you build one. September 6 commits added bounded Gaussian capture playback with native provenance and validation.

MidStream inspects a model's output while it is still streaming

— Project releases

MidStream analyses LLM output in flight rather than after the fact. A September change adds the aidefence-core Rust crate over shared pattern packs, with pack driven prompt injection detection for tool invocation and exfiltration URLs.

MoE Foundry splits one big model into focused specialists

— Experimental

A Mixture of Experts model contains many expert networks. MoE Foundry measures which experts a workload actually uses, keeps a selected group per layer, and exports a smaller model to test — with a CLI, an MCP server and a published validation receipt.

ruv-drone gains a predictive advisory for fleet coordination

— Project releases

ruv-drone coordinates cooperative UAV fleets above the per vehicle autopilot: formation keeping, Raft consensus and a hard flight safety boundary. A September 2 commit wires RuForecast in as a predictive advisory.

SPARC 1.0 turns a methodology into an enforced state machine

— Tutorials

SPARC was a way of working: specification, pseudocode, architecture, refinement, completion. Version 1.0 makes the five phases executable in TypeScript with deterministic gates, signed receipts, npm skills and both local and remote MCP transports.

RVM 1.7.0: a ruv:// name is a name, not permission

— Project releases

RVM is a runtime built for agents rather than for processes. Release 1.7.0 adds the ruv:// context namespace with a deliberate separation: resolving a name grants nothing, and the right to act arrives separately as a capability handle.

OpenAVO: open autonomy, fixed authority

— Experimental

OpenAVO is the project home for the open MetaHarness implementation of an Autonomous Variation Operator. The agent gets room to experiment; the harness keeps permissions, budgets, invariants, promotion, rollback and evidence.

Autogenous turns spatial state into bounded, derived observations

— Experimental

Autogenous converts runtime failures into verified, reversible software adaptations. The radio-moe 0.3.1 release checkpoints an adapter that turns authorized Cognitum Spaces state into bounded observations, rejecting raw sensing fields outright.

AgentiCow queries across a branched memory natively

— Project releases

AgentiCow branches a vector memory in about half a millisecond regardless of base size, so parallel agents can work on isolated copies. A late June change added native approximate search across the branch boundary.

Repo Explainer turns a repository into a page people understand

— Tutorials

Most READMEs are written by people who already understand the project for people who already understand the project. Repo Explainer builds a visual explainer page instead, and a late June commit wired the full pipeline together.

rupixel searches documents by how the page looks

— Project releases

Text search cannot read a scanned table or a chart. rupixel screenshots each page and searches the pictures. A late June commit replaced every stub with real MiniLM embeddings and a live in browser demo.

PhotonLayer trains a multi plane optical front end end to end

— Experimental

PhotonLayer is a deterministic optical AI front end in pure Rust, where a learned phase mask compresses light before any electronics are involved. A June 18 commit trained a multi plane diffractive cascade end to end.

A founding advisory team for sovereign agentic AI

— Cognitum

Cognitum One named its founding advisory team and investors, drawn from enterprise software and silicon leadership, as it brought sovereign agentic AI to enterprise infrastructure.

SAFLA hardens the self aware feedback loop before anything else

— Project releases

SAFLA is a self aware feedback loop architecture with an MCP integration for Claude Code. Its most recent substantive change is unglamorous and correct: replace MD5 with SHA-256, fix argument injection and remove a tracked environment file.

Back to the newsroom