Skip to content

Instantly share code, notes, and snippets.

View ruvnet's full-sized avatar
💭
hacking the multiverse.

rUv ruvnet

💭
hacking the multiverse.
View GitHub Profile
@ruvnet
ruvnet / README.md
Last active November 4, 2026 02:25
Voyager inspired offline command security lab: fictional model, five tests, animated SVG

Can Voyager 1 Be Hacked?

A 1977 computer is still accepting commands across interstellar space. Could someone other than NASA take control?

Voyager 1 has a Command Computer Subsystem that decodes commands, sequences operations, and performs fault protection. NASA describes six onboard computers with roughly 32,000 words of memory combined. This is a remarkable study in the difference between physical inaccessibility and cryptographic security.

The short answer

Unauthorized spacecraft commanding is possible in principle, but there is no publicly demonstrated takeover of Voyager 1. The absence of publicly documented modern encryption is not proof that arbitrary radio transmissions would work. Specialized command formats, receiver constraints, immense distance, and secure ground operations matter.

@ruvnet
ruvnet / ruflo-3.71.0.md
Created November 1, 2026 20:19
ruflo-mods 0.15.0: automatic context delegation for big file reads (opt-in)

ruflo-mods 0.15.0 (ruflo 3.71.0): big file reads, handled automatically

In short

  • When Claude reads a large file just to answer a question, ruflo can now step in automatically. A cheap model (Haiku) reads the file, picks out the lines that matter, and Claude sees only those lines, each one checked against the real file.
  • On a test of 6 real files, all answers stayed correct and the total cost fell by about 22%, with the Haiku worker's cost included.
  • It's off by default. Nothing changes until you turn it on.

How to turn it on

It ships in the ruflo-mods plugin (0.15.0) from the ruflo plugin marketplace. Update the plugin in Claude Code with /plugin, then set the plugin option contextDelegation:

@ruvnet
ruvnet / README.md
Last active November 1, 2026 20:02
Ruflo shipped the core Super Agent architecture 508 days before DeepSeek public prototype | verified timeline and animated SVGs

DeepSeek Super Agent? Ruflo was doing the core of this 508 days earlier.

DeepSeek did not introduce the idea of one intelligent orchestrator coordinating many AI workers. Ruflo, originally Claude Flow, was already shipping that foundation in June 2025.

DeepSeek's public Super Agent research prototype arrived on October 31, 2026. By then, Ruflo already had almost 17 months of public development behind its multiagent architecture.

That is not an opinion. It is documented in Git commits and published software.

Animated timeline of Ruflo's lead over DeepSeek's public prototype

@ruvnet
ruvnet / ruflo-3.70.0.md
Created November 1, 2026 19:19
ruflo 3.70.0: governed context delegation (opt-in), with checked citations

ruflo 3.70.0: hand big reads to a cheaper worker, keep only checked citations

In short

  • Opt-in context delegation. When an agent needs to read a lot of code just to answer one question, ruflo can hand that read to a cheaper worker model. The main agent gets back only short answers, and each answer must quote the exact source lines it relies on.
  • Every quote is checked. ruflo verifies each one against a fingerprinted copy of the file. A made-up quote, a wrong line number or the wrong file is rejected, and the agent falls back to reading the source itself.
  • Off by default. Nothing changes unless you turn it on.

Install: npx ruflo@3.70.0 (or @latest).

How to try it

@ruvnet
ruvnet / ruflo-3.69.0.md
Created November 1, 2026 18:00
ruflo 3.69.0 + agentdb 3.3.0: edit what your agents remember

ruflo 3.69.0 and agentdb 3.3.0: edit what your agents remember

In short

  • Agents could already save past work in agentdb and search it. Now they can also open a record by its id, list records, correct one in place and delete one. The record keeps its id when you edit it.
  • This release also brings more than a dozen fixes from a review of open issues and nightly "Dream Cycle" reports. 4 of them are security fixes.

Install: npx ruflo@3.69.0 (or @latest).

Why editing matters

When a ruflo session ends, it saves a short summary as an "episode". Until someone checks the result, that episode is stored as unverified (reward 0). Before this release there was no way to mark it later as "this worked" or "this was wrong". Your only option was to delete it and save a new one, which also changed its id.

@ruvnet
ruvnet / ruflo-3.68.0-memory.md
Created November 1, 2026 04:40
ruflo 3.68.0: agents remember past sessions (agentdb 3.2 keyword + vector search)

ruflo 3.68.0: your agents now remember past sessions, and find them by exact names

In one paragraph

When a ruflo session ends with a summary, ruflo now saves that summary as an "episode". Later, any agent can ask "what happened last time we worked on this?" and get back the most relevant past episodes. The search understands meaning: "running out of memory while resizing pictures" finds a session about a thumbnail out-of-memory bug. It also catches exact names that meaning-only search tends to miss: ask about ZX-4471 and you get the session that wrote migration ZX-4471, not a session that merely sounds similar.

Why this matters

Before this release, ruflo looked as if it was saving sessions into agentdb's episode memory. It wasn't. The code called two functions (startEpisode and endEpisode) that agentdb has never had, so the step was quietly skipped every time. Nothing in ruflo ever read episodes back either, including the ones memory distill creates. 3.68.0 fixes both halves.

@ruvnet
ruvnet / ruflo-3.65.0-release.md
Last active October 30, 2026 13:18
ruflo 3.64.0 → 3.65.0: verified learning + deeper Claude Code mods integration (ADR-502..510)

ruflo 3.64.0 → 3.65.0: verified learning and deeper Claude Code mods integration

Date: 2026-10-29 / 30 · Tracking issue: #4581 · ADRs: ADR-502 to ADR-510 · Packages: @claude-flow/cli, claude-flow, ruflo · Plugin: ruflo-mods 0.8.0 → 0.12.0

Status: both releases are published and verified.

  • 3.64.0: published, tag v3.64.0.
  • 3.65.0: published 2026-10-30.
@ruvnet
ruvnet / ruview-c5-qualification.md
Last active October 29, 2026 18:50
RuView ESP32-C5 quickstart, guarded installation and measured transport limits

RuView ESP32-C5 quickstart and measured limits

Firmware v0.8.17-esp32 is published as an experimental prerelease. Its final bundles passed checksum, version/chip and exact-run checks; the release C5 images were installed and tested on COM8. Registry harness 0.9.3 is not yet published because npm denied CI authorization (issue #2390). Use the checked-out main source installer below meanwhile: https://github.com/ruvnet/RuView/releases/tag/v0.8.17-esp32

Identify and preserve the board

Confirm the exact board model and port. The measured device was production ESP32-C5HR8 revision 1.2, with 8 MB flash and 8 MB embedded PSRAM. A USB bridge listing alone does not identify the ESP chip.

In an isolated Python environment:

@ruvnet
ruvnet / rufield-announcement.md
Created October 14, 2026 14:49
RuField 0.2.0 candidate: checking radio signal quality before trusting sensing results

RuField: checking the signal before trusting the sensor

I’ve published a RuField 0.2.0 release candidate that adds a way to inspect radio signal quality alongside our sensing work.

A software defined radio turns radio waves into samples that software can examine. This update connects diagnostics from the W1700K SDR project to RuField through RuView tooling.

Think of it as checking the microphone before trusting a recording. If samples go missing or the signal suddenly changes, the software should make that visible.

RuField can now read those diagnostic records, flag missing data and receiver quality problems, and compare signal power against a short reference baseline. If the configuration changes or records are skipped, it discards that baseline rather than carrying an unreliable comparison forward.

@ruvnet
ruvnet / README.md
Last active October 14, 2026 06:42
RuVector Python native attention: published 0.2.0 runnable quickstart and release recovery evidence

RuVector Python: native attention quickstart

Use published ruvector-attention==0.2.0 to blend synthetic embeddings with scaled dot product attention. This example calls the RuVector Rust runtime and checks its output against an independent NumPy reference.

Install and run

Python 3.10 or later; the native ruvector==0.1.3 wheel must support your platform. Save this README and attention_quickstart.py together.

uv venv .venv