Forkcast Hacker News #17: mini-AGI, office365_flake, cua, lossless-memory, jevals, foremerge, c8-ass
Episode 17 covers 30 open-source projects trending on Hacker News, spanning AI training on consumer GPUs, Linux Office 365 integration, computer-use agent
Watch Episode
Episode Summary#
Episode 17 covers 30 open-source projects trending on Hacker News, spanning AI training on consumer GPUs, Linux Office 365 integration, computer-use agent models, lossless AI memory, deterministic LLM evaluation, parallel coding agent conflict detection, and tools for code review, robotics serving, document export, voice-controlled AI, and distributed SQLite.
Repositories Covered#
mini-AGI#
mini-AGI - is a continual learning byte-level language model that assembles its own architecture, trains from scratch on a single 8 GB VRAM GPU, and keeps learning from everything it reads. It stores its weights as ordinary files on disk and pages them onto the card as it needs them, so the para
office365_flake#
A February 2026 report in the Bottles tracker of Office 365 x64 running on Wine 10.20 with; The classic ruados / eylenburg Office-on-Wine notes (Direct2D registry tweak, copying the; The September 2026 Bottles announcement that Microsoft 365 installs, signs in (with 2FA) and runs
cua#
Cua Fleets: Provision a Linux desktop, run a command, and save a screenshot.; CUA-S1: Explore small, specialized models for computer-use decisions.; Cua Driver: [Operate Calculator and verify its result](https://cua.ai/do
lossless-memory#
Keep every line. Raw conversation logs are stored in full. Nothing is summarized, ever. Summaries are a map; the log is the territory.; Timestamp everything. Every record — utterance, action, document chunk — carries a timestamp, and every index is built on top of that time axis. We call thi
jevals#
Evals and guardrails for agents, using Jev-style decision models instead of an LLM judge. All the evals for a trace go out as one request that costs a few thousandths of a cent and comes back in a few hundred milliseconds, so you can run them on every trace and inside the agent loop.
foremerge#
c8-ass#
- ; [ ] Add support for the
INCBINdirective?
jeff#
Catch code issues before they catch you.
InstinctFlash#
[2026/09/17] RTX 5090 support. Deploy on your workstation with the same Runtime API used on Jetson Thor. Setup · Reproduce.; [2026/09/16] RTX 4090 support. Desktop inference and WebSocket serving with dedicated installation profiles. [Setup](
gdocs-me-up#
GDocs-Me-Up: A Google Docs High-Fidelity Exporter#
Autobot#
AutoBot: a self-improving agentic harness that makes frontier AI better at finishing complex knowledge work.
npunlock#
Intel ships programmable SHAVE cores inside its NPUs, but the public stack
exposes only graph-level programming. npunlock reconstructs the missing path
from custom C code to a runnable NPU kernel.
oj#
OJ is a Rust-native build tool for React apps.
frost-icon-theme#
One glass language everywhere — the same two shapes and a glyph for folders, places, apps and files.; File types that say what they are — JSON is { }, Markdown is M↓, a lock file is a padlock, a; Fixes for developer files — GNOME types bun.lock, .env, Dockerfiles and .gitignore
Seal#
Sealed envelopes for the people you leave behind.
graf#
Graphify started with a great idea and became popular fast. The problem is that its Python/NetworkX architecture does not scale well. It installs about 30 direct dependencies, and the CLI reloads the entire graph into memory for every query. On a large repo, that can make each search slow enough to
omcp#
omcp specification; omcp protocol schema; omcp documentation
swarm#
Swarm builds, visualizes and exports family trees, entirely on your Mac. Trees are plain GEDCOM files in a folder you can open in Finder: no accounts, no cloud, no telemetry, no AI.
all-your-agents#
Watch every coding agent on this machine, and inspect the sessions they leave behind.
local-coder#
Which models fit this machine’s memory, GPU, disk, and useful context budget?; Which model should explore, plan, implement, verify, research, or review?; Can each selected model produce real structured tool calls, not tool-shaped text?; Can the coding model read, edit, and reread a file correctly?;
hibi#
hibi#
ambits#
Reading — search, callers and symbol lookup that answer in symbols rather than files. The agent asks for App/process_compaction, not for 2,000 lines of app.rs.; Remembering — ambits records every symbol the agent reads, at what depth, throughout the session. After a compaction it hands t
agent-chaperone#
Not a sandbox. It cannot stop a server from doing something the call did not describe. Pair it with OS-level isolation for that.; Not a replacement for the client’s permission prompts. It is a second opinion that sees the content, which permission prompts do not. It never auto-approves anyth
try-omarchy-windows#
The full Omarchy 4.0.3 desktop on new or reset guests: Hyprland, the bar, notifications, all 22 the
cli#
Tell you exactly what will break before a change lands.; Enrich AI prompts with project structure, dependencies, and blast-radius analysis.; Validate architecture with circular dependency checks, orphan detection, and compiler-backed validation.; Enforce merge policies in GitHub Acti
jev-lint#
Check JavaScript and TypeScript against project conventions written in plain English. Define inline or Markdown rules; Jev reports possible violations. Built for coding agents, with compact output by default and optional pretty or JSO
reader-source#
Native reading: import and read PDF, EPUB, and plain-text books in a; Contextual AI: chat about a selected passage, the current page, the book; Inline understanding: rewrite a difficult passage in clearer language; Highlights and notes: preserve selections, annotations, and reading;
blackgit#
Cli Partial clone + sparse, per-user views on the client (git black, Python) — wraps stock git:; Cli Agent; Server path-level blob authorization
jev-graph-search#
Jev Graph Search helps an AI agent find useful notes and passages in a local Obsidian vault or a folder of Logseq Markdown notes. It makes a shortlist on your machine, then uses Jev to rank those candidates against your question. The result includes the original passage and source reference, so the
sqlite-modal#
from_name creates or looks up App sqlite-modal-{name} (create_options → @app.server); One SyncServer container per name. max_containers is fixed at 1.; connect(path) opens a local connection and waits until the Server is up; Sync is explicit (push / pull). Conflicts are last-push-win
Watch#
- Video: https://www.youtube.com/watch?v=ODV34T_iYPU
- Cover: https://i.ytimg.com/vi/ODV34T_iYPU/hqdefault.jpg
Notes#
Transcript and notes will be added from Forkcast output artifacts.