← All transcripts

GitHub Trending Today #43: Kakehashi, qm, gander, SimpleEnglish, skill-recorder, humanizer-cli, crm Transcript, AI Summary & Key Points

Github Awesome · 12 days ago · Science & Technology · 15:04 · EN-US

🧠 AI Summary

Thirty-five trending open-source GitHub projects cover low-level computing, offline-first software, AI agents and model inference, developer workflows, creative tools, scientific visualization, hardware, and data systems. Several projects emphasize local execution, sandboxing, durable state, explicit human approval, provenance, and avoiding unsupported guesses. The roundup also highlights unusual technical approaches, including running large models from disk, transferring files through animated QR codes, reconstructing 3D scenes from ordinary video, and booting directly into a local language model.

🔑 Key Points

  • Kakehashi runs arm64 macOS command-line binaries on arm64 Linux by loading Mach-O files directly and translating BSD system calls in user space.
  • WASTE streams selected mixture-of-experts weights from NVMe and is reported to run the full 2.78 trillion parameter Kimi K3 on a 64-GB MacBook Pro at 0.45 to 0.62 tokens per second.
  • Kimi K3 in C runs Moonshot's 2.78 trillion parameter base model on one x86-64 CPU, using 8.24 GB of RAM at its smallest preset but taking 32.69 seconds per token and requiring roughly 1.7 TB of fast storage.
  • Several projects prioritize local or offline operation, including Gander, Humanizer CLI, Graybox, Nightrun, DSCode, Trace file lineage, and Scopey.
  • Vercel's marketing team Eve template routes campaign work among five specialist agents and pauses broadcasts and destructive actions for human approval.
  • Skill recorder converts recorded work sessions into reusable agent procedures, but analysis sends screenshots, URLs, clipboard previews, and narration to GitHub's cloud.
  • CRM uses a durable research agent and explicitly prohibits guessing personal information; weak evidence becomes a suggestion for a human to resolve.
  • The roundup includes safety and provenance mechanisms such as network-blocked sandboxes, auditable change ledgers, machine-checkable Lean certificates, file-lineage evidence, and role-based permissions.

✅ Actionable items

  • Deploy Inkstone to a personal Cloudflare account by forking the repository.
  • Keep secrets out of Skill recorder sessions because analysis sends screenshots, URLs, clipboard previews, and narration to GitHub's cloud.
  • Use human approval for broadcasts and destructive actions in agent workflows.
  • Treat weak evidence as a human-review suggestion rather than allowing an agent to guess.
  • Use strict mode in Ratchet when only definite code-quality findings should block changes.
  • Run commands inside a network-blocked workspace sandbox by default.
  • Use checksums before saving files reconstructed from QR streams.
  • Separate recorded proof from guesses when tracing file lineage.
  • Keep open derm clear of medical diagnosis claims and account for impact, pinch, electrical, and laser hazards.

🧰 Tools & AI usage

AI is used for

  • Workplace coordination and campaign production — QM and the marketing team Eve template isolate agents, route work among specialists, and preserve shared context.00:37
  • Procedure recording — Skill recorder reconstructs the intent and ordered steps of a real work session for reuse as an agent skill or scheduled automation.02:47
  • Writing analysis — Humanizer CLI identifies mechanical signs of machine-written prose and provides editing guidance.03:17
  • Code-change review — Ratchet checks agent-produced edits for unnecessary dependencies, duplicate helpers, thin wrappers, and other patterns.03:44
  • Large-model inference — WASTE and Kimi K3 in C stream model components from storage and use caches to run models without conventional GPU requirements.04:34
  • Research and CRM maintenance — CRM uses a durable research agent to process queued database rows while preventing unsupported personal inferences.05:56
  • Coding assistance — Soul advisor, DSCode, Scopey, and related projects coordinate coding agents, isolate work, checkpoint changes, or detect scope drift.06:24
  • Business-context retrieval — AWS's context ontology accelerator converts databases and files into validated ontology and knowledge-graph context for agents.13:22

🔗 Links mentioned

📄 Transcript

Searchable transcript of GitHub Trending Today #43: Kakehashi, qm, gander, SimpleEnglish, skill-recorder, humanizer-cli, crm — Github Awesome (15:04). Search for a phrase, then click its timestamp to jump straight to that moment in the video.

Captions sourced from the original video on YouTube, published by Github Awesome. The video, its captions and all related intellectual property remain the property of their respective owners; AINotes claims no ownership. Provided for research, accessibility and search — see the Transcript Notice and Copyright Policy.

00:00 Welcome back to GitHub Awesome. This is GitHub trending today number 43. 35 trending open source projects on GitHub right now. Let's get into it. Kakehashi runs arm64 macOS command line binaries on arm64 Linux without a virtual machine or jit compiler. Because the CPU architecture already matches, it loads Mach-O files directly, provides a freestanding lib system, and translates BSD system calls in user space.

00:26 The project currently demonstrates working 7-Zip, curl, threading, and clang probes across Docker and Linux virtual machines. QM turns a workplace agent into a shared system for an entire startup. Each person, channel, and project gets isolated memory, files, credentials, permissions, scheduled jobs, and a durable sandbox, while Slack and the web share the same identity.

00:51 Teams can swap between code X and Claude code without rebuilding the core. Admins choose approval and content screening policies, but its dangerous mode removes both. Inkstone is a markdown notebook you deploy to your own Cloudflare account by forking the repo, with notes kept as plain text throughout. Wiki links, backlinks, block references, and a relationship graph put it in Obsidian territory, with full text search indexed in D1, including Chinese.

01:22 Offline writes queue in the browser and reconcile with optimistic concurrency, keeping conflict copies rather than overwriting. Gander is an Android viewer that opens documents, spreadsheets, slides, media, markdown, and code entirely offline. It requests no app permissions and doesn't even declare internet access. Files arrive through Android system picker, while office formats render inside a lockdown web view using bundled libraries.

01:49 You also get document search, recent file thumbnails, and deep zoom for large photos. Simple English gives coding agents a testable writing standard for technical documentation instead of the vague instruction to write clearly. It applies simplified technical English rules such as short sentences, active voice, consistent terminology, and conditions before commands.

02:14 The skill also adapts the method for errors, runbooks, incident reports, and release notes. Vercel's marketing team Eve template splits campaign work across five specialists instead of asking one general agent to fake every discipline. A lead loads shared brand context, then routes positioning, long-form content, social posts, SEO, or email to a fresh specialist session.

02:38 Deliverables flow into Notion, Typefully, and Resend, while broadcasts and destructive actions pause for human approval. Skill recorder turns a real work session into a reusable procedure for an AI agent. Record a task once, and it captures clicks, window switches, visited pages, and optional narration before Copilot reconstructs the intent and ordered steps.

03:00 You can review the result, then export it as a skill.md file or scheduled automation. Recording stays local until analysis, but that step sends screenshots, URLs, clipboard previews, and narration to GitHub's cloud, so keep secrets out. Humanizer CLI puts 33 common signs of machine-written prose in a local terminal reference with before-and-after examples for every pattern.

03:24 Its check command scans drafts for 13 mechanical tells, including padded phrases, chatbot leftovers, rhetorical openers, and overused punctuation. Search and show commands let you inspect the wider catalog while prompt copies the full editing guide for use elsewhere. Ratchet closes the loop between coding agent rules and the edits they produce. A post-tool hook examines each change for new dependencies, duplicate helpers, thin wrappers, and handwritten versions of standard library features.

03:56 It also tracks budgets for files, dependencies, and net added lines with an auditable ledger. Findings are graded by certainty, and strict mode blocks only definite cases. QR Ferry moves a file between two devices as an animated QR stream with no network involved on either side. The clever part is fountain coding. Source and repair symbols interleave, so the receiving phone can join mid-cycle, drop blurred frames, take them out of order, and still rebuild the file once it has enough unique symbols.

04:29 Four checksums verify before saving. WASTE runs mixture of experts models larger than available memory by keeping the shared trunk in RAM and streaming only selected experts from NVMe. A bounded cache reuses hot weights, while router lookahead starts disk reads before each expert it is needed. The maintainers report the full 2.78 trillion parameter Kimi K3 running on a 64-GB MacBook Pro at 0.45 to 0.62 tokens per second.

05:02 Anatomy, branded in the app as Anatomy Atelier, turns nine human organs into interactive 3S study objects. You can rotate and zoom each model, open labeled hotspots, isolate structures, and switch to cross-section or layer views. The surrounding interface adds key facts, organ comparisons, common conditions, and short quizzes. Kimi K3 in C runs Moonshots 2.78 trillion parameter base model on one x86-64 CPU by treating memory as a budget, not a fixed requirement.

05:32 The C99 engine streams a packed dense trunk from disk and keeps routed experts in an adjustable LRU cache with no GPU, BLAS, or inference framework. The maintainer measured 8.24 GB of RAM at the smallest preset, but also 32.69 seconds per token. You'll still need roughly 1.7 TB of fast storage. This CRM puts a durable research agent at the center with the database serving as its notebook.

06:03 It runs on its own deployment and schedule, leasing rows from a work queue. So, closing the browser doesn't stop it. The rule it never breaks is that nothing about a person is guessed. No tool accepts a confidence score since a model grading its own certainty will flatter itself. Weak evidence becomes a suggestion a human settles. Soul advisor turns Codex sub-agents into explicit delivery lanes instead of letting one session do everything.

06:31 The primary Soul session writes a five-part specification, routes routine work to Luna, and riskier changes to Terra, then verifies the resulting diff. A fresh Soul reviewer must return ship before completion. The workflow checks role files and runtime routing rather than silently falling back. 10 proofs turns 10 reported advances in mathematics and theoretical computer science into machine-checkable Lean certificates.

06:58 Instead of asking readers to trust prose alone, each result has a formalization that can be compiled against Lean 432.0 and Mathlib. The set spans sphere packing, non-sofic groups, circuit complexity, quantum games, and several long-standing conjectures. Termixer Unix terminal into a two-deck DJ mixer for live sets and title cycle sessions. It controls MPV and super collider sources with per channel faders, three-band EQ, filters, and a central crossfader.

07:29 A 4x4 sample pad grid includes its own sequencer. And the app can discover pulse audio, pipewire, and jack sources automatically. Inktide is a browser boat racing game where every mesh, texture, animation, and sound is generated in code. Six Gerstner waves drive an infinite ocean, and the same wave field feeds both rendering and six-point hull buoyancy, so boats actually pitch and roll with the water.

07:55 Cell shading combines hull outlines with a Sobel post-process for interior lines. Vidmap reconstructs a sparse 3D map and camera trajectory from ordinary video by using temporal structure directly in the reconstruction. It combines temporal feature tracks, loop closures, metric depth, and global optimization to estimate camera poses and intrinsics offline.

08:17 Results export as a COLMAP model with optional interactive HTML or rerun visualizations. The complete shelf packs an interactive seven-volume library into one HTML file. Each cloth-bound book is built from separate boards, hinges, page blocks, foil, and procedural textures, then opens into a responsive inspection view with orbit controls and curved page turns.

08:41 A small state machine keeps shelf-to-detail transitions continuous, while accessible HTML controls sit above the WebGL scene. Rabbit is a from-scratch C++ 3D engine built around a sparse set entity component system and an OpenGL renderer. Its Forge editor has dockable panels, live asset thumbnails, scene gizmos, and play controls that snapshot the scene, then restore it when you stop.

09:06 Assets use stable IDs and readable JSON, while Lua behaviors can hot reload during development. Graybox turns loose notes into a local cross-linked Markdown wiki without requiring a vector database. Captures stay verbatim in an immutable inbox, while an organizer extracts people, tasks, projects, and decisions into typed pages with backlinks. Questions search those pages, follow one graph hop, and cite the original notes instead of guessing.

09:38 Nightrun boots a computer directly into a local language model with no conventional operating system underneath. The UEFI application draws its own terminal, loads a quantized model fully into RAM, and then seals storage so generation performs no disk or network access. It runs on x86-64 PCs and Raspberry Pi 5 with output checked against llama.cpp. Cat assistant gives an MCP compatible agent read-only access to your Brazilian open finance data through pluggy.

10:06 It can group spending, inspect individual transactions, summarize credit card bills, and list investment positions across connected accounts. Merchant category corrections are learned locally and applied retroactively while diagnostics expose consent and sync status. GL4SS turns a map coordinate, a year, and an hour into a speculative historical scene, then optionally animates it with sound.

10:33 Its timeline uses 284 unevenly spaced stations while 55 prepared journeys jump to moments such as ancient Uruk or Apollo 11. You can pin two generated frames and wipe between them, and completed scenes stay in browser storage. Open derm is a research system that moves a camera over the body to create repeatable, high-resolution 3D maps of skin. A multi-axis gantry follows the subject's contour, maintains a configured 110 mm standoff, and captures images for reconstruction.

11:05 Later scans of the same site can be registered and compared for change. It is not a medical device and cannot diagnose cancer, and the robot presents impact, pinch, electrical, and laser hazards. DSCode is a local-first coding agent with deep seek defaults, but it can switch across Codex, Anthropic, and Grok without changing tools or sessions. It stores conversations as local JSON L trees, gives parallel implementers isolated Git work trees, and creates durable checkpoints after successful patches.

11:39 Commands run inside a network blocked workspace sandbox by default. Singular gives offline-first apps a real local SQLite database while keeping the server authoritative. Browser clients persist through OPFS, native clients use SQLite, and writes enter an optimistic outbox before reconciling against one ordered server commit log. The protocol has separate TypeScript and Rust cores checked by the same conformance fixtures, plus web and mobile bindings.

12:11 YouPool replaces hand-editing configuration files when switching model providers across Claude Code, Codex, Hermes, and Open Code. It stores endpoints, keys, and model choices in one desktop app, probes reachability without generating tokens, and updates only the config fields it owns. Atomic writes and adjacent backups reduce the chance of damaged settings.

12:33 BC7215AC Matter turns an infrared air conditioner into a Matter device using an ESP32. It learns the protocol from the existing remote, then exposes temperature, fan, and power controls to Apple Home, Google Home, Alexa, or Home Assistant without modifying the AC's circuitry. Commands from the original remote can also sync back to the phone. Headroom puts your AI coding quotas and deployment health in one glance instead of making you hunt through billing pages and dashboards.

13:04 A local Python host reads existing Claude, Codex, Cursor, and other tool sessions, then feeds a macOS menu bar app, iPhone widgets, Apple Watch complications, or an optional ESP32 desk display. It also surfaces failed builds and service alerts. AWS's context ontology accelerator gives agents structured business context instead of dumping more documents into a prompt.

13:29 It scans databases and unstructured files, models their relationships as an ontology and knowledge graph, then serves validated context through Sparkle queries and MCP tools. Namespace isolation and role-based permissions keep teams separated. Trace file lineage answers the awkward question, which script, notebook, or agent created this file? It scans code, document metadata, and Git history to rank likely sources while clearly separating guesses from recorded proof.

14:00 Wrap a command and it captures every changed file, builds a dependency view, and flags outputs made stale by updated inputs. Everything stays local. Scopey watches coding agent sessions for the moment they quietly drift away from your request. The Rust CLI converts each new prompt into a current scope, evaluates tool activity in bounded background jobs, and can inject a correction or send a notification when work goes off track.

14:26 It supports Claude Code, Codex, and Open Code while ignoring child agent sessions by default. Bindwidth sizes private model infrastructure around the constraint that stops your workload rather than defaulting to GPU memory alone. It calculates limits for KV cache, serving capacity, and runtime concurrency, then compares eligible hardware, sovereign rentals, and enterprise subscriptions. Every profile is labeled as measured, published, community reported, or estimated, and everything runs in your browser.