Forkcast Hacker News #5: bsharp, make-look-scanned, recall, talos, ember-hackernews, shumai

Episode Summary#

Welcome back to Forkcast. This is Hacker News show number 5. 35 open-source projects that have recently been trending on Hacker News, including a GPU-accelerated Git client hitting 120 frames per second on a 37,000-line file. Let’s get into it.

Repositories Covered#

bsharp#

Perfect pitch is usually considered an innate gift, but research shows children can develop it with the right training method — and there’s no accessible app for that. Eguchi chord identification method with 9 white-key and 5 black-key chords, adaptive weighting algorithm, multi-user profiles, ~20-2

make-look-scanned#

You need a PDF to look like a physical scan of a printout — skew, grain, warm paper tone — but existing tools are clunky desktop software. CLI + browser WASM, deterministic output via content-hash seed, 8 tunable knobs (skew, noise, blur, edge shadow, jpeg quality), ~9 commits, Go + MuPDF, browser u

recall#

Every Claude Code session starts cold — you waste tokens re-explaining your project from scratch. Fully local, zero API calls, TF-IDF + TextRank summarizer, resume from ~1-2K token context.md, auto-save on session end, history.md append-only log, MIT license

talos#

WebAssembly needs formal verification but existing interpreters are separate from proof tools — you can’t prove theorems about the same code that runs. Wasm interpreter in Lean 4, weakest precondition calculus, executable semantics = proof object, 109 commits, 3 Lake packages (Interpreter → CodeLib

ember-hackernews#

Most Hacker News readers are web-based, not native, and accessibility is treated as an afterthought. Native SwiftUI iOS/macOS app, zero third-party dependencies, 24 commits, VoiceOver with meaningful labels, color-blind friendly cues, offline disk cache, threaded comment rendering, MIT license

shumai#

Frame.io is the standard for video review but it’s closed-source, expensive, and you can’t self-host it. S3-compatible storage, frame-by-frame annotations, AI agent with custom skills, Gemini metadata autofill, semantic search via vector embeddings, distributed transcoding via Temporal, sandboxed ag

bun-sqlgen#

Raw SQL in Bun is fast but untyped — one misspelled column and you get runtime errors instead of compile-time safety. Type generator for Bun.sql queries, validates against real Postgres/SQLite schema at build time, no ORM, no Docker needed, CI check flags, null-safe types

pagecast#

Coding agents generate HTML reports and dashboards but sharing them requires deploying to a server or sending files. Publishes HTML/Markdown to Cloudflare Pages from terminal, password protection via edge functions, expiring URLs, live goal progress pages, agent skills for Codex and Claude Code, dep

kyde#

Full IDEs are overkill when you mostly just need commit, diff, and browse views — and you want it fast. GPU-accelerated Git client in Rust + gpui (Zed’s GUI framework), ~120fps on 37k-line file, embedded PTY terminal, tree-sitter syntax highlighting, side-by-side diff with word-level highlights, mac

ymawky#

Web servers are built on layers of abstraction — nobody writes them in raw assembly anymore, but doing so reveals exactly what the kernel provides. Pure ARM64 assembly, syscall-only (no libc), fork-per-connection, 569 stars, GET/PUT/DELETE/POST/CGI, path traversal prevention, Slowloris protection, 2

Tiny#

You want a fast scripting language but also real concurrency across CPU cores — not an event loop. Bytecode VM in Go, JIT compilation to WebAssembly, OS-level threads via spawn/await, structural typing, pattern matching, packed arrays, VS Code LSP, MIT

claude-pulse#

Claude Code runs in the terminal — you can’t see token usage, approve tool calls from your phone, or recover crashed sessions. Local zero-dependency dashboard, phone approvals via ntfy, session recovery, live token cost tracking, full-text search across all sessions, Codex support added, 100% local,

selector-forge#

Writing resilient CSS selectors for web scraping and testing is tedious — one DOM change and everything breaks. Browser extension (Chrome + Firefox), AI-generated selectors re-verified against live DOM, single and list selection modes, Manifest V3, built with WXT + React, MIT

y#

Coding agent apps have fixed UIs — you can’t ask the app to reshape itself to fit your workflow. Self-modifying Electron desktop app, Modify rail edits live UI with diff review, protected kernel vs mutable userland, runs Claude Code + Codex locally, isolated parallel workspaces, rollback snapshots

buzz#

Team workspaces treat AI agents as second-class citizens — bots get restricted permissions, separate APIs, and no audit trail. Nostr relay where agents and humans share the same rooms, signed events for everything, git events via NIP-34, YAML workflows, Tauri desktop app, buzz-cli for agents, Apache

crawlie#

SEO audit tools are either expensive SaaS subscriptions or command-line tools that produce raw data without explanations. 49 audit rules with plain-English fixes, CLI + MCP server + desktop app, dual Health + GEO scores, Rust engine, agent skills for autonomous audits, MIT

halo#

AI agents fail in production but debugging them means reading thousands of trace lines manually — you need systematic failure analysis. RLM-based agent optimizer, collects OpenTelemetry traces, identifies systemic failure modes, generates harness changes, desktop app + Python CLI, 946 stars, gpt-5.4

Next-Move-Theory-Canon-and-Skills#

Product decisions are treated as intuition contests — there’s no structured algorithm for choosing what to build next. Advanced JTBD v3.4 framework rebuilt on neuroscience, Next Move Theory with RAT + ABCDX segmentation + Theory of Constraints, 7 Claude Code skills (market research → value prop → PR

crespo#

Pasting entire codebases into LLMs wastes tokens on boilerplate — you need a compact structural view. Tree-sitter AST parsing to XML blueprints, 10 languages, ~86% average token reduction, 3 modes (structure/summary/concat), Groq-powered AI summaries, MIT

pii-gui#

Redacting PII from documents before sharing them with AI tools is a chore — and you don’t want your data leaving your machine. Tauri 2 desktop app, local ONNX models for PII detection, true PDF redaction with opaque rectangles, OpenAI privacy-filter + BardsAI EU PII backends, ~1.8K tok/s on M1 Max,

glueRun-go#

Running AI coding agents in parallel on a repo is chaotic — you need leases, gates, audits, and worktree isolation. 3-tier scheduling (L0/L1/L2), git-worktree isolation per task, durable JSON leases, detached dispatch, auditor model validates output, decider fast-path for failures, GPL-3.0

Debug-Brief#

Debugging sessions produce no documentation — what you learned stays in your head instead of becoming a PR comment. Records terminal commands + notes, builds evidence-backed Markdown report, tracks red-to-green window, recognizes 10+ test runners, secrets redacted, pure Python stdlib, MIT

react-native-boost#

React Native’s standard Text and View components have hidden performance costs that compound in complex UIs. Babel plugin auto-swaps RN components for native primitives, one-line config, 505 stars, measurable FPS improvements, Expo-compatible, 24 releases, MIT

SQLJoiner#

Writing complex JOIN queries by hand is error-prone — you need to see the table relationships visually. Drag-and-drop canvas for SQL JOINs, import existing SQL to reverse-engineer onto canvas, saves contexts, paginated results grid, PHP backend + vanilla JS frontend, Electron desktop app, GPL-3.0

wsl-dashboard#

Managing WSL instances means typing wsl.exe commands — there’s no polished GUI for distro management. Rust + Slint + Skia rendering, ~10MB RAM in tray, 35+ languages, port forwarding + firewall, USB device management via usbipd-win, distro migration and export, 2.8K stars, GPL-3.0

quikdown#

Markdown editors are either bloated or incomplete — you want bidirectional HTML conversion, rich fences, and a tiny footprint. ~17KB core, zero dependencies, bidirectional Markdown↔HTML, drop-in editor with Mermaid/MathJax/STL/GeoJSON fences, MCP server with 24 tools, 99.3% test coverage, BSD-2-Clau

code-mapper#

Reading code you didn’t write — or code an LLM wrote for you — means manually tracing imports and data flows. Static import graph + LLM semantic annotations, function-level call graph, Obsidian-linkable Markdown vault, local Ollama support, Bun runtime, MIT

graphsignal-profiler#

GPU inference profiling is either NVIDIA Nsight (complex) or nothing — you need production-grade visibility without the overhead. Continuous CUDA profiling via CUPTI, sidecar process for zero workload impact, vLLM + SGLang + PyTorch support, LLM generation tracing with per-step timing, AI optimizati

domvault#

Your second brain tool owns your data — you want plain Markdown you control, with AI surfacing connections. Self-hosted knowledge graph, plain Markdown vault, agent-driven synthesis loop (sparks → neurons), opt-in Constellation publishing with ed25519 signatures, agent contract CLAUDE.md, PolyForm N

nanoeuler#

Understanding how LLMs actually work means reading PyTorch abstractions — you want to see every matrix multiply and every gradient. ~116M-parameter GPT-2 style in C/CUDA, hand-written forward+backward passes, custom FlashAttention kernel (~3x speedup), byte-level BPE tokenizer, pretrain→SFT pipeline

Cachet#

LLM API calls cost money on every retry and re-run — even when you’re asking nearly the same question. Drop-in proxy in Rust, dual-layer cache (exact hash + semantic embedding), local lexical embedder, streaming SSE tee+replay, live savings dashboard, single binary, MIT

prismag#

Different LLM tasks need different models — but switching means juggling multiple chats and copy-pasting context. Per-block model routing via @@tags, Go CLI, chained execution with context flow, supports Anthropic/OpenAI/Ollama/vLLM, Cursor + Claude Code integration, REPL with transcripts

ze#

Standard directory jumpers decay with real time — go on vacation and all your scores reset to zero. Event-clock-based exponential moving sum scoring, fork of z.sh, no decay during idle periods, half-life ~87 cd actions, bash/zsh/ksh93/mksh compatible

duckle#

ETL tools are either cloud-locked SaaS or heavy Java servers — you want drag-and-drop pipelines that run locally. ~65MB single-file desktop app, 290+ connectors, 50+ transforms, local AI assistant (Qwen 2.5 Coder 1.5B via llama.cpp), DuckDB execution, git-friendly JSON workspaces, MCP server, 60 UI

memoryops-ai#

AI assistants have memory features but no governance — no audit trail, no deletion guarantees, no tenant isolation. Enterprise memory governance layer, 10 enforced invariants (tenant isolation, deletion guarantee, provenance, policy-before-storage), typed memory (episodic/semantic/procedural), FastA

Watch#

• Video: https://www.youtube.com/watch?v=pdc1r7pHo6E
• Cover: https://i.ytimg.com/vi/pdc1r7pHo6E/hqdefault.jpg

Notes#

Transcript and notes will be added from Forkcast output artifacts.