🛠️ Hacker News Tools

[Other] Chat-tails: Throwback terminal chat, built on Tailscale

[Other] Show HN: Deterministic PCIe Diagnostics for GPUs on Linux I built a small Linux tool to deterministically verify GPU PCIe link health and bandwidth.<p>It reports: - Negotiated PCIe generation and width - Peak Host→Device and Device→Host memcpy bandwidth - Sustained PCIe TX&#x2F;RX utilization via NVML - A rule-based verdict derived from observable hardware data only<p>This exists because PCIe issues (Gen downgrades, reduced lane width, risers, bifurcation) are often invisible at the application layer and can’t be fixed by kernel tuning or async overlap.<p>Linux-only: it relies on sysfs and PCIe AER exposure that Windows does not provide.

[Other] Show HN: I built the fastest RSS reader in Zig Well, I certainly tried. I had to, because it has a certain quirk inspired by &quot;digital minimalism.&quot;<p>The quirk is that it only allows you to fetch new articles once per day (or X days).<p>Why? Let me explain...<p>I want my internet content to be like a boring newspaper. You get it in the morning, and you read the whole thing while sipping your morning coffee, and then you&#x27;re done! No more new information for today. No pings, no alerts, peace, quiet, zen, etc.<p>But with that, I needed it to be able to fetch all articles from my hundreds of feeds in one sitting. This is where Zig and curl optimisations come in. I tried to do all the tricks in the book. If I missed something, let me know!<p>First off, I&#x27;m using curl multi for the network layer. The cool thing is it automatically does HTTP&#x2F;2 multiplexing, which means if your feeds are hosted on the same CDN it reuses the same connection. I&#x27;ve got it configured to handle 50 connections total with up to 6 per host, which seems to be the sweet spot before servers start getting suspicious. Also, conditional GETs. If a feed hasn&#x27;t changed since last time, the server just says &quot;Not Modified&quot; and we bail immediately.<p>While curl is downloading feeds, I wouldn&#x27;t want CPU just being idle so the moment curl finishes downloading a single feed, it fires a callback that immediately throws the XML into a worker thread pool for parsing. The main thread keeps managing all the network stuff while worker threads are chewing through XML in parallel. Zig&#x27;s memory model is perfect for this. Each feed gets its own ArenaAllocator, which is basically a playground where you can allocate strings during parsing, then when we&#x27;re done, we just nuke the entire arena in one go.<p>For parsing itself, I&#x27;m using libexpat because it doesn&#x27;t load the entire XML into memory like a DOM parser would. This matters because some podcast feeds especially are like 10MB+ of XML. So with smart truncation we download the first few X mb&#x27;s (configurable), scan backwards to find the last complete item tag, cut it there, and parse just that. Keeps memory usage sane even when feed sizes get massive.<p>And for the UI I just pipe everything to the system&#x27;s &quot;less&quot; command. You get vim navigation, searching, and paging for free. Plus I&#x27;m using OSC 8 hyperlinks, so you can actually click links to open them on your browser. Zero TUI framework needed. I&#x27;ve also included OPML import&#x2F;export and feed groups as additional features.<p>The result: content from hundreds of RSS feeds retrieved in matter of seconds, and peace of mind for the rest of the day.<p>The code is open source and MIT licensed. If you have ideas on how to make it even faster or better, comment below. Feature requests and other suggestions are also welcome, here or GitHub.

[Other] No Graphics API

[Other] The GitHub Actions control plane is no longer free

[Other] GitHub will begin charging for self-hosted action runners on March 2026

[Other] Show HN: Ducktape – a tiny HTTP/2 wrapper around DuckDB's Appender API Hi HN! I’m an engineer at Artie where we do real-time data replication.<p>We were adding MotherDuck as a destination and the first version just used DuckDB’s Go driver directly. It worked great on my machine… until we wired it into our Transfer service (<a href="https:&#x2F;&#x2F;github.com&#x2F;artie-labs&#x2F;transfer" rel="nofollow">https:&#x2F;&#x2F;github.com&#x2F;artie-labs&#x2F;transfer</a>).<p>Because the driver requires CGO, our cross-compiles to amd64 and arm64 started failing, we lost our easy static binaries, and our Docker images had to pull in C toolchains and system libraries just to support one dependency. We tried isolating the CGO bits in a separate module, but it still caused CI failures and forced us to rewrite chunks of our build pipeline. At that point it was clear we didn’t want CGO anywhere near our main service.<p>So I built ducktape: a tiny standalone microservice that wraps DuckDB’s Appender API behind HTTP&#x2F;2 streams. Clients stream NDJSON over HTTP&#x2F;2, and ducktape appends directly into DuckDB on the other side. No CGO in the main codebase, and we keep our cross-platform, pure-Go build story.<p>The overhead was surprisingly low in benchmarks: ~757 MiB&#x2F;sec over HTTP&#x2F;2 vs ~848 MiB&#x2F;sec in-process — about 90% of native performance but over the network.<p>ducktape is open source and MIT licensed: <a href="https:&#x2F;&#x2F;github.com&#x2F;artie-labs&#x2F;ducktape" rel="nofollow">https:&#x2F;&#x2F;github.com&#x2F;artie-labs&#x2F;ducktape</a><p>I’d love feedback, especially if you’ve tackled CGO isolation differently or have ideas to squeeze out more performance!

[CLI Tool] AIsbom – open-source CLI to detect "Pickle Bombs" in PyTorch models

[Other] Gh-actions-lockfile: generate and verify lockfiles for GitHub Actions

[Database] Biscuit is a specialized PostgreSQL index for fast pattern matching LIKE queries

[Other] Show HN: Misata – synthetic data engine using LLM and Vectorized NumPy Hey HN, I’m the author.<p>I built Misata because existing tools (Faker, Mimesis) are great for random rows but terrible for relational or temporal integrity. I needed to generate data for a dashboard where &quot;Timesheets&quot; must happen after &quot;Project Start Date,&quot; and I wanted to define these rules via natural language.<p>How it works: LLM Layer: Uses Groq&#x2F;Llama-3.3 to parse a &quot;story&quot; into a JSON schema constraint config.<p>Simulation Layer: Uses Vectorized NumPy (no loops) to generate data. It builds a DAG of tables to ensure parent rows exist before child rows (referential integrity).<p>Performance: Generates ~250k rows&#x2F;sec on my M1 Air.<p>It’s early alpha. The &quot;Graph Reverse Engineering&quot; (describe a chart -&gt; get data) is experimental but working for simple curves.<p>pip install misata<p>I’d love feedback on the simulator.py architecture—I’m currently keeping data in-memory (Pandas) which hits a ceiling at ~10M rows. Thinking of moving to DuckDB for out-of-core generation next. Thoughts?

[Other] Rust GCC backend: Why and how

[CLI Tool] Show HN: Open-source Markdown research tool written in Rust – Ekphos Hi! I just made an obsdian alternative in terminal after searching for an Obsidian like TUI and got nothing. So I built one myself.

[Other] A proof of concept of a semistable C++ vector container

[Other] VS Code deactivates IntelliCode in favor of the paid Copilot

[Other] Show HN: Autograd.c – a tiny ML framework built from scratch built a tiny pytorch clone in c after going through prof. vijay janapa reddi&#x27;s mlsys book: mlsysbook.ai&#x2F;tinytorch&#x2F;<p>perfect for learning how ml frameworks work under the hood :)

[Other] A kernel bug froze my machine: Debugging an async-profiler deadlock

[Other] I used Claude Code to write a piano web app

[Monitoring/Observability] Show HN: Cordon – Reduce large log files to anomalous sections Cordon uses transformer embeddings and density scoring to identify what&#x27;s semantically unique in log files, filtering out repetitive noise.<p>The core insight: a critical error repeated 1000x is &quot;normal&quot; (semantically dense). A strange one-off event is anomalous (semantically isolated).<p>Outputs XML-tagged blocks with anomaly scores. Designed to reduce large logs as a form of pre-processing for LLM analysis.<p>Architecture: <a href="https:&#x2F;&#x2F;github.com&#x2F;calebevans&#x2F;cordon&#x2F;blob&#x2F;main&#x2F;docs&#x2F;architecture.md" rel="nofollow">https:&#x2F;&#x2F;github.com&#x2F;calebevans&#x2F;cordon&#x2F;blob&#x2F;main&#x2F;docs&#x2F;architec...</a><p>Benchmark: <a href="https:&#x2F;&#x2F;github.com&#x2F;calebevans&#x2F;cordon&#x2F;blob&#x2F;main&#x2F;benchmark&#x2F;results&#x2F;README.md" rel="nofollow">https:&#x2F;&#x2F;github.com&#x2F;calebevans&#x2F;cordon&#x2F;blob&#x2F;main&#x2F;benchmark&#x2F;res...</a><p>Trade-offs: intentionally ignores repetitive patterns, uses percentile-based thresholds (relative, not absolute).

[IDE/Editor] Show HN: I built an open-source alternative to the "brainrot IDE" that YC funded VSCode extension with games (Snake, Plinko, slots), HN&#x2F;LessWrong reader, and Pomodoro timer for AI coding wait times. Fake currency only.<p>Marketplace: <a href="https:&#x2F;&#x2F;marketplace.visualstudio.com&#x2F;items?itemName=TouchGrassIDE.touch-grass-ide" rel="nofollow">https:&#x2F;&#x2F;marketplace.visualstudio.com&#x2F;items?itemName=TouchGra...</a><p>Built in a week with Claude Code after seeing Chad IDE get ratio&#x27;d for integrating real gambling. Happy to answer questions or get roasted on the code.