Hetzner inference - free OpenCode with Qwen3.6-35B

Hetzner now offers experimental (but free) OpenCode inference running Qwen3.6-35B. Handy for hobby projects once your free OpenCode usage runs out.

Hetzner inference (Sliplane)

DeepSeek via Claude Code

DeepSeek V4 is a pretty good coder, and you can plug it into Claude. It was great back when the tokens were ultra-cheap; still worth a try.

DeepSeek for Claude Code

Picking model vs. reasoning effort

A useful heuristic: if the problem feels like it needs a genius, reach for a bigger model; if it needs a pen and lots of paper, crank up the reasoning effort; if it needs a genius sitting down with a pen and paper, turn both up to 5/5.

Threads post

Context bombs - stopping AI attackers in their tracks

Tracebit’s research on “context bombs” - a class of attack where malicious content planted in an agent’s context detonates later rather than immediately.

Context bombs (Tracebit Research) tracebit-com/context-bombs

meat - review only the parts that matter

Humans still need to review agent-written code in critical systems, but not for style, nil-checks, or imports - the review that matters is concepts, algorithm choices, and architecture. meat uses a model to reduce a diff down to exactly those parts and show you “the meat”. A different take than hunk, which also shows up a lot in my feeds.

meat on GitHub hunk on GitHub

papercuts - a complaint box for coding agents

A tiny CLI that gives coding agents a complaint box: a JSONL file in the repo where they log the friction they hit (dead-end tool calls, broken links, missing helpers) before pushing through. Interesting, have not tried it though. Other people had the same idea with frog.

papercuts on GitHub frog on GitHub

Proactive agents

Boris Tane (of loggingsucks fame) wants LLM-backed agents to take over on-call. This requires a lot of plumbing (which his company is building), but fundamentally you also go from passive agents (they only work when you set a goal) to proactive, reacting on e.g. drops in P99 response times.

Proactive agents (Polylane)

Each AI agent gets its own GitHub identity

I need to get this set up: Give every AI agent its own GitHub identity so bot commits are attributed per-agent rather than all lumped together.

Each AI agent gets its own GitHub identity

MAUI AI design skill

A MAUI app-design agent skill - a SKILL.md capturing the design guidance you would otherwise have to paste into every prompt. I have dabbled a bit in MAUI for Android app development recently.

MAUI AI design skill SKILL.md

Why human-level AI won’t be enough

Adam Tornhill argues that even human-level AI will not be enough, since AI produces so fast, even small quality issues will become problematic.

Why human-level AI won’t be enough

Anders Hejlsberg on 10x TypeScript and AI

Anders Hejlsberg recommends an approach I have also had success with: Instead of asking coding agents to do large-swath changes, use them to produce deterministic code, e.g. scripts, to do the change. Then application becomes deterministic. They touch shortly on the issue of replacing junior devs with agents - where will the next seniors come from then?

Anders Hejlsberg: 10x Faster TypeScript & why AI won’t replace SWEs (YouTube)

Two new Steve Yegge essays

Agents are persons, and you need to treat them well? Not sure what to say, I had a few discussions on this one. I am still on team “it is a probabilistic next-token predictor”, but I guess if you forget that and focus solely on the emergent properties of what it does it might make sense to view it that way? I love Yegges new AI comics, inspired by the crazy happenings in wheelhouse.

The Shape of Things to Come Model Welfare Fences, not Sandboxes

Bridging the AI adoption gap

You think you are falling behind, because you are not running agent swarms or a software factory like wheelhouse or Gas Town. But in reality, you are probably top 1% adopter if you are actively using a coding agent daily.

Bridging the AI adoption gap AI Adoption is a myth

You should deploy directly to prod

They are not wrong. It is like the old XP mantra: If it hurts, do it more often (to force improvement). But you will need solid, dependable instrumentation, and deployment automation.

You Should Deploy Directly to Prod

Varlock Azure Key Vault plugin

Varlock’s Azure Key Vault plugin - “SOPS on steroids, with schema”. A declarative, schema-validated approach to managing Key Vault secrets that looks more ergonomic than the usual custom tooling.

Varlock Azure Key Vault plugin

NuGet issue #8287 - seven years and counting

Lock files prevent publishing when <RuntimeIdentifiers> is set. Seven years and counting - this bit me this month.

NuGet/Home#8287

mise on Windows

I mentioned mise last month; worth noting it now installs cleanly on Windows via winget install jdx.mise, and the dev-tools page is a good overview of what it can manage.

mise dev tools

inshellisense

Microsoft’s terminal autocomplete for shells (bash, zsh, fish, PowerShell, and more) - IDE-style completions right in the terminal. Unfortunately it messed up my NixOS terminal, caveat emptor.

inshellisense on GitHub intro

WinDirStat

WinDirStat now supports fast NTFS scanning (the trick WizTree popularized), and it is open source / free for non-commercial use. winget install windirstat.windirstat.

WinDirStat

WSL Dev Drive performance

I have been working on sandboxing coding agents in WSL. Conclusion: Do not share a repo between Windows and WSL-Linux. You can choose which side is going to be dog-slow (wsl.localhost / /mnt/c), and you are going to have locking conflicts, line ending conflicts … just don’t. I think a sandbox VM is the way to go, and with NixOS you can make such a VM shareable in a team.

WSL Dev Drive performance

Microsoft developer changelog

Microsoft’s developer changelog - the firehose of everything shipped across their developer products. With so much churn happening, not sure if I can take this in too.

Microsoft developer changelog

Lightpanda - a curl-like browser

Lightpanda is a headless browser you can use like curl, with JavaScript evaluation and markdown conversion built in. Nice for turning pages into LLM-friendly text without a full browser stack.

lightpanda-io/browser

Cache break plugin

When you do things in Copilot that breaks the token cache, that costs you. This plugin will notify you when that happens, so you can develop a feel for what not to do.

Cache break plugin (gist)

Not holding back the ocean

An ode to coding going away, with devs moving up the stack.

Not holding back the ocean

Varar - from the creator of Cucumber

From the creator of Cucumber: Varar reimagines BDD with markdown “oaths” referencing “steps” (code). You can then check if the implementation breaks an oath. I never jumped on the Cucumber train, it is too technical for the business people I work with, and too untyped for myself.

varar.dev

On building scalable control planes

Werner Vogels on how AWS evolved the EC2 control plane to scale. A good systems read on building control planes that do not fall over under load.

On building scalable control planes

Soulver

Soulver is a “notepad calculator” - you type calculations mixed with text and it evaluates them. Mac/iOS only, unfortunately, but interesting concept.

Soulver

How contactless tap & pay works

A good explainer on how contactless tap & pay actually works under the hood - EMV, NFC, and the cryptography that makes it safe enough to wave your card at a terminal.

Contactless tap & pay - how it works

TMOG - Task Manager OG

Dave Plummer (the original Windows Task Manager author) is back with TMOG - Task Manager OG - a from-scratch, deeply instrumented task manager for macOS, Windows, and Linux. Free public beta. I love the neon default look!

TMOG

Thimbleweed Park 2

Ron Gilbert announced Thimbleweed Park 2 - “not really a sequel, not really a prequel”, just more Thimbleweed Park, due early 2028 on PC/Mac/Linux. The original team (Mark Ferrari, Gary Winnick, David Fox, and others) is back, and there is a new dev blog to follow. I discovered Thimbleweed Park very late (post Return to Monkey Island), but it is a great game.

Thimbleweed Park 2 announcement TWP2 dev blog

M.2 - NVMe and SATA

I was today years old when I learned that some motherboards only support NVMe M.2 cards - not SATA ones. Setting up my new Medion S06e for the home lab, the M.2 SATA stick I bought was not recognized at all.

I then choked on my morning coffee after seeing current NVMe prices - they, too, have been hit by the RAMpocalypse, with prices basically doubling year over year. Fortunately I could shuffle an NVMe drive out of an old laptop, which happily accepted the SATA replacement in its M.2 slot.

Medion S06e product detail

Omarchy Quattro

Besides dabbling with NixOS, I also installed Omarchy on an old laptop.While it will take some getting used to to work productively with the tiling paradigm, the UI consistency and eye candy is unparalleled in Linux, rivaling MacOS. With the passion of DHH (and 8 million USD) behind it, this is the first Linux distro that makes me consider switching my daily driver laptop…

Omarchy Quattro