So much has happened in the time I was busy carving wooden spoons, that I felt it simpler to just ignore the hype.

It will pass, just like the NFTs and the dApps and microservices, I don’t need to care about it. Or that’s how I thought.

One morning, in this seemingly never ending winter, a friend told me excitedly how his employer got them Claude subscriptions and encouraged them to try it for work. And his reaction was something along the lines: “it’s crazy! I love it! we’re doomed…”

I felt that was an overreaction, I tried LLMs for coding before and aside from small snippets, it couldn’t do more than 50 lines without stumbling on itself.

But it was a really gray hazy winter morning, and I didn’t feel like splitting wood in the quiet of the house, so I made some coffee and got my laptop out.

Stages

I gave it the usual task that I really want to see done but I’m always dreading to start: Stages in rcmd

rcmd is a macOS app switcher that allows you to hold Right Command, and press the first letter of the app name. So press V for VSCode, G for Ghostty, C for Chrome: instant app switching without cycling through app icons.

Stages was dreamed up as a way to switch workspaces with the same paradigm. So press C for Coding where VSCode and Ghostty would focus and everything else was hidden, D for Design and so on.

But the functionality is way harder than it seems, the app needs to record the on-screen windows, and be able to restore them when the computer is restarted.

The task was complex, just to outline some things Claude needed to do:

• understand the pre-existing app codebase with all its UI quirks
• refactor the UI to be able to present sets of apps and windows
• find ways to record custom window data like:
  • the project opened in VSCode
  • the working directory in Ghostty
  • the current URL in Chrome
  • files/folders/projects opened in other apps like Finder, Pixelmator, Sketch, Xcode etc.
• find ways to re-open windows at that specific project/file/folder
• figure out a way to hide individual windows on macOS

It’s weird, but there’s no instant hide/show for windows on macOS. You can minimize them, but that has slow animations and unwanted side effects.

I mentioned to Claude that I knew a window manager called Aerospace which had a workaround for this.

I gave Claude a 200 line stages.md, let it loose on the rcmd code, and stared at it crunching stuff for a few minutes. I went to make another coffee, and when I came back it was… done?

I hit Build in Xcode and was prepared to see a barrage of errors, that’s how it always was. I got ready to screenshot the errors and the stupid code diff and show my friend how “no, we’re not doomed, it just got lucky on that specific task”.

No errors though, the app compiled and launched like nothing happened. Uhh.. now what? The code diff was immense, I guess I should just see if anything actually works anymore.

I hid every app except VSCode and Ghostty, pressed rcmd + lalt - C and it said Stage C was assigned. No way this works… I brought back all apps on screen, quit VSCode, pressed rcmd - C and instantly VSCode and Ghostty were the only apps visible and focused, in the correct folder and project.

I looked through its thinking process and it even did a web search for Aerospace macOS window manager, found its Github, reverse engineered its workings until it found the window hiding workaround I mentioned in passing and implemented it from scratch in rcmd.

Astonished is a small word for what I felt.

Of course, this was not a feature ready for release. There was so much code review and so much polishing work to do (not all apps have ways to restore their state after quitting, some focus flickering happened, and Stage creation should be more intuitive).

But the fact that an LLM could do such complex work, changed everything.

Picture-in-picture

🔗 Pipiri is a macOS app that can show any window in a floating picture-in-picture window

I had this barebones app I did as a standalone Swift file, for creating a Picture-in-Picture view of a window where there’s data I’d like to keep an eye on.

I mostly used it to watch long-running terminal commands while I continued coding, and for keeping Xcode logs visible while I tested parts of my app.

I thought it was too niche to release, so I didn’t have the motivation to do the work on creating an UI, licensing, webpage, icon etc.
I saw this as an opportunity to see if Claude can do it. After all, I still felt this to be a useful idea after years of using it.

I created a blank Xcode macOS app project and gave Claude minimal instructions:

• adapt pip.swift to a menubar app with a SwiftUI architecture
• add a Settings window where I can adjust the trigger hotkey, the position of the PiP panel etc.
• implement auto-updates with Sparkle
• add license verification through Gumroad
• link to a Contact form and a privacy policy, similar to how I do it in my other apps

What can I say? It did everything. Granted, the main PiP functionality had already been written manually by me years ago, this was just doing the things that every macOS dev hates to do but has to repeat on every app.

Then onto the webpage. Manually wrote a list of features with explanations in a README.md because I prefer to address future users of my apps personally, then told it to build the landing page by taking inspiration from my other app pages, using what I wrote in the README.

I was more than relieved to see 95% of the webpage look exactly as I would have done it myself. Not a single word changed, only HTML structure, colors, image paths etc. Adapting old webpage structure for a new app was not something I enjoyed doing.

I also got a little carried away and asked Claude to implement a VM bytecode interpreter and write the offline license validation in that bytecode. I didn’t expect it to actually do it! It even added a license generator CLI on top.

I launched a v1.0 and it was received well enough. Got 2 feature requests from the first day:

• Multi-window mode (allow keeping multiple PiP panels visible from different)
• Idle/Change detection (get notified when the window contents haven’t changed in a while, or if they did change after being still for a while)

These are features I would have liked myself, but was too busy to start working on them before.

Now, Claude was not able to do these properly by simply mentioning the features. I had to explain the features in detail, how I want the multi-window PiPs to appear on top of each other, what APIs it should use, describe how the idle notification should look and that it’s not a system notification.

In the end, the code needed a lot of review, testing and manual fixing, but just getting me started on the features was enough to motivate me to finish and release them.

Getting a semi-functional feature in readable code that I can polish is much more motivating than getting a huge refactor and full implementation that doesn’t even compile.

After releasing the change detection feature, I got a nice message from someone working at a hospital telling me how they use it to know when a patient has been added to the admission list without having to check it very second.

This is always humbling to me when it happens, reminds me how naive and limited we are as devs thinking that we considered all use cases. When in fact we should probably just build the tools, and people will find use for them in their own way.

Event-based automation

🔗 Crank is a macOS app that can run scripts and Shortcuts on event triggers

My little brother was telling me about how he’s trying to find ways to make money without having to work long hours in coffeeshops and restaurants.
These AI tools are an opportunity for people like him, who know what they’d like to be doing, but can’t because they have to sell their time for a meager salary.

He has already started a few ideas like websites for small local shops that want an online presence, copywriting, product photo shoots.

So I thought why not also throw in a simple macOS app. I can bring in the ideas that I’ve been sitting on for the past 5 years, help with code review and fixing, and he can just prompt Claude and test the app until it’s working as expected.

The main theme was simple, low maintenance apps only.

I had this idea for an event-based automation app: just the basic parts of what has already been done in Keyboard Maestro or Shortery, at a much cheaper price and free for basic needs.

In support emails and chats, I’ve had a lot of people ask for some very niche functionality in my other apps. Usually, it’s some kind of automation, things like:

• “can I disable the MacBook screen only when I connect my Dell monitor at home?”
• “how can I make all screens black when I want to audition a piece in my recording studio?”
• “can optimised images in ~/Blog be renamed to YYYY-MM-dd-name and converted to webp?”
• “please add an option to toggle HDR when I open IINA or any other movie app”

So I wanted to give those people the ability to resolve their frustrations without having to wait for a dev to do it.

I hopped on a screen sharing session with my brother and let him ask Claude to build this app bit by bit. The first thing that jumped out is: you have to know the terminology to get quality results.

He’s not a dev, he doesn’t know what a debounce or throttle is, so incidentally, he doesn’t know to ask Claude to implement it for events that arrive in bursts. He has no idea what a private API is so he can’t guide Claude to use one where we need to access low-level system APIs. And obviously concepts like dry run or event log are not familiar, but definitely useful in an app like this.

Interestingly though, he was able to get Gemini to write a prompt for Claude that included a lot of these features that you’d find in an automation app. I wouldn’t have thought of that.

But dev intervention was still necessary. Some added features were just dangling in the UI without any effect internally, and it was impossible for a non-dev to know what they were for.

We started that way, then he continued on his own while I worked on my own apps. Even without my assistance, the result was still a very functional app that could run scripts or launch apps automatically on event triggers.

And the thing is, people could have used this as is, it was useful enough to be released. But I knew it wouldn’t be received well, people expect a certain polish for apps that have paid tiers.

The dev touch

To be really useful, the app needed some triggers that Claude did not know to implement on its own because such code uses obscure private APIs and IOKit data. Things like:

• ambient light data through the MacBook’s integrated sensor (ALS.swift)
• if camera is turned on or off (IsCameraOn.swift)
• Focus mode or Do Not Disturb changes

Thankfully I had already implemented those in mac-utils, but adapting them to Crank’s event trigger architecture would be quite the work. By now, I learned that this is where Claude shines: once it has working code to reference, it does a really good job at adapting that code to another project’s structure.

After I pointed Claude to the files and told it to create triggers out of them, I needed less than half an hour of testing and fixing logic bugs on my part to get everything in good order. It can write quite readable code when it wants.

I remembered seeing this crazy app that told you the MacBook lid angle, so I thought what the heck, someone might find a use for that. I told Claude to implement a lid angle trigger as well based on that code. I’m thinking one might want to disconnect stubborn Bluetooth peripherals when the lid is closing, and you only have a chance to do that before the lid was fully closed.

Open source becomes incredibly more useful with an LLM to understand the code in seconds instead of days. It also feels like stealing, not gonna lie. If I did the research myself, it wouldn’t have felt like that, and I can’t figure out why.

Cherri

Some actions simply can’t be done through scripts. One example of that is enabling Do Not Disturb (or changing the Focus Mode). There’s no CLI that can do it because the system does not expose those APIs.

But it can easily be done through Shortcuts, which has a Turn Do Not Disturb On/Off action built-in. But (big but…) a named Shortcut containing that action needs to already be created by the user, we can’t call individual actions. And I can’t have Crank ask the user to create Shortcuts and drag actions around for such basic things.

Shortcuts can be pre-created as file.shortcut, and installed by simply opening them, so I could provide those Shortcuts bundled with the app. But doing a rough calculation, I would have to create about 100 individual Shortcuts by hand, using mind numbingly slow drag and drop.

Programmer laziness got the best of me and I went looking for a quicker way. Somehow I stumbled on this gem: Cherri - a programming language that compiles directly to a signed Shortcut

I tried to get Claude to write a few basic Shortcuts with Cherri, but it clearly didn’t know the syntax. So I told Claude to just learn it from the website, and gave it the whole list of actions I need.

In less than 10 minutes it:

• learned the Cherri programing language syntax
• learned to use the cherri compiler
• got action definitions with cherri --action
• wrote and compiled 98 individual Shortcuts, fully functional with inputs and outputs

damn

The scripts problem

The problem with creating an automation app, is that you usually have to spend most of your time building a huge library of built-in actions. I didn’t want to go that route.

That’s why Crank’s actions seem pretty limited:

• Shell script
• AppleScript
• Shortcuts
• Launch App
• Open URL
• Open File
• Notification

But in my opinion, this set of actions allows you to do most common things on a Mac, if you know what to use. However, not all users of this app will know what a script is, or how to write one, so I thought I should lower the barrier of entry as much as I can.

If LLMs can write such good code, why not have them write the automation scripts for us? (well, because they could delete all our files with a misplaced semicolon, that’s why…) But non-devs would try an LLM anyway, so I thought I could at least make their life easier and add some safety guards.

In this test, I wanted to see if an LLM understands how to make another less capable LLM get good results.

So I told Claude to create a context.md describing common CLI utilities that would help in automating a Mac, and express that as a prompt for Gemini.

I added a Generate Code button in the script editor and tried to make the free Gemini tier write short useful scripts, instead of hundred-lines hallucinations on CLIs that don’t exist.

It’s not perfect, but when it works, it feels magical.

Turns out, LLMs know how to talk to other LLMs. The generated context.md tells Gemini things like:

You are a macOS shell automation expert.
…
Output ONLY the script code — no markdown fences, no preamble, no explanation

Claude even went as far as to customize the context based on the Mac it runs on:

Assume the script will run on \(ARCH) Macs with macOS \(MACOS_VERSION)

Use the event variables set by Crank if relevant to the task

Tool Availability on this System:
… basic stuff like ffmpeg, jpegoptim etc. …

It’s not complex code or writing, but not very creative either. And I like that I can delegate the boring parts to a smart text generator so I can focus on the harder reverse engineering and making Macs do things they weren’t supposed to do.

At least until it learns to work with Frida and a disassembler and I’m fully replaced.

Consequences

I started this exploration as an AI skeptic, or better said, ignorant. But things are progressing faster than I can adapt to them. So while I’m no longer a skeptic, I’ll probably stay ignorant in the long term.

I see Claude Code as a valuable tool for experienced devs, just like I saw syntax highlighting, autocomplete, IDEs and other dev tools when they appeared. It enables me to spend less time in front of a screen without neglecting the people using my app, which is my end goal.

But it is frightening to think; what would my mind go through if I was a junior dev in college and seeing a $20/month algorithm do what I’d need to learn in years?

And now I’m thinking “I’m fine, Claude is smart but it’s still far from being able to do all the complex work I do for creating a truly useful app”. But far is relative, and such a different measure in the world of AI.

I’m probably in denial, I remember laughing at LLMs hallucinating non-existent Swift syntax less than 3 months ago.

To calm some of the people using my apps who might also read this: I’m not becoming a vibe coder and giving Claude free reign on Lunar and Clop.

At least at the moment, the codebases of those apps are still too messy and complex for an LLM, I don’t want even the remote possibility of introducing a critical bug. rcmd still has a small codebase and doesn’t work with hardware and sensitive files, so it’s safer to try it there, but every line is still reviewed and validated by hand.

I’ll probably try to get Claude to help me on things I started but never had the time to finish, like:

• building a custom filesystem index and fuzzy search algorithm for Cling
• adding natural-language calendar event scheduling in Grila
• fixing long-standing bugs in StartupFolder and IsThereNet

But it will probably function more as a talking rubber duck, than as an employee I can trust to do my work for me.

Which is very similar to my experience when trying to find the most suitable wood finish for my hand carved wooden spoons and coffee cups.

It’s hard to find a finish that:

• cures fast (in less than 2 days)
• is food safe (can be used on things you eat or drink from, like spoons and cups)
• is free of solvents (doesn’t smell of nasty chemicals that make you dizzy every time you apply it)
• and with the hard requirement of making wood hydrophobic and leaving a visually pleasant sheen without creating a plasticky layer

Natural drying oils

Sure, we have pure tung oil which doesn’t need any solvent, has a pleasant nutty smell, and cures into a food-safe hydrophobic polymer inside the wood fibers. But it takes 2 to 4 weeks for the curing to happen depending on oxygen intake, and the finish is quite matte and boring.

While tung oil is natural, that doesn’t mean you can ingest it. It is a powerful irritant in its liquid form, it is only food-safe after curing into its inert polymer state.

There’s also the more rare polymerized (or heat-bodied) tung oil, which is a honey-like viscous oil that cures in a few days because it was heated to very high temperatures in the absence of oxygen. That would be ideal except I can’t find it anywhere around me to buy, and it would most likely need thinning with citrus oil solvent (or D-limonene).

Citrus oil has a nice but way too potent smell of orange peel. It might seem pleasant at first compared to the chemical smell of naphtha and white spirits, but it can be too much if you finish small wooden objects often like I do.

And there are plenty other drying oils, the most well-known being linseed oil, which is indeed edible but has the unfortunate disadvantages of yellowing over time, taking ages to cure and having a persistent oily-grassy flavor. Polymerized linseed oil (aka stand oil) cures faster but still retains the nasty flavor.

Careful not to confuse it with boiled linseed oil or BLO, which uses metallic drying agents like cobalt or manganese salts.

The boiled name comes from the old practice of heating the linseed oil with lead oxide inside, effectively shortening the oxidation process. But nowadays there’s no heat involved in BLO.

More details in The lowdown on BLO

The oil painting world has experience with many drying oils like safflower oil, walnut oil, poppy seed oil, hemp seed oil. Unfortunately they cure slower and into a less durable film while also being rare and expensive so they aren’t good alternatives for wooden utensils.

Hardwax oils

For a good few months I used Osmo Polyx Oil which is a solvent-based hardwax oil. Its safety datasheet and full declaration of ingredients mention about 50% of it being white spirit solvent, and the rest is composed of natural oils processed to become drying oils, and a small amount of hard carnauba wax.

It is fast curing in less than a day, leaves a beautiful shimmer on the wood surface, it is food safe after curing, but smells awful because of the solvents.

In the winter months I carve indoors and have to finish the pieces indoors as well, and the horrible solvent smell fills my house for a whole day. I seem to have become sensitized to the smell and now I have almost a full can of Osmo that I can’t use without a mask, even outdoors.

There’s also Rubio Monocoat and other two-component hardwax oils where the base component is usually a solvent-free blend of drying oils and waxes, and the accelerator component is Hexamethylene diisocyanate or HDI. The base component can cure on its own in about 3 weeks and the accelerator shortens the curing time to less than a day.

This alternative cures fast into an inert and food safe polymer, smells like apple pie, and the finish has a slight satin effect. But it’s really hard to mix properly and apply on small wooden objects like spoons and cups. I almost always use too much accelerator, the mixture ends up too viscous and hard to spread on the wood and the constant opening of the bottles and cans makes the components dry prematurely in their containers.

However, this idea of blending drying oils with hard wax seems to be promising and it’s the path I continued on developing.

Notable mentions

Food-safe epoxy is another alternative, but I hate both the idea of using synthetic plastics and their plasticky look takes away from the experience of using wooden utensils.

Why even use wood if you’re going to cover it in a layer of clear plastic?

Many people also recommend other edible oils like coconut oil, olive oil etc. But these don’t cure and don’t provide any protection for hot liquids. They either go rancid inside the wood fiber or get washed out after using the wooden object.

Some recommend non-edible petrol-based mineral oil (aka liquid parrafin) because it doesn’t go rancid, but has the same effect of not actually doing much for protection and will leak into hot liquids.

Beeswax is often used by kuksa makers (traditional Scandinavian wooden cups) but wax molecules are too large to penetrate wood fibers well, and it has a low melting point of 60℃ (140°F) so hotter liquids are out of the question.

Carnauba wax has a higher melting point at 80℃ (176°F) and is harder than beeswax, but it has a glossy finish, having a similar look to epoxy resin after buffing.

The ideal finish should combine the durable hydrophobic properties of fiber-penetrating tung oil with the fast drying properties of a top wax layer, in an easy to apply but solvent-free blend.

Urushi

Some carvers use urushi lacquer which is the sap from a tree common to Japan. The result is a hard and glossy finish that can resist temperatures of over 300℃ (572°F), but the look and process of application is far from what I’m looking for.

To ensure curing is complete, it needs 10 to 20 very thin coats over many months in a warm and humid environment. The wood then takes this dark brown color, regardless of what wood you started with, which is the opposite of what I want.

I work with fruit and flower woods that are less common in woodworking (apple, fig, jasmine etc.) and I want people to know that and experience what that specific wood looks and feels like. I want to keep their original hues and tints and accentuate the wood grain of each piece.

Someone from Israel bought a spoon I made from fig tree wood, but still asked about it a bit incredulously. It had greenish light wood grain, with a bit of untouched bark kept on the handle and a wonky curve that followed the shape of the tree branch it was made from.

They have many fig trees there but had never seen such a thing made from their wood, so they felt a connection to their motherland in that usual object.

My mother-in-law cut a large lilac tree and gave me some thick branches. It’s a wood that cracks very easily while drying, and I couldn’t use most of it. But the few things I made from it, I gifted back to my wife’s mother and sisters in the form of cooking spoons and are still cherished by them today.

Its pastel purple grain accents remind them clearly of the tree that used to flower every spring in their yard.

A brown lacquer would never be able to create this effect, and that is why I’m still trying to find the perfect natural oil based finish.

Hassui Ceramic (or liquid glass)

This is basically polysiloxane dissolved in alcohol and isoparaffin solvent. After 3 weeks it cures into a food-safe, hydrophobic layer which is durable, but not very good looking in my eyes.

It leaves the wood looking bare, without the color enhancing and rippling effect of oil. Some might prefer that, especially people that worked hard on getting a very shiny knife finish on their carving.

In my case, this is not the finish I’m looking for because of the 3-week curing time, the solvent-based solution, the colorless look and the fact that this finish is very expensive and hard to find outside Japan.

Here’s some more information from a woodturner’s experience with the finish: Liquid Glass Wood Finish – Update

Natural hardwax oil

For the impatient, here’s a comparison video on how the hardwax oil blends look and perform in a water test on some plum wood blanks.

My experiments started simple, with melting about 10g of carnauba wax flakes into 40g of pure tung oil.

Most people recommend a double-boiler or bain-marie approach to melting the wax. That’s because the heat source would be a container of boiling water which can never get above 100℃ (212°F), a safe temperature to avoid reaching the flash point of these oils.

I just drop a spoonful of carnauba flakes into a glass jar of tung oil and place it in the microwave for a minute until everything becomes liquid.

For small quantities, the microwave works wonders, and there’s no danger of overheating the oil or gelifying it because it cools off rapidly.

experiment #1

Ingredients:

• 40g pure tung oil
• 10g carnauba wax flakes

Process:

• pour the tung oil into a glass jar or ceramic container
• drop the carnauba wax flakes into the oil and stir to disperse the flakes
• place in the microwave for 30 seconds then stir again for a few seconds
• repeat until the wax melts and everything is liquid
• let it cool down until the mix is solid

Result:

Brittle paste that’s hard to scoop from the jar and apply on the wood. Doesn’t spread easily, feels sticky.

After the whole wood surface was covered and left for 10 minutes, the excess was wiped with a paper towel. After wiping, the finish is very matte, which means that not much of the wax remained, and the oil didn’t penetrate the fiber enough because of the consistency.

Smells nice though.

Improving the consistency

The consistency changes a lot based on the oil-to-wax ratio.

Too much wax and the blend cools into a brittle paste that’s very hard to scoop from the jar. Too little and the blend is too liquid with not enough wax remaining on the wood surface to form a film. A 4:1 oil-to-wax ratio is good, having enough wax for the top layer, but in this simple formulation it is still too hard and brittle.

Instead of increasing the oil quantity which could defeat the purpose of the wax layer, I added a bit of beeswax to soften the blend.

This works, the paste becomes softer but it makes the top layer of wax not durable enough because of its lowered melting point, so I needed an alternative.

My idea is to have the wood absorb the tung oil from the blend while leaving the wax molecules on the wood surface. The wax should be soft enough to not create a glossy layer and not be brittle, but hard enough to protect the wood for the time it takes the tung oil to cure.

experiment #2

Ingredients:

• 40g pure tung oil
• 7g carnauba wax flakes
• 2g beeswax
• 1g lanolin

To soften the blend without having to add too much beeswax, I found lanolin to be a good addition. It is a highly hydrophobic wax extracted from sheep wool that even in tiny amounts can make the paste consistency softer and easier to apply.

Process:

Similar to #1, just immerse the waxes into the oil in a glass jar and microwave in increments of 30 seconds until everything is liquid. Leave to cool until solid.

Result:

Softer and more homogenous paste that can be scooped easily from the jar with a finger or rag. Spreads relatively easy onto the wood surface, feels like a more viscous hand cream.

Smells amazing, nutty, vanilla and honey scent, faint enough to not bother anyone.

After covering the whole wood surface, I used a heat gun to liquify the finish and enhance wood fiber penetration. The paste melts very easily and the oil in it gets absorbed quickly in the wood pores, leaving the wax to solidify.

After the wood has cooled, I applied another layer because it seemed that some parts of the wood had little to no wax remaining. I removed the excess with a paper towel and buffed the wax to a satin look.

small issues

curing speed

This finish is not fast curing by any means. It is fast drying, as the oil gets absorbed and the wax on top solidifies into a very thin medium-hard layer.

But the melting point of the wax is about 75℃ (167°F) so still not perfect for very hot liquids. You definitely can’t brew your tea directly into a cup finished this way, not until the oil cures completely after 2-4 weeks.

I tested it by brewing a cup of espresso which seems to be fine, the temperature of the coffee is far below 75℃ (167°F) when it reaches the cup.

Using a spoon for eating soup also works fine, because the soup needs to be below 45℃ (113°F) for me to be able to eat it.

After usage, the wooden spoon and cup could be easily washed with warm water and soap, the wood didn’t stain at all. The wax still worked as I could see the water droplets sliding off the wood.

sheen and appearance

Visually, it doesn’t leave a very strong sheen. It has a soft waxy look, instead of the shimmer of oiled wood.

That’s because wax doesn’t follow the micro porous surface of the wood, it solidifies into a smooth surface. So instead of scattering light you get mirror-like reflections creating that glossy wax look.

Commercial finishes use flattening agents like fumed silica to get around this, but I’m trying to use as little additives as possible for now.

The look also comes from the carnauba wax being softened by the beeswax and lanolin which also translates into a softer top layer.

natural resins

The woodworking world has moved on almost completely from natural resins. Almost every finish is either based on polyurethane or epoxy resin if you’re looking for a harder top layer. At best, you’re looking at a shellac layer which doesn’t mix with oils.

Thankfully oil painting forums and blogs still have a trove of information on tree resins, drying oils and siccatives.

That’s how I found out about damar resin, a semi-hard resin extracted from Dipterocarpaceae trees growing in southeast Asia. It is a clear resin, melts at about 120℃ (248°F) which is easily attainable in a microwave, and has a bit of flexibility so that it doesn’t create a cracking layer on wood.

It can be found in the form of crystals or pellets in art shops and it’s pretty cheap. People dissolve it in turpentine to create varnishes, but I wanted to incorporate it into my oil-wax blend through heat instead of solvents.

experiment #3

Ingredients:

• 50g pure tung oil
• 10g carnauba wax flakes
• 2g beeswax
• 2g lanolin
• 3g damar resin
• 5g coconut oil
• 2 drops of MnZrCa drier
• 2 drops of vitamin E

To increase sheen and make the top layer stronger, I melted damar resin crystals directly into oil, without any solvent. I also added a very small amount of manganese, zirconium and calcium drier to make the curing faster, and some vitamin E to increase shelf life.

Process:

Add the carnauba wax, beeswax and lanolin into the tung oil in a glass jar and microwave in increments of 30 seconds until everything is liquid.

Add the damar resin crystals into the coconut oil and microwave in increments of 10 seconds until everything is liquid. Let the mix cool down for about 30 seconds then add the drops of MnZrCa drier and vitamin E. Stir until the mix is homogenous.

Pour the damar resin-oil blend into the tung oil blend and stir until homogenous. Let cool down until solid.

Result:

Medium-soft paste that can be scooped relatively easily from the jar with a finger or rag. Spreads easy enough onto the wood surface.

Smells the same as the last experiment: nutty, vanilla and honey scent, faint enough to not bother anyone. The nasty smell of the drier doesn’t transfer into the blend because its quantity is too small. The resin has no smell in its solid form.

After covering the whole wood surface, I used a heat gun to liquify the finish and enhance wood fiber penetration. The paste melts very easily and the oil in it gets absorbed quickly in the wood pores, leaving the wax and resin to solidify into a very thin layer on top.

After that there was almost no excess left and I buffed the top layer to a satin look using a paper towel.

Curing:

The tung oil inside the wood cured in about 48 hours.

This can be verified by melting the wax layer with a heat gun then removing the liquid wax with a rag. The bare wood then can be sanded with 240 grit to check if the wood dust falls off easily. If the wood dust seems wet and gums on the sandpaper, the oil is still not completely cured.

Look:

The satin sheen looks better because of the resin. Regions with medullary rays look wavy and move in the light, regions with straight fibers look exactly between glossy and matte.

Osmo Polyx Oil still has a prettier shimmer that this finish doesn’t achieve, I’ll look into adding trace amounts of fumed silica to see if it makes the finish settle better into the wood fiber microstructure and scatter light instead of being more on the smooth glossy side.

The fumed silica might also help with adding shear thinning capability so that the paste is even easier to scoop from the jar without softening it.

notes on added ingredients

metallic driers

The metallic driers are probably the first thing that jump out, because we always read about how bad they are when seeing people talk about natural wood finishes. Not all driers are bad, and it depends on things like quantity, leaching probability and type of metal used.

Sure, lead and cobalt driers are out of the question, I don’t want them anywhere near my food and beverages no matter how small the risk of leaching. They contain heavy metals that can have catastrophic effects if they accumulate in your body.

I use a drier containing manganese neodecanoate to speed up surface-drying, and zirconium and calcium salts for through-drying.

Zirconium and calcium are known to be safe. Zirconium is the same metal that dental implants are usually made of, even ceramic knives are made using zirconium, it seems to be a pretty inert metal when it comes to contact with our body.

Manganese is toxic if inhaled or ingested in high amounts, but the quantity of manganese in 2 drops of drier is so incredibly low you could take them orally and not feel any effect other than the horrible solvent taste. That, combined with the fact that the drier disperses throughout the oil makes the quantity too low to matter.

The strictest Tolerable Upper Intake Level for manganese is set to 8 mg/day. In fact we get some of that quantity daily through food that is enriched with vitamins and minerals. If the drier were to ever leach completely from the finish, one would get less than 0.1mg per lifetime of object.

You’d probably have to eat the spoon completely for that to happen, in which case the quantity of ingested wood might become more of a problem.

But the whole food-safe standard on drier based finishes is actually based on the fact that a cured finish traps that metal into an inert polymer which can’t leach into food. Just like Teflon pans or aluminum and copper pots, everything is safe until the layer chips off somehow.

I compromised on having a quantity small enough to not pose any safety risk and that wouldn’t shorten shelf life by making the oil cure inside the jar, but high enough to make the oil cure in 48 hours instead of 2-4 weeks. This, in my opinion, makes the finish safer by being more predictable in curing time, and minimizing the probability of food contacting uncured tung oil.

coconut oil and vitamin E

Because damar resin melts at 120℃ (248°F), tung oil is not suitable for melting the resin directly into it because tung oil can gelify at high temperatures, hindering its curing capabilities. That’s why I melt the resin crystals in a small amount of coconut oil, which has a high enough smoking point of 180℃ (356°F), smells amazing and doesn’t go rancid easily.

I added 2 drops of vitamin E (aka tocopherol) because it is a powerful antioxidant and will lower the chance of ingredients like coconut oil going rancid. It can also lower the chance of the tung oil curing inside the container by inhibiting the autoxidation reaction.

This is the current formulation that I’m using for finishing my wooden spoons and cups, and which I’m selling on my online shop in Romania.

Comparison video

In the following video, I applied the 3 experiments on some plum wood blanks, sanded to 600 grit and water popped 3 times.

In total I had 7 blanks:

1. experiment #1
2. experiment #2
3. experiment #3
4. unfinished wood
5. Osmo Polyx Oil
6. raw tung oil
7. raw tung oil with drier

After applying the finishes, I left them for 24 hours to cure indoors at 22℃ (72°F) at 50% humidity, then sprayed them with water to see how the water beaded up on the surface.

Conclusions:

Osmo Polyx still creates the highest sheen, but felt a bit too glossy on this specific wood, not sure why. I actually liked the damar resin and drier blend the most.

Osmo contains titanium dioxide to whiten the finish a bit, and it also contains more refined oils, so the wood colors looked less saturated. I guess that’s good if you want to keep a light wood from getting a yellow tint, but on my spoons and cups I very rarely need that, so I like the colour imparted by the tung oil.

Water beaded up on all finished blanks and was absorbed by the unfinished wood instantly as expected. Osmo and our #3 blend had the largest water beads.

After cleaning up the water with a paper towel and letting the wood dry, the first two blends lost their sheen and colour almost completely and felt less smooth to the touch. The damar resin and drier really did their job in the third blend.

The tung oil only blank felt the roughest after drying, the oil only created a very weak hydrophobic layer because it couldn’t cure after 24 hours. The blank with the drier added felt smoother, but still absorbed a bit of water. It needed more layers and more time to get full fiber saturation.

experiment #4

to be continued…

I also remember seeing my father use a very coarse stone wheel placed on a motor shaft which threw many sparks when he sharpened some large axe for splitting wood.

I had neither of those around anymore in my rented place in the city so I jumped headfirst in the mind numbing and sometimes esoteric art of getting a sharp blade.

Carving knives

The first blade type I had to sharpen was for my BeaverCraft carving knives.

They fortunately came with a strop, basically a plywood base in the form of a paddle, with leather stuck to it on both sides, and a green waxy bar. Unfortunately I had no idea what to do with it.

Stropping is, at the most basic level, dragging the blade back and forth on a semi-hard surface (like leather). Clean leather won’t make your blade sharper though, and that’s why you have that green bar which you need to rub onto the leather. It contains a fine abrasive that slowly removes tiny amounts of steel from the very tip of your blade to make it sharper.

Reading about stropping and watching videos about it gets into a lot of debate on:

• how often to strop
• on what angle should the blade be held on the leather
• should you move from the heel of the blade to the tip, or the reverse
• can you strop too much and round the apex
• what is the green compound for and how should it be applied

And so on. It’s hard to find much exact science on this, everyone figures it out as they go, and experts share their beliefs based on their extensive experience.

I settled on stropping after every 15 minutes of constant carving, but I had no idea what I was doing or if I got any result.

That’s one thing I learned that would have relieved me of much frustration at the start: find a way to test the outcome of your motions, otherwise you’ll just keep repeating things that don’t work

And it also applies to programming: so many times I’ve seen colleagues writing code but with fear and uncertainty as they didn’t know that what they did would have the expected result. And only because they didn’t know how to test how the code would work.

For example they were backend programmers, that only knew how to run their backend python server, but had no idea how to run the whole stack to check if the front end did work correctly with their changes.

Or app devs who didn’t know how to use a local database, and instead always feared release day because of how their changes will impact the production db.

Nowadays I know how to test for sharpness.

If I want a basic working blade I test if the knife cuts printer paper easily and without tearing it. If I want a carving blade, I check if it pops hairs off my arm easily. And I check that often so I don’t do hundreds of blade passes for nothing.

Anyway, back to my carving knives. For a long time I did a pretty bad job at keeping them sharp.

I would carve for a while then feel the need to push too much into the blade or get tear out in the wood, then I would move the blade back and forth on the strop a few dozens of times, then getting back to carving I would notice a slight improvement which went away after just a few cuts, then back to stropping a thousand times for nothing.

Stropping is not sharpening, and after so many cuts those blades needed serious sharpening which a piece of leather doesn’t do.

Well, actually, stropping is sharpening, but it’s a very very fine kind of sharpening. It’s similar to using a super fine grit stone, something between 14000 and 100000 grit.

The green paste is a super fine abrasive that does remove metal from the apex of the blade, but so little at a time that I would need to do thousands of passes to get the same result that a coarse grit sharpening stone would do in a few passes.

I started looking into sharpeners, and because “stones” felt like something old which only my mother used because she didn’t have any better method, I would look into “modern” sharpeners like:

• pull-through V-blades
• honing rods
• powered belt grinders
• precision sharpening rigs
• rolling sharpeners

I tried a few of those expensive methods, ruined a few blades with them, wasted days on reading reviews and watching tutorials… none of those methods felt like a deterministic way to get a sharp blade every time.

The precision sharpening rig is probably the closest to that, but I don’t want to waste precious workbench space on that. Also, I can’t take it with me when traveling.

Wood planes and chisels

Once I got into larger projects, I discovered how to use wood planes (which seems to be thought of as an antiquated method in Romania, manual woodworking is dead here). I also discovered they need frequent sharpening and their blade shape seemed to be perfect for using a flat surface sharpening method like, you know, a stone.

Every woodworker on YouTube showed how to sharpen planes on stones so I thought I should probably get one eventually. But which type?

There are whetstones, natural stones, ceramic stones, oiled stones, diamond stones, you name it.

After trying a ceramic stone which needed to be soaked in water for 5 minutes before doing any sharpening, and constantly wetted between every few passes, and after making a wet mess and chipping the stone and dreading to sharpen because I had to keep the stone in water, I started looking into drier methods.

Diamond stones seemed to be more up my alley: a modern method based on an old, tried and true idea.

Whetstones and oiled stones are great as well, I just wanted something with less maintenance and more resistant to abuse.

At this point I got sick of wasting money on this hobby, so I got the cheapest set of Chinese diamond plates online. It came in a set of 4 thin steel plates with foam backing, each plate having a different grit: 400, 600, 1000 and 1200. Let’s not get into grit standards, it’s probably FEPA but who knows.

And what do you know, it worked, it was easy to use, cheap and no special instructions needed.

I also got lucky in finding out about the OUTDOORS55 YouTube channel and the science of sharp website who do microscope analysis of blades and methods of sharpening.

They’re getting as close as possible to actually knowing what the heck happens when you move the blade on top of a stone, leather, denim or any type of material people have tried sharpening on.

This helped me cut through the bullshit fast and get to a simple working method of getting any blade from rusty to shaving sharp in a few minutes.

My own sharpening blocks

At this point I had these diamond plates scattered on my workbench, I had a leather strop, a green stropping compound bar that painted my fingers green every time I used it, a round diamond file for my hook knives (because you can’t sharpen the inside of curved blades on a flat stone) and a piece of flexible leather for stropping those hook knives.

It worked, but it was a mess. And I also work on the go a lot, I always do some wood project at my parents house, carve some branch at a walk in the woods or need to sharpen both a kitchen knife indoors and a chisel outdoors. I wanted to simplify this, but could not find any ready made product I wanted.

So I got to doing what I know best, half-assing an improvised product that works for me but would be ashamed of showing it to anyone.

First iteration

I cut a piece of 18mm thick beech wood into the shape and size of a diamond plate, then stuck a 600 grit plate on one side, and the leather on the other side.

This not only got the stone and strop higher up from the table which helped a lot with sharpening long knives, but also put them in the same package which I could carry with me.

I called this a sharpening block in my mind.

To keep the block from slipping while I moved the blade over it, I cut a piece from a non-slip silicone baking mat. Even the small pressure of the blade makes the silicone adhere well to both the block and whatever surface it’s sitting on.

I still didn’t know what to do if I needed a coarser/finer grit, what to do about the green waxy mess of the stropping compound, and how to carry the hook knife sharpeners.

Second iteration

This summer, my sister-in-law visited from Italy and asked if I could make a coffee table for her. Timing was tight because I had to make it in a week so she could transport it from Romania back to Italy by car.

I had no time to find wood slabs so I got two oak wood panels from a big-box store, glued them one on top of the other to make the table-top thicker, and figured I’ll find some table legs afterwards.

I found some 15cm diameter smooth beech logs at a firewood seller nearby. Turns out they were leftovers from a veneer factory that can take a really long and thin log slice automatically, leaving the log core as smooth as a turned piece.

The table turned out pretty nice and solid, her children love sitting on top of it or hiding under it with their toys. I guess it could also be used for holding coffee cups, they didn’t get a chance to try that yet.

So this left me with some 36mm thick oak wood that I thought could make for a better sharpening block.

This time, I glued two coarse and fine grit plates back to back (either 400/1000 or 600/1200) and embedded some neodymium magnets inside the wood to keep the diamond steel plates firmly attached but able to flip easily.

For the strop, I didn’t have any leather left and it was crazy expensive, so I looked into alternatives.

You can basically use any semi-hard porous surface for a strop, even plain thick cardboard. It just needs to hold the fine particles of the compound and not flex too much when pressing the blade into it. People seem to use leather, denim, cotton, felt, cardboard, balsa wood etc.

After building my workbench, and the leg vise for it which needed some rubberized cork for the vise faces, I now found myself with 3m² of that rubberized cork on my hands because I could only buy it in bulk here.

I tested the cork for stropping and I was amazed to see it’s even better than leather:

• it holds more compound which makes it last more
• the bars of waxy compound break down easier when dragging them across cork, almost like drawing with wax crayons. On leather I needed to heat the bars often.
• it doesn’t get cut as easily. I would often drag the blade wrongly across the leather, cutting into it. The cork is like a self-healing material where cuts disappear
• it’s cheaper
• it can be shaped by sanding, which allows me to give the strop rounded edges that are great for stropping recurve blades (those knives which curve to the inside and annoyingly only make contact with the edge of the strop)
• it doesn’t dry and curl like leather

So I cut and glued a 3mm thick piece of that cork on the other side of the oak base and loaded it with pink corundum stropping compound.

Yes, this is the time I discovered the even more esoteric world of stropping, lapping and honing compounds which are not green bars of waxy stuff. We’ll get to that.

The 36mm thick oak gave me enough space to drill 20mm diameter holes with a forstner bit, where I could place a wooden dowel lined with that same cork for honing the hook knives. Another hole was for the stropping compound.

I finally reached an all-in-one sharpening block, super stable, with coarse enough grit for getting a bad blade into shape fast, fine grit for sharpening, enough strop for a few years and a way to sharpen recurves and hook knives.

It was a bit too thick though…

Stropping compounds

The green stuff

Most people get a green bar of compound and use it all their life and never think about it. But nooo, I had to do research and see what that compound contains and why is it so waxy and is there better stuff?

My understanding is that the green stuff is made of very fine particles of chromium oxide, which is also used as green pigment in cosmetics and painting, and is why it makes everything in your life green when touched.

I had a Dialux Vert bar, and searching for its material safety data sheet surfaces the following:

Composition/information on ingredients:
    mixture of fatty acids and paraffins, aluminum oxide, chromium oxide

So it is made of very fine particles of chromium and aluminum oxide suspended in paraffin wax.

Does that mean it can be melted and poured into cylindrical molds? Yes it does! Adding some green flakes in a small cup with walnut oil in it allows it to be melted in a microwave oven.

Any type of oil works, I tested sunflower oil, olive oil, coconut oil, mineral oil etc. you just need some kind of non-solid fat to make the paste less viscous when melted.

After lining the 20mm hole with wax paper and pouring the melted mix, it solidified as a green cylinder with a chapstick-like consistency, that I could easily get out and apply to the cork.

Diamonds

In the last decade, diamond pastes and sprays started becoming popular for stropping.

Diamond powder was widely used for lapping gemstones for a long time, and diamond paste was being used in the dental industry for polishing implants. Eventually people figured they can suspend the very fine diamonds in a sprayable emulsion, advertise it for stropping and ask a ton of money for it.

Diamond sprays are not accessible here in Romania, but I was able to get some lapping pastes of 3 microns and 0.25 microns particle size to test.

For reference, chromium oxide green polishing compound is usually formulated with 0.5 microns particle size, but that doesn’t mean that all particles are of that size.

That’s like the average particle size which gives you an idea of what kind of polish you can expect from it.

I admit, I like diamonds. They’re easy to apply by squirting the paste from a syringe, I can throw it in a bag without making a green mess, it does seem to cut metal very slightly faster and lasts a bit longer between applications. But a tiny syringe that lasts 3 months costs more than a green bar that lasts me years. I can’t make this compromise.

There are people that bought the diamond powder directly and made their own pastes and emulsions but it all seems too complicated for little gain.

I did try some cheap Chinese AliExpress diamond pastes, but for the life of me I can’t figure out how they calculate the grit and particle size. I bought the finest I could find which is listed as W0.5 micron mesh. I don’t know what means but it doesn’t get to a shaving sharp blade easily.

Cubic boron nitride

Looking into other abrasives, there’s another loved stropping compound: cubic boron nitride or CBN emulsions.

It’s impossible to buy that here, but it’s way too expensive anyway and I’m pretty sure the difference would be marginal.

People tout that CBN and diamonds are actually necessary for harder steels, but I’ve yet to see that tested. My guess is that it’s just people’s way of justifying the purchase of a new and expensive toy, we’re all guilty of that.

Wikipedia says that Cr₂O₃ has a hardness of 8 to 8.5 Mohs which is far higher than plain steel at 4 Mohs and even higher than tungsten at 7.5 Mohs. If it can scratch tungsten, I’d say it can hone a steel knife.

Corundum

I eventually found a pale pink compound that has the same fine grit as the green one, but without the color problem.

It’s made from corundum, which is aluminum oxide without the chromium. It’s harder than the green stuff (at 9 Mohs), doesn’t leave its color everywhere, and it’s dirt cheap. I bought the Lea Chromax brand and I’m very happy with it so far. I paid €10 for a huge 800g brick that will last me a lifetime.

I break small chalk-like pieces from it that I pocket or leave throughout the house to ensure I always hame some on hand. It applies easily to both cork and leather and doesn’t flake off or stick to the blade.

I gifted sharpening blocks along with pieces of stropping compound to many friends and relatives and the pink Chromax block still looks as large as when I bought it.

Third iteration

I am working on a way to merge the all-in-one quality of the thick oak block with the pocketability of the thin beech block.

I also got my hands on some narrow 20mm width diamond plates of very fine 3000grit. I find them useful for when I need a very sharp blade for doing finishing cuts on a spoon or when planing dense wood. I’d like to integrate this plate somehow in the block.

It could maybe be glued to the side of a 20mm thick block of wood, and used hand held instead of on a table.

I did try that on the beech block I already had, and it seems to work nicely. It’s not that hard to sharpen freehand and handheld as long as it’s just for doing the last fine honing on an already sharp blade.

The wax paper method for pouring melted compound is also not ideal. I’m thinking that pouring the compound into a chapstick tube would make it easier to use, and be thinner so it can fit inside a hole in the wood block.

I also have these round diamond files which are great for sharpening round blades, serrated knifes, drill bits, even some rip cut saws. They come as a double-ended rod with a conical file on one end, and a round + flat file on the other end, and they’re meant to be placed in a pencil-like holder.

I cut the rod in half with an angle grinder as I can’t fit the whole length of the rod inside the wood block. With only half of it, I can drill a 6mm hole in the block to embed the file.

I can’t fit the cork lined wooden rod for honing though. I only have 3mm thick cork for now which is way too thick for this use case.

There’s this thing called Nanocloth which is like a thin microfiber cloth specifically for CBN and diamond emulsions. It’s hella expensive and I would never buy such a thing, but it gave me an idea: I could try wrapping some thin felt cloth around a 10mm diameter wooden dowel. It should hopefully fit in the block and hold enough compound for stropping.

And I guess that’s it, that would be my ideal sharpening method:

• a compact block of solid wood
• with 2 wide diamond plates on one face (coarse and fine grit)
• a cork strop on the other face
• a narrow diamond plate on the side (very fine grit)
• an embedded round diamond file
• an embedded strop rod
• and a stropping compound, inside the wood block somehow, that’s easy to get out and apply

And all that in a cheap package that should cost less than $30 to assemble.

I’m still working on it, I prepared some new 18mm oak blocks and I’ll update this post with the results.

@solvaholic: Sorry I missed your comment of many months ago. I no longer build software; I now make furniture out of wood. The hours are long, the pay sucks, and there’s always the opportunity to remove my finger with a table saw, but nobody asks me if I can add an RSS feed to a DBMS, so there’s that :-)

I say legendary because it has over 9000 reactions and most are positive. There’s a reason why so many devs resonate with that comment.

A lot of us said at some point things like “I’m gonna throw my laptop out the window and start a farm”. Even my last team leader sent me a message out of the blue saying “I think I’ll run a bar. I want to be a bartender and listen to other people’s stories, not figure out why protobuf doesn’t deserialize data that worked JUST FINE for the past three years”.

You know the drill, sometimes the world of software development feels so absurd that you just want to buy a hundred alpaca and sell some wool socks and forget about solving conflicts in package.json for the rest of your life.

I went through those stages too: when the Agile meetings at my last job got so absurd that we were being asked to estimate JIRA task time in T-shirt sizes, I took the decision to quit that comfy well paying job for the uncertainty of making a living from macOS apps. I had only one app that didn’t even work on the latest Apple Silicon chips, and it was making $0, so I really took a bet with it.

Recently, when people started coming with so many unrealistic and absurd expectations and demands about what my apps should do, I started thinking if it would be possible to leave software development for a more physical trade.

A bit of history

Most of my pre-college time was spent on things I didn’t want to do.

I had a bit of childhood, but then I started going to school 6 hours per day, with 1-2 hours spent on commute after 5th grade. I only liked the 10-minute breaks between classes where I played basketball or practiced parkour.

Every day after I came back from school, I had to work in agriculture, either out in the field with crazy winds and sun and UV radiation, or inside a 100-meter long greenhouse where it’s either a 50°C sauna or a muddy rainforest. I was very bad at every job I was given, but it’s what my parents did for a living and I had to help them, no questions asked.

The few hours that remained, usually very late at night, tired both physically and mentally, I spent practicing acoustic guitar, doing bodybuilding exercises, writing bad poetry or drawing graphite portraits.

I almost never did homework or memorize whatever had to be memorized for the next day of school. I just couldn’t justify spending those few hours I had left on even more stuff I did not want to do.

When I found my liberty in college, hundreds of kilometers away from my parents, it’s like something clicked. I suddenly became incapable of doing work that I found meaningless.

Failing classes became acceptable, quitting jobs was something I did with little remorse if I felt I wasn’t helping anyone with the work I was assigned, and bureaucracy became a disease I had to avoid at all costs.

I still washed the dishes though. Cleaning and other “chores” never felt meaningless for some reason.

The first wood thing I did

… was a chess board and piece set. With magnets inside them. Where the pieces look nothing like ordinary chess pieces.

I was trying to get the pieces to snap into place in a satisfying way, and make sure the game stays that way when kids or dogs inevitably bump the table where the board sits.

You know how Magnus Carlsen always adjusts his pieces so meticulously before a game? Well I have half of that obsession as well so I wanted to avoid doing that.

How it was done

I started with a cheap but hefty pine board which I rounded with a lot of sandpaper. Then I asked my wife to help me colour in the darker squares because I’m pretty bad at colouring inside the edges (both literally and figuratively). We used some wood floor markers for that and the colour seems to be holding well.

Most chess board builds you see on YouTube are done by gluing squares of different wood species with alternating colors, but I had neither the skill nor the tools to do that.

Then I drilled holes for the super strong neodymium magnets from the underside of the board, having to get really close to the top side without passing through. I failed on two squares, but some wood putty took care of that.

I spent a few sunny days on the balcony sculpting the pieces with a badly sharpened knife and my Dremel. This was quite satisfying, there’s something really nice about seeing a non-descript rectangle take the shape of a little horse in your hands. I mean knight, but in Romanian that piece is called “horse”, and I really don’t see any knight there.

Regarding the design, I got some inspiration after seeing these modernist chess sets, which not only looked beautiful in my eyes, but also had these geometric shapes that didn’t need that much sculpting to replicate. I found ready-to-buy spheres and cubes of wood at a craft shop around me (which took care of pawns and rooks), and the rest were carved out of rectangles and cones of wood.

Kaval

Two Octobers ago, a Romanian music band called Subcarpați was holding a free “make a Kaval with your own hands” course, where a flute artisan taught the basics of his trade for a week.

The Kaval or “caval” is a long flute with 5 holes and a distinct lower register where notes can sound melancholic and coming from far away, as opposed to the thin cheerful sound of the small shepherd flute.

Ever since I bought my first Kaval, I always wanted to learn how to build one myself. It’s one of those trades where there’s very little info on the internet, so it feels almost mystical compared to what I’m used to in programming. I would also have the chance to walk home with the finished flute, so of course I went to the course.

I loved the fact that we worked in teams of two, and that everything had to be done by hand with no power tools. Even the long bore through the 70cm branch of elder tree had to be done with a hand drill, taking turns to rest our hands.

The artisan had been a shepherd himself since childhood, and taught himself with a lot of trial and error about how to build good sounding flutes and how to make the holes so that the flute stays in tune. But he didn’t know why the holes should be at those specific distances or why the wood tube should be of that specific length for each scale.

I wanted to know those things, because I had an idea of making a universal Kaval that can play in any scale.

You see, if you want to play on top of songs in various scales, you need a Kaval made for each specific scale. So you’ll need an A minor flute, and a B minor one and a C minor one and so on, for a total of 12 different flute lengths.

I eventually found info on how a flute works by thinking about it as an open or closed tube where the vibrating air creates nodes and antinodes that should coincide with the hole position. At the moment I’m still studying this and working towards my “universal flute” goal.

What does this have to do with software?

For the past 10 years I lived in rented apartments, usually at the 3rd or 4th story with no access to a courtyard. I was never able to get used to that, given that all my childhood I lived and played in a 2000m² courtyard, on a road where there were more slow horse carriages than noisy cars.

This year I moved into a rented house with a tiny but welcoming garden and a bit of paved court and only now I notice the effect this has had on my mind and behaviour.

I develop macOS apps for a living, and there are some unhealthy things in this field that piled up over the years. I get a lot of messages in a demanding and negative tone, and because walking outside the apartment meant unbearable car noise, obnoxious smells and zero privacy, I always defaulted to simply acting on the feedback, putting up with it and working long hours into the night, instead of going for a walk to calm down.

A few months ago, the most absurd demands started coming up for my apps: things like “why does your app not control the volume of my <weird sound device>? why don’t you just do it, people pay you for it” when the app in question is Lunar, an app for controlling monitor brightness, not sound devices.

Or “why do you disable your apps from working on Windows?”, or “make Clop compress text and copy it to clipboard” (where Clop is my app that automatically compresses copied images, videos and PDFs, I have no idea what compressing text even means in that context).

But this time, I was able to simply walk out the front door, grab a branch of beech wood, and, because I remembered my wife saying we forgot to package the french rolling pin when moving, I took out my pocket knife and started carving a simple rolling pin for her. It was so liberating to be able to just ignore those messages for a while and do something with my hands.

I understand that those people don’t know better, and they would have no idea that there’s no checkbox where you can choose whether an app works on macOS, Windows or Linux. I understand how if the app does something with audio volume or compression, some think that it should do everything related to those workloads, even if it’s completely outside the scope of the app.

But the combination of the negative tone and getting message after message, some people being so persistent that they insist on sending me those messages through all possible mediums (email, Discord, Twitter, contact form, they’ll find me everywhere), makes it hard to just ignore them.

There’s also this oily smell of AI and machine learning in the tech atmosphere, where I no longer feel relevant and I seem to have stopped caring about new tech when I noticed that 8 in 10 articles are about some new LLM or image generation model. I guess I like the smell of wood better.

Side tangent on privileges of being a software dev

I know I’m privileged to even be able to have the choice of what to do with my time. I got lucky when I chose a computer science university at the right time which allowed me to progress towards a huge semi-passive income in the last 10 years. that doesn’t mean I didn’t work my ass off, but luck plays a huge role too

I got “lucky” to have my mind traumatised into some kind of OCD-like state where I hate leaving a thing unfinished. So I plow through exhaustion, skip meals, miss house chores and annoy dear people around me because I know “I just need to fix this little thing” and I’ll finish this app/feature/task I started. Even though I also know there’s no real deadline and I can leave it half-finished for now.

But even if it sounds annoying for a person like me to whine about how I don’t feel good or I feel burnt out, the privilege doesn’t negate the feelings. The regression to the norm will make everyone, rich or poor, get used to the status quo and complain about every thing that’s just a little worse than their current state. That’s happiness and sadness in a nutshell.

I’m also vaguely aware that software dev as we know it is about to disappear soon, and I got tired of learning the newest thing just to have it replaced next year. I got tired of back pain and chronic finger pain from so many hours of sitting and typing, I’d rather have pain from work that also builds some muscle.

And I got so tired of everything being online, immaterial, ephemeral and lonely, like indie development tends to be.

Woodworking with cheap tools and free wood

This house we rented is small and the owners had to fit the bedroom upstairs. I really don’t like climbing stairs up and down, especially when I have to let my dog out three times per night. So we gave up a room and started furnishing our own bedroom downstairs.

I didn’t want to buy bedside tables for the price of the bed itself, so I thought I could maybe make by own. I’m not yet skilled enough to build my own bed though, so we bought that

One day, while walking with my dog, I noticed that some trees were getting trimmed in the vicinity of our house and there were a lot of white birch branches on the side of the road. I said why not?, grabbed some branches and walked like a lunatic with white long sticks dangling up and down and a black dog zig-zagging left and right, all the way home.

I had another small pine panel left from that chess project so I started thinking about the simplest way to turn what I have into a bedside table.

I used low-grit sandpaper to give the board some nice round corners because I love squircles, swallowed about a spoonful of sawdust because I couldn’t find any breathing mask left, criss-crossed 4 branches in a way that would give a stable base, and screwed them to the underside of the board with long wood screws.

The legs would wobble around though, so I drilled small 3mm holes into each branch where they met in the middle, and weaved a florist wire through them to keep the table steady.

The laptop bed table

After I’ve shown the bedside table to a friend of mine, he said he also needed a laptop table for those mornings when he’d rather not get out of bed. I wanted to say that’s not very healthy, but what got out instead was sure thing, I’ll do it!. Oh well..

I still had the large desk top I glued from smaller beech boards, on which I worked for the past 4 years. It stayed unused currently, so I cut part of it and built this cute thing:

You’ll notice three defining features that every laptop table should have:

• a hole for a charging cable
• a carved coaster for the coffee cup
• a mildly surprised face? 😦

To tell the truth, all those are side effects of me drilling holes where there should be no hole, and dropping the board on the ground multiple times because my workbench was not large enough. All the things that could go wrong, went wrong with this table.

I hid the defects by turning them into features.

The whole truth actually is that the table looks nothing like what I planned. I bought these nice hidden brass cylindrical hinges to make the table foldable. That way, you could fold the sides flat inside and use it as some kind of armchair desk if you wanted.

I wasn’t able to drill the correctly sized or positioned holes for the hinges because I still lack a lot of knowledge and skill in working with wood. So after losing my temper with the frickin’ hinges that still didn’t fit after a full day of drilling and chiseling, I glued the sides and inserted 2 trusty long wood screws per side, which I patched with a glue gun that made the screw holes look like eyes.

After I also carved the handles, the table grew kind of a personality of its own, as you can see above.

Why didn’t I do some wood joints, like a dovetail instead of ugly screws and glue?

Because I had no idea they existed. Also, I wasn’t able to fit a simple hinge, I would have probably never finished this table if I tried learning wood joinery on it.

This reminds me of how whenever I did pair programming with a colleague, I noticed how they were doing some “nonoptimal” action and I would say:

Why don’t you just use ripgrep instead of sifting through all these files?

Because they don’t know it exists, stupid. Or because they just want to get this thing done and move on, they don’t grep files all day like you do.

But in my ignorance, I seem to have chosen a good enough joining method. As you can see in this wood joinery comparison, 5cm (2inch) screws can hold more than 50kg (110lbs) of force, and I used even longer screws so I think it’s going to hold a 3kg laptop just fine.

Oh right, forgot about this little detail.. I also added a cork pocket for holding a notebook, tablet, phone etc. which I lined with a microfiber cloth on the inside for strength and sewn to the wood with that leftover alpaca wool for style.

The bookshelf without books

While we were stuck in the apartment in the 2020 pandemic, me and my wife bought a lot of stuff that we thought would help us learn new things and start new hobbies. I thought I’m going to build smart LED lighting all my life and my wife would become a professional wool knitter. We were losing our minds, for sure.

So now we were stuck with crates of stuff we haven’t used in years, and didn’t want to start unpacking them around the house. The clutter that followed after the pandemic, tired our minds just as much as the lockdown itself.

We dumped the crates on an unused stairway spot, and I thought that a bookshelf as large as that spot would clear the clutter.

But I could not find any bookshelf that large, certainly not for cheap. So I traced a few lines in Freeform, took some measurements, and ordered a bunch of large pine boards and a ton of long screws.

I also ordered the cheapest portable workbench I could find ($30) that had a vise, so I can stop making sawdust inside.

A few days later, I got to sawing the shelves with my cheap Japanese pull saw I bought from Lidl years ago.

Hint: Hand sawing a long wood board with no skill will certainly end up with a crooked edge. Stacking up 5 boards one on top of the other will still end up crooked.

Uhm, I guess the hint is, buy a track saw, or make sure the crooked edge isn’t visible.

My wife helped a lot with measuring and figuring out where to drill holes and place the screws, while my dog inspected the work regularly to make sure the defects were hidden correctly.

It took two days of screwing.. erm.. driving screws, I mean. But in the end we got the result we wanted! And I got sores in my right arm for days, driving those long screws is harder than I thought.

The desk that became a workbench

In the thumbnail of this post you can see the current “workbench” I use, which is basically that $30 vise workbench I bought for the bookshelf, with the top of my previous “coding desk” attached in the front.

In the image you can see (bottom-left to top, then right):

• the cheapest block plane I could find ($8)
• a red no-name plane I found in the shed of that 100-year old house that we never finished rebuilding
• an axe I found rusted and partly broken in the same shed, on which I learnt how to sharpen and restore axes
• a folding japanese pull saw that I take everywhere with me
• some grip blocks on which I place boards for sanding
• a bottle of Osmo Polyx oil I use for finishing
• a set of carving knives from Beavercraft (really good and they were available at a nice discount)
• a combination square (tucked somewhere at the top of the bench)
• a branch of elder tree, which is prepared for drilling a hole through it for making a kaval

I also own some no-name chisels that work well enough and some card scrapers that I still struggle sharpening.

The only power tools I have are a Makita drill and a random orbit sander on which I did spend some money, an old circular saw I found in that old shed (it was good enough to cut miters on that laptop table) and a Dremel I use rarely because I don’t like its power cord. I prefer battery powered tools.

The window bench

Our dog Cora loves sitting at the window, growling at old people and barking at children passing around. Yeah, she’s terrified of children for some reason.

But the window sill is not wide enough and her leg kept falling with a “clang” on the radiator below. So I widened it by placing two glued up boards of pine on top of the radiator, that I planed and smoothed beforehand.

This is when I learned that a hand plane is not some antique tool that nobody uses anymore, but a quite versatile piece that can easily smoothen grain where I would waste 5 sheets of sandpaper and choke on sawdust.

I had to still let the heat radiate somehow, so I drilled large holes with a forstner bit, but I also blew the grain fibers on the underside because I had no idea of this possible problem. Turns out there is a simple solution to drilling large holes without ripping the fibers:

1. Drill a small 3-6mm hole in the center with a normal wood drill, all the way to the other side (this will help see where the forstner bit should be place from both sides)
2. Place the forstner bit in the hole (this also helps with keeping the bit centered) and drill the large hole, stopping midway through the board
3. Turn the board around and repeat step 2 until you meet the other end of the hole

We also wanted to sit with Cora and there was not much space between the bed and the radiator, so I built a narrow bench. I used another two pine boards of the same size, but this time glued them on the side to create a wider board.

For the legs, well the tree trimming continued near us, so one day I found some thick cherry branches which I brought home, scraped the bark from them, then attached them to the bench using screws from the top side.

I was ok with a rustic look so I didn’t spend much on finishing, patching holes, or even proper wood drying. I did use the hand plane to chamfer the edges though, I love taking those thin continuous wood shavings from the edge.

The trunk coffee table

We recently visited my parents, and loved how the grass finally started growing in some spots where their house and court renovation was finished and was no longer spewing cement dust. It was an abnormally sunny April and I wanted to chat with them at a coffee outside in the early morning before they started the field work, but there was nowhere to place the coffee outside.

First world problems right? If you read about The tail end, you might already understand why a trivial thing like coffee time with my parents feels so important to me.

So one day, while walking on a gravel road near their house, I noticed one neighbour had these huge logs of beech that were recently cut. I thought that would be easy to turn into a small exterior coffee table, so I went to ask if I could buy one.

Well I kind of had to yell “HELLO!” at their gate because I didn’t know their name, and did that a few times until a seemingly sleepy old man appeared at the front door (it was 5 PM) asking what I want. I asked how much he’d want for one of those logs, but he just said to get one, no money needed. Ok, there’s no point in insisting, I chose a wide enough but not too wide log, because these things are heavy and I wasn’t sure I could lift it, and rolled it slowly back home.

I didn’t have my usual tools at my parents house, so I improvised. I found a battered cleaver which my dad used for chopping kindling for the barbecue. I sharpened it as well as I could, then used a hammer to roll a burr on the back of the cleaver that I could use for scraping.

Beech wood has such a smooth hard wood under the bark that it didn’t even need sanding. I used my dad’s power planer to smooth out the top and make a quasi-flat surface then finished it with some walnut oil and it was (almost) ready!

Because the wood was so green, it was certain that it will crack and roughen as it dried. So I cut a groove and wrapped a flat iron band around the top to keep it from moving too much. The bottom can expand as much as it wants, I’m actually quite curious to watch the table morph throughout the summer as we use it.

The orchard bench

Because we were born in villages that aren’t that far apart, me and my wife always visit both our parents in the same trip. This time when I got to my parents-in-law, I took a stroll through their little orchard. They added new trees this year! I can’t wait to taste the large apricots.

What struck me as odd about the orchard was that there was no patch of grass to lay on. They like digging up the soil every year, and leaving it like that: an arid looking patch of land made of dry dirt boulders. I thought a bench would be a good solution and what do you know, there was an old broken door thrown in the firewood pile just outside the orchard, that had the perfect length and width for a bench.
I forgot to take a photo of the door, but it looked kind of like this one, only worse and with a large rhomboid ◊ hole at the top.

I got to work immediately, dismantling the door piece by piece and pulling out nail after nail (they really liked their nails in those old times). I was left with two long and narrow wooden boards, a pile of rotten wood and two pocketfuls of rusted nails.

I sawed the broken ends of the boards, then I used my father-in-law’s power planer to remove the old gray wood from the top, bottom and sides to get to the fresh wood below. There were a lot of holes and valleys so I had to scrape them by hand with sandpaper rolled around a screwdriver. This took a few more days than I expected, but I eventually got two cleanish boards of.. fir? pine? No idea.

I used a velcro sandpaper attachment for the battery powered drill to sand out the rotten sides and give the boards a curvy and smooth live edge.

For the legs, I stole some more firewood from their pile, where I found some thick branches of unidentified species that were roughly the same length. Stripping the bark with an axe made them look good enough so I screwed them at the four corners of the board. The bench was wobbly with just the legs, so I strengthened it sideways by adding shorter and thinner branches of more unidentified wood between the legs and the center of the board.

I had to do something with the rhomboid ◊ hole, so I filled it with a square 4-by-4 salvaged from a recently dismantled shed, and now the bench has 5 legs. Instead of sawing the leg to size, I left it protruding above the bench and placed another thick salvaged board on top of it to serve as an arm rest, or coffee table, or a place for the bowl of cherries.

For the finish, I burned the bench and the bottom of the legs to get a honey-brown aspect and to make it water resistant. I put a very thin layer of whatever wood lacquer I found in my in-laws shed, just for resistance because I don’t like glossy wood.

Other small wood things

Water glass shelf

We don’t have much space on the current eating table, so I built a two-shelf stand where we can place the always present water filter jug and the glasses and free up some of the center space.

It’s incredible how strong just a few screws can be.

Kaval stand

I thought I should finally do something about the kavals always rolling around on some table or couch throughout the house, so I made a stand from long thin wood boards glued on the side, and finished it with sunflower oil to give it a golden/orange colour.

This way I can always expand it by adding more boards to the side if I want to add more flutes.

Sharpening block

I need to sharpen blades almost daily, be it the pocket knife, axe, plane blade or chisels. So I made a custom sharpening block with the perfect tools for my sharpening technique.

It has a $5 diamond plate with 600 grit on one side and a $5 leather strop (a piece of leather belt might work just as well) on the other side. I attached the leather with two small screws at the top so I can take it out easily if I need a flexible strop for my carving gouge for example. It is loaded with diamond paste which can be found for cheap at gemstone cutting online stores (the knife-specific pastes are a lot more expensive and I’m not sure why).

To be honest, a $0.5 green compound (chromium oxide) works just as well for stropping, that’s what I used before and still use for my detail carving knives. It gives a smoother edge than the diamond, the disadvantage being that it needs to be re-applied more often on the leather and that you need a bit more blade passes to get the same result. The diamonds seem be cutting faster, but really not much faster.

A bit of a tangent on the sharpening topic

I went through all the phases with sharpening tools. I’ve used water stones, natural stones, ceramic stones, pull-through carbide sharpeners (don’t use these), powered belt sharpeners, wheel sharpeners.

Aside from the pull-through sharpeners and the steel rods, all the others work just as well with the right technique. I settled on the diamond plate because they’re cheap, stay flat, need zero maintenance, and can cut through any type of metal. Paired with a leather strop, for me it’s the simplest way to sharpen.

I recommend this OUTDOORS55 video for a no-bullshit sharpening tutorial and the Science of Sharp blog if you’re curious what the different sharpening techniques do to an edge under a microscope.

I had no idea what that is, but boy, did it sound cool?…

I fondly remember how at that time we weren’t chasing high salaries and filtering jobs by programming languages and frameworks. We just wanted to learn something.

As students, we needed money as well of course, but when I got the job for 1750 lei (~€350), I suddenly became the richest 18 year old in my home town, so it wasn’t the top priority.

And we learnt so much in 2 years.. obscure things like AOP, a lot of x86 assembly, reverse engineering techniques which dumped us head first into languages like Java, .NET, ActionScript? (malware authors were creative).

But most of all, we did tons of Python scripting, and we loved every minute of it. It was my first time getting acquainted with fast tools like Sublime Text and FAR Manager. Coming from Notepad++ and Windows Explorer, I felt like a mad hacker with the world at my fingertips.

I’m known as a macOS app dev nowadays, but 9 years ago, I actually started by writing petty Python scripts which spurred the obsessive love I have nowadays for clean accolade-free code and indentation based languages.

What does all that have to do with static websites though?

Pythonic HTML

Well, 5 years ago, when I launched my first macOS app, I found myself needing to create a simple webpage to showcase the app and at the very least, provide a way to download it.

And HTML I did not want to write. The XML like syntax is something I always dreaded, so overfilled with unnecessary </> symbols that make both writing and reading much more cumbersome. I wanted Python syntax for HTML so I went looking for it.

I went through pug…

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12
13
14
15
16
17
18
19
20
21
22
23

doctype html
html
  head
    title Lunar - The defacto app for controlling monitor brightness
    meta(itemprop='description' content='...')
    style.
      a.button {
        background: bisque;
        padding: 0.5rem 1rem;
        color: black;
        border-radius: 0.5rem;
      }
      body {
        display: flex;
        flex-direction: column;
        align-items: center;
        justify-content: center;
        text-align: center;
      }
  body
    h1(style='color: white; font: bold 3rem monospace') Lunar
    img(src='https://files.lunar.fyi/display-page.png' style='width: 80%')
    a.button(href='https://files.lunar.fyi/releases/Lunar.dmg') Download

pretty, but still needs () for attributes, and I still need accolades in CSS and JS

then haml…

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24

!!!
%html{lang: "en"}
  %head
    %meta{content: "text/html; charset=UTF-8", "http-equiv" => "Content-Type"}/
    %title Lunar - The defacto app for controlling monitor brightness
    %meta{content: "...", itemprop: "description"}/
    :css
      a.button {
        background: bisque;
        padding: 0.5rem 1rem;
        color: black;
        border-radius: 0.5rem;
      }
      body {
        display: flex;
        flex-direction: column;
        align-items: center;
        justify-content: center;
        text-align: center;
      }
  %body{style: "background: #2e2431; min-height: 90vh"}
    %h1{style: "color: white; font: bold 3rem monospace"} Lunar
    %img{src: "https://files.lunar.fyi/display-page.png", style: "width: 80%"}/
    %a.button{href: "https://files.lunar.fyi/releases/Lunar.dmg"} Download

even more symbols: %, :, => and / for self-closing tags

…and eventually stumbled upon Slim and its Python counterpart: Plim

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12
13
14
15
16
17
18
19
20
21
22

doctype html
html lang="en"
  head
    title Lunar - The defacto app for controlling monitor brightness
    meta itemprop="description" content="..."
    -stylus
      a.button
        background bisque
        padding 0.5rem 1rem
        color black
        border-radius 0.5rem
      body
        display flex
        flex-direction column
        align-items center
        justify-content center
        text-align center

  body style="background: #2e2431; min-height: 90vh"
    h1 style="color: white; font: bold 3rem monospace" Lunar
    img src="https://files.lunar.fyi/display-page.png" style="width: 80%"
    a.button href="https://files.lunar.fyi/releases/Lunar.dmg" Download

ahhh.. so clean!

Here’s how that example looks like if I would have to write it as HTML:

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29

<!DOCTYPE html>
<html>
    <head>
        <title>Lunar - The defacto app for controlling monitor brightness</title>
        <meta itemprop="description" content="...">
        <style>
            a.button {
                background: bisque;
                padding: 0.5rem 1rem;
                color: black;
                border-radius: 0.5rem;
            }

            body {
                display: flex;
                flex-direction: column;
                align-items: center;
                justify-content: center;
                text-align: center;
            }
        </style>
    </head>

    <body>
        <h1 style="color: white; font: bold 3rem monospace">Lunar</h1>
        <img src="https://files.lunar.fyi/display-page.png" style="width: 80%">
        <a class="button" href="https://files.lunar.fyi/releases/Lunar.dmg">Download</a>
    </body>
</html>

not particulary hard to read, but writing would need a lot of Shift-holding and repeating tags

The thing I like most about Plim, and why I stuck with it, is that it can parse my other favorite symbol-hating languages without additional configuration:

• Python for abstracting away repeating structures
• Stylus for writing style tags
• CoffeeScript for the script tags
• Markdown for long text content

Here’s a more complex example to showcase the above features (might require sunglasses):

example of writing a HDR page section, similar to the one on lunar.fyi

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48

---!
  # use Python to generate the dynamic image sizes for the srcset attr

  WIDTHS = [1920, 1280, 1024, 768, 640, 320]

  def srcset(image, ext, page_fraction=1.0):
    return ','.join(
      f'/img/{image}/{width}.{ext} {width // page_fraction:.0f}w'
      for width in WIDTHS
    )

doctype html
html lang="en"
  head
    -stylus
      # use Stylus to do a readable media query that checks for wide color gamut

      @media screen and (color-gamut: p3)
        @supports (-webkit-backdrop-filter: brightness(1.5))
          section#xdr
            -webkit-backdrop-filter: brightness(1)
            filter: brightness(1.5)

  body
    section#xdr
      picture
        source type="image/webp" srcset=${srcset('xdr', 'webp', 0.3)}
        source type="image/png" srcset=${srcset('xdr', 'png', 0.3)}

      -md
        # write markdown that renders as inline HTML

        Unlock the full brightness of your XDR display

        The **2021 MacBook Pro** and the **Pro Display XDR** feature an incredibly bright panel *(1600 nits!)*,
        but which is locked by macOS to a third of its potential *(500 nits...)*.

        Lunar can **remove the brightness lock** and allow you to increase the brightness past that limit.

    -coffee
      # use CoffeeScript to detect if the browser might not support HDR

      $ = document.querySelector
      safari = /^((?!chrome|android).)*safari/i.test navigator.userAgent

      window.onload = () ->
        if not safari
          $('#xdr')?.style.filter = "none"

And best of all, there is no crazy toolchain, bundler or dependency hell involved. No project structure needed, no configuration files. I can just write a contact.plim file, compile it with plimc to a readable contact.html and have a /contact page ready!

So that’s how it went with my app: I wrote a simple index.plim, dropped it on Netlify and went on with my day.

Complexity Cost

• 1 pip install for getting the Plim CLI
• 1 npm install for getting stylus and coffeescript (optional)
• 1 build command for generating the HTML files

Complex simplicity

The app managed to get quite a bit of attention, and while I kept developing it, for the next 4 years the website remained the same heading - image - download button single page. It was only a side project after all.

Working for US companies from Romania made good money, but it was so tiring to get through 3h of video meetings daily, standups, syntax nitpicking in PR review, SCRUM bullshit, JIRA, task writing, task assigning, estimating task time in T-shirt sizes??

In April 2021 I finally got tired of writing useless code and selling my time like it was some grain silo I could always fill back up with even more work…

I bet on developing my app further. Since my college days, I always chose the work that helps me learn new concepts. At some point I had to understand that I learnt enough and had to start sharing. This time I really wanted to write software that helped people, and was willing to spend my savings on it.

Comically Stuffed Stylesheets

A more complete app also required a more complete presentation website, but the styling was getting out of hand. You would think that with flexbox and grids, you can just write vanilla CSS these days, but just adding a bit of variation requires constant jumping between the CSS and HTML files.

A presentation page is usually only 10% HTML markup. The rest is a ton of styling and copy text, so I wanted to optimize my dev experience for that.

There’s no “go to definition” on HTML .classes or #ids because their styles can be defined ✨anywhere✨. So you have to Cmd-F like a madman and be very rigorous on your CSS structure.

The controversial but very clever solution to this was Tailwind CSS: a large collection of short predefined classes that mostly style just the property they hint at.

For example in the first code block I had to write a non-reusable 5-line style to center the body contents.

1
2
3
4
5
6
7

body {
  display: flex;
  flex-direction: column;
  align-items: center;
  justify-content: center;
  text-align: center;
}

With Tailwind, I would have written the body tag like so:

1

body.flex.flex-col.justify-center.items-center.text-center

That might not seem like much, some would argue that it’s even a lot less readable than the CSS one. Can’t I just define a .center class that I can reuse?

Well, think about a few things:

• this might repeat on many sections of the page, but with slight variations (what if I want a centered row, or longer paragraphs of text aligned to the left)
• responsive sections might need to alter layout (e.g. vertical on mobile, horizontal on desktop) and media queries will quickly blow up the style size
  • .md:flex-row.flex-col is what you would write in Tailwind
• adding dark/light mode support is yet another media query
  • .dark:bg-white.bg-black looks simple enough
• interactions like hover effects, complex shadows and filters like blur and brightness is syntax that’s often forgotten
  • .shadow.hover:shadow-xl creates a lift off the page effect on hover by making the shadow larger
  • .blur.active:blur-none un-blurs an element when you click on it
• choosing colors and reusing them needs a lot of attention
  • .bg-red-500.text-white sets white text on saturated red
  • red-100 is less saturated, towards white
  • red-900 is darker, towards black

Sure, long lines of classes might not be so readable, but neither are long files of CSS styling. At least the Tailwind classes are right there at your fingertips, and you can replace a -lg with a -xl to quickly fine tune your style.

Complexity Cost:

• 1 command added for building the minimal CSS from the classes used
• 1 npm install for getting the Tailwind CLI
• 1 config file for defining custom colors, animations etc. (optional)

Responsive images

So many people obsess over the size of their JS or CSS, but fail to realize that the bulk of their page is unnecessarily large and not well compressed images.

Of course, I was one of those people.

For years, my app’s website had a screenshot of its window as an uncompressed PNG, loading slowly from top to bottom and chugging the user’s bandwidth.

I had no idea, but screenshots and screen recordings are most of the time up to 10x larger than their visually indistinguishable compressed counterparts.

I even wrote an app to fix that since I’m constantly sending screenshots to people and was tired of waiting for 5MB images to upload in rapid chats.

It’s called Clop if you want to check it out.

Yes, just like that famous ransomware, it wasn’t that famous at the time of naming the app.

I needed a lot more images to showcase the features of an app controlling monitor brightness and colors, so I had to improve on this.

Delivering the smallest image necessary to the user is quite a complex endeavour:

1. Optimize the image using ImageOptim
2. Resize it to fit multiple screen sizes using vipsthumbnail
3. Figure out what fraction of the page width will be occupied by the image
4. Write a suitable srcset attribute to load the suitable image
5. Optional: convert the image to formats like webp, avif or JPEG XL for smallest file size

I did so much of that work manually in the past… thankfully nowadays I have imgproxy to do the encoding, optimization and resizing for me.

I just have to write the srcset, for which I defined Plim and Python functions to do the string wrangling for me.

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12
13
14

-def image(img, ext='png', factor=0.4, mobile_factor=1)
    picture
        -call img=${img} ext=${ext} factor=${factor} mobile_factor=${mobile_factor} self:sources
        img srcset=${srcset(img, ext, factor)}

-def sources(img, ext='png', factor=0.4, mobile_factor=1)
    source type="image/avif" srcset=${srcset(img, ext, mobile_factor, convert_to="avif")} media="(max-width: 767px)"
    source type="image/avif" srcset=${srcset(img, ext, factor, convert_to="avif")} media="(min-width: 768px)"

    source type="image/webp" srcset=${srcset(img, ext, mobile_factor, convert_to="webp")} media="(max-width: 767px)"
    source type="image/webp" srcset=${srcset(img, ext, factor, convert_to="webp")} media="(min-width: 768px)"

    source type="image/${ext}" srcset=${srcset(img, ext, mobile_factor)} media="(max-width: 767px)"
    source type="image/${ext}" srcset=${srcset(img, ext, factor)} media="(min-width: 768px)"

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12

WIDTHS = [1920, 1280, 1024, 768, 640, 320]

def imgurl(image, width, ext="png", convert_to=""):
    conversion = f"@{convert_to}" if convert_to else ""
    return f"https://img.panaitiu.com/_/{width}/plain/https://lunar.fyi/img/{urllib.parse.quote(image)}.{ext}{conversion}"


def srcset(image, ext="png", factor=1.0, convert_to=""):
    return ",".join(
        f"{imgurl(image, width, ext, convert_to)} {width // factor:.0f}w"
        for width in WIDTHS
    )

Complexity Cost

• 1 imgproxy server that needs to run somewhere publicly available, be kept alive and secure
• some Python and Plim code for generating srcsets

Hot reloading

After 2 weeks of editing the page, Cmd-Tab to the browser, Cmd-R to refresh, I got really tired of this routine.

I worked with Next.js before on Noiseblend and loved how each file change automatically gets refreshed in the browser. Instantly and in-place as well, not a full page refresh. I got the same experience when I worked with React Native.

There should be something for static pages too, I thought. Well it turns out there is, it’s called LiveReload and I had to slap my forehead for not searching for it sooner.

After installing the browser extension, and running the livereloadx --static file watcher, I got my hot reloading dev experience back.

Actually now that I think about it, Hugo has super fast hot reloading, how does it accomplish that? Yep, turns out Hugo uses LiveReload as well.

Complexity Cost

• 1 more command to run in 1 more terminal panel, multiplex helps with that
• 1 browser extension to install and hope it’s not compromised or sold to a data thief
• 1 npm install for getting the livereloadx CLI

Contact pages

After releasing the new app version, many things were broken, expectedly.

People tried to reach me in so many ways: Github issues, personal email, through the app licensing provider, even Facebook Messenger. I had no idea that including an official way of contact would be so vital.

And I had no idea how to even do it. A contact form needs, like, a server to POST to, right? And that server needs to notify me in some way, and then I have to respond to the user in some other way… sigh

I thought about those chat bubbles that a lot of sites have, but I used them on Noiseblend and did not like the experience. Plus I dislike seeing them myself, they’re an eyesore and a nuisance obscuring page content and possibly violating privacy.

After long searches, not sure why it took so long, I stumbled upon Formspark: a service that gives you a link to POST your form to, and they send you an email with the form contents. The email will contain the user email in ReplyTo so I can just reply normally from my mail client.

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25

form action="https://submit-form.com/some-random-id"
    label for="name" Name
    input#from hidden="true" name="_email.from" type="text"
    input#name name="name" placeholder="John Doe" required="" type="text"

    label for="email" Email
    input#email name="email" placeholder="john@doe.com" required="" type="email"

    label for="subject" Subject
    input#email-subject hidden="true" name="_email.subject" type="text"
    input#subject name="subject" placeholder="What's this message about?" required="" type="text"

    label for="message" Message
    textarea#message name="message" placeholder="Something about our apps perhaps" required="" type="text" rows="6"

-coffee
    # Custom subject and name: https://documentation.formspark.io/customization/notification-email.html#subject

    nameInput = document.getElementById("name")
    fromInput = document.getElementById("from")
    nameInput.addEventListener 'input', (event) -> fromInput.value = event.target.value

    subjectInput = document.getElementById("subject")
    emailSubjectInput = document.getElementById("email-subject")
    subjectInput.addEventListener 'input', (event) -> emailSubjectInput.value = event.target.value

Complexity Cost

None, I guess. I just hope that the prolific but unique Formspark dev doesn’t die or get kidnapped or something.

And you call this “simple”?

It’s not. Really. It’s crazy what I had to go through to get to a productive setup that fits my needs.

One could say I could have spent all that time on writing vanilla HTML, CSS and JS and I would have had the same result in the same amount of time. I agree, if time would be all that mattered.

But for some people (like me), feeling productive, seeing how easy it is to test my ideas and how code seems to flow from my fingertips at the speed of thought, is what decides if I’ll ever finish and publish something, or if I’ll lose my patience and fallback to comfort zones.

Having to write the same boilerplate code over and over again, constant context switching between files, jumping back into a project after a few days and not knowing where everything was in those thousand-lines files.. these are all detractors that will eventually make me say ”f••k this! at least my day job brings money”.

Reusability

So many JS frameworks were created in the name of reusable components, but they all failed for me.

I mean sure, I can “npm install” a React calendar, and I am now “reusing” and not “reimplementing” the hard work of someone better than me at calendar UIs. But just try to stray away a little from the happy path that the component creator envisioned, and you will find it is mind-bendingly hard to bend the component to your specific needs.

You might raise a Github issue and the creator will add a few params so you can customize that specific thing, but so will others with different and maybe clashing needs. Soon enough, that component is declared unwieldy and too complex to use, the dev will say “f••k this! I’d rather do furniture” and someone else will come out and say: here’s the next best thing in React calendar libraries, so much simpler to use than those behemoths!

I never had this goal in mind but unexpectedly, the above setup is generic enough that I was able to extract it into a set of files for starting a new website. I can now duplicate that folder and start changing site-specific bits to get a new website.

Here are the websites I’ve done using this method:

• lowtechguys.com / where I publish my apps
• subsol.one / app I made for my brother to publish parties on
• robert.panaitiu.com / my brother’s personal website
• lunar.fyi / the app I have been talking about above

And the best thing I remember is that for each website I published a working version, good looking enough, with a contact page and small bandwidth requirements, in less than a day.

How does this solve the problem of straying away from the happy path? Well, this is not an immutable library residing in node_modules, or a JS script on a CDN. It is a set of files I can modify to the site’s needs.

There is no high wall to jump (having to fork a library, figuring out its unique build system etc.) or need to stick to a specific structure. Once the folder is duplicated, it has its own life.

For those interested, here is the repo containing the current state of my templates: github.com/alin23/plim-website

I don’t recommend using it, it’s possible that I’m the only one who finds it simple because I know what went into it. But if you do, I’d love to hear your thoughts.

Gatsby? Jekyll? Hugo?

Weirdly, this website I’m writing on is not made with Plim. At some point I decided to start a personal website, and I thought it probably needs a blog-aware site builder.

At the time, I didn’t know that RSS is an easily templatable XML file, and that all I need for a blog is to write Markdown.

I remember trying Gatsby and not liking the JS ecosystem around it. Jekyll was my second choice with Github Pages, but I think I fumbled too much with ruby and bundle to get it working and lost patience.

Both problems stemmed from my lack of familiarity with their ecosystems, but my goal was to write a blog, not learn Ruby and JS.

Hugo seemed much simpler, and it was also written in Go and distributed as a standalone binary, which I always like for my tools.

I marveled at Hugo’s speed, loved the fact that it supports theming (although it’s not as simple as it sounds) and that it has a lot of useful stuff built-in like syntax highlighting, image processing, RSS generator etc. But it took me sooo long to understand its structure.

There are many foreign words (to me) in Hugo: archetypes, taxonomies, shortcodes, partials, layouts, categories, series. Unfortunately, by the time I realized that I don’t need the flexibility that this structure provides, I had already finished this website and written my first article.

I also used a theme that uses the Tachyons CSS framework, for which I can never remember the right class to use. I thought about rewriting the website in Plim but converting everything to Tailwind or simple CSS would have been a lot of work.

I eventually started writing simple Markdown files for my notes, and have Caddy convert and serve them on the fly. Helps me write from my phone and not have to deal with Git and Hugo.

I still keep this for longform content, where a laptop is usually needed anyway.

You notice you never use the MacBook built-in display anymore, and it nags you to have it in your lower peripheral vision.

Closing the lid is not an option because you still use the keyboard and trackpad, maybe even the webcam and TouchID from time to time. So you try things:

• you try turning off the display by lowering the brightness completely. 🤔 hmm ok but now:
  • your mouse wanders to that screen sometimes
  • some windows get lost in there
  • ..and you still waste GPU cycles for rendering 6 million unused pixels
• you mirror the monitor to the built-in screen. nice, this solves the first two issues!
  • ok but why did the resolution change? do I have to change it back every time I do this??
  • wait, why don’t I get notifications anymore?! oh. there’s a setting for that
• you walk away from the desk, the screen goes to sleep
• you come back, the screen is now on something like 6% brightness, not completely off anymore
  • ok, press Brightness Down again, I can live with that
  • oh, mirroring got disabled as well.. at least there’s Cmd+Brightness Down

Why isn’t there a way to actually disable this screen?

BlackOut

Because a lot of users of my 🌕 Lunar app told me about their grievances with not being able to turn off individual displays in software, I went down the rabbit hole of display mirroring and automated all of the above.

Now someone can turn off and on any display at will using keyboard shortcuts, and can even automate the above MacBook + monitor workflow to trigger when an external monitor gets connected and disconnected.

But it’s still nagging me that somehow macOS can actually disable the internal screen completely, but we’re stuck with this zero-brightness-mirroring abomination.

Clamshell Mode

When closing the MacBook lid while a monitor is still connected, the internal screen disappears from the screen list and the external monitors remain available.

This function is called clamshell mode in the laptop world. Congratulations, your $3000 all-in-one computer is now just an SoC with some USB-C ports. Ok, you also get the speakers and the inefficient cooling system.

In the pre-chunky-MacBook-Pro-with-notch era, the lid was detected as being closed using magnets in the lid, and some hall effect sensors. So you were able to trick macOS into thinking the lid was closed by simply placing two powerful magnets at its sides.

With the new 2021 design, the MacBook has a hinge sensor, that can detect not only if the lid is closed, but also the angle of its closing. Magnets can’t trick’em anymore.

But all these sensors will probably just trigger some event in software, where a handler will decide if the display should be disabled or not, and call some disableScreenInClamshellMode function.

So where is that function, and can we call it ourselves?

The software side

Since Apple Silicon, most userspace code lives in a single file called a DYLD Shared Cache. Since Ventura, that is located in a Cryptex (a read-only volume) at the following path:

/System/Cryptexes/OS/System/Library/dyld/dyld_shared_cache_arm64e

Since that file is mostly an optimised concatenation of macOS Frameworks, we can extract the binaries using keith/dyld-shared-cache-extractor:

1
2

mkdir -p ~/Temp/dyld && cd ~/Temp/dyld
dyld-shared-cache-extractor /System/Cryptexes/OS/System/Library/dyld/dyld_shared_cache_arm64e $PWD

Let’s extract the exported and unexported symbols in text format to be able to search them easily using something like ripgrep.

I’m using /usr/bin/nm with fd’s -x option to take advantage of parallelisation. I like its syntax more than parallel’s since it has integrated interpolation for the basename/dirname of the argument (note the {/})

1
2
3
4
5
6
7
8

mkdir symbols private-symbols

fd --maxdepth 1 -t f \
    . ./System/Library/*Frameworks/*.framework/Versions/A/ \
    -x sh -c 'nm --demangle --defined-only --extern-only {} > symbols/{/}'
fd --maxdepth 1 -t f \
    . ./System/Library/*Frameworks/*.framework/Versions/A/ \
    -x sh -c 'nm --demangle --defined-only {} > private-symbols/{/}'

Searching for clamshell gives us interesting results. The most notable is this one inside SkyLight:

1
2
3
4

~/Temp/dyld ❯ rg -i clamshell

symbols/SkyLight
1710:00000001d44bce70 S _kSLSDisplayControlRequestClamshellState

SkyLight.framework is what handles window and display management in macOS and it usually exports enough symbols that we can use from Swift so I’m inclined to follow this path.

Let’s see if the internet has anything for us. I usually search for code on SourceGraph as it has indexed some large macOS repos with dyld dumps. Looking for RequestClamshellState gives us something far more interesting though:

Looks like Apple open sourced the power management code, nice! It even has recent ARM64 code in there, are we that lucky?

Here’s an excerpt of something relevant to our cause:

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58

SLSDisplayPowerControlClient *gSLPowerClient = nil;

enum {
    kPMClamshellOpen          = 1,
    kPMClamshellClosed        = 2,
    kPMClamshellUnknown       = 3,
    kPMClamshellDoesNotExist  = 4
};

void handleSkylightCheckIn(void)
{
// ...
    // create ws power control client
    NSError *err = nil;
    gSLPowerClient = [[SLSDisplayPowerControlClient alloc] initAsyncPowerControlClient:&err notifyQueue:_getPMMainQueue() notificationType:kSLDCNotificationTypeNone notificationBlock:^(void *dict) {
        if (dict != nil) {
            handleSkylightNotification(dict);
        } else {
            ERROR_LOG("Received a nil dictionary from WindowServer callback");
        }
    }];
// ...
}

void requestClamshellState(SLSClamshellState state)
{
    /* Forward clamshell state to WindowServer
     A) a request with a clamshell state of close in interpreted as a turn off clamshell display (clamshell close)
     B) a request with a clamshell state of open in interpreted as a turn on internal and ANY external displays (clamshell open)
     */

    if (!gSLCheckIn) {
        ERROR_LOG("WindowServer has not checked in. Refusing to change clamshell display state");
        return;
    }

    NSError *err = nil;
    NSMutableDictionary *request = [[NSMutableDictionary alloc] initWithCapacity:1];
    NSNumber *ns_state = [[NSNumber alloc] initWithUnsignedChar:state];
    [request setValue:ns_state forKey:kSLSDisplayControlRequestClamshellState];
    SLSDisplayControlRequestUUID uuid = [gSLPowerClient requestStateChange:(NSDictionary *const)request error:&err];
    if ([err code] != 0) {
       ERROR_LOG("Clamshell requestStateChange returned error %{public}@", err);
    } else {
       INFO_LOG("requestClamshellState: state %u, Received uuid %llu", state, uuid);
        struct request_entry *entry = (struct request_entry *)malloc(sizeof(struct request_entry));
        entry->uuid = uuid;
        entry->valid = true;
        STAILQ_INSERT_TAIL(&gRequestUUIDs, entry, entries);
    }
    if (request) {
       [ns_state release];
       [request release];
    }
    if (err) {
        [err release];
    }
}

So it’s instantiating an SLSDisplayPowerControlClient then calling its requestStateChange method. SLS is a prefix related to SkyLight (probably standing for SkyLightServer), let’s see if we have that code in our version of the framework.

I prefer to do that using Hopper and its Read File From DYLD Cache feature which can extract a framework from the currently in-use cache:

Ok the class and methods are there, let’s look for what uses them. Since it’s most likely a daemon handling power management, I’ll look for it in /System/Library.

And looks like powerd is what we’re looking for, containing exactly the code that we saw on SourceGraph.

1
2
3
4
5

❯ rg -uuu requestClamshellState /System/Library/ 2>/dev/null

/System/Library/CoreServices/powerd.bundle/powerd: binary file matches (found "\0" byte around offset 4)

❯ hopperv4 -e /System/Library/CoreServices/powerd.bundle/powerd

Writing the code

To link and use SLSDisplayPowerControlClient we need some headers, as Swift doesn’t have the method signatures available.

Looking for SLSDisplayPowerControlClient on SourceGraph gives us more than we need.

Let’s create a bridging header so that Swift can link to Objective-C symbols, and a Swift file to where we’ll try to replicate what powerd does.

1
2

mkdir clamshell && cd clamshell
touch Bridging-Header.h Clamshell.swift

Bridging-Header.h

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12

#import <Foundation/Foundation.h>

@interface SLSDisplayPowerControlClient {}

- (id)initAsyncPowerControlClient:(id*)arg1 notifyQueue:(id)arg2 notificationType:(UInt8)arg3 notificationBlock:(void (^)(NSDictionary*))notificationBlock;
- (id)initPowerControlClient:(id*)arg1 notifyQueue:(id)arg2 notificationType:(UInt8)arg3 notificationBlock:(void (^)(NSDictionary*))notificationBlock;

- (unsigned long long)requestStateChange:(id)arg1 error:(id*)arg2;
@end

extern NSString* kSLSDisplayControlRequestClamshellState;
UInt8 kSLDCNotificationTypeNone = 0;

Clamshell.swift

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33

import Foundation

enum ClamshellState: Int {
    case open = 1
    case closed = 2
    case unknown = 3
    case doesNotExist = 4
}

var err: AnyObject?
let skyLightPowerClient = SLSDisplayPowerControlClient(powerControlClient: &err, notifyQueue: DispatchQueue.main, notificationType: kSLDCNotificationTypeNone) { dict in
    print(dict as Any)
}

func requestClamshellState(_ state: ClamshellState) {
    // Send the request
    let request: [AnyHashable: Any] = [
        kSLSDisplayControlRequestClamshellState: NSNumber(value: state.rawValue)
    ]

    var err: AnyObject?
    let uuid = skyLightPowerClient!.requestStateChange(request, error: &err)

    // Check the response
    if (err as! NSError?)?.code != 0 {
        print("Clamshell requestStateChange returned error", err?.localizedDescription ?? "")
    } else {
        print("requestClamshellState: state %u, Received uuid %llu", state, uuid)
    }
}

print(skyLightPowerClient!)
requestClamshellState(.closed)

Compiling…

To compile the binary using swiftc we have to point it to the location of SkyLight.framework which is located at /System/Library/PrivateFrameworks.

We then tell it to link the framework using -framework SkyLight and import our bridging header. Then we run the resulting binary.

I prefer to run this using entr to watch the files for changes. With the code editor on the left and the terminal on the right, I can iterate and try things faster by just editing and saving the file, then watch the output on the right.

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12
13
14
15

swiftc \
    -F/System/Library/PrivateFrameworks \
    -framework SkyLight \
    -import-objc-header Bridging-Header.h \
    Clamshell.swift -o Clamshell
./Clamshell

# For faster iteration, watch file changes with entr:

echo Clamshell.swift Bridging-Header.h | entr -rs '\
	swiftc -F/System/Library/PrivateFrameworks \
		-framework SkyLight \
		-import-objc-header Bridging-Header.h \
		 Clamshell.swift -o Clamshell \
	&& ./Clamshell'

Well.. it’s not working. The error is not helpful at all, there’s nothing on the internet related to it.

1
2

<SLSDisplayPowerControlClient: 0x600000fb8720>
Clamshell requestStateChange returned error The operation couldn’t be completed. (CoreGraphicsErrorDomain error 1004.)

Looking for errors

Maybe the system log has something for us. One can check that using Console.app but I prefer looking at it in the Terminal through the /usr/bin/log utility.

1

log stream --predicate 'eventMessage contains "Clamshell"'

Something from AMFI about the binary signature. CMS stands for Cryptographic Message Syntax which is what codesign adds to a binary when it signs it with a certificate.

1
2
3

kernel: (AppleMobileFileIntegrity) AMFI: '/Users/alin/Temp/dyld/clamshell/Clamshell' has no CMS blob?
kernel: (AppleMobileFileIntegrity) AMFI: '/Users/alin/Temp/dyld/clamshell/Clamshell': Unrecoverable CT signature issue, bailing out.
tccd: [com.apple.TCC:access] AUTHREQ_ATTRIBUTION: msgID=60667.1, attribution={responsible={TCCDProcess: identifier=kitty, pid=19959, auid=501, euid=501, responsible_path=/Applications/kitty.app/Contents/MacOS/kitty, binary_path=/Applications/kitty.app/Contents/MacOS/kitty}, requesting={TCCDProcess: identifier=Clamshell, pid=60667, auid=501, euid=501, binary_path=/Users/alin/Temp/dyld/clamshell/Clamshell}, },

I have GateKeeper disabled and running the binary from a terminal that’s added to the special Developer Tools section of Security & Privacy, so this shouldn’t cause any problems.

I checked just to be sure, and signing it with my $100/year Apple Developer certificate gets rid of the CMS blob error but doesn’t change anything in the result.

Phew, let's take a break

I just arrived after a long train ride at the house I'm rebuilding with my wife, and wanted to share this nice view with you 😌

It's January, but the sun is warming our faces and the hazelnut trees are already producing their yellow catkins.

Ten years ago, the children of the house's previous owners were walking in knee deep snow and coasting downhill on their wooden sleds, hurting a few young fir trees on the way down. 🌲

Seasons are changing.

Digging deeper

Some system capabilities can only be accessed if the binary has been signed by Apple and has specific entitlements. Checking for powerd’s entitlements gives us something worrying.

The binary seems to use com.apple.private.* entitlements. This usually means that some APIs will fail if the required entitlements are not present.

1
2
3
4
5
6
7
8
9

> codesign -d --entitlements - /System/Library/CoreServices/powerd.bundle/powerd

Executable=/System/Library/CoreServices/powerd.bundle/powerd
[Dict]
	...
	[Key] com.apple.private.SkyLight.displaypowercontrol
	[Value]
		[Bool] true
	...

We can try to add the entitlements ourselves. We just need to create a plist file and use it in codesign:

Entitlements.plist

1
2
3
4
5
6
7
8

<?xml version="1.0" encoding="UTF-8"?>
<!DOCTYPE plist PUBLIC "-//Apple//DTD PLIST 1.0//EN" "http://www.apple.com/DTDs/PropertyList-1.0.dtd">
<plist version="1.0">
    <dict>
        <key>com.apple.private.SkyLight.displaypowercontrol</key>
        <true/>
    </dict>
</plist>

Sign the binary with entitlements and run it:

1
2
3

❯ codesign -fs $CODESIGN_CERT --entitlements Entitlements.plist Clamshell
❯ ./Clamshell
Job 1, './Clamshell' terminated by signal SIGKILL (Forced quit)

Looks like we’re getting killed instantly. The log stream shows AMFI is doing that because we’re not Apple and we’re not supposed to use that entitlement.

1
2
3
4
5
6
7
8

kernel: mac_vnode_check_signature: /Users/alin/Temp/dyld/clamshell/Clamshell: code signature validation failed fatally: When validating /Users/alin/Temp/dyld/clamshell/Clamshell:
  Code has restricted entitlements, but the validation of its code signature failed.
Unsatisfied Entitlements: com.apple.private.SkyLight.displaypowercontrol

kernel: (AppleSystemPolicy) ASP: Security policy would not allow process: 57234, /Users/alin/Temp/dyld/clamshell/Clamshell
amfid: /Users/alin/Temp/dyld/clamshell/Clamshell not valid: Error Domain=AppleMobileFileIntegrityError Code=-413 "No matching profile found" UserInfo={NSURL=file:///Users/alin/Temp/dyld/clamshell/Clamshell, unsatisfiedEntitlements=<CFArray 0x155e1b600 [0x1f0e613a8]>{type = immutable, count = 1, values = (
    0 : <CFString 0x155e12db0 [0x1f0e613a8]>{contents = "com.apple.private.SkyLight.displaypowercontrol"}
)}, NSLocalizedDescription=No matching profile found}

AMFI

What’s this AMFI exactly and why is it telling us what we can and cannot do on our own device?

The acronym stands for Apple Mobile File Integrity and it’s the process enforcing code signature at the system level.

By default, the OS locks these private APIs because if we would be able to use them, a malware or a bad actor would be able to do it as well. With it locked by default, malware authors are deterred from trying to use these APIs on targets of lower importance as this would usually need a 0-day exploit.

In the end it’s just another layer of security, and if in the rare case someone needs to bypass it, Apple provides a way to do it. The process involves disabling System Integrity Protection and adding amfi_get_out_of_my_way=1 as a boot arg.

1
2
3
4
5
6
7
8
9

# Inside a Recovery terminal (to disable SIP)

> csrutil disable
> reboot

# Inside a normal terminal after disabling SIP

> sudo nvram boot-args="amfi_get_out_of_my_way=1"
> sudo reboot now

I don’t recommend doing this as it puts you at great risk, since the system volume is no longer read only, and code signatures are no longer enforced.

I only keep this state for research that I do in short periods of time, then turn SIP back on for normal day to day usage.

In case you need to revert the above changes:

1
2
3
4
5
6
7
8

# Inside a normal terminal before enabling SIP

> sudo nvram boot-args=""

# Inside a Recovery terminal (to enable SIP)

> csrutil enable
> reboot

No more AMFI?

Unfortunately even after disabling AMFI, we’re still encountering the CoreGraphicsError 1004. It’s true, AMFI is not complaining about the entitlements anymore, they’re accepted and the binary is not SIGKILLed.

But we still can’t get into clamshell mode using just software.

Frida

If you haven’t heard of it, Frida is this awesome tool that lets you inject code into already running processes, hook functions by name (or even by address), observe how and when they’re called, check their arguments and even make your own calls.

Let me share with you another macOS boot arg that I like:

1

sudo nvram boot-args=-arm64e_preview_abi

This one enables code injection. Now we can use Frida to hook the SkyLight power control methods to see how they are called as we close and open the lid:

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12
13
14
15
16
17

> sudo frida-trace -t SkyLight -m '-[SLSDisplayPowerControlClient *]' powerd

// Closing the lid
           /* TID 0x5427 */
  4617 ms  SLSDisplayControlRequestClamshellStateKey: 2
  4617 ms  -[SLSDisplayPowerControlClient requestStateChange:0x13cf06a60 error:0x16b8ca828]
  4628 ms     | -[SLSDisplayPowerControlClient service]
  4628 ms     | -[SLSDisplayPowerControlClient sendStateChangeRequest:0x13cf06a60 uuid:0x16b8ca7e0]
  4628 ms     |    | -[SLSDisplayPowerControlClient service]

// Opening the lid
           /* TID 0x8a17 */
 10537 ms  SLSDisplayControlRequestClamshellStateKey: 1
 10537 ms  -[SLSDisplayPowerControlClient requestStateChange:0x13cc1e1c0 error:0x16b9567a8]
 10538 ms     | -[SLSDisplayPowerControlClient service]
 10538 ms     | -[SLSDisplayPowerControlClient sendStateChangeRequest:0x13cc1e1c0 uuid:0x16b956760]
 10538 ms     |    | -[SLSDisplayPowerControlClient service]

We got our confirmation at least. powerd is indeed calling SLSDisplayPowerControlClient.requestStateChange(2) when closing the lid.

Let’s check what happens when we try to call that method in Clamshell.swift.

We first add the line readLine(strippingNewline: true) at the top of the Clamshell.swift file to make the binary wait for us to press Enter. This is so that we have a running process that we can attach to with Frida.

1
2
3
4
5
6
7
8

> sudo frida-trace -t SkyLight -m '-[SLSDisplayPowerControlClient *]' Clamshell

           /* TID 0x103 */
  1475 ms  SLSDisplayControlRequestClamshellStateKey: 2
  1475 ms  -[SLSDisplayPowerControlClient requestStateChange:0x600001d64510 error:0x16d8d3c90]
  1479 ms     | -[SLSDisplayPowerControlClient service]
  1479 ms     | -[SLSDisplayPowerControlClient sendStateChangeRequest:0x600001d64510 uuid:0x16d8d3a10]
  1479 ms     |    | -[SLSDisplayPowerControlClient service]

Everything looks the same, seems that we’re not looking deep enough.

The request method seems to access the service property which is an SLSXPCService. XPC Services are what macOS uses for low-level interprocess communication.

A process can expose an XPC Service using a label (e.g. com.myapp.RemoteControlService) and listen to requests coming through, other processes can connect to it using the same label and send requests.

The system handles the routing part. And the authentication part.

Looks like an XPC Service can also be restricted to specific code signing requirements, is it possible that this is what we’re running into here?

Let’s trace SLSXPCService methods as well using Frida:

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42

> sudo frida-trace -t SkyLight -m '-[SLSDisplayPowerControlClient *]' -m '-[SLSXPCService *]' powerd

// Closing the lid while observing powerd
           /* TID 0x518b */
  3029 ms  -[SLSDisplayPowerControlClient requestStateChange:0x139621c60 error:0x16f0c2828]
  3029 ms  SLSDisplayControlRequestClamshellStateKey: 2
  3043 ms     | -[SLSDisplayPowerControlClient service]
  3043 ms     | -[SLSDisplayPowerControlClient sendStateChangeRequest:0x139621c60 uuid:0x16f0c27e0]
  3043 ms     |    | -[SLSDisplayPowerControlClient service]
  3043 ms     |    | -[SLSXPCService sendXPCDictionary:0x13a913be0]
  3043 ms     |    |    | -[SLSXPCService reinitConnection]
  3043 ms     |    |    |    | -[SLSXPCService enabled]
  3043 ms     |    |    |    | -[SLSXPCService enabled]
  3043 ms     |    |    |    | -[SLSXPCService connected]
  3043 ms     |    |    | -[SLSXPCService connection]
  3452 ms  -[SLSXPCService handleXPCEvent:0x13ad0ea40]
  3452 ms     | -[SLSXPCService enabled]
  3452 ms     | -[SLSXPCService cfStringToCStringPtr:0x1f3133020]
  3452 ms     | -[SLSXPCService connected]

> sudo frida-trace -t SkyLight -m '-[SLSDisplayPowerControlClient *]' -m '-[SLSXPCService *]' Clamshell

// Trying to send the clamshell request in software
  1435 ms  -[SLSDisplayPowerControlClient requestStateChange:0x6000014d4030 error:0x16b123c90]
  1435 ms  SLSDisplayControlRequestClamshellStateKey: 2
  1444 ms     | -[SLSDisplayPowerControlClient service]
  1444 ms     | -[SLSDisplayPowerControlClient sendStateChangeRequest:0x6000014d4030 uuid:0x16b123a10]
  1444 ms     |    | -[SLSDisplayPowerControlClient service]
  1444 ms     |    | -[SLSXPCService sendXPCDictionary:0x600003ec4000]
  1444 ms     |    |    | -[SLSXPCService reinitConnection]
  1444 ms     |    |    |    | -[SLSXPCService enabled]
  1444 ms     |    |    |    | -[SLSXPCService connected]
  1444 ms     |    |    |    | -[SLSXPCService autoreconnect]
  1444 ms     |    |    |    | -[SLSXPCService enabled]
Process terminated

// ...we're missing this stuff
//  3043 ms     |    |    | -[SLSXPCService connection]
//  3452 ms  -[SLSXPCService handleXPCEvent:0x13ad0ea40]
//  3452 ms     | -[SLSXPCService enabled]
//  3452 ms     | -[SLSXPCService cfStringToCStringPtr:0x1f3133020]
//  3452 ms     | -[SLSXPCService connected]

Great! or not?

I’m not sure if I should be happy that we found that our clamshell request doesn’t work because we don’t have an XPC connection, or if I should be worried that this means we won’t be able to make this work with SIP enabled.

I guess it’s time to go deeper to find out.

XPC Services

Now that we have access to Frida, we can use the handy xpcspy tool to sniff the XPC communication of powerd.

I’m thinking maybe we can find the endpoint name of the XPC listener and just connect to it and send a raw message directly, instead of relying on SkyLight to do that.

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12
13
14
15

> sudo xpcspy --parse powerd

// Closing the lid

xpc_connection_send_message
<OS_xpc_connection: <connection: 0x13a808820> { name = (anonymous), listener = false, pid = 30630, euid = 88, egid = 88, asid = 100014 }>
<OS_xpc_dictionary> { count = 4 contents =
    "PayloadType" => <OS_xpc_uint64: <uint64: 0x81917509705f5717>: 3>
    "Command" => <OS_xpc_uint64: <uint64: 0x81917509705f572f>: 4>
    "Payload" => <data> { length = 92 bytes, contents = {
	    SLSDisplayControlRequestClamshellStateKey = 2;
	}
    "UUID" => <OS_xpc_uint64: <uint64: 0x81917509705f5647>: 41>

}

So we have name = (anonymous), listener = false, pid = 30630.

An anonymous listener, can it get even worse? The PID coincides with WindowServer --daemon so it’s definitely the message we’re also trying to send. But with an anonymous listener, we’re stuck to relying on SkyLight’s exported code to reach it.

I guess we need to go back to do some old-school assembly reading.

Filling empty spaces

After renaming some sub-procedures in Hopper, looking at the graph reveals the different code paths that powerd and Clamshell are taking through SLSXPCService.reinitConnection.

powerd

1. sees that the service’s enabled and connected properties are true
2. so it gets out of reinitConnection
3. and straight into sending the XPC dictionary through the available connection.

Clamshell

• sees that enabled, connected and autoreconnect are false
  • so it fails with a CGError
• if those properties were true it would go on the right-side code path which
  • checks if properties at 0x20 and 0x28 are non-zero
  • then proceeds to reconnect.

Adding some Memory.readPointer calls inside __handlers__/SLSXPCService/reinitConnection.js shows us what SkyLight is expecting to see at 0x20 and 0x28:

Two NSMallocBlocks right after the OS_xpc_connection and the OS_dispatch_queue_serial properties.

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12
13
14
15
16

  5502 ms   -[SLSXPCService reinitConnection]
  5502 ms       arg0 obj: <SLSXPCService: 0x11f50f130>

  5502 ms       Memory.readPointer 0x8 0x101

  5502 ms       Memory.readPointer 0x10 0x11f50bb10
  5502 ms       SLSXPCService at 0x10 <OS_xpc_connection: <connection: 0x11f50bb10> { name = (anonymous), listener = false, pid = 396, euid = 88, egid = 88, asid = 100014 }>

  5502 ms       Memory.readPointer 0x18 0x11df05970
  5502 ms       SLSXPCService at 0x18 <OS_dispatch_queue_serial: Power Management main queue>

  5502 ms       Memory.readPointer 0x20 0x11f50a740
  5502 ms       SLSXPCService at 0x20 <__NSMallocBlock__: 0x11f50a740>

  5502 ms       Memory.readPointer 0x28 0x11f50a770
  5502 ms       SLSXPCService at 0x28 <__NSMallocBlock__: 0x11f50a770>

Judging by the contents of SLSXPCService.h, those are the closures for errorBlock and notificationBlock:

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12
13
14

@interface SLSXPCService : NSObject <SLSXPCServiceProtocol> {
	char _enabled;
	char _connected;
	char _setTarget;
	char _autoreconnect;
	NSObject*<OS_xpc_object> _connection;
	NSObject*<OS_dispatch_queue> _notifyQueue;

	/* This would be 0x20 */ id _errorBlock;
	/* This would be 0x28 */ id _notificationBlock;

	/*^block*/id _clientErrorBlock;
	/*^block*/id _clientNotificationBlock;
}

I’m inching closer to the good code path but I seem to never get there.

So here’s what I did so far in Clamshell.swift before calling requestClamshellState:

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12

guard let service = skyLightPowerClient.service else {
    print("SLSXPCService is nil")
    exit(1)
}

service.autoreconnect = true
service.errorBlock = { err in
    print("service.errorBlock", err)
}
service.notificationBlock = { notification in
    print("service.notificationBlock", notification)
}

After calling requestClamshellState, the code crashes with SIGSEGV inside createNoSenderRecvPairWithQueue:errorHandler:eventHandler: because it branches to the 0x0 address.

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12

// The crashing instruction is:
ldr    x8, [x19, #0x28]

// Because the memory at [x19, #0x28] contains 0x0

// And register x19 contains:
(lldb) po $x19
<__NSMallocBlock__: 0x600000c08330>
 signature: "v8@?0"
 invoke   : 0x1a081a958 (/System/Library/PrivateFrameworks/SkyLight.framework/Versions/A/SkyLight`__75-[SLSXPCService createNoSenderRecvPairWithQueue:errorHandler:eventHandler:]_block_invoke)
 copy     : 0x1a05e0a18 (/System/Library/PrivateFrameworks/SkyLight.framework/Versions/A/SkyLight`__copy_helper_block_e8_32b40r)
 dispose  : 0x1a05e09d4 (/System/Library/PrivateFrameworks/SkyLight.framework/Versions/A/SkyLight`__destroy_helper_block_e8_32b40r)

Giving up (for now)

Unfortunately I’m a bit lost here. I’ll take a break and hope that the solution comes in a dream or on a long walk like in those mythical stories.

The article is already longer than I’d be inclined to read so if anyone reaches this point, congrats, you have the patience of a monk.

If there are better ways to approach a problem like this one, I’d be glad to hear about it through the contact form.

I’m not always happy to learn that I’ve wasted 4 days on a problem that could have been solved in a few hours with the right tools, but at least I’ll learn how not to bore people with writings on rudimentary tasks next time.

Not really, no. Not without annoying workarounds and a confusing user experience.

Another email, another annoyed user: Firefox not loading websites when launched through rcmd! It works when launched from Alfred.. Please fix ASAP!! I’m gonna fix this Firefox issue once and for all!

Launch Xcode, open the rcmd project, check the launchApp function code, it’s just a NSWorkspace.open call on Firefox.app, what does Alfred do differently?

Disassemble Alfred.app in Hopper, look for NSWorkspace.open, of course it’s there, it’s the exact same thing.

Try open /Applications/Firefox.app in a terminal, it works, websites load as expected.

Breakpoint on launchApp, check the debugger again, let’s be rigorous, what am I really calling open on?

Argument is /System/Volumes/Data/Applications/Firefox.app which is just a symlink to /Applications/Firefox.app right? .. or was it the other way around? Anyway let’s just try it for the sake of it, I’m desperate.

Run open /System/Volumes/Data/Applications/Firefox.app, huh?? no websites load? THAT WAS IT?!

Add path.replacingOccurrences(of: "/System/Volumes/Data", with: ""), build, run, hold Right Command, press F, Firefox launches and holy cow everything works!!

I don’t even care why anymore, let’s just release this fix on the App Store.

And while I’m at it, why not try to add that window switching capability that people have been asking about?

I remember something about Accessibility permissions not being available in the sandbox, but I just used an App Store app that was able to request the permissions so there has to be a way, how hard could it be?

Well it turns out it’s pretty darn hard, and I’m still working on this window switching thing to this day.. sigh.. let me tell you about it.

Apps vs windows

There’s an important distinction between switching windows and switching apps on the Mac. As opposed to Microsoft Windows where you just Alt-Tab through .. well, windows, on macOS you Command Tab through apps by default. When an app with multiple windows is focused, Command backtick will cycle through the windows of that app.

Six years ago I was a Windows power user, and when I got my first Mac, Command Tabbing through apps felt very weird. Suddenly I was closing all windows of Sublime but its icon was still there in the Command Tab list, or I would minimize Chrome and focusing its icon didn’t unminimize it. The app vs window distinction just didn’t exist in my mind.

Now, after 6 years, the macOS way feels a lot more intuitive:

• I mostly switch between apps with a single window (browser, terminal etc.)
• There’s a very small subset of apps where I might have more than one window (code editor, image/PDF viewer)
• I might want to keep my code editor app running even after I closed all its windows, so I can have it load instantly when opening a new window to edit a file/project
• Minimizing windows that are irrelevant at the moment allows me to cycle through the relevant ones with Command backtick
• No need to see a thumbnail of each window, when almost all apps are single window

Of course it might just be the power of habit, after all I was able to be just as productive with the Windows way in the past ¯\_(ツ)_/¯

Command Tab Tab Tab Tab…

The app centric approach is nice but having to switch between 10 different apps at a time gets annoying fast.

Pressing Tab 5 times in a row to get to the app I want could be categorized as a first world problem and I should just get used to it. But doing that 50 times a day and having to always visually check if I chose the right icon, tends to break my flow of thinking, and makes me get tired faster because of all the context switching.

That’s the main reason I created rcmd, to switch apps without thinking about switching apps.

My right thumb rests nicely on the Right Command key and I barely use that easy to reach key. So I turned it into a dedicated app switching key.

Dynamic assignments

I decided to dynamically assign each app the first letter of its name so that I don’t have to try to remember what key did I assign to Xcode?. I just hold Right Command and press X without any mental effort because I know I have no other app starting with X.

And if I forgot that Xcode is not already running (or if it crashes in the background like it sometimes does), rcmd launches it automatically (since I clearly wanted it running if I tried to focus it).

Static assignments

Xcode is a happy case though. I have so many apps starting with S that I decided custom assignments might be a better fit for that. I left Sublime Text for the S key since it’s my most used app, and then assigned mnemonic keys for others:

• O for Soulver
• P for Spotify
• E for Sketch (because K is taken by the Kitty terminal)
• B for Safari browser
• Other rarely used apps (SF Symbols, Slack, Sublime Merge) will be reachable by cycling using rcmd-rshift-s (it’s good enough for me as I rarely have those open)

Seek and hide

Often I need to check the status of an app briefly and then get back to what I was doing. Some examples

• check a long running task in the terminal
• check if I got an email I’m waiting for while notifications are paused
• see what’s this dope song that started playing from my Discover Weekly playlist

That’s why I added the Hide action in rcmd.

Now I just hold Right Command and press K to check the Kitty terminal, then, without lifting any finger, press K again to hide it and get back to what I was doing.

This also allows the system to activate App Nap for the hidden app and put it into a lower energy usage state until I need it again.

Is window switching even needed?

Unfortunately yes, there are many cases where an app might have a lot of windows open:

• Separate projects/folders open in Sublime Text
• Multiple documents in Pages or Microsoft Word
• Lots of PDFs open for referencing in Preview

Available solutions

1. App Expose: Command Tab allows pressing the ↓ Down Arrow key with the app icon selected, to expose all the windows of that app for visual selection.

   • It’s nice and useful for when you churn windows a lot, but it’s way too slow for cases when you mostly have the same windows open.

2. Command backtick `: this native macOS hotkey will cycle through the windows of the current app but we’re back to square one where you have to visually analyze each window to see if you got the right one in focus.

3. Alt-Tab: this is a really nice open source app which replicates the Microsoft Windows way of selecting windows by thumbnails.

   • It’s what I used for a long time, until I got too frustrated with the fact that all my seven Sublime Text windows look exactly the same and I have to also read the whole window title to find the one I want to focus.

4. Contexts.co: a fuzzy searcher for window titles. I’ve used it in the past and it was definitely faster than the rest but it still required more key presses than I wanted

   • I don’t really need to search the whole window title, just the project name.

5. Stage Manager: the new addition in macOS Ventura, which in its current state is just discoverable Spaces.

   • That’s the feeling I got from using it: stages are just like spaces, but more visible (through the left sidebar) and easier to reach for (by clicking on them or by focusing a window in a specific stage).
   • It still doesn’t provide any keyboard control and moving specific windows in and out of the stages requires too much work with the mouse.
   • At least for Spaces I had yabai to provide keyboard shortcuts for moving the current window to whatever space I wanted to.

My preferred solution: the Right Option key

It’s a sunny day in Brașov, I’m on my balcony taking in the sun, testing and perfecting XDR Brightness to make working in direct sunlight easier on my MacBook 14” while also rewriting parts of the Lunar UI in SwiftUI.

I’ve already written a lot of SwiftUI boilerplate in my other projects, so I’m mostly copy pasting stuff between Sublime Text windows. I also have three Sublime windows with disassembled macOS private frameworks to look for the hidden functions I need to improve the XDR Brightness curve and responsiveness.

Juggling with all these windows suddenly became very frustrating.

Why can’t I focus exactly the window I want with one hotkey just like I focus apps with rcmd?

I’m probably going to have the same set of windows for the next few days, I know the names of the projects I have open in them, I could use the first letter of the project name to reference a specific window.

The Right Command key is taken, but right beside it stands another rarely used key: the Right Option key (ralt for short)

I want to be able to press ralt-r to focus the Sublime window containing the rcmd project, ralt-l to focus the Lunar project, ralt-v for the Volum project, ralt-p to get to the PrivateFrameworks folder and so on.

The plan seems simple enough:

• get the list of windows and their title from the current app
• extract the first letter of the project name
• assign Right Option + letter to some focusWindow function
• get back to the real work

Oh right … the sandbox

It’s not like the above hasn’t been done before, there are plenty of window switcher and snap/resize examples on macOS, some of them are even open source:

• yabai
• Amethyst
• Rectangle

One window snapping tools is even on the App Store: Magnet

But why are there no window switchers on the App Store?

Well, for app switching, Apple provides a really nice API to enumerate and activate running apps without needing any intrusive permissions: NSRunningApplication

Finding Xcode and focusing it

1
2
3
4
5

let apps = NSWorkspace.shared.runningApplications
let xcode = apps.first { app in
    app.bundleIdentifier == "com.apple.dt.Xcode"
}
xcode?.activate()

But there’s no such thing for enumerating the windows of those running apps. All of the apps that work with app windows, need to tap into the Accessibility API, the one that gives you full access to extract and modify the contents of everything visible and invisible.

And so, window enumeration becomes possible, by fetching the array of UI elements under the AXWindows attribute of an app.

But since a window is like just any other UI element, then there’s no focus or activate method, so how do these apps manage to focus a window?

Take a look at this nice and intuitive snippet extracted from yabai:

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12
13
14

static void window_manager_make_key_window(ProcessSerialNumber *window_psn, uint32_t window_id)
{
    uint8_t bytes1[0xf8] = { [0x04] = 0xf8, [0x08] = 0x01, [0x3a] = 0x10 };
    uint8_t bytes2[0xf8] = { [0x04] = 0xf8, [0x08] = 0x02, [0x3a] = 0x10 };

    memcpy(bytes1 + 0x3c, &window_id, sizeof(uint32_t));
    memset(bytes1 + 0x20, 0xFF, 0x10);

    memcpy(bytes2 + 0x3c, &window_id, sizeof(uint32_t));
    memset(bytes2 + 0x20, 0xFF, 0x10);

    SLPSPostEventRecordTo(window_psn, bytes1);
    SLPSPostEventRecordTo(window_psn, bytes2);
}

Even though I knew that key window meant focused window in macOS terminology, it still took me a while to land on this code and start believing that this is really focusing a window.

In the end, what that code represents is message passing to the SkyLight private framework, the one that handles the macOS window management, Dock, Spaces and a ton of other stuff. I’m guessing someone sneaked in a VM debugger or looked through the assembly code to find the right bytes to send.

Ok, enumeration and focusing is doable, what else do we need? Right, Accessibility permissions. Here comes the biggest hurdle.

How do you escape the macOS sandbox?

You don’t.

On macOS, an app can be run:

• within a sandbox
  • where it has its own limited view of the file system and limited access to privileged APIs
• outside the sandbox
  • where it has access to everything that’s not guarded by SIP (System Integrity Protection)

App Store apps can only run inside the sandbox, and within that, an app can’t ask for Accessibility permissions. The API for that just throws a silent error and does nothing.

But then how does Magnet do it, and a few other apps as well like Peek or PopClip for example?

Turns out, these apps have a special exception from Apple, mostly because they were on the App Store before the sandbox has become mandatory: objective c - How to use Accessibility with sandboxed app? - Stack Overflow

I can barely get my apps to not be rejected by the App Store reviewers, I’m not going to get an exception just so that rcmd can focus specific windows. So now what?

Workarounds

I thought, if there was an app running outside the sandbox and listening for rcmd’s listWindows and focusWindow commands, I might be able to get this working.

I remembered Hammerspoon having a really complete window management support and it also being scriptable with Lua made it the perfect choice.

HTTP would probably be overkill for this, I knew Hammerspoon had an inter-process communication (IPC) API built-in so I tried to use that.

1
2
3

static NSString *portName = @"Hammerspoon";
CFMessagePortRef messagePort = CFMessagePortCreateRemote(NULL, (__bridge CFStringRef)portName);
// messagePort is NULL here

Well nope, the sandbox doesn’t allow that.

What about the hs CLI that Hammerspoon provides, I knew that you could send arbitrary IPC messages using that, right?

Nope again, any process run by a sandboxed app will inherit that sandbox limitations.

Ok fine, HTTP it is! Thankfully Hammerspoon provides an HTTP server and I just need to register a callback and make it listen on a port. Since we’ve already reached this madness, let’s go straight to websockets.

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12
13
14
15
16
17
18
19
20

function rcmdCallbackWS(msg)
    local params = hs.json.decode(msg)
    local response = "{}"

    if params.cmd == "listWindows" then
        response = hs.json.encode(hs.window:allWindows())
    elseif params.cmd == "focusWindow" then
        hs.window.get(params.window):focus()
    end

    return response
end

server = hs.httpserver.new(false, false)
server:setName("rcmd-hammerspoon")
server:setInterface("localhost")
server:setPort(3094)

server:websocket("/ws", rcmdCallbackWS)
server:start()

Alright, this seems to work. I can connect to the Hammerspoon websocket, get all windows, and focus windows by their IDs.

Now how do I explain to rcmd users that in order to focus windows, they need to:

• download a zip file from GitHub releases
• install another app with no affiliation to rcmd
• give that app Accessibility permissions
• install a Lua script in the ~/.hammerspoon directory
• ensure Hammerspoon is kept running all the time

Automating the workarounds

The App Store guidelines explicitly forbid an app from installing another app or binary to enhance its capabilities.

2.4.5 Apps distributed via the Mac App Store have some additional requirements to keep in mind:

(iv) They may not download or install standalone apps, kexts, additional code, or resources to add functionality or significantly change the app from what we see during the review process.

So I can’t install Hammerspoon automatically (it would be a bad idea anyway, this is malware behavior), but I can try to automate most of the stuff and present it as a 1-button install action.

So I wrote a function to download Hammerspoon.zip, unzip it in a temporary folder, move it to /Applications, write init.lua and rcmd.lua inside the ~/.hammerspoon directory, launch Hammerspoon and wait for the websocket to be available.

The user only has to click an Install window switcher button, no big deal.

Quarantine says “not so fast”

You see, when a sandboxed app downloads a file, the system automatically adds the com.apple.quarantine extended attribute to the file.

1
2

> xattr -l ~/Downloads/Hammerspoon.zip
com.apple.quarantine: 0083;62ea4f5c;Safari;3A6D521B-5E0D-4202-80C4-A5EB567DC246

This means that macOS GateKeeper will prevent you from launching any downloaded app or running any binary directly from code.

Even if the user tries to launch the downloaded app manually afterwards, it will still fail with the App can’t be opened error.

No amount of xattr -cr Hammerspoon.app will fix this if run from the sandbox.

Great. Scrap the download and install part, split the button into two buttons:

1. Install Hammerspoon which only shows text instructions on how to download and install the app manually
2. Install custom script which writes the Lua script files to disk

I’ve streamlined this process as much as the sandbox allows me, and after giving the app to some beta testers, every single one of them found it so confusing that they said they would not use it.

And who can blame them, I myself find it too convoluted whenever I test it.

So is this on the App Store?

Yes, surprisingly. It passed App Review without a single rejection.

I hid the feature behind a Try experimental window switching red button to deter support emails on the subject, but it’s there for anyone to try and use.

After the initial setup, it actually works pretty reliably, and the websocket connection to Hammerspoon is so fast that I don’t ever notice this happens over the network. It feels like a native window switcher to me.

But I wasn’t able to create a seamless experience like I did for app switching.

Oh well, at least I solved my own problem and can get back to what I was doing.

One month later.

Exactly 3 months and a day after placing an order through a Romanian Apple reseller, I finally got my 14-inch M1 Max.

Well, actually.. I first got the wrong configuration (base model instead of CTO), had to return it to them after wasting a day on migrating my data to it, they sent my money back by mistake, had to pay them again, and after many calls and emails later the correct laptop arrived.

As soon as these devices were in the hands of users, requests started coming in for Lunar to provide an option to get past the 500 nits limit for everyday usage

Over the last week I tried my best to figure out how to do this, but it’s either impossible to raise the nits limit from userspace, or I just don’t have the necessary expertise.

I’ll share some details that I found while reverse engineering my way through the macOS part that handles brightness.

Testing the system

Playing a HDR video

I first started by playing this HDR test video (open it in latest Chrome or Safari for best results): hdr-test-pattern.webm

Which resulted in a blinding white at 1600 nits:

This generated the following logs in Console.app:

1
2
3

WindowServer    Display 1 setting nits to 888.889
corebrightnessd SDR - perceptual ramp clocked: 227.095169 -> 252.268112 - 49.169426% (239.142059 Nits)
WindowServer    Display 1 commitBrightness sdr: 211.603, headroom: 4.20075, ambient: 4.3396, filtered ambient: 13.6333, limit: 1600

SDR cap in normal lighting

After setting the display brightness to max, I could see in the logs that SDR (Standard Dynamic Range) was being capped at 400 nits:

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12
13
14
15
16

WindowServer    Display 1 setting nits to 1600
WindowServer    Display 1 setting display headroom hint to 7.56866
WindowServer    Display 1 commitBrightness sdr: 216.548, headroom: 7.38865, ambient: 4.24854, filtered ambient: 13.3472, limit: 1600
corebrightnessd PCC: Set PCC: Factor:=1.0496 CabalFactor:=0.0033 time=2.000000 Lux:=13.6080 Nits:=229.1757 result=1 error=(null)
WindowServer    Display 1 commitBrightness sdr: 301.188, headroom: 5.3123, ambient: 4.24854, filtered ambient: 13.3472, limit: 1600
WindowServer    Display 1 setting nits to 1602.03
corebrightnessd levelPercentage 0.334298, level = 4.967383 (nits/pwm), lux = 15.000000
WindowServer    Display 1 commitBrightness sdr: 301.571, headroom: -1, ambient: 4.79275, filtered ambient: 15.0569, limit: -1
WindowServer    Display 1 setting display headroom hint to 5.27556
WindowServer    Display 1 commitBrightness sdr: 321.478, headroom: 4.97701, ambient: 4.79275, filtered ambient: 15.0569, limit: 1600
WindowServer    Display 1 commitBrightness sdr: 340.675, headroom: 4.69655, ambient: 4.79275, filtered ambient: 15.0569, limit: 1600
WindowServer    Display 1 commitBrightness sdr: 377.322, headroom: 4.24041, ambient: 4.79275, filtered ambient: 15.0569, limit: 1600
corebrightnessd PCC: Set PCC: Factor:=1.0340 CabalFactor:=0.0023 time=2.000000 Lux:=15.0569 Nits:=377.3223 result=1 error=(null)
WindowServer    Display 1 setting nits to 1600
WindowServer    Display 1 setting display headroom hint to 4
WindowServer    Display 1 commitBrightness sdr: 400, headroom: -1, ambient: 4.96577, filtered ambient: 15.6004, limit: -1