Something like https://graphite.com/ to create stacked PRs that are reviewable probably would have helped. Can be replicated with local LLMs or remote AI providers with locally configured agentic workflows. Never used graphite personally, but I’ve seen some open source maintainers use it to split up large PRs.
With these sorts of tasks models really seem to suffer from not knowing what packages or conventions have been deprecated. This is really obvious with an immature ecosystem like nix.
This is where custom setups will start to shine.
https://github.com/upstash/context7 - Pull version specific package documentation.
https://github.com/utensils/mcp-nixos - Similar to above but for nix (including version specific queries) with more sources.
https://github.com/modelcontextprotocol/servers/tree/main/src/sequentialthinking - Break down problems into multiple steps instead of trying to solve it all at once. Helps isolate important information per step so “the bigger picture” of the entire prompt doesn’t pollute the results. Sort of simulates reasoning. Instead of finding the best match for all keywords, it breaks the queries down to find the best matches per step and then assembles the final response.
https://github.com/CaviraOSS/OpenMemory - Long conversations tend to suffer as the working memory (context) fills up so it compresses and details are lost. With this (and many other similar tools) you can have it remember and recall things with or without a human in the loop to validate what’s stored. Great for complex planning or recalling of details. I essentially have a loop setup with global instructions to periodically emit reinforced codified instructions to a file (e.g., AGENTS.md) with human review. Combined with sequential thinking it will identify contradictions and prompt me to resolve any ambiguity.
The quality of the output is like going from 80% to damn near 100% as your knowledge base grows from external memory and codified instructions in files. I’m still lazy sometimes and will use something like Kagi assistant for a quick question or web search, but they have a pretty good baseline setup with sequential thinking in their online tooling.
Like, from Yellow Dog Linux? Was my first Linux distro as a kid (grew up with an Apple computer). Red Hat based so checks out. TIL.
It’s really not that different from a traditional web search under the hood. It’s basically a giant index and my input navigates the results based on probability of relevance. It’s not “thinking” about me or deciding what I should see. When I say a good assistant setup, I mean I don’t use Gemini or ChatGPT or any of the prepackaged stuff that tries to build a profile on you. I run my own setup, pick my own models, and control what context they get. If you check my post history I’m heavily privacy conscious, I’m not handing that over to Google or OpenAI.
The summary helps me evaluate if my input was good and the results are actually relevant to what I’m after without wading through 20 minutes of SEO garbage to get there. For me it’s like getting the quality results you used to get before search got enshitified. It actually surfaces stuff that doesn’t even show up on the front page of a traditional search anymore.
I’m in software development and land on both sides of this argument.
Having to review or maintain AI slop is infuriating.
That said, it has replaced traditional web searching for me. A good assistant setup can run multiple web searches for me, distill the useful info cutting through the blog spam and ads, run follow up searches for additional info if needed, and summarize the results in seconds with references if I want to validate its output.
There was a post a couple days ago about it solving a hard math problem with guidance from a mathematician. Sparked a discussion about AI being a powerful tool in the right hands.
neovim user (inside zellij) and same. More of a full blown IDE than an editor.
Also for the keybind memory impaired like myself:
Totally agree with your overall point.
That said, I have to come to the defense of my terminal UI (TUI) comrades with some anecdotal experience.
I’ve got all the same tools in Neovim as my VSCode/Cursor colleagues, with a deeper understanding of how it all works under the hood.
They have no idea what an LSP is. They just know the marketing buzzword “IntelliSense.” As we build out our AI toolchains, it doesn’t even occur to them that an agent can talk to an LSP to improve code generation because all they know are VSCode extensions. I had to pick and evaluate my MCP servers from day one as opposed to just accepting the defaults, and the quality of my results shows it. The same can be done in GUI editors, but since you’re never forced to configure these things yourself, the exposure is just lower. I’ve had to run numerous trainings explaining that MCPs are traditionally meant to be run locally, because folks haven’t built the mental model that comes with wiring it all up yourself.
Again, totally agree with your overall point. This is more of a PSA for any aspiring engineers: TUIs are still alive and well.
There is a praise/validation kink for that.
Firefox Nightly + arkenfox userjs + uBlock Origin + Bitwarden as my daily driver.
Been a couple years since I checked up on arkenfox still being good. I get flagged as a bot all the time and constantly get popups about WebGL (GPU fingerprinting) so I assume its working as intended for my threat model.
Tails when I really care.
Mullvad VPN as my regular VPN with ProtonVPN for torrents.
GrapheneOS / NixOS as my OS.
Proton Visionary for most cloud services except passwords and I don’t really use Proton Drive. I do use ProtonPass for unique emails to every provider.
Kagi for searches / AI.
Etesync for contacts because Proton didn’t sync with the OS last I checked.
Backblaze B2 for cloud storage with my own encryption via rclone (Round Sync on GrapheneOS)
Keypass for a few things like my XMR wallets and master passwords I don’t even trust in Bitwarden.
https://jmp.chat/ for my mobile provider.
Pihole with encrypted DNS to Quad9.
https://onlykey.io/ for the second half of my sensitive passwords (Bitwarden, LUKS, Keypass, OS login). First half memorized.
Its a lot. I burned myself out a couple years ago keeping up with optimizing privacy and this setup has served me well for 2 years without really changing anything. The cloud services are grey areas in terms of privacy but the few ads that leak through uBlock have zero relevance to anything about me.
deleted by creator
So, I don’t use OpenWRT (for main router), but generally in each vlan you will need:
When cutting I don’t factor in burned calories as extra calories I can eat.
When doing heavy lifting trying to build muscle and bulk up I’ll add some extra calories. I don’t do this often anymore.
My normal is more of a leangains style where on high activity days I’ll eat at 10% deficit up to maintenance, and on low activity days I’ll eat a 20% deficit.
If you’re doing a recomposition (burn fat and build muscle) then yeah, you are losing fat but gaining muscle so the scale can be misleading. Body fat % or pant size become better indicators. You may see what is referred to as a “whoosh” where every few weeks you seem to drop a few Kg over night (water weight).
When I am doing a recomp I tend to do weigh-ins every 2-4 weeks. When cutting I’ll do it weekly.
I’m sure that I’m in a caloric deficit. Eating more raw fruits and fibrous food like beans, more fish and less meat and fats.
Have you done any tracking? It is a pain to do all the time but worth maybe doing it for a week just to be sure. I struggle with knowing when I am full and can easily put myself over on something high calorie like nuts.
If no progress for a few months, lab testing may be worth it. Thyroid issues can impact weight gain and loss.
Finally there are medications like GLP-1 (expensive in some areas) that mess with insulin production and the hunger signals in your brain that have been known to be effective.
BSP Tree (with custom nodes).
With a vanilla BSP-tree you can accomplish your diagram. Simply reordering your splits can do it by making the footer and main content areas first. Better approach is to support splits on non-leaf nodes. In the example below just split the root where all its children go to the new top node and a new bottom bar leaf node is created.
Root (split: vertical, ratio: 0.6)
├── Left child (Window A)
└── Right child (split: horizontal, ratio: 0.5)
├── Top child (Window B)
└── Bottom child (Window C)
To access neighbors you’ll need your nodes to track their parents (double linked). Then you can traverse until all edges are found. Worst case its O(height+num neighbors that exist) if I am remembering.
Depending on how efficient you want it to be, there are speed ups. It has been awhile, but I do remember keeping the tree as balanced as possible makes the search around log(n). Each split node keeping an index of all children in its sub-tree also reduces how much traversing is needed when you need all children after a split.
Can get a little complicated but it is doable. That said, how many splits will a TUI have? This may be preemptive to do.
Custom nodes is where you support some patterns that could use further optimizations. Tables that will always be a grid. Tab bars that are a 1xn grid could be a further specialized node.
This is all about layout. Fixed/Dynamic width/height windows, padding and margins, borders, are all render processing and don’t effect the layout (unless you want reactivity). By that I mean you have windows that will split differently when the viewport is portrait or landscape and it dynamically adjusts to the window size. Sometimes with different “steps” like a square viewport may be different from both portrait or landscape or 4:3 could be treated different from 16:9.
TUIs are not my day job but I’ve made a few in my day. Above are just my opinions from experiences. There is no “right” answer but hopefully some of this helps your journey.
TypeScript is my day job and using a custom JSX Factory makes it pretty easy to define HTML-like interfaces for devs that can support mixing layout, render attributes, content, and app logic.
Explicit BSP splits:
<Split type="vertical" ratio={0.6}>
<WidgetA />
<Split type="horizontal" ratio={0.5}>
<WidgetB />
<WidgetC />
</Split>
</Split>
Custom nodes:
<Container>
<TabBar>
<Tab>Tab 1</Tab>
<Tab>Tab 2</Tab>
</TabBar>
<StatusBar />
</Container>
Not sure your stack but throwing it out there as something I’ve used successfully.
https://www.zimaspace.com/ - Low end Zimaboard is $90.
https://www.hardkernel.com/shop/odroid-h4/ - Cheapest seems to be H4 at $99.
These are from the last time I researched as I was in the same boat.
I had specific PCI requirements, so there may be cheaper.
Soma. Its like sleepytime ecstasy that lasts 20-30min.
high levels of serotonin in the body - agitation, hallucinations, fever, sweating, shivering, fast heart rate, muscle stiffness, twitching, loss of coordination, nausea, vomiting, diarrhea.
100% the same boat when I started! I actually deleted a paragraph that said when I was young I wanted to learn how to do things the right way. You eventually learn there is no right way us devs will agree upon. You won’t have opinions on the better way until you’ve done it an objectively bad way.
Reading code helps with this. I’m always looking at my library’s and framework’s implementations. Half my dependencies I swear I patch to make them “better” (subjectively).
Items are in a hash table using color/material type/shape as the hashing method optimized for human pattern recognition providing O(1) access. The table is smaller than the number of items causing some collisions. Those items are in a randomly sorted vector. Average case is still around O(1) with an O(n) worst case.