mathstuf

Does it need to be proxied through the `claude` tool still? Or is it authenticating directly? From what I can see (for Amp at least), it's all being proxied through a `claude` binary. Perhaps I'm seeing outdated info from the announcement though?

> Neither of these updates explained what happened or what the fix was.

"Bug fixes and performance improvements "

Even worse than the "reformat" commit message that your bisect landed on.

Heh…I once was in a state-level coding event (it was a small portion of a larger competition) where half of the test was turning in code on a CD during the competition, with the written half during the event. My CD was deemed unusable for whatever reason (it had worked on XP and Fedora 6 or 7 at home) and didn't count towards my score. I still got second in the event. I declined to continue because I couldn't trust that the judges would be able judge my submission fairly and that with half of my score missing I still got second that I didn't need to prove anything else at the cost of more after-school practice hours and wrecking my perfect attendance record during my senior year to travel to nationals.

FD: I am a CMake developer.

Yes, config packages are better. But I think doing find_package everywhere is better. Assuming you install an SDK for others to use your project. If you're a "product", vendor away. The issue comes when you want to vendor X and Y and both vendor Z independently. Then you're stuck de-vendoring at least one and figuring out how to provide it yourself internally. IMO, better to just let Z make its own install tree and find it as a package from there.

One can write good Find modules, but there is some "taste" involved. I wish we had more good examples to use as templates.

> 1) There could probably be tax credits or deductions for SWEs who 'volunteer' their time to work on these projects.

Why exclusive to SWEs? They tend to be more time-restricted than financial-restricted (assuming the "SWE" comes from a job description). I'd be more interested in making sure that those with less well-paying jobs are able to access such benefits rather than stacking it onto those already (probably) making 6-figures.

Of course, the problems arise in the details. Define "volunteer": if $DAYJOB also uses it (in a way related to my role), is it actually, instead, wage theft? Also, quantifying the benefit is a sticky question. Is maintaining 10k emoji packages on NPM equivalent to volunteer work on libcurl? Could it ever be? Is it volunteer work if it ends up with a bug bounty payday? Google's fuzzing grant incentives?

I had an '89 Cherokee to 235k and sold it for ~60% of my purchse price after 6 years after garages only quoted insanity for the smallest things (Dad is a mechanic, but the commute there for repairs isn't feasible on the regular and apartment living is not conducive to the required garage/tools).

Dad has seen AMC I-6s go 400k before the transmission died and ended its run.

I'm sure someone can make a verification parrot too. Why pull yourself up by the bootstraps when you can just strap a rocket engine on your boots instead?

I've thought about making a "word search" and embedding the passphrase in it using a pattern (e.g., a subset of a Knight's tour, a space-filling curve overlay, or some other sampling algorithm).

I agree that the YAML can get out of hand. We use the `extends` keyword to put together jobs from pieces so that details can live in one place and the job bits and graph description can live in another. The way we've done our pipelines are very difficult to do with GHA as we build a DAG (with splits and forks) that are greatly aided by artifacts being integrated into GitLab-CI instead of separate piecemeal actions.

We also need custom runners anyways because macOS and Windows are important and getting those with graphical session access and/or CUDA hardware in the cloud is either $$$$ or severely limited. Even with our setup, we split the build and test phases so that CUDA hardware slots aren't wasted on running compilers. It also lets us test a single build under different environments easily.

So, yeah, I can see fighting with the feature spectrum, but you need to restrict yourself in most other cases with that kind of stuff too. But at least what we do is possible with GitLab-CI.

Why does it always feel so black-and-white? On our projects we do rebases all the time so that topic history is clean. But we merge into the target branch(es) (our merge bot supports merging to old release branches from a single MR). This gives us:

- clean integration branch histories (series of merge commits) - merge commits can contain metadata (topic-level descriptions, trailers for who reviewed, merge request links, test results, etc.) - you can be (pretty) sure that `git bisect --first-parent` will not run into any compilation problems (logical conflicts occur, but are fairly rare; use merge queues to be sure) - none of the "you merged main into your topic" "backwards merges" to deal with too

Merging and rebasing each have their pros and cons, so why not use the pros of each and mitigate a lot of the cons at the same time.

I have tried many times and I have been very unable to make any progress. In The Witness game, I ended up just guessing at the hardest whistle puzzles because remembering high/low order is very difficult at real-time speeds (mostly I can only compare "higher/lower than previous" so something like 2-1-3 is very indistinguishable from 2-1-2). Absolute positioning is even worse. The "end game" tanker puzzle is still unsolved because I don't want to cheat at that one but I have never been able to get the audio part down. I believe this is because my short-term memory "ingests" via audio (100% aphantasia, so no visuals to help out), so trying to "store" more notes erases anything longer than a note or two back (or I remember those notes and miss new notes coming in). Contrast this with Simon-like games (original, Brain Warp, Bop-It, etc.) where I was routinely able to get 12+ because I could "trace" a pattern in the lights/motions and "offload" the audio bits quickly enough to follow along and "compress" the sequence.

In this app, I ended up hunting for the first note and by the time I found it, I couldn't remember "how far" up/down the next note was (and by even 3 notes I often lost the direction from 2 to go). I /could/ figure out the key positions and go from the staff, but that goes through the FACE/EGBDF path of what notes the lines on the staff mean and is 100% non-aural for me. It's also extremely slow.

I also have a horrible time finding melodies and rhythms in music (kind of like Steve Martin in The Jerk, just more…physically coordinated at least). I end up "tapping" out every note and lyric syllable instead of being able to tease anything apart (separate instruments help, but then I lose "the rest"). Rarely was I able to effectively utilize those "song tapper" apps in the 00's to find anything I wanted.

I do remember playing the recorder in class in elementary school and at least "Horse With No Name" on a guitar in…some higher grade (7th? 8th?), but I believe those were very physical-oriented finger memory rather than anything to do with my ear helping out. New songs were basically from-scratch practices rather than being able to build on prior experience.

So while my ear might not be "tin", so to speak, my mind is not well-wired for ir.

Per-display DPI settings. No snooping on input without permission. Awareness of the lock screen (the compositor can know that the lock screen is active and provide alternate keybindings instead of having to configure the lock application as well). Locking is not blocked by context menus being open.

I ran XMonad for 15 years, but recently switched to river and am loving it.

Look into river. It has the window management and keybindings able to be offered by other tools (I have an idea to implement one using XMonad's layouts).

It also vastly improved battery on my Dell Pro laptop. 58% battery used in 7h45m (light compilation day, but no suspend).