It's sort of hard to draw the line between what should be a skill and what should be an MCP server. Most things shouldn't be an MCP server. A local interface to a local code index stored in sqlite shouldn't be an MCP.
I've come to a conclusion on which I think is better now. And in a move that will surprise anyone that knows me, it's different from the most popular opinion.
A thing that bothers me some times about the specific subgenres of software engineering I've chosen is that everything is very transient. No code I write or decisions I've made or technical steering will survive me or even last a couple years after I leave a job. My footprint on the world from my career is basically nonexistent from any sort of longevity perspective.
I recently came across a popular franchise and was looking forward to trying it. After installing 150gb I booted it up and it asked me to login to their account. No steam linking or anything. Is hard to find games that don’t do this but it’s an instant delete for me.
I don’t want to out the person but a blog post came across my feed the other day in the genre of “people who criticize <thing> probably misunderstand it”. This is one of the worst genres.
cargo-audit answered this question for me, and I see that it's tracked in rustsec, just maybe a little less formally than I imagined or desired since there's no direct link between a yank and a rustsec issue.
I'm a little surprised that there doesn't seem to be any disciplined tracking on why crates were yanked. When something disappears I'd sort of like to know if it implies I was exposed to a supply chain attack or whether it's just down to the whims of the developer.
I asked it why it couldn’t just write the freepascal one correctly instead of going through this indirection and it said it’s because python is a more dynamic language.
This is one of the most interesting LLM-isms I've seen. I asked MiMo to create a zmachine in freepascal and as part of that it wrote a complete one in python to help it debug the freepascal version.
@romlib_ You want pi’s tree then I believe. The thing that seems to be the biggest annoyance with context to me is that if you selectively trim it to remove old or outdated information, you invalidate kv cache.
@rovarma I do suspect that “good” automatic compaction beats tree in the general case. But everyone has organically developed vastly different workflows some of which fall more naturally into tree than others.
@rovarma I have tried Pi but not really as a daily driver. My initial conclusion, which deserves some reconsideration when I have time, is that it is poor for the things I need to do.
A lot of context is garbage, and it's way easier for garbage to accumulate in a larger window. Plus that garbage costs money on every turn. Maybe $0.01/M, but you're still paying for very low value.