Right now there's a huge friction with IDEs, that if I have it open while claude is working on something, then any caches it has get stale very quickly, and it churns processing trying to keep up with edits.

Any leverage that claude could benefit from LSPs or language/AST MCPs is undone by that tooling not being able to keep up or stay fresh, so the IDE just "runs hot" for no purpose.

Which means I'm incentivised to close my IDE while claude is working, and now I've got used to not looking at the code.

Like you, I really want a system that works better. I want claude to search and edit through better tooling, it would be faster, more accurate, and should in theory be able to leverage compilers to apply edits better and more accurately.

I don't want claude wasting an awful lot of time grepping for things that a compiler cache can answer, but if that compiler cache can't keep up then it's actively harmful.

I'm sure there must be a better way, there must be a way to have smarter more change-aware cache invalidation, and smarter more cache-aware claude tool use, but so far none has happened.

Long term I'm sure the pairing will emerge, but unfortunately I don't think that's going to come from the IDE vendor side, it'll have to come out of some harness research lab, where Anthropic or a competitor finds a way to reliably leverage this stuff for a speed and accuracy boost.

It's completely bonkers to me that an LLM can almost "natively" read profiler output, but doing so involves vast amounts of text and context ingestion, when there are real tools sitting there that it ignores in preference for either reading vast amounts or iffy python one-shot scripts to parse stuff.