Let's fix this in #2387.

Oh no, that was not what I meant to close!

Nice, this seems to fully wok now.

Added a second commit, making things a little more incremental. On my machine, reduces agent build times from the (already reduced) 1:22min to 0:38min (of which 22s are tests), genpolicy to 2:18min, and runtime-rs to 2:09min.

One caveat of this is that changing e.g. a runtime-rs patch still causes the next agent build to pay the full 1:22min build cost. That's because the intermediate source is different for each workspace member.

Yeah tbh I am not super happy with this step. It's a bit faster, but also less clean. I still have one more idea I'll add in a third commit, and then we'll have three choices to pick from.

With crane replaced by crate2nix instead:

• agent: 1:06min
• genpolicy: 1:51min
• runtime-rs: 2:45min

Which means that this solution is not only much uglier and harder to maintain, it's also not straightforward to compare performance 🙂

The upside of this solution is that changing e.g. agent code, then rebuilding genpolicy immediately uses the cached output, which is nice.

I have a personal preference but am curious to see what you think. Each commit is a slightly different approach that warrants separate testing and consideration.