Skip to content
Notifications
Clear all

Hot take: The Claw family's aggressive model prefetching is the root of most perf issues

1 Posts
1 Users
0 Reactions
1 Views
(@emilyt)
Estimable Member
Joined: 1 week ago
Posts: 98
Topic starter   [#16772]

Alright, I'm diving into this with a mix of love and frustration. 😅 I've been a huge fan of The Claw's ecosystem—the autocompletion is *magic* when it's working. But after months of wrestling with my editor feeling sluggish and watching my memory usage climb, I've started pointing the finger at one thing: the family's aggressive model prefetching.

I'm talking about `clangd`, `rust-analyzer`, `pyright`, and others that follow a similar "load everything up front" philosophy. On my M2 MacBook (Ventura), with Neovim and a typical web dev project (Node, Python, a bit of Rust), I'll see my RAM spike on startup as all the language servers pre-index the entire workspace. It's great for latency later, but the initial hit is brutal. If I have a few tabs open or my team's project management tools running (shoutout to Linear and Notion!), the whole system chugs.

Here's my typical plugin list that seems to clash:
- nvim-lspconfig (with the usual Claw suspects)
- nvim-cmp (for completion)
- telescope.nvim
- A few lightweight utilities for git and navigation

The pattern I see:
* Editor startup slows from ~200ms to over a second.
* Memory stays elevated even when I'm not actively using that language.
* The worst conflicts happen when another plugin (like a database client or a file watcher) also tries to claim resources during startup.

I'm starting to think the "smart" prefetch is too greedy for most real-world, multi-language projects. Has anyone found a reliable way to throttle this behavior? Or are we better off switching to language servers that are more lazy-load friendly?

Happy benchmarking!


Always testing.


   
Quote