Codex Is Live in Zed
zed.dev
zed.dev
Though I don’t understand what you mean by markdown autocomplete.
Their extensions API is still a bit lacking, but Jupyter support in the style of vscode should be possible with the current capabilities.
This would unblock people to write their own Jupyter integration for example, or whatever else they want. There's load of cool stuff like Argus https://github.com/cognitive-engineering-lab/argus that rely on creating buffers with custom UI, and Flowistry https://github.com/willcrichton/flowistry that rely on graying out some code, and I want this stuff on Zed too
we all want something.
Second, in the case of people making feature requests, it could be a net-societal-gain [2] if feature requesters made some kind of binding commitment. (See also the hold-up problem [3].) Perhaps a potential customer would commit to "if/when feature X gets added, I will commit to using the product for 2 hours." or "... I will spend $10 on the associated cloud services." (The question of what happens if the customer reneges also has to be agreed upon up front.)
[1]: https://en.wikipedia.org/wiki/Revealed_preference
[2]: known as social welfare (not to be confused with welfare programs -- this is the neoclassical economic framework after all!): https://en.wikipedia.org/wiki/Social_welfare_function
[3]: this paper discusses the hold-up problem in the context of vaccine investment and development: https://www.nber.org/system/files/working_papers/w28168/w281...
This is actually super interesting. Thank you so much for sharing.
So...where does that leave the Zed team? If existing LSPs aren't good enough, that's not a Zed problem: they're building an editor, not LSPs for your favorite language.
It was super slow (thought I think that applies to CLI Codex too), it wasn't outputting any text explaining what it was trying to achieve, and it started off down a path that made no sense. Claude Code in Zed has some rough edges but it's at least usable.
In terms of GUI agents, Cursor is still a lot nicer experience, IMO. Though I do still prefer just using Claude Code cli, personally.
This team ships so much that if they sold an LTS product, it means they'd support the release for 24 hours.
An example is that when I have a module like Namespace::SuperAbcModule in a file at namespace/super_abc_module.rb and I rename the file to namespace/super_module.rb, Cursor will immediately suggest to change the module name to `Namespace::SuperModule`, Zed won't.
Also Cursor will suggest updates to lines throughout a file whereas Zed sometimes doesn't even look ahead 1-2 lines.
Having Claude Code and Codex built into the sidebar is hardly better than having them running in a terminal. I wish they'd invested all this time and effort improving the inline suggestions.
Secondly, it was just one example that came to me from comparing this the other day. You could compile a long list of examples where Cursor gives better completions than Zed does.
I know that days of yak-shaving with LSP and emacs only gets me to a janky imitation of Visual Studio/XCode semantic search on my C++ work codebase. "Fuck it, let an LLM auto-complete based on vibes" has some appeal when you just get sick of trying to arm-wrestle clangd into ... whatever XCode or Visual Studio are doing to have functional semantic search across the project.
Although I have to say LLMs were a disaster at vibe-auto-completing in VSCode. So I mostly stick with semantic search in the IDE and editing in emacs like I always have.
It's human nature to start trusting AI suggestions just because they look good enough without actually checking them. That's going to lead to massive issues down the line the more it goes on.
Snippets are more useful.
If you're doing something repetitive, a vetted snippet does wonders. Learning how to make your own snippets with appropriate tab stops is a seriously underrated skill.
High competence in regex search-and-replace, multi-cursor, and snippets is an amazing combination.
But currently I sadly have to say the model's "help" is often a net negative.