The guy TJ De Vries works at Sourcegraph + contributes to neovim and he’s building sg.nvim - a plug-in to hook Sourcegraph AND their code assistant Cody to neovim LSP.
"Please parse <dict name> and re-key it by <field>", making sure to remove entires where <x> is Blue and converting <y> to Yellow."
and it'll dump out a 40 line (working!) parser in ~3 seconds I can then further customize. It's honestly remarkable as you can then interactively ask it to update/adjust "Can you make that a reusable function where I pass in X,Y, and Z?" "Can you convert that to a tuple comprehension?", "Can you unroll that loop and add inline comments explaining the regex?" are all futher-drilldowns I'd use and expect good responses to.
I do have ShellGPT setup to give me (similar generic) responses in the terminal; but I haven't found a good way to let it see my code yet to let it parse my data structures as fluidly.
We're primarily focusing on VSCode, IntelliJ and Neovim for Cody. Of course I'll be working on an Emacs version, but that's kinda best-effort for now.
As for the new crop of codegen models, they seem to be getting to parity with GPT/Claude/Bard-class models for code autocompletions, but not so much for other tasks.
We're working on incorporating OSS models, but I'd be surprised if they're ready for prime-time this year. I think next year they'll be huge.
Just my $0.02, take with a grain of salt. Shit moves fast.