Surely this is how all async communication with a lawyer works though? Or are gmail drafts privileged if you can make a case that it was going to be sent to your lawyer? Is a letter at your house privileged if you can make a case that it was going to be sent to your lawyer?

You need the LLM to be able to respond with tool use requests, and then your local harness to process them and respond to it. You can read how tool calling works with eg Claude API to get the idea: https://platform.claude.com/docs/en/agents-and-tools/tool-us...

Under the hood something like Claude Code is calling the API with tools registered, and then when it gets a tool use request it runs that locally, and then responds to the API with the result. That’s the loop that enables coding.

Integrating with an IDE specifically is really just a UI feature, rather than the core functionality.

Still needs to be published first, but looks like it automates all the annoying UI things you mentioned.

The problem would be new versions that fix security issues though, and because this is all open source as soon as you publish the fix everyone knows the vulnerability. You wouldn’t want everyone to stay on the insecure version with a basically public vulnerability for a week.

I wrongly thought that the verified provenance UI showed a package has a trusted publishing pipeline, but seems it’s orthogonal.

NPM really needs to move away from these secrets that can be stolen.

So I think they’re injecting this as a tip on using Copilot, that just happens to be their integration with Raycast.

I have no idea what their actual partnership with Raycast looks like, maybe this is part of what they offered them? But it’s not a traditional link to another product ad like it appears to be from Raycast being a link.

Now I'd probably just get Claude Code to write something locally, but there might be some value in a version of pyground that can use an LLM to write the Python code for you (and give it the data shape) but keeps the local execution setup the same.