This is a prototype of a copilot-based codefix that's the equivalent of Typescript's Infer From Usage codefix. Unlike that codefix, it uses only the body of the function (and any surrounding code) to infer types. However, it creates new interfaces with reasonable names and uses them where appropriate, which the old codefix cannot.

I have a lot of open questions about how this codefix should work, so I want to get it into Insiders, perhaps with an attached survey. I found that the copilot chat extension already logs telemetry about which button users press on the editorChat window, which might be good enough.

Some open questions:

• Will enough people see this experiment if it's behind the typescript.experimental.aiQuickFix setting?
• Is it better to show this as a separate codefix or a followup? As a refactor? Right now it's a separate codefix, because (1) the existing codefix can interfere with the LLM-based inference (2) the results are quite different, so it makes sense to present as different to the user.
• Could the codefix could be skipped if there are lots of types around already?
• Could the codefix prompt by modified to omit "Please generate new interfaces" if there lots of interfaces already declared?
• This branch is based on code by people on the vscode team, but I don't see any commits but mine. Are the other experiments merged to main already?

More work needed:

• The codefix is offered in JS files but always produces Typescript types, not JSDoc ones. I haven't found a prompt that works, at least.