I'm curious if this type of convention might cause confusion or hurt performance since I'm assuming models haven't been trained on functions being named this way? (And might also be easy to get wrong for humans, rather than say __ which does at least exist in code.)

Personally I've also run into model errors before about the existence of precise punctuation patterns in code or data, so it seems like this might confuse tokenization and/or remove semantic value?

I feel ultimately the guidance of that final projection, from MCP-tool-data to client-model, is wholly in control of the client and is complex; depending on use case, prompting and model capability. So isn't something we should proscribe but rather various client share their advice?

So isn't something we should proscribe but rather various client share their advice?

Yes, the newer versions of this documentation update offer a few suggestions rather than prescribing a specific approach, for the reasons you mentioned.

@olaservo At least with Claude 3.7+, the ___ works fine - Amazon Q CLI does this and I haven't had any issues with it to date, but some of this probably has to do with what the tool prompt looks like as well.

Added another URI-based suggestion because it keeps coming up in the committee - I have more concerns about that in this context, but having more options doesn't hurt (mostly).

@nickcoai Just made the phrasing even less prescriptive, in case it still sounded that way.

Looks good to me!