While this is a common view, recently I’ve begun to wonder if it may be secretly an antipattern. I’ve run into a number of cases over the years where additional fields don’t break parsing, or even necessarily the main functionality of a program, but result in subtle incorrect behavior in edge cases. Things like values that are actually distinct being treated as equal because the fields that differ are ignored. More recently, I’ve seen LLMs get confused because they hallucinated tool input fields that were ignored during the invocation of a tool.

I’m a little curious to try and build an API where parsing must be exact, and changes always result in a new version of the API. I don’t actually think it would be too difficult, but perhaps some extra tooling around downgrading responses and deprecating old versions may need to be built.

If you’re writing a client, I could see this being a problem.

If you’re writing a server, I believe the rule is that any once valid input must stay valid forever, so you just never delete fields. The main benefit of DisallowUnknownFields is that it makes it easier for clients to know when they’ve sent something wrong or useless.