You cannot modify unexported fields from separate packages. This applies to composite literals too. The error seems clear enough to me.

I'm not referring to the error description. I'm suggesting literal assignments should just be banned, or when an unexported field exists, initialised to zero value when attempting to do literal assignment (since you can't access the unexported field anyway).

Please reopen issue.

I don't understand what you mean, but I'm going to reopen this for now.

Also, please note that language changes should be proposals, as they need careful consideration. You should try to make your proposal as detailed and clear as possible.

In your example, the solution would be to use a keyed literal.

Note that using an unkeyed composite literal of a struct type defined in another package is already explicitly recommended against, to preserve compatibility when such types add fields - the Go 1 compatibility promise specifically excludes such unkeyed literals, for example. There's even a proposal to disallow them altogether, and they're already a warning in vet.

@pjebs
If you are the maintainer of the library package, then you can put a zero-size non-exported field in your struct type to avoid using non-keyed literals in user code.

@mvdan Minor point: please don't add a NeedsDecision label to Go 2 issues. We're using that as a marker to be applied only after review. Thanks.

@pjebs Like @mvdan , I don't understand what you are proposing. You are describing how the language works today, and you say "perhaps it should just be disallowed," but I don't know what you mean. What should be disallowed? Thanks.

1. I was suggesting that literal assignments cause more trouble than they are worth and should be disallowed.
   OR
2. Literal assignments are allowed but if a field is not accessible (i.e. external pkg + unexported), then the literal assignment would operate as if the unexported fields are non-existent, but upon creation of object, they are zero-valued.

I showed the 2 different error messages to highlight that the status quo seemed inconsistent, since the first error message says too few values which explains the error perfectly but seems to also imply that the situation can be remedied by adding extra values. ....but adding extra values produces implicit assignment of unexported field, which also explains the situation perfectly. You simply can't win!

Hence the proposal. Remove or rectify.

The solution is to use a keyed literal. The vet -composites check can help with this.

We aren't going to require people to use keyed literals even for structs defined in the same package. And we don't want to have different rules for literals defined in the same package and literals defined in a different package. We could perhaps improve the compiler error messages for these cases. I'll repurpose this issue for that.