I think I like adding APIs to x/crypto/nacl/box better, if nothing else because the real box API is awkward, and adding the much more useful anonymous options next to it would help users who are confused about how to use box.

Sounds good to me. In that case, I propose adding two new functions to the box package

func SealAnonymous(out, message []byte, peersPublicKey *[32]byte) []byte {}
func OpenAnonymous(out, box []byte, privateKey *[32]byte) ([]byte, bool) {}

The nonce generation will be calling into blake2b functions that may return errors. In practice, errors are only returned from those functions if a bad hash size is specified. Since we know the hash size will be valid, I propose panicking if an error is encountered, rather than surfacing them to the caller.

This actually might be a bit more complicated. The ephemeral keypair needs to be generated randomly. So, SealAnonymous should also probably take a rand io.Reader argument. Since reading this can fail, I suppose SealAnonymous should also return an error? It's rather unfortunate to add this complexity to what should be a simple API.

Alas, there's a lot of unfortunate things in that API: bool instead of error, array pointers, append instead of inlined allocation... but I guess we should be consistent, so yeah.

func SealAnonymous(out, message []byte, recipient *[32]byte, rand io.Reader) ([]byte, error)
func OpenAnonymous(out, box []byte, key *[32]byte) (message []byte, ok bool)

I am tempted by the more idiomatic SealWithoutSender name, but I guess we should be consistent with libsodium's naming.

A nil rand should default to crypto/rand.Reader.

Sounds good. Do I need to write a design doc for this? Do I need to wait for an official approval here, or can I go ahead with a PR?

Feel free to open a CL, we should leave this issue open for a week or so before marking it accepted to give time to the community to comment.

For what it's worth, please let the proposal review meetings add Proposal-FinalCommentPeriod so that anything in final comment period is listed in the minutes (golang.org/s/proposal-minutes). We'll list it this week. Thanks.

Based on the discussion above this seems like a likely accept.
Will add to the proposal minutes and leave open for another week for comments.

Something I didn't realize is that the OpenAnonymous API requires the local public key to generate the nonce. It can be generated from the private key, but that's quite the performance overhead. It's also too late to make the private key carry both.

https://go-review.googlesource.com/c/crypto/+/205241/1/nacl/box/box.go#156

libsodium takes the public key from the caller.

int crypto_box_seal_open(unsigned char *m, const unsigned char *c,
                         unsigned long long clen,
                         const unsigned char *pk, const unsigned char *sk);

I kind of hate that they are two *[32]byte that are easy to swap, but the API cleanliness probably doesn't justify an extra scalar multiplication at every decryption, basically doubling the fixed overhead.

func OpenAnonymous(out, box []byte, publicKey, privateKey *[32]byte) (message []byte, ok bool)

The only alternative I see is to fork the API entirely, with a new GenerateKey that returns a bundled private+public key, at that point maybe in a new package, like golang.org/x/crypto/box/anonymous? It'd be a good opportunity to use slices rather than byte arrays, return slices instead of appending, and return error values, too.

Opinions?

(@rsc, yep, will do that going forward. I think I marked this one before we discussed it.)

My vote would be to take the public key as an argument. We could even make it optional and calculate the public key if it isn't provided. It seems less than ideal to have a separate package with a new key type that isn't interoperable with the existing keys.

Change https://golang.org/cl/205241 mentions this issue: nacl/box: support anonymous seal/open

Alright, let's go for the libsodium-like API.

func SealAnonymous(out, message []byte, recipient *[32]byte, rand io.Reader) ([]byte, error)
func OpenAnonymous(out, box []byte, publicKey, privateKey *[32]byte) (message []byte, ok bool)

No change in consensus, so accepted.