[alexjurkiewicz]

(2024)

My favourite part of these tools is the zany use of numbered file descriptors. `keypair` outputs the public key on fd 5 and secret key on fd 9. But signing reads the secret key on fd 8, while verification reads the public key on fd 4! Why aren't they the same?? I have to read the manpage every time.

[Retr0id]

I'm curious, what do you actually use it for?

I'd have otherwise guessed that this tool mainly exists just to test lib25519. Personally I'd only ever want a library, or some higher-level tool. A CLI tool that just does raw signing feels like a weird (and footgun-shaped) middle ground.

[tptacek]

This mostly exists to test lib25519 and ostensibly to build systems with shell scripts (though: few people would do that). It is a weird and footgun-shaped middle ground.

[XorNot]

It's why no one has succeeded in replacing GPG: you need a lot of systems to work in order to have an actual viable one, the ability to spit out signatures from keys is required but not sufficient.

[adastra22]

GPG is pervasive for the same reason git is pervasive: network effects. There are plenty of better alternatives.

[XorNot]

Such as? I need an alternative which supports commutative trust relationships of some sort which are revocable.

[adastra22]

You (knowingly?) picked the one counter example, lol. Web of trust is the one application of PGP/GPG for which there isn’t a product ready replacement tool to point towards. GPG is built around web of trust, but this is generally believed to have been a very, very bad idea and the source of innumerable security problems for nearly every application that has tried to make use of it. The GPG replacements I would point to are purpose-built for specific domains and eschew web of trust:

https://soatok.blog/2024/11/15/what-to-use-instead-of-pgp/

That said, you might find what you are looking for in the Rebooting Web of Trust project, and the various decentralized identity (DID) implementations that have come out of it:

https://www.weboftrust.info/

[XorNot]

No I picked the case I'm dealing with most commonly: which is establishing trust. X509 certs will also do this.

I have numerous criticisms of the GPG system but it's not a solution to just not implement any solution at all: I.e. I need revocation lists, I need intermediate keys, I need the ability to establish alternate chains of trust or promote a chain to trusted. Some of this is very hard to do with x509 even or not will supported.

[C4K3]

Keybase or any of the tools inspired by keybase (foks.pub etc)

[adastra22]

Isn’t keybase to GPG what github is to git?

[Fnoord]

> I'm curious, what do you actually use it for?

FTA:

> These tools allow lib25519 to be easily used from shell scripts.

I've never used ed25519-cli, but not having to use a library is nice for someone who isn't a programmer.

[tptacek]

The Venn diagram of "not a programmer" and "can safely use Ed25519" is two non-overlapping circles.

[PunchyHamster]

"this app needs me to generate a key and point to it in config" is plenty of overlap

[Retr0id]

If you just want a raw ed25519 private key then `head -c32 /dev/urandom` does the job. But usually you want a DER/PEM wrapper or similar, which the openssl cli tools handle nicely.

[kfreds]

I don't consider myself a programmer and I can use Ed25519 safely. I do however understand computing fairly well.

[Retr0id]

I consider myself a programmer and ed25519-understander, but the idea of using it directly within a shell script terrifies me.

[alexjurkiewicz]

Simply combine this tool with `openssl enc` and your shell script is as secure as any shell script could be

[loeg]

Someone writing shell scripts is a programmer, for better or worse.

[gnull]

That's such a user-hostile design decision. I can't fathom what justifies it (other than kinky taste).

Makes your commands unreadable without a manual, leaves a lot of room for errors that are quietly ignored. And forces you into using a shell that comes with its own set of gotchas, bash is not known to be a particularly good tool for security.

And to those who stay this adds flexibility: it doesn't. Those file descriptors are available under/dev/fd on linux, with named options you can do --pk /dev/fd/5. Or make a named pipe.

[minitech]

> Those file descriptors are available under/dev/fd on linux, with named options you can do --pk /dev/fd/5.

If you have a procfs mounted at /proc and the open syscall to use on it, sure (and even then, it’s wasteful and adds unnecessary failure paths). Even argument parsing is yet more code to audit.

I think the design is pretty good as-is.

[gnull]

It's 2025, dude. You can't be seriously telling me how difficult it is to parse arguments. It may be difficult in C, but then we're down another sick rabbit hole of justifying bad interface with bad language choice.

One open syscall in addition to dozens already made before your main function is started will have no observable effect whatsoever.

[minitech]

The context is what’s essentially a shell-accessible library for a minimal set of cryptographic primitives. It’s very reasonable to want it to be as lightweight, portable, and easy to audit as possible, and to want it to run in environments where (continuing on Linux for example) the open syscall to /dev/fd/n -> /proc/self/fd/n will not succeed for whatever reason, e.g. a restrictive sandbox.

Not involving argument parsing simplifies the interface regardless of how easy the implementation is, and the cost is just having to look up a digit in a manual that I certainly hope anyone doing raw ed25519 in shell is reading anyway.

[gnull]

Make a named pipe then. Shells have built-in primitives for that. I.e. <() and >() subshells in bash, or psub in fish. Or have an option to read either a file descriptor or a file.

I can't understand why you keep inflating the difficulty of simple commandline parsing, which the tool needs to do anyway — we shouldn't even be talking about it. Commandline parsing code is done once (and read once per audit) while a hostile user interface that bad commandline creates takes effort to use each time someone invokes the tool. If the tool has 1000 users, then bad interface's overhead has 1000× weight when we measure it against the overhead of implementing commandline parsing. This is preposterous.

> Not involving argument parsing simplifies the interface

From interface perspective, how is `5>secretkey` simpler than `--sk secretkey`? The latter is descriptive, searchable and allows bash completion. I'll type `ed25519-keypair`, hit tab and recall what the argument called.

You can't justify poorly made interface that is unusable without opening the manual side by side. Moreover, the simplest shell scripts that call this tool are unreadable (and thus unauditable) without the the manual.

  ed25519-keypair 5>secretkey 9>publickey

You see this line in a shell script. What does it do? Even before asking some deeper crypto-specific questions, you need to know what's written in "secretkey" and "publickey" files. You will end up spending your time (even a minute) and context-switch to check the descriptor numbers instead of doing something actually useful.

[minitech]

> which the tool needs to do anyway

It doesn’t. The tool has no command-line arguments.

Please learn how the various shell concepts you’re referencing (like <()) actually work and get back to me if you still need to after that.

In any case, I’m well aware of the readability benefit of named arguments, and was when I made the original comment. So as you can imagine, I maintain that it’s a more than reasonable tradeoff, and I’ve covered the reasons for that. If you have nothing (correct) to add beyond hammering on this point, save it.

[minitech]

(nicer reply to this)

Yes, I’m aware of the readability benefit of named arguments, and made the original comment with that awareness too.

> Make a named pipe then. Shells have built-in primitives for that. I.e. <() and >() subshells in bash,

That’s /proc/self/fd again. But okay, you can make a named pipe to trade the procfs mount and corresponding open-for-read permission requirement for a named pipe open-for-write permission requirement without receiving the other benefits I listed of just passing a FD directly.

> I can't understand why you keep inflating the difficulty of simple commandline parsing

Not only have I not “kept inflating” this, I barely brought up the related concept of it being unnecessary complexity from an implementation side (which it is).

> which the tool needs to do anyway

It doesn’t. The tool has no command-line arguments.

> From interface perspective, how is `5>secretkey` simpler than `--sk secretkey`? The latter is descriptive, searchable and allows bash completion. I'll type `ed25519-keypair`, hit tab and recall what the argument called.

Not introducing More Than One Way To Do It after all (“Or have an option to read either a file descriptor or a file”) here is a good start, but it’s hard to beat passing a file descriptor for simplicity. If the program operates on a stream, the simplest interface passes the program a stream. (This program actually operates on something even simpler than a stream – a byte string – but Unix-likes, and shells especially, are terrible at passing those. And an FD isn’t just a stream, but the point is it’s closer.) A file path is another degree or more removed from that, and it’s up to the program if/how it’ll open that file path, or even how it’ll derive a file path from the string (does `-` mean stdin to this tool? does it write multiple files with different suffixes? what permissions does it set if the file is a new file – will it overwrite an existing file? is this parameter an input or an output?).

Your attached arguments seem to be about convenience during interactive use, rather than the kind of simplicity I was referring to. (Bonus minor point: tab completion is not necessarily any different.)

> Moreover, the simplest shell scripts that call this tool are unreadable (and thus unauditable) without the the manual.

That might be a stretch. But more importantly, who’s trying to audit use of these tools without the manual? You can be more sure of the program’s interpretation of `--sk secretkey` (well, maybe rather `--secret-key=./secretkey`) than `9>` if you know it runs successfully, but for anything beyond that, you do need to know how the program is intended to work.

Finally, something I probably should have mentioned earlier: it’s very easy to wrap the existing implementation in a shell function to give it a named-parameter filepath-based interface if you want, but the reverse is impossible.

[PunchyHamster]

it being option can be nice if you don't want your keys touching disk and need to pass it over to other apps.

it being default is insanity

[jedahan]

I was wondering the same thing. My best guess is that is to guard against operator misuse. Like usb-a only plugging in one way. Anything that is secret will never accidentally print to stdout. String interpolation in bash with `—option $empty` might be safer than `8<$empty`. Have to explore more but yeah, this is a new pattern for me as well.

[yellowapple]

Another possible factor driving the decision to use numbered file descriptors: the logic to validate that a file exists (or can exist) at a given path, is readable/writable, etc. gets punted to the shell instead of being something the program itself has to worry about.

[gnull]

Those descriptors like 5 could be mapped to anything, including descriptor 1, stdout.

[chuckadams]

What a strange convention. I'm partial to minisign, which works on plain old files.

[tptacek]

This little CLI is not meaningfully an alternative for signify/minify. Here's a good piece on signify from its author (who also comments here):

https://www.openbsd.org/papers/bsdcan-signify.html

[alfiedotwtf]

I’m guessing it’s to support the test framework it’s built with?

[PunchyHamster]

support is fine. Being default is crazy

[pseudohadamard]

It's djb's web site so it's a djb design. With great genius comes great different thinking.