keyfunc/README.md

# keyfunc

`keyfunc` is a Go command-line tool by [@sneak](https://sneak.berlin) — its
license is not yet chosen
([#14](https://git.eeqj.de/sneak/keyfunc/issues/14)) — that turns a BIP-39
mnemonic into SSH keys, age identities and child mnemonics, each of which can be
recreated from that mnemonic at any time. The same mnemonic, key type and index
always give the same key.

It uses the BIP-85 entropy deriver from `git.eeqj.de/sneak/secret/pkg/bip85` and
takes the same steps as that repository's `agehd` package.

Commands are grouped by what is derived: `keyfunc ssh ...` for ed25519 SSH keys,
`keyfunc age ...` for age identities and for encrypting and decrypting with
them, and `keyfunc mnemonic ...` for child mnemonics derived from the main one.

## Getting Started

Build from a clone and run the binary:

```
git clone git@git.eeqj.de:sneak/keyfunc.git
cd keyfunc
make build
./keyfunc --version
```

`make build` produces `./keyfunc`. Every deriving command needs a mnemonic; see
[Giving it the mnemonic](#giving-it-the-mnemonic) for where it is read from, then
for example:

```
./keyfunc ssh pub -n 0 --mnemonic-command 'secret get foo'
```

## Rationale

A key you can derive again never has to be backed up. One mnemonic, kept safe
once, stands behind every key this tool produces: lose a laptop and the SSH key,
the age identity and any child mnemonic on it come back from the mnemonic alone,
at the same index, byte for byte. Nothing else has to be written down, copied
between machines, or stored in a secret manager, because it can always be
derived again.

## Design

The entry point is a thin `cmd/keyfunc/main.go` (what `make build` builds) that
calls into `internal/`. The packages there are:

- `internal/derive` turns a mnemonic into the 32 bytes a key is made from: it
  walks BIP-39 seed, BIP-32 master key and BIP-85 entropy, and holds the shared
  constants (the byte count and the largest key index).
- `internal/mnemonic` finds the mnemonic to work from — a command, an
  environment variable, or a terminal prompt — and refuses one that fails the
  BIP-39 checksum.
- `internal/sshkey` turns the derived bytes into an ed25519 SSH key
  (`sshkey.go`) and serves that key from an in-process SSH agent on a private
  unix socket, keeping it out of any file (`agent.go`).
- `internal/agekey` turns the derived bytes into an age identity and encrypts
  and decrypts with it.
- `internal/childmnemonic` derives a child mnemonic from the main one using
  BIP-85's own mnemonic application.
- `internal/cli` builds the cobra command tree and runs it. Under it,
  `cli/options` holds the flags every command shares, and `cli/ssh`, `cli/age`
  and `cli/mnemonic` are the command groups.

### Adding a key type

Adding a key type is one package under `internal/` that turns the 32 derived
bytes into that type's key, plus one cobra subcommand under `internal/cli/` that
groups its commands.

## Derivation

The mnemonic is turned into a key like this:

1. mnemonic -> BIP-39 seed (empty passphrase);
2. seed -> BIP-32 master key;
3. master key -> 64 bytes of BIP-85 entropy at the path below;
4. entropy -> BIP-85 DRNG (SHAKE256); read 32 bytes;
5. those 32 bytes become the key in the way the key type needs.

The path is:

```
m/83696968'/<app>'/<n>'
```

- `83696968` is the fixed BIP-85 purpose.
- `app` is the application number of the key type. There is no vendor id: the
  path is meant as a standard any implementation can follow, not something tied
  to one tool. Each key type's number is spelled the way BIP-85 spells its own
  RSA application (`828365` is the ASCII codes of `R`, `S`, `A` written out):
  SSH is `838372` (`S` `S` `H`), age is `657169` (`A` `G` `E`).
- `n` is the key index: flag `--index` / `-n`, default `0`.

## Giving it the mnemonic

The mnemonic itself is never a command-line argument. It is looked for in this
order; the first one found wins:

1. `--mnemonic-command <command>`: a shell command, run with `sh -c`, whose
   standard output is the mnemonic. Example: `--mnemonic-command 'secret get
   foo'`. Whitespace around the output is dropped. If the command exits with a
   non-zero status, the tool prints its standard error and exits with status 1.
2. Environment variable `KEYFUNC_MNEMONIC_COMMAND`: the same, as a shell
   command held in the environment.
3. Environment variable `KEYFUNC_MNEMONIC`: the mnemonic itself.
4. A prompt on the terminal with echo turned off.

If none of these is available and standard input is not a terminal, the tool
refuses and exits with status 1. A mnemonic that fails the BIP-39 checksum is
refused with a message saying so.

`KEYFUNC_MNEMONIC` and `KEYFUNC_MNEMONIC_COMMAND` are removed from the
environment before the system `ssh` (`keyfunc ssh to`) and `sftp`
(`keyfunc ssh install`) are started, so the mnemonic is never handed on to
them.

Every command takes `--index` / `-n` and `--mnemonic-command`, and has `--help`.
`keyfunc --version` prints the version. `make build` stamps it; a binary
installed with `go install` reports the module version instead.

## SSH keys: `keyfunc ssh`

Only ed25519 keys are produced. The application number is `838372`, so the
path is `m/83696968'/838372'/<n>'`. The 32 bytes from step 4 are the ed25519
seed.

Test vector, mnemonic
`abandon abandon abandon abandon abandon abandon abandon abandon abandon abandon abandon about`:

```
index 0: ssh-ed25519 AAAAC3NzaC1lZDI1NTE5AAAAIJZOtOczrc/7CQytcuFwt7s4r8KjkZWkwjLZWBaFKD+7
index 1: ssh-ed25519 AAAAC3NzaC1lZDI1NTE5AAAAIOEWY8+/gmHYVC4u0Y0I4FKs+eVUulTPHfk9VtXw1tMF
```

### `keyfunc ssh pub`

Prints one `authorized_keys` line to standard output:

```
ssh-ed25519 <key> <comment>
```

The comment defaults to `keyfunc/ssh/<n>`; change it with `--comment`.

### `keyfunc ssh priv`

Prints the unencrypted private key in OpenSSH format (the
`-----BEGIN OPENSSH PRIVATE KEY-----` block that `ssh` reads) to standard output
and nothing else, so it can be redirected into a file. The key's comment is the
same as for `pub`.

### `keyfunc ssh install <[user@]host> [-- sftp options...]`

Adds the `pub` line to `~/.ssh/authorized_keys` on the host. No command is run
on the host: the file is fetched, changed here, and written back with the
system `sftp` client in batch mode.

The first connection lists `~/.ssh` and then fetches
`~/.ssh/authorized_keys` from it. The file reads as empty in two cases only:
`sftp` reported `~/.ssh` itself as not being there, or the listing came up and
the file was not in it. Any other outcome of that connection fails the run — a
`~/.ssh` that is there but cannot be entered, an `authorized_keys` that is there
but cannot be read, or a connection that did not come up — and the tool prints
what `sftp` said and exits with status 1 without writing anything, rather than
put a file back holding the new key alone. The listing is what tells a missing
directory from one shut to the user, which `sftp` reports on a fetch the same
way; the wording of a missing file elsewhere does not count either, since `ssh`
writes `No such file or directory` about an `-i` it cannot find on a session
that then authenticates through the agent. If an identical line is already in
the file, the tool prints `already present` and connects no further. Otherwise
the line is added (after a newline, if the file did not end with one) and a
second connection:

- makes `~/.ssh` and sets it to mode `0700`, but only when the first connection
  found none; a `~/.ssh` that was already there keeps the mode it had;
- uploads the new file as `~/.ssh/authorized_keys.keyfunc-<random>` and sets it
  to mode `0600`;
- renames that file over `~/.ssh/authorized_keys`.

The tool then prints `added`. So a run that adds a line connects twice. The
rename is the step that either happens or does not: the file on the host is
never half-written. `sftp` does it in one step against servers that offer
OpenSSH's POSIX rename extension, as OpenSSH's own server does; a server
without it may refuse to rename onto a file that is already there.

If a step fails, the tool prints what `sftp` said, removes nothing, and exits
with status 1. It names the uploaded file only when the step that failed was
the upload or one after it, which is where a file of that name can be on the
host; a failure before the upload names none. Everything `sftp`
writes goes to standard error, so the tool's own standard output is only
`added` or `already present`.

Anything after `--` is passed to `sftp` unchanged, which is where the port goes
(`-P 2222`, not `-p`). How the connection authenticates is up to the user's
normal `ssh` setup, except that batch mode does not prompt: a key or an agent
has to do it, not a typed password.

### `keyfunc ssh to <host> [ssh arguments...]`

Derives the key, serves it from an SSH agent that runs inside the tool on a unix
socket in a new private `0700` temporary directory, then runs the system `ssh`
with `-o IdentityAgent=<that socket>` followed by the host and all remaining
arguments unchanged. The tool exits with `ssh`'s exit status and removes the
socket and directory on the way out. The private key is never written to disk. A
SIGINT, SIGTERM or SIGHUP ends `ssh` and still removes the socket and directory,
and the tool then exits with status 1 unless `ssh` reported one of its own.

## age identities: `keyfunc age`

The application number is `657169`, path `m/83696968'/657169'/<n>'`. The 32
bytes from step 4 are clamped as X25519 requires and become an age identity,
the same steps `sneak/secret` takes in its `agehd` package. `secret` derives at
a vendor-specific path today; for its keys to equal this tool's it moves to
this path, which is a change in `secret`, not here.

Test vectors, mnemonic
`abandon abandon abandon abandon abandon abandon abandon abandon abandon abandon abandon about`:

```
recipient index 0: age1xwdy9y6ckyfsgjc8k02e9uhsf3fmjy0ufysewlj68kmx5n67e3nsg2mftq
recipient index 1: age1pmm92sxaf5mazjwvjph7dx2zq9r5p8l3rarfgqm7hmakqhvgyy4q5p3w7j
identity index 0: AGE-SECRET-KEY-19QKK2P38598XLXMQFFU3P7J9PLDD7527T70JDHGDJ7AMNF3XT44S00JFU5
```

### `keyfunc age pub`

Prints the recipient, the `age1...` public key, on one line.

### `keyfunc age priv`

Prints the identity, the `AGE-SECRET-KEY-1...` line, and nothing else.

### `keyfunc age encrypt [-n N] [--to <recipient>...] [-o <file>] [<file>]`

Encrypts the file (or standard input) with age. The recipients are the derived
identity's own recipient, plus any given with `--to`, so the same mnemonic can
always decrypt what it encrypted. Output goes to `-o` or standard output;
`--armor` writes the text form. Nothing is written except the output.

### `keyfunc age decrypt [-n N] [-o <file>] [<file>]`

Decrypts the file (or standard input) with the derived identity. Output goes to
`-o` or standard output. If the identity is not one of the recipients, the tool
says so and exits with status 1.

## Derived mnemonics: `keyfunc mnemonic`

### `keyfunc mnemonic [-n N] [--words 12|18|24]`

Prints a child mnemonic derived from the main one, using BIP-85's own mnemonic
application (number `39`, English, path
`m/83696968'/39'/0'/<words>'/<n>'`, entropy taken as the specification says,
not through step 4). Default 12 words. A child mnemonic is a full mnemonic in
its own right: it can seed another `keyfunc`, another wallet, or `secret`, and
it never has to be written down, since it can be derived again.

Test vector: the child-mnemonic step is checked against BIP-85's own published
vectors, which derive from the specification's master key
`xprv9s21ZrQH143K2LBWUUQRFXhucrQqBpKdRRxNVq2zBqsx8HVqFk2uYo8kmbaLLHRdqtQpUm98uKfu3vca1LqdGhUtyoFnCNkfmXRyPXLjbKb`.
At key index 0 the 12-word English child mnemonic is:

```
girl mad pet galaxy egg matter matrix prison refuse sense ordinary nose
```

## Errors

Errors go to standard error and the exit status is 1, except for `ssh to`,
which passes through `ssh`'s own exit status.

## Entrypoints

The repo adheres to the
[Scripts to Rule Them All](https://github.com/github/scripts-to-rule-them-all)
standard: most Makefile targets are thin shims over an executable in
`script/` (`build` and `clean` are the exceptions).

- `script/bootstrap` installs everything needed to build and develop (git, make,
  Go), idempotently, from nix, apt, brew or apk; it does not install the linter,
  which only runs inside Docker.
- `script/setup` prepares a fresh clone: it runs `bootstrap`, then installs the
  git pre-commit hook.
- `script/projectname` prints the project name; other scripts call it so they
  stay identical across repos.
- `script/test` runs `go vet` and then the test suite, rerunning verbosely if a
  test fails.
- `script/lint` runs the linter inside the image built from `Dockerfile.lint`
  (which pins the linter by hash), so a complaint fails the build and leaves no
  container behind.
- `script/fmt` formats the Go source in place.
- `script/fmt-check` checks that formatting without writing, failing if anything
  is unformatted.
- `script/check` runs `test`, `lint` and `fmt-check` and changes no files.
- `script/docker` builds the Docker image tagged with the project name.
- `script/cibuild` is the CI build the Gitea workflow calls: it runs the linter,
  then `docker build`.
- `script/precommit` is what the git pre-commit hook runs: `go mod tidy` and
  `go fmt`, failing if `go.mod` or `go.sum` changed, then `check`.
- `script/install-precommit` installs the git pre-commit hook that runs
  `script/precommit`.

## TODO

The open issues that stand between the tree and a 1.0 release:

- [#14 Choose a license and add LICENSE](https://git.eeqj.de/sneak/keyfunc/issues/14)
- [#15 Decide the Go module path before 1.0](https://git.eeqj.de/sneak/keyfunc/issues/15)

## License

Not yet chosen. The license is the owner's decision, still open on the tracker
([#14](https://git.eeqj.de/sneak/keyfunc/issues/14)); the `LICENSE` file is added
when that issue is answered.

## Author

[@sneak](https://sneak.berlin).