ikiwiki/doc/todo/separate_authentication_from_authorization.mdwn

[[plugins/openid]] and the new [[plugins/emailauth]] both assume that
the user's authentication identity (how they log in) is suitable as
an authorization identity (who they are when they have logged in). In
particular, the OpenID or email address goes into the git history.

Relatedly, I'm not sure I'd be comfortable with enabling [[plugins/emailauth]]
on my own sites while it writes users' email addresses into the git history:
non-technical people (and many technical people for that matter) get
possessive about who can know their email address. The usual expectation for
email addresses on websites seems to be that they will be used by the site
owner (and maybe their outsourced service providers), but not available
to random third parties. The principle of least astonishment would suggest
that we should do the same here.

(The expectation of privacy for direct git commits is rather different:
I think we can expect direct git committers to know that they
should either set a plausible non-email-address in their git identity,
like I used to use my OpenID, or have good spam filtering.)

If we present email-based users in the "web UI" using only the local-part
of their address, we also have a potentially confusing situation where
`chris@example.com` and `chris@other.example.net` both contribute to a wiki:
if I'm reading the code right, they'd both be presented as `chris`, with no
way to change that other than using a different email address?

Here is a sketch of a different account model that would address that:

* An account has a username, e.g. `smcv`. It normally matches some regexp that
  includes neither @ nor / (to rule out collisions with email addresses
  and OpenIDs).

  * We currently allow qr{-[:alnum:]+/.:_} by default, so passwordauth
    accounts can in principle collide with OpenIDs. That would probably be
    worth fixing (for new account creation) anyway - I don't think we want
    users with / in their names, which would make their user-page into a
    subpage?

    > I have fixed passwordauth to not let urls be registered. It seems this
    > was not quite a security hole; it didn't let registering a username that
    > already existed, so if an openid was an admin, as long as the user logged
    > in using that openid, someone else couldn't come along and passwordauth
    > collide with it. (Might be exploitable if you could guess an openid that
    > was going to be added as an admin later though.) --[[Joey]]

* If passwordauth is enabled, accounts may have a password. Users can
  authenticate to an account that has a password by entering that password.
  The username is always the account name (because there's little reason
  to do anything else).

* If httpauth is enabled, anyone who can authenticate to the web server
  automatically gets access to the account of the same name in the wiki.
  (Or we could consider having a configurable map
  { web-server-level username => wiki account } but the default would be
  an identity mapping.)

* If OpenID is enabled, accounts may have an OpenID.
  The owner of that OpenID can log in, and gets logged-in to that account.
  Either reusing the same OpenID for multiple accounts is not allowed, or
  if the same OpenID is attached to more than one account the user can choose
  (as an extra step). Optionally, more than one OpenID could be allowed.

* If emailauth is enabled, accounts may have an email address.
  Users can authenticate to an account that has an email
  address (and is not a grandfathered OpenID) by using the token challenge.
  (passwordauth accounts could already do a password-reset, so this is not
  any less secure.)

* Creating an account from an email address (maybe also OpenID?) has a new
  step: choosing a username, with some text about "this name will appear
  in recentchanges and in the wiki's history". The default would be the
  local-part (user) from the email address.

* Grandfathered OpenID support: every existing account that looks like an
  OpenID has that OpenID associated with it, and it cannot be changed or
  removed. The displayed form is openiduser().

* Grandfathered emailauth support, if required (but it might not be required
  if we implement this model before the next ikiwiki release): every existing
  account that looks like an email address has that email address associated
  with it, and it cannot be changed or removed. The displayed form is
  emailuser() but we should maybe change that to output something more
  like `smcv@…`.

* Hypothetically, an account could also have a https client certificate
  for a new client-certificate plugin, or a Google account for a new OAuth2
  plugin, or whatever, and all of the above applies equally.

* Unlike the current OpenID support, if the user's authentication provider
  goes away (or if Google stops doing OAuth2 and moves on to the next big
  thing), they can associate a different authentication identity with
  their existing wiki account, and continue.

This is basically the same model that Mozilla Persona encourages,
except using emailauth ourselves instead of outsourcing (the equivalent
of) that step to Mozilla.

Thoughts?

--[[smcv]]

> I always find it a little ackward that i have two different accounts on this wiki: one for OpenID, and the other (regular account) for email notifications (because of [[bugs/notifyemail_fails_with_some_openid_providers/]]). It seems to me those accounts should just be merged as one, ie. I was expecting to be able to choose a username when i registered with openid.
> 
> Also, when you talk about "separating authentication from authorization", i immediately thought of [[todo/ACL/]] and [[todo/Zoned_ikiwiki/]], so i thought i could mention those... having stability in the usernames would help in the design of those... --[[anarcat]]

> I'm not opposed to this, but I don't anticipate having resources to do any
> work on it either. (I do hope to obscure email addresses from git
> commits.) --[[Joey]]