No way to find out if user is authorized to write to a container

[rhiaro]

No way to find out if user is authorized to write to (/delete from) a container without actually trying to write (/delete) something.

[dmitrizagidulin]

As Henry indirectly alluded to in the solid-spec/#54, there's a way to solve this using Access-Control-Request-Method headers.

Specifically, by sending a "Preflighted request", an "exploratory" HTTP OPTIONS request where you pass in your credentials and the verb you actually want to perform on the resource in the Access-Control-Request-Method header.
Something like:

OPTIONS /documents/doc1 HTTP/1.1
...
Access-Control-Request-Method: DELETE

The server would then return the verbs you're authorized for in Access-Control-Allow-Methods: PUT, GET, OPTIONS, DELETE etc.

[rhiaro]

Sounds good, +1 to implement

[bblfish]

I am not sure I alluded to using "Access-Control-Request-Method" in issue 54.

In the CORS specification I think this won't actually work, as you are not meant to be authenticated when sending the preflight request (please verify). The browser will pass the Origin header but that is the identity of the JS, not the identity of the user. CORS is about identifying the agent in the browser, rather than identifying the JS agent itself.

I have written up a wiki page for this issue Client-Authentication Discovery. Interested in feedback.

[dmitrizagidulin]

you are not meant to be authenticated when sending the preflight request (please verify).

I double-checked the spec, there's no requirement prohibiting the passing of credentials to preflight requests.

[dmitrizagidulin]

The Client-Authentication Discovery proposal that you link to sounds like decent capability to have on the client library side.

It won't address this particular issue here, since the client won't always have read-access to a resource's ACLs.

[bblfish]

you are not meant to be authenticated when sending the preflight request (please verify).

I double-checked the spec, there's no requirement prohibiting the passing of credentials to preflight requests.

The spec is really not easy to read. I know this from experience. I think the relevant part of the spec is 7.1.5 Cross-Origin Request with Preflight. It says in point 1 there

Exclude user credentials.

Concerning your second point:

The Client-Authentication Discovery proposal that you link to sounds like decent capability to have on the client library side.

It won't address this particular issue here, since the client won't always have read-access to a resource's ACLs.

That wiki page also considers what one can do when the ACL is not made visible, in the first section Allow Headers

[csarven]

Transferring this issue to solid/specification for further consideration. See also #170 as one solution.

[csarven]

@bblfish Thanks for writing up the Client-Authentication Discovery for this issue. Still relevant today.

As @dmitrizagidulin mentioned, the client may not be able to read the ACL or even be able to discover the location of the ACL. This is reflected in the current state of the specification:

To discover, read, create, or modify an ACL auxiliary resource, an acl:agent MUST have acl:Control privileges per the ACL inheritance algorithm on the resource directly associated with it.

Even if all clients were to read the ACL, it raises the question of whether the policy they are granted to read is a subset of the information in the ACL. I don't think this flexible or at least requires more from servers - changing the representation based on agent. Alternatively, it is unclear if the link relation should refer to a different ACL resource describing only their readable bits. This duplicates the information and whether the resource is ephemeral or not is unknown.

While Allow includes HTTP methods as value, it is mapped from ACL. The WAC-Allow approach is also based on ACL, and so I presume the cost to calculate per request to be about the same. For an approach based on Link header (using URIs from the ACL vocab) would be in the same category.

Do we have other approaches documented? What implementation experiences can we take into account to make a decision here?