The key words "MUST", "MUST NOT", "REQUIRED", "SHALL", "SHALL NOT", "SHOULD", "SHOULD NOT", "RECOMMENDED", "NOT RECOMMENDED", "MAY", and "OPTIONAL" in this document are to be interpreted as described in [BCP 14] when, and only when, they appear in all capitals, as shown here.
# 0 Abstract
[User-Controlled Authorization Network (UCAN)][UCAN] is a trustless, secure, local-first, user-originated authorization and revocation scheme. This document describes a container format for transmitting one or more UCAN tokens as bytes, regardless of the transport.
# 1 Introduction
The UCAN spec itself is transport agnostic. This specification describes how to transfer one or more [UCAN] tokens bundled together, regardless of the transport.
UCAN tokens, regardless of their kind ([Delegation], [Invocation], [Revocation], [Promise]) MUST be first signed and serialized into DAG-CBOR bytes according to their respective specification. As the token's CID is not part of the serialized container, any CID returned by this operation is to be ignored.
All the tokens' bytes MUST be assembled in a [CBOR] array. The ordering of tokens in the array MUST NOT matter. This array SHOULD NOT have duplicate entries.
That array is then inserted as the value under the `ctn-v1` string key, in a CBOR map. There MUST NOT be other keys.
To serialize the container into bytes, the inner CBOR structure MUST then be serialized into bytes according to the CBOR specification. The resulting bytes MAY be compressed by a supported algorithm, then MAY be encoded with a supported base encoding.
The CBOR bytes MUST be prepended by a single byte header to indicate the selection combination of base encoding and compression. This header value MUST be set according to the following table:
Several attacks are possible if UCAN tokens aren't validated. If CIDs aren't validated, at least two attacks are possible: [privilege escalation] and [cache poisoning], as UCAN delegation proofs depends on a correct hash-linked structure.
By not including the CID in the container, the recipient is forced to hash (and thus validate) the CIDs for each token. If presented with a claimed CID paired with the token bytes, implementers could ignore CID validation, breaking a core part of the proof chain security model. Hash functions are very fast on a couple kilobytes of data so the overhead is still very low. It also significantly reduces the size of the container.
Compression is a relatively demanding operation. As such, using it is a tradeoff between size on the wire and CPU/memory usage, both when writing and reading a container. The transport itself can make compression worthwhile or not: for example, HTTP/2 and HTTP/3 headers are already compressed, but HTTP/1 headers are not. This being highly contextual, the choice is left to the final implementer.
While it is tempting to write a single implementation to read and write a container, it is RECOMMENDED to separate the implementation into a reader and a writer. The writer can simply accept arbitrary tokens as bytes, while the reader provides a read-only view with convenient access functions.
Many thanks to all the [Fission] team and in particular to [Brooklyn Zelenka] for creating and pushing [UCAN] and other critical pieces like [WNFS], and generally being awesome and supportive people.
