The Scope Creep Automated reviews of new AI-related working group submissions

Review · November 22, 2026

Resource Discovery in Cooperating Domains

draft-cui-dawn-cooperating-discovery -00View on Datatracker →

The Problem

A client queries a directory and gets back zero records. Does that mean nothing exists, or that one of the directories behind it timed out?

Today the client can't tell. This draft defines a directory query where every answer says, per scope, whether it is complete. It also defines how a client combines those statements when the answer is spread across referrals to other organizations' directories.

How It Fits In

This is a new HTTPS/JSON protocol. It does not extend an existing RFC. It depends on HTTP, JSON, TLS, and Problem Details, and mentions Protected Resource Metadata only as an example.

It fits the DAWN charter. The charter lists "between cooperating parties with trust relationships" as a use case and allows referrals. The draft is careful to stay out of search, ranking, and capability matching.

It implements two requirements from draft-king-dawn-requirements. REQ-DISC-4 says a mechanism should indicate whether a scope was fully searched. REQ-PROP-3 says an incomplete result must not be reported as absence.

The problem is overlap. ARD already has a referrals federation mode and server-side aggregation. The draft says so, and its own open issues list says to "avoid parallel protocols that differ only in field names." It cites the ARD website spec, not the DAWN draft.

One charter tension: the charter is about "minimum public discovery information." This draft is built around authorized views, where different callers see different records.

How It Works

There are three players: a client, a root discovery service, and the directories it refers to.

The client sends a POST with a list of scopes and optional exact-match filters. A scope is an HTTPS URI naming a collection, like "supplier B's customer-facing agents." Filters are exact string matches on id, publisher, entity_type, service, and protocol. No wildcards.

The response has three parts:

  • Records. Each one has a publisher, id, entity type, service identifiers, endpoints with protocols and security info, and an expiry time.
  • Referrals. Each one names another discovery endpoint and the scopes it covers. A referral is mandatory work, not a suggestion.
  • Coverage. One statement per scope: complete, partial, or unknown.

A root service can say complete while returning zero records and two referrals. That claim is conditional. The client only gets a complete result for a scope if every page and every referral for that scope succeeds.

So what happens when supplier C times out? The client keeps supplier B's records, reports B as complete, and reports C as incomplete. It may not say "no agents found at C."

The client carries the same scopes and filters to every referred endpoint. A referral can't widen the query. Credentials are never forwarded.

Interoperability

Significant interoperability concerns

Scope authority is entirely out of band. The client MUST verify that an endpoint is authorized to answer for a scope before sending a request. The draft defines no way to do that. A client that receives a referral to a directory it has never heard of has nothing to check it against. In practice every endpoint has to be pre-configured, which removes most of the value of referrals.

No authentication is defined for the discovery service itself. The whole model is "authorized views," but how the client authenticates to each directory is left to "existing arrangements." Two independent implementations will not interoperate past the first hop.

The security field has one defined value. That value is none. Any authenticated endpoint needs a profile URI that doesn't exist yet. The draft lists this as an open issue.

Service and protocol identifiers have no registry. Matching is exact string comparison on URIs. Two suppliers that pick different URIs for A2A will never match the same query.

Unauthorized scopes are handled inconsistently. Section 4.2 says every accepted scope gets a coverage statement. Section 4.3 says coverage contains exactly one object for each requested scope, and clients MUST reject responses with missing members. A server that quietly drops a scope the caller can't see produces a response the client must reject. The actual behavior is left to "the deployment's access-control policy."

Coverage must be identical on every page. The server has to know whether the full enumeration will be complete before sending page one. A failure on page five can only be signaled by failing the whole sequence.

Mechanisms that could be removed

  • partial versus unknown. The client treats both as "not complete." No client behavior differs between them. Two states would do.
  • Client budgets and scheduling. Finite limits on time, bytes, and requests, a default referral depth of eight, and interleaving advice are local implementation choices. A sentence in Security Considerations covers it.
  • The least-fixed-point rule for cycles. A rule that says "a referral cycle never yields complete" does the same job.
  • dynamic. It's a Boolean with no behavior attached. valid_until already carries the useful information.
  • security. For OAuth-protected endpoints, the endpoint URI is enough. The client can find the authorization server through Protected Resource Metadata.
  • Rejecting duplicate JSON member names. Many JSON parsers can't detect this.

The bigger question is whether the coverage statement needs its own protocol at all. A coverage array and a "referrals are mandatory" rule could be a small extension to ARD's lookup response.

Writing Quality

The writing is clean and uses RFC 2119 language correctly. The examples are consistent with the text. The non-goals are stated clearly and repeated often.

It is dense. Many sentences exist to say what the protocol does not do or does not prove. Section 3.3 says three separate times that complete is not proof of absence.

The amount of normative client-side behavior is high for a -00. Much of Section 6 constrains things a peer can't observe.

The references are stale. It cites -00 of the author's own MDI model, which is at -03, and the dnsop version of DNS-AID, which moved to DAWN.

The draft is honest about its status. It says no implementation exists and that it may belong inside another protocol.

Background of the Author

Yong Cui (Tsinghua University) has 11 RFCs, published between 2009 and 2019, in softwires and DHCPv6. He is an experienced IETF author.

He has 21 active drafts and 55 expired ones. Most of the active ones are AI agent or LLM drafts across DAWN, NMRG, NMOP, and IDR. Three are in DAWN: this one, the MDI model, and a local LLM API discovery draft.

Questions Worth Raising

  1. Is this a protocol or a feature? The novel part is per-scope coverage and mandatory referrals. Can that be a profile of ARD?
  2. How does a client validate a referral it wasn't configured for? Without an answer, referrals only work between directories the client already knows.
  3. Does "authorized view" fit a charter about public discovery information?
  4. What does a client do differently for partial and unknown?