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

Review · November 22, 2026

Federation Between Capability Indexes Summary

draft-artamokhov-dawn-index-federation -01View on Datatracker →

The Problem

Agent and tool directories are run by many independent operators. Each one lists only what its operator has seen. A client sees one catalogue, and a provider has to register with every index separately.

The discovery landscape survey calls this out as a gap: no interoperable way for directories to federate. This draft proposes the simplest possible answer. Indexes crawl each other over HTTPS, and each operator decides which peers' catalogues to index.

How It Fits In

This is outside the DAWN charter, and the draft says so. The charter excludes catalogs of capabilities, indexing across the wider Internet, and trust management. This draft touches all three. The author offers it as input to the gap analysis, not for adoption.

It is the second half of a two-draft set. Everything depends on the author's marketplace discovery draft, which defines the discovery document at /.well-known/agent-marketplace, the signed catalogue, and key pinning. That is the only normative dependency besides HTTP, JSON, and URI. It cites revision -01 of that draft, which is now at -03.

There is no OAuth or WIMSE dependency. Nothing in the draft authenticates anyone.

It adds one new relationship: index to index. It defines one HTTP request header, one JSON message, and a two-state model for peers.

High-Level Summary

Crawl. An index starts from a configured seed list and fetches each peer's discovery document and catalogue. It follows the peers list in each document to find more. Crawling runs on a schedule with depth and count limits. Queries are answered from stored copies, never by contacting peers live.

Reciprocal identification. A crawling index sends its own URL in an Agent-Marketplace-Crawler header. The index being read can log it and learn who is reading it. The value is an unauthenticated claim and grants nothing.

Announcement. An index can POST a small JSON object to another index to say "I exist". The signature is optional and replayable. The draft says it must not influence admission.

Admission. Every known peer is either pending or admitted. Pending means known but not indexed, not queryable, and not crawled further. Admitted means the operator chose to index its catalogue. Closed admission is mandatory: only the operator creates pending entries. Open admission is optional: strangers can create pending entries through an announcement or the crawler header.

Preview. An operator can look at a pending peer's catalogue before deciding. Previewed entries stay out of query results.

The security section is mostly about SSRF. Every mechanism makes an index fetch a URL someone else chose, so the draft requires blocking private address ranges, re-checking at connect time, and not following redirects.

Significant Interoperability Concerns

The announcement endpoint has no defined location. The draft says it is deployment-specific and names the reference implementation's path. Two independent implementations cannot announce to each other.

The draft never says what state a crawled peer lands in. Section 2 says the crawl follows peers and stores each catalogue. Section 5.2 says a peer becomes pending only by operator action under closed admission. So is a peer found through an admitted peer's peers list pending, admitted, or ignored? This is the core of the federation model and it is undefined.

What federation delivers to a client is out of scope. "How the catalogue is then used is outside this document." The draft doesn't say whether an index re-exports admitted peers' entries in its own catalogue, how duplicates are handled, or how conflicting listings resolve. Two conforming indexes can federate and give clients completely different results.

The signed string doesn't match the message. The JSON members are index_url and discovery_url. The canonical string signs hub_url and well_known_url. The draft explains this as "the deployed wire format". That is a sign the spec is documenting one implementation.

The crawler header is underspecified. There is no field syntax definition and no Structured Fields type. It is required on discovery document fetches only, with nothing said about catalogue fetches. The comparison rule for self-crawl detection says "normalise scheme, host case and path" without pointing at a normalization algorithm.

Key rotation requires every operator to act by hand. A changed key is never accepted automatically. With 21 peers that works. With 2,000 it doesn't.

There is no rejected state. A peer is pending or admitted. A peer the operator has refused stays pending and counts against the pending limit, or gets deleted and can announce again.

Mechanisms That Could Be Removed

The announcement signature. It is optional, replayable, and forbidden from affecting any decision. It proves only that the sender holds the key it just sent. Drop it, along with signer_public_key and capabilities_count.

The new header. An unauthenticated URL identifying a crawler is what User-Agent and From already carry. If the goal is authenticated crawler identity, the Web Bot Auth work on HTTP Message Signatures is the place to look.

Most of the admission model. Preview display rules, recording which automatic checks ran, operator reversal, and storage separation are all internal to one index. None of it is visible on the wire. The one admission rule that matters for interoperability is rule 5: a pending peer must not appear in the published peers list. That rule changes the meaning of peers to "peers I have admitted", and it belongs in the discovery document draft.

What's left after trimming is small: a peers list with defined semantics, crawl limits, and SSRF rules. That could be a section of the companion draft.

Writing Quality

The writing is good. It is 13 pages, uses BCP 14 keywords correctly, and justifies most requirements in a sentence. The SSRF section is specific and practical. The author is candid about weaknesses: the replayable announcement, the missing endpoint location, split views, and the lack of independent implementations.

The naming is inconsistent. The text says "capability index", the header and well-known URI say "agent marketplace", and the signed string says "hub".

The Implementation Status section is honest about maturity. Six public indexes run this, all operated by the author or affiliates. A second implementation in Go is by the same author. No unrelated party runs a federating index.

Revisions -00 and -01 were posted a day apart. The -01 changes are editorial.

Background of the Author

Aleksandr Artamokhov (Independent) has no RFCs. They have two active Internet-Drafts, this one and its companion discovery draft, both first posted in late October 2026. They are a new IETF author. The companion draft has gone through four revisions in a few days, so they are actively iterating.

Bottom Line

This is a clean write-up of a working system, offered as gap-analysis input for work DAWN has explicitly deferred. It is useful as a data point that crawl-based federation with operator admission can work at small scale. As a spec, it leaves the two questions that matter for interoperability unanswered: what happens to transitively discovered peers, and what a federated catalogue looks like to a client.