[kperry-godaddy]

Normative text for the change feed proposed in #96. One optional endpoint, GET /changes: an ordered, cursor-driven, replayable feed of the identifiers a registry added, updated, or removed, and nothing else. A consumer fetches each entry from the identifier's authority (the <publisher> domain, resolved as §5.1 directs) and confirms every delete there. The feed says what exists and what changed; the authority says what the entry says.

What changes

• spec/ard.md
  • New §5.6 Change Feed (GET /changes) — Optional: request, response, change object, cursor and delivery rules, the retention window, producer rules, consumer rules, security, errors. Placed after §5.5 Observability (merged from main) so it does not renumber §5.3.x (PR Add direct catalog entry lookup by identifier #74 adds there).
  • §5 intro names the new layer; §5.2 gains a "Change-Feed Subscription (Optional)" pipeline bullet; §5.4 gains one paragraph naming query-time and index-time federation and stating that federation: none keeps its meaning.
  • Appendix B adds 410 CURSOR_EXPIRED and names 501 NOT_IMPLEMENTED (Explore already returns 501 without a code).
  • D.3 lists GET /changes among the OpenAPI-defined interfaces. D.4 is untouched: conformance tooling is a follow-up PR.
• spec/schemas/ard.openapi.yaml: /changes path (cursor, limit; 200/400/401/410/429/501/500), a ChangeFeedResponse schema referencing the new JSON Schema the same way SearchResultItem references ard-entry.schema.json, and a 410CursorExpired response. info.version left for the publishing commit.
• spec/schemas/ard-change.schema.json (new): Cursor, Change, ChangeFeedResponse. Validated against the 2020-12 metaschema and against the §5.6.2 example plus negative cases.
• spec/schemas/ard.cddl: cursor, change, changes-response.
• adr/0011-index-time-federation-change-feed.md (new): context, decision, why identifiers rather than entries, alternatives, exclusions, consequences. Numbered after the merged ADR-0010; renumber if Add direct catalog entry lookup by identifier #74 lands first.

Decisions taken on the open questions in #96

Question	What this PR does
Q1 identifiers only	Yes. A change is {cursor, op, time, identifier}; the change object is closed (additionalProperties: false).
Q2 optional vs required	Optional with 501, as for Explore; a registry operated as an upstream for other registries SHOULD implement it. One sentence; easy to drop if the group prefers plain optional.
Q3 retention	Sliding window that MUST contain every held identifier, met either by indefinite retention (never 410, retention omitted) or by re-emitting an identifier's latest change before it ages out. Floor: SHOULD P7D. Cursor older than the window → 410 CURSOR_EXPIRED → staged replay; the live index is never emptied.
Q4 unreachable authority	A delete is never applied while the authority cannot be reached; how long an entry then survives is the same question §5.2 leaves open for crawled entries, and the text says so rather than inventing a lifetime here.

Deliberately not in this PR

• Identifier forms whose authority is not a publisher domain (the two reserved identifier segments discussed in Design: index-time federation via a change feed of identifiers (GET /changes) #96). They depend on the §4.5.1 alignment definition in §4.5.1 says the trust domain MUST align with the publisher and never defines align #94 / PR spec: define 'align' for publisher authority binding (§4.5.1) #98 and should be validated against real records first; the text keeps the authority concept general so a follow-up can add profiles without rewriting §5.6.
• Inline entries, digest, origin, hop path, reason, filters, push transports, publisher-served feeds, a DNS record for the feed (recorded in the ADR with the condition for adding each).
• Conformance tooling (producer vectors, consumer fixtures) and any change to auto (hop limit, partial-answer marker).
• Reaching entries that are not resolvable from their publisher domain (in-page markup on a page the consumer does not know). The text states this limit in the §5.6 intro; such entries keep being indexed by crawling.

Checks

• OpenAPI parses; every $ref (internal and to both JSON Schema files) resolves.
• JSON Schema passes the Draft 2020-12 metaschema; the §5.6.2 example validates; unknown op, extra properties, bad cursors, bad durations, and >1000 changes are rejected.
• Conformance demo still passes (it does not read the changed files; included as a smoke test).

Closes nothing on its own; tracks #96. Related: #63, #87, #88, #94, PR #74.

🤖 Generated with Claude Code

[rvguha]

This is too much of a burden to put on ARD endpoints.

…

On Mon, Oct 12, 2026 at 12:14 PM Keith Perry ***@***.***> wrote: @kperry-godaddy <https://github.com/kperry-godaddy> requested your review on: ards-project/ard-spec#108 <#108> spec: add an optional change feed (GET /changes) for index-time federation (#96 <#96>) as a code owner. — Reply to this email directly, view it on GitHub <#108?email_source=notifications&email_token=ABICKCXVE76TYR43ZGMCNBT5TUUPPA5CNFSNUABQM5UWIORPF5TWS5BNNB2WEL2JONZXKZKFOZSW45CON52GSZTJMNQXI2LPNYXTGMZQG42DENZRGI3TNJTSMVQXG33OWBZGK5TJMV3V64TFOF2WK43UMVSKKZLWMVXHJLDGN5XXIZLSL5RWY2LDNM#event-33074271276>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/ABICKCUFMXYINY4VFUGHJRT5TUUPPAVCNFSNUABGKJSXA33TNF2G64TZHMYTENBTG4ZTSMBXGI5US43TOVSTWNJYGE4TCMZQG4ZDLILWAI> . You are receiving this because your review was requested.Message ID: ***@***.***>

[kperry-godaddy]

@rvguha - Thanks for the feedback.

• This endpoint is optional. Implementers of ARD don't need to implement it.
• Systems like AGNTCY and ANS already provide some type of sync/event feeds. This just puts them in a standard format everyone can consume. This can be the bridge between different systems.
• Fan-out at query time is not scalable.