Repository navigation
Open
Labels
claude-triagedIssue has been triaged by the Claude Code triage routine. Remove to re-triage.Issue has been triaged by the Claude Code triage routine. Remove to re-triage.media-buyIssue concerns the media-buy protocol domainIssue concerns the media-buy protocol domainneeds-wg-reviewBlocked on a working-group decision — surface in WG meeting agendasBlocked on a working-group decision — surface in WG meeting agendasspec / protocol
Milestone
Description
Implementing the experimental collection slices on the reference training agent (#8691, part of #8181) turned up seven places where the spec is silent or ambiguous. Each one notes what the training agent does now. They're collected here for the WG's end-to-end review of the 3.3 experimental slices.
- Bulk-change threshold for small catalogs.
collection_catalog.bulk_changefires above "5% or 100 cards, whichever is smaller". For any catalog under 20 cards that's less than one card. Agent: a minimum of one card. Question: should the spec state a floor? - Cards on proposal reads.
request_proposalsandrefine_proposalsproducts carry no collection selectors, soinclude_collectionscould never return anything there. Agent: adds the references only when the flag is set. Question: should proposal products always carry collection selectors? - Follower
as_of.as_ofmeans when the count last changed (feat(media-buy): experimental list_collections and collection card webhooks #8279), but a sweep only observes a change after it happens. Agent: uses the observation time. Question: say so explicitly? - Card subscriptions seen by older readers. A
sync_accountsnotification config registered at 3.3 withcollection_card.*event types is echoed on later 3.2 account reads, whose enum lacks those types. Question: filter them for callers below 3.3, or rely on the rule that clients tolerate unknown enum values? - "Verified" with one
adagents.jsonfor several publishers. Verification is defined against each publisher's ownadagents.json. The training agent's single shared file speaks for several fictional publishers, so a real buyer resolving each publisher domain would treat those copies as unverified. Itsadagents.jsonis also unversioned, so the experimentalbrand_refreaches every reader. Question: how should a multi-publisher sandbox or test seller present verifiable declarations? changed_fieldswhen a removed card returns. Agent: omitschanged_fieldson thecollection_card.updatedfor a re-added card. Question: should the spec say so?list_collectionswith an account the seller can't resolve. Agent: answers as public. Question: error (e.g.ACCOUNT_NOT_FOUND) or a public answer?
Related: #8183 (design), #8279 (spec), #8141 (response collections), #8098 (follower counts), #8181.
Metadata
Metadata
Assignees
Labels
claude-triagedIssue has been triaged by the Claude Code triage routine. Remove to re-triage.Issue has been triaged by the Claude Code triage routine. Remove to re-triage.media-buyIssue concerns the media-buy protocol domainIssue concerns the media-buy protocol domainneeds-wg-reviewBlocked on a working-group decision — surface in WG meeting agendasBlocked on a working-group decision — surface in WG meeting agendasspec / protocol
Activity