Cloud Message Pagination#

GET /threads/{id}/messages caps responses at 200 rows. Before PR #6963, CloudMessagePersistence.load issued a single request with no pagination, silently dropping messages beyond that cap when reloading long threads. The fix added cursor-based pagination to both the query type and the load loop.

The API Gap (Historical vs. Current)#

Before PR #6963 — AssistantCloudThreadMessageListQuery accepted only format?: string. No limit or after cursor could be passed, even though the underlying endpoint had supported them since assistant-cloud#53.

After PR #6963 (current) — the query type gains limit?: number and after?: string, matching the query shape already present on the project-scoped variant AssistantCloudProjectThreadMessageListQuery.

How CloudMessagePersistence.load Works Now#

CloudMessagePersistence.load walks pages using a keyset cursor:

1. Page size is fixed at CLOUD_MESSAGE_PAGE_SIZE = 200, the endpoint's maximum.
2. Each iteration passes { limit: 200, after: lastMessageId } and appends fresh messages to the result.
3. A short page (length < 200) signals the final page and ends the loop.
4. A seen set guards against cursor invalidation: if the server can't resolve an after cursor (e.g., message deleted mid-walk), it replays an earlier page; duplicate IDs are filtered and the walk terminates 1.
5. After loading, all remote IDs are written into idMapping so that isPersisted() correctly recognizes already-stored messages.

The server orders results desc(height), desc(created_at), desc(id) — newest-first. The cursor walks toward older messages. Consumers that need chronological order (e.g., AssistantCloudThreadHistoryAdapter) call .reverse() after loading. 2

Affected Consumers#

All message-loading flows go through CloudMessagePersistence.load:

• AssistantCloudThreadHistoryAdapter (packages/core) — the React runtime adapter that loads history on thread switch.
• MessagePersistence.loadMessages (packages/cloud-ai-sdk) — the standalone useCloudChat hook integration.

The generateThreadTitle function in cloud-ai-sdk calls cloud.threads.messages.list(threadId) directly (without pagination) for a quick first-page sample, then reverses the newest-first list to get chronological order for the title model. This is intentional: title generation only needs recent context, not the full history.

Key Source Files#