The caller's reports, newest first, one page at a time.

agent_id

Narrows the list to the reports cut from one agent, which is how the agent page's Reports tab reads. Optional, and its absence means "everything I own" -- the two callers want genuinely different lists rather than one being a special case of the other.

Filtered in SQL rather than by fetching a page and dropping the rows that do not match. That distinction is not premature: this list is keyset-paged, so a client-side filter would return a page of twelve mixed reports and then show however many of them happened to belong to the agent, which reads as "this agent has three reports" when it has thirty. The predicate has to be inside the window, not outside it.

An unknown or unowned agent_id is not an error. It returns an empty page, because user_id is already in the WHERE clause and no row can escape it: asking about somebody else's agent is answered with "you have no reports from that", which is true, and which leaks nothing about whether the agent exists.

No index is added for this. ix_agent_reports_agent_id already exists on the column, and the account-scoped row count here is in the tens.

Why this grew a cursor

It used to select every report the account owned, with no limit and no cursor, and render all of them. That is fine at nine reports and it is a page that gets slower every time somebody uses the feature -- the failure mode is not a crash, it is a screen that takes a little longer to open each month until somebody calls it broken.

Keyset, not offset

Same shape as the two public directories: order by (created_at DESC, id DESC) and resume from a cursor naming a position in that order. id is the tie-break, so the order is total and a page boundary cannot sit in the middle of a group of reports created in the same second -- which is exactly what happens when somebody generates two from the same agent.

Offset paging would re-read and skip rows, and would also drop a report from page two the moment a new one is written between requests. A cursor names a row, so a report arriving while the reader pages cannot shuffle the page they have not read yet.

The shape changed

This used to return a bare array. It now returns {items, next_cursor}, matching the directories, because a list with no envelope has nowhere to put "there is more". next_cursor is null on the last page.