Why it matters: AI-generated integrations often fetch page one perfectly, return clean results, and then quietly stop forever, leaving a sales team wondering why the customer count never moves.
The Details:
AI assistants pattern-match against training data dominated by single-request, single-array API examples. The resulting code is genuinely correct for small datasets, which is exactly why it passes tests: fixtures rarely hold enough records to force a second page, so no assertion ever checks that all records arrived, only that some did.
Pagination styles fail in different shapes. Offset pagination breaks under concurrent writes, silently skipping or duplicating rows. Cursor pagination can loop forever if an API signals the end with an empty string instead of null. Link-header pagination gets missed entirely because the next-page pointer sits in HTTP headers, not the JSON body an AI parser scans.
Some APIs cap page size without an error. Ask for 500 records and get 100 back with no warning, and unless requested size is compared against returned size, that cap vanishes into the logs unnoticed.
Three defenses catch what the sync code cannot see about itself. A reconciliation job compares source and destination counts on its own schedule, independent of the sync logic. Structured page-shape logging records requested size, returned size, and cursor state per page, turning a week of guesswork into one query. A canary record with a known ID gets checked after every run; if it arrives intact, loss is limited to the tail, and if it doesn't, the whole sync is suspect.
Bottom Line: A pagination failure that never throws an error is not a smaller bug. It is a bigger one, because nothing tells you it happened.
Enjoy this article?
Listen to the Claude Code Conversations radio show or join the community.
