Recover cleanly when a continuation cursor expires — what should the operator record?
An expired continuation is an operational boundary, not evidence that the remaining source records do not exist. The useful response is to preserve the accepted portion, record why continuation stopped and decide whether a newly authorized retrieval is needed. Never imply that a local checkpoint extends the service's cursor lifetime.
Preserve accepted evidence and branch a restarted retrieval instead of splicing incompatible cursor chains.
Record the returned expiration condition and last accepted page sequence without copying the opaque cursor into a broadly visible incident note. A cursor-expired response is different from an invalid-cursor response; preserve the actual category so recovery does not incorrectly blame the operator's query construction.
State which admitted rows came from each chain and whether any gap remains unresolved; do not claim the restart recovered identical ordering. If full recovery cannot be established, deliver the accepted bounded table with stopped-expired and the unresolved scope rather than describing the restart as an exact replay.Capture the actual failure
Record the returned expiration condition and last accepted page sequence without copying the opaque cursor into a broadly visible incident note. A cursor-expired response is different from an invalid-cursor response; preserve the actual category so recovery does not incorrectly blame the operator's query construction.
Check the originating tool
Use current documentation to distinguish research continuation from workspace continuation; do not send one tool's cursor to another tool. Current Montage search documentation describes more_results using meta.next_cursor, while workspace continuation uses signals_next; check the actual originating response before choosing the continuation tool.
Keep the partial snapshot
Mark the chain stopped-expired and preserve its accepted identifiers under the original query revision and receipt times. Accepted rows retain their original receipt timestamps even if a later restart returns matching keys, because the second observation may carry updated descriptive fields.
Branch any authorized restart
Give a restarted retrieval a new chain_key, then compare overlaps rather than pretending it is the next page of the expired snapshot. The restarted chain can overlap, omit or reorder earlier rows; reconcile those differences as separate observations rather than patching the old sequence numbers to look continuous.
Explain the resulting scope
State which admitted rows came from each chain and whether any gap remains unresolved; do not claim the restart recovered identical ordering. If full recovery cannot be established, deliver the accepted bounded table with stopped-expired and the unresolved scope rather than describing the restart as an exact replay.
Keep the boundaries clear.
- Preserve accepted evidence and branch a restarted retrieval instead of splicing incompatible cursor chains.
- Operational record: Capture the actual failure. Context badges do not certify access.
Start with a read-only check: ask your agent to report actual source coverage before collecting records or exporting an audience.
Primary references
Documentation reviewed 2026-11-15. Product capabilities, client availability and plans can change. Third-party marks identify their products, not an affiliation.