Checkpoint each accepted page before continuing — what should the operator record?
A long evidence table becomes fragile when the only record of progress is an open assistant conversation. A page checkpoint gives another operator a clear resume boundary without guessing which rows were already accepted. Build the checkpoint from actual returned metadata and preserve partial coverage alongside the accumulated records.
Keep page receipts and accepted identifiers recoverable when a multi-page result is interrupted.
Assign chain_key and page_sequence locally, then copy the originating query reference and returned continuation state into an access-controlled checkpoint. The first page is sequence one only in your local ledger; do not infer any undocumented server page-number parameter from that convenient local numbering.
Provide the last committed sequence and current stop reason, not an unsupported assertion that the next operator can resume an expired continuation. A receiver should be able to identify the safe stopping point without seeing an opaque token; request-context details should remain in their access-controlled location.Record the starting response
Assign chain_key and page_sequence locally, then copy the originating query reference and returned continuation state into an access-controlled checkpoint. The first page is sequence one only in your local ledger; do not infer any undocumented server page-number parameter from that convenient local numbering.
Separate received from committed
Mark a page received only after a response arrives, and committed only after its admitted identifiers are stored in the working table. Store accepted_key_digest or an equivalent permitted membership reference with the commit marker so receipt arrival and table admission can be checked independently.
Test an interrupted append
Use a permitted local copy to simulate stopping between receipt storage and table append; the ledger should reveal the incomplete commit. After the simulated stop, reconciliation should append missing admitted keys once and leave already committed keys unchanged, without rewriting the original response envelope.
Reconcile the resume edge
Compare the last committed identifier set with the next received page before appending, retaining any overlap as retry evidence. Track duplicate_received and newly_admitted separately; a retry page can contain both categories and should not be discarded or fully appended as one undifferentiated block.
Hand over the checkpoint
Provide the last committed sequence and current stop reason, not an unsupported assertion that the next operator can resume an expired continuation. A receiver should be able to identify the safe stopping point without seeing an opaque token; request-context details should remain in their access-controlled location.
Keep the boundaries clear.
- Keep page receipts and accepted identifiers recoverable when a multi-page result is interrupted.
- Operational record: Record the starting response. 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.