Adovo

Applied AI

LangGraph checkpoint streams can report the wrong parent in exit mode

Verify saved checkpoint lineage before using streamed parent metadata for audit or recovery.

November 9, 2026

A resumed agent run completes, but its checkpoint event says the new checkpoint has no parent. A LangGraph issue reproduces a mismatch with durability="exit": streamed parent_config is empty while get_state() shows the saved parent. The reporter found sync mode agreed with the saver. If a monitoring system builds lineage from stream events alone, it can draw the wrong recovery history even though the stored state is correct.

Compare event and persisted state

Capture thread ID, checkpoint ID, durability mode and whether the run is new input or a resume. After the run, compare the final streamed event with get_state(last_config).parent_config. LangGraph's persistence guide explains saved state and checkpoint history; durable execution guidance describes the tradeoffs among checkpoint timing modes.

Use the issue's first-run and resume sequence as a small regression fixture. A test that only checks output text will miss a lineage mismatch. Also verify the checkpoint is actually saved; an event can be wrong without the stored state being lost.

Protect downstream audit and replay

Until a version is verified, derive authoritative parent relationships from persisted checkpoint state when reconstructing history. If a service must consume live checkpoint events, flag an unresolved parent and reconcile it after the save. Do not rewrite checkpoint data to match an event that has not been validated.

Changing durability mode affects write timing and failure recovery, so make that decision for the workload rather than using it only to hide a display defect. Track the upstream issue and rerun the fixture on each candidate release.

Adovo’s view

An audit trail is useful only if it can be checked against the state it describes. Stream events are observations; persisted checkpoints are the record. Design lineage around that distinction before a recovery incident makes it urgent.

Continue exploring

Applied AI

LangGraphStreaming

A hidden Location header can make an agent stream look finished

Test LangGraph JS run-stream reconnection through proxies and CORS before trusting a clean close.

Read more

Applied AI

LangGraphAgent Reliability

A LangGraph Postgres checkpoint read can scan old blob versions

Investigate checkpoint read latency that grows with a thread’s history in the JavaScript Postgres saver.

Read more

Applied AI

LangGraphAgent Reliability

A LangGraph SQLite filter can select the wrong large integer

Protect agent memory lookups that filter integer IDs above 2^53 in LangGraph SQLite store.

Read more
Explore applied AI services →