mirror of
https://github.com/langchain-ai/langgraph.git
synced 2026-10-14 12:15:25 +02:00
Elior Nataf LackritzandGitHub
6aa0afba68
`InMemorySaver.get_delta_channel_history` looked its target up by
`config["configurable"].get("checkpoint_id", "")`, so a config without a
`checkpoint_id` matched no checkpoint, and every channel came back with
no seed and no writes, which a `DeltaChannel` reads as empty. The base
implementation, `PostgresSaver` and `SqliteSaver` read from the thread's
latest checkpoint instead, as `get_tuple` does.
A missing id now resolves to the namespace's newest checkpoint id, with
the same `max` that `InMemorySaver.get_tuple` uses. Graphs always pass a
checkpoint id here, so only direct callers of the saver API were
affected, and a call with an id does the same work as before.
Thanks @Hotragn, who pinned this down on #8242, where @longquanzheng's
branch had already fixed it.
JS twin: langchain-ai/langgraphjs#2979
## Tests
- `test_get_channel_writes_without_checkpoint_id_reads_the_latest` fails
on main (`{'messages': {'writes': []}}`).
- The `checkpoint`, `checkpoint-conformance`, `checkpoint-sqlite`,
`checkpoint-postgres` (Postgres 16), `prebuilt` and `langgraph` suites
pass, the last two without their Redis cases. The DeltaChannel
conformance suite also passes against `InMemorySaver`; CI skips that run
in `libs/checkpoint`, where the conformance package isn't installed, so
the new test lives in `test_memory.py`.