mirror of
https://github.com/langchain-ai/langgraph.git
synced 2026-08-31 20:29:46 +02:00
Nick Hollon
1bf0f1b7f8
Commit 133082af preserved the full checkpoint ns on v2 message
events so subgraph chat-model tokens emit at their full path (matching
JS's handleChatModelStart). That promoted root-level chat tokens from
the empty tuple to depth 1 (e.g. ("call_model:<task>",)), but the root
MessagesTransformer was scope_exact=True at scope=(), so every
streamed message was dropped — run.messages yielded zero streams for
any graph calling model.invoke() / stream_v2().
Mirror JS's namespaces=[[]], depth=1 filter: accept events at the
transformer's scope or exactly one segment deeper. SubgraphTransformer
still forwards deeper events to the matching child mini-mux, whose
own MessagesTransformer applies the same scope+1 rule.
Also update stale test assertions: ChatModelStream.output.content is
now always a list of v1 content blocks by construction, so tests
compare via str(stream.text) (sync) or an explicit block-shape
assertion on the assembled AIMessage (async, since AsyncProjection
isn't str-able).