A LangChain chain stops streaming token by token if any step in it is not streaming-capable

Streaming through the expression language works by each runnable forwarding chunks downstream. Any component that must see its complete input before producing output, such as a parser that validates a whole JSON document or a custom function that takes a string, buffers the stream and then emits one chunk. The chain still works and still calls the stream method, so the failure is invisible except as latency: nothing appears until everything is ready.

Output parsers that operate on partial input, including the streaming JSON parser, are built to avoid this and yield progressively.

When only the final text is streaming but you also need to observe tool calls, retriever hits or intermediate steps, use astream_events with version set to v2 rather than trying to thread callbacks through. It emits typed start and end events for every runnable in the graph, each tagged with the run id and any tags you attached, which is also the cleanest way to attribute cost to a specific step.