mirror of
https://github.com/langchain-ai/langgraph.git
synced 2026-08-26 17:42:24 +02:00
Nick Hollon
28ce32edc7
- EventLog(maxlen=N) caps retention. When the buffer is full, push evicts the oldest item and advances an absolute _first_seq so cursors can detect they've fallen off the back. A lagging cursor raises BufferOverflowError on its next read — mirrors the restored=false signal from the reconnection scenario (§06). - New cursors start at the current head of the buffer, not seq 0. For unbounded logs this is indistinguishable from the old behavior; for bounded logs, new consumers see whatever is still retained. - StreamChannel(name, *, maxlen=N) forwards maxlen to its inner log. - StreamMux(..., max_events=N) sets a default maxlen for every log / channel it binds (main event log plus each transformer projection). Explicit per-log maxlen wins over the mux default. - StreamingHandler.stream() / astream() expose max_events: caller sets the run-wide memory budget; transformer authors can override per-log when they know better. Default unbounded, matching §15 Q3.