GitHub user purushah added a comment to the discussion: [Discussion] Context compaction for Flink Agents
Thanks, this is exactly the walk-through I needed. Agree on all of it. For the group A PR I'll add a restart test for the round cap, since that is the one that changes control flow. Masking and truncation should keep replay matched as long as ids and order don't change. I'll add a test that proves that too. For C and D, noted: each summariser or judge call gets its own stable id, and a trigger is either deterministic or records its decision so replay reaches the same point. A budget on usage across rounds will keep its own recoverable counter. A current-prompt estimate can be recomputed. On the Trace Log, yes. One entry per compaction step with which messages were masked or truncated, the measured sizes, and any token estimate labelled as one. Cache behaviour I'll measure separately from the provider's cached token counts. I'll take you up on the group C replay tests when we get there. Is it ok if I tag you on the follow-up issue? GitHub link: https://github.com/apache/flink-agents/discussions/1202#discussioncomment-19060201 ---- This is an automatically sent email for [email protected]. To unsubscribe, please send an email to: [email protected]
