Do we still want to keep semanticCompact?

[Astro-Han]

#981 introduced semanticCompact as an experiment in improving model attention during a long active turn. #986 implemented it, but kept it off by default because we did not yet have a controlled on/off evaluation showing that it improved task results.

The surrounding architecture has changed since then. Standard Context Compaction now handles mid-turn capacity recovery, while semanticCompact still maintains its own policy, rolling block, controller state, validation, telemetry, and Agent Loop integration.

I am looking at simplifying Context Compaction into one path for manual, pre-turn, mid-turn, and overflow use. Before removing semanticCompact from that architecture, I want to check whether we still intend to pursue it as a separate attention experiment.

@likun666661, do you still want to keep working on this idea? Is there a known use case or planned evaluation that standard capacity compaction does not cover?

If we keep it, I think it needs an owner, a controlled semantic-on/off evaluation, and a clear result that would justify the extra mechanism. Otherwise, I would remove it for now. The canonical RuntimeEvents remain intact, so we can bring the experiment back later if we have stronger evidence or a concrete use case.

Related: #981, #986, #882, #481

Drafted with Codex. I reviewed the repository evidence and own the proposal.

[likun666661]

Thanks for checking. I am fine with removing semanticCompact for now.

The current historyCompact path already provides LLM-backed semantic compression (text summaries for normal providers and provider-native compaction for Codex), and it now covers manual, pre-turn, mid-turn, and overflow recovery. We do not currently have a concrete use case or controlled evaluation showing that the separate attention-first semanticCompact path improves task outcomes enough to justify its additional policy, controller state, validation, telemetry, and Agent Loop integration.

Please go ahead and remove it. Keeping the canonical RuntimeEvents intact gives us a clean way to revisit the experiment later if stronger evidence or a concrete use case emerges.

[Astro-Han]

Thanks, @likun666661. We have agreement to remove semanticCompact for now and reconsider it only if a concrete use case and controlled evaluation emerge. The removal is tracked in #3550 and will be included in #3545, so I’m closing this discussion.