Summary

Context compaction currently feels too abrupt in long-running sessions: when compaction triggers, a large middle chunk can be summarized in one pass, which may drop too much in-flight task context at once.

This is especially noticeable in messaging-platform workflows (e.g. Feishu group chats) where users cannot easily reset with /new and expect continuity during active tasks.

Problem

When context pressure reaches threshold, compaction can remove a broad middle range in one step. Even with summary handoff, this can cause:

• sudden loss of nuanced task state
• assistant responses that feel disconnected from just-executed work
• poor UX during active multi-step execution

Expected behavior

Compaction should be smoother/rolling so that recent in-flight context is preserved longer, instead of one-shot middle truncation.

Proposal

Introduce rolling compaction:

• cap one compaction pass to a bounded number of oldest turns (instead of summarizing the entire middle in one pass)
• allow multiple smaller compaction passes over time
• keep the most recent active-task turns intact as long as possible

Repro pattern

1. Run a long multi-turn session with frequent tool calls and long outputs.
2. Reach ~85-95% compaction pressure.
3. Observe that one compaction pass summarizes a large middle span.
4. Continue task and compare response continuity before/after compaction.

Reference implementation

I opened a draft PR from a fork with a rolling-compaction approach + test:

• chore: initialize fork branch #9550

If helpful, I can split that into a focused PR scoped only to compaction behavior.