Many long‑horizon agents compress their context according to a global rule—usually a token budget—without regard to the task they are performing. We ask whether the agent's recent behaviour can forecast when such compression will be harmful. The TRACE public corpus contains 590 AppWorld compaction boundaries; each boundary is replayed from a re‑executed prefix under both the pre‑compaction context and the summary, and the burden of subsequent actions—calls that error or repeat a previous call—is recorded. Results show that pre‑boundary history predicts post‑compaction harm only weakly. An internally prespecified contrast by prefix placement yields a wide null, and the naive “has‑written” label actually measures trajectory phase rather than harm. The best extension‑protocol trigger attains AUROC 0.66 (0.64 on its own label) compared with a same‑boundary replicate of 0.72; the best frozen, interpretable trigger avoids 21% of harmful boundaries while retaining 84% of compaction opportunities, exceeding random‑rule expectations on count but not on burden mass (a post‑hoc comparison). Whether this trigger outperforms a token‑budget rule at matched retention cannot be evaluated with the current release. We outline which corpora should be shipped to enable that comparison.
Review: The paper offers a careful quantification of compression effects, highlights the limited predictive power of recent history, and proposes an interpretable trigger that balances harm avoidance with opportunity retention, pointing to concrete data‑release recommendations for future work.