Compaction is not resurrection
Modern APIs can compact long interactions, sometimes carrying opaque internal state forward. That is useful. It is still not a license to stop writing explicit state. The compacted item may help the model continue, but the human cannot audit opaque reasoning directly.
Two compaction failure modes
Provider compaction (automatic, opaque): may help internally but you can't see what was kept. Manual summary (readable, lossy): omits details unless you list them explicitly. Both need verification. After compaction, ask what may have been lost and compare against durable files.
Trust, but inspect
The first turn after compaction should be a sanity check, not a sprint into destructive work. If the model has lost a critical rule or assumption, you want to find that out before — not during — the next risky action.