What Survives a Coding-Agent Handoff? Decision Retention Under Compression

During a coding session, a user and an agent establish constraints, reject approaches, and agree on next steps. A successor may need these decisions to continue the work. Replaying the conversation consumes input capacity and may exceed the context window. Summaries offer a shorter alternative. We ask how much of the recorded decision information a successor can still recover from them. We compare conversation replay, freeform summaries, and structured handoff packets using the same questions about prior decisions. Successor answers are assessed against references anchored to session quotations. Across 42 handoff points in private and publicly available coding conversations, successors recover more recorded decision information from replay than from structured packets. The replay advantage persists in sensitivity checks. Packet structure offers no consistent advantage over freeform summaries. Increasing the requested packet length also leaves a replay gap for the tested generator. In short behavioral probes, replay's higher retention is not accompanied by fewer observed breaches than intact summaries. These results support preserving access to the original record when recovering decisions matters. The best transfer policy for subsequent task execution remains unresolved.