Use in-memory event history for condenser replay

[neubig]

No description provided.

[github-actions]

🟡 Acceptable — the in-memory replay approach is sound, but two things need attention before this merges.

This review was generated by an AI agent (OpenHands) on behalf of the repository maintainers.

---

Was this automated review useful? React with 👍 or 👎 to this review to help us measure review quality.
Workflow run: https://github.com/neulab/agent-data-protocol/actions/runs/26795767723

[github-actions]

🟠 Important: _ensure_agent_ready() is a private (underscore-prefixed) SDK method. Coupling to internal implementation details is fragile — the SDK can rename or remove it without a major version bump and this will break silently. If there is no public API for "initialize without sending a message", that gap should be raised with the SDK team. At minimum, add a comment explaining why this private method is called here and what it guards against.

[github-actions]

🟡 Suggestion: conversation.state.events is no longer read anywhere in this function after the refactor — event_history is the canonical source for all condensation logic. Yet TrackingSDKEventBuilder.append still calls super().append() (line 115), which writes every event to conversation.state.events; and here the condensation is also written there explicitly to keep the two lists in sync.

This dual-write pattern obscures what the true source of truth is. If conversation.state.events is being kept in sync intentionally (e.g. as a guard against unknown SDK side effects that read it internally), add a brief comment saying so. If it is not needed, remove the super().append() call in TrackingSDKEventBuilder and this explicit append, which would make the migration complete and the code self-documenting.