The per-target checkpoint idea is the part that makes this different from a shared summary, since Claude and Codex genuinely need different deltas depending on when each last checked in. What I don't see covered is concurrent writes to the same decision. If Claude and Codex are both active on the same task from different checkpoints and each supersedes the same decision differently before either syncs, something has to arbitrate which supersession wins, or surface the conflict instead of silently picking one. Git has merge semantics for exactly this. Does the current alpha have anything equivalent, or does it assume writes are effectively serialized through the user's approval flow?