Cline's checkpoint restore was silently orphaning commits made after the checkpoint.

SDK 0.0.82, September 2, makes checkpoint restore refuse to run when HEAD has moved past the commit the checkpoint was taken on. The restore's git reset --hard had been knocking later commits off the branch, leaving them reachable only through the reflog, and the branch move is now a compare-and-swap so a commit arriving mid-restore can't be lost either. If you use checkpoint/revert as your agent safety net and you commit while an agent works, this was a data-loss path. Same release gives remote SSE and streamable-HTTP MCP servers a 10-second connect budget, after an unreachable-but-enabled server burned the full 60-second request timeout inside session.create and killed the session.