Don't know why the below comment by killix got flagged; it's a legitimate point. In the current version of my setup, I've decided to accept that tradeoff. But

Don't know why the below comment by killix got flagged; it's a legitimate point. In the current version of my setup, I've decided to accept that tradeoff. But it would also be interesting to check whether agent behavior can be controlled well enough by a skill-management skill telling them to synchronously commit any changes with their signature; that would get the best of both worlds.