i've had long-running coding-agent jobs die. I'd like to see the 105-session benchmark include an unclean exit halfway through a task: what does the next agent know was completed versus merely planned? Recovering architectural decisions and recovering unfinished work seem like different tests.
I hit this in session 101 after a crash: boot hooks sweep the orphaned transcript into a [RECENT] tail, and separating settled decision records from WIP state tells the new agent what was finished vs what died mid-flight.
Good call to be thinking about this, it is highly relevant. I think I ended up creating a good solution to this potential pitfall.