This is a known cloud-agent reliability issue we’re actively working on. The agent’s machine can become unreachable mid-run, and in some cases it gets recreated afterward, which is why yours came back about an hour later. Your conversation and history are saved on our side, but file changes made inside the machine that weren’t committed and pushed to GitHub don’t carry over when the machine is recreated. That’s why the in-progress work was lost.
Until the fix lands, the best way to protect long runs is to have the agent commit and push incrementally instead of only at the end. Adding an instruction like “commit and push after each step” to the agent (or as a project rule) means finished work lands on your branch as it goes, so a machine restart costs you the current step at most rather than several hours.
On the model: Composer 2.5 runs in its faster mode by default, so when the run restarted it re-selected that default. It wasn’t a deliberate downgrade. If you want the non-fast variant, pick it explicitly when starting the agent.
I’ll keep tracking the unreachable-environment issue and follow up here once there’s an update.