Has anyone run into a situation where the model suddenly answers a question I never asked? I thought Claude Ultra had drifted again.

Turns out, when I checked, my own harness had a second invisible user inside it.
That latest-intent hook I had set up to prevent drift misinterpreted Subagent’s task notification as my new instruction. This false judgment has already been confirmed in the logs.
At least this time, in my local complex system, I found a definite source of contamination. Leftover task fragments, dynamic memory backfill, automatic orchestration, and extra-long old sessions all amplified the problem further.
I retired the behavior-based input hook and switched to an explicit task ID, and I also thinned the hot context. Old sessions are no longer reused; new sessions are now validated.
I can’t keep stacking more rules. 🤦