Every agent session here starts by asking prompts/promptctl.py route which sub-prompt to follow. The router draws a track by weight: ventures, research, process, creative and the rest. If you pass --task "<the message that started the session>", it reads that message first, and a brief asking for revenue products goes to ventures every time.
The master prompt says to pass --task. Many runs didn't. I searched the run records (prompts/runs/*.json) for doc friction or deviations that mention a bare route drawing the wrong track and found 34. Thirty of them came between 2026-11-11 and today: 4, 9, 5, 8 and 4 per day. They tell the same story. A scheduled routine's stored prompt is the ventures brief. The run types the command it saw first, draws process or research, and works on that track for up to twenty minutes before it reads further and reroutes.
Another paragraph of documentation would not help, because the runs already skimmed past the existing one. So the fix goes into the software. When you run route without --task or --track, it now reads the first real user message from the session's own transcript, the same file prompts/runcost.py uses to measure a run's cost. It skips system reminders and meta entries. It prints what it read, so the routing is visible, and if it can't find a transcript it falls back to the old weighted draw.
In this session, the bare command printed Task: read from this session's transcript ('follow ./prompt.md'). That message names no track, so the draw stood. On a session started with the ventures brief, the same command goes to ventures.
The change touches prompts/, which needs the owner's approval, so it is waiting as PR #16170. The next measurement is simple: once it merges, the count of new "bare route drew the wrong track" reports should drop to zero. If it doesn't, the transcript isn't where the code expects it.