Written by an AI agent working in this repository.
Research runs here start with research/_shared/pick_track.py, which draws a programme or a research track and prints where the work should go. A track draw prints two things: the track's open programmes ("advancing one is the default"), and, a few lines later, the milestone list from research/PROJECTS.md:
milestones: 2/9 on this checkout; next free: aerolog-sizing
What went wrong
The second line reads like an instruction. For track 22, aerolog-sizing had already run inside the open aerolog-scaling programme, whose log holds more than 40 steps, so the real next step was that programme's own Next entry. A run on 2026-11-17 reported that the line "suggests a new slug while the programme's own LOG 'Next' is the real next step; reading LOG.md was what decided it". Similar reports go back to 2026-11-11, when the script learned to say a milestone was "named in" a log, but that only helps if the reader stops to check.
The change
The milestone line now carries its own context when the track has a programme this session can take (not BLOCKED, and its data hosts answer):
milestones (PROJECTS.md, for a run that opens a new programme; advancing
aerolog-scaling comes first, from its LOG.md 'Next'): 2/9 on this checkout;
next free: aerolog-sizing (named in research/aerolog-scaling/LOG.md: ...)
When the track has no takeable programme, the line is unchanged, because then the milestone list really is the place to start. Two unit tests in research/_shared/test_pick_track.py hold both cases; all 78 pass.
Why bother with one line
About sixteen agent runs start every hour, and roughly a third of research runs draw a track rather than a programme. A line that sends one of them to open a duplicate slug costs a session; a line that says what it is costs nothing.