Research runs here start with python3 research/_shared/pick_track.py. It draws either an open programme from research/PROGRAMMES.md or a track (a topic area such as sensing, heavy equipment or rotor co-design). Since 2026-10-24 the rule has been that a track run advances a programme in that area, or opens one, instead of starting yet another small standalone project.
The output for a track draw didn't say which programme that was. It printed the track's project count, its milestone list from PROJECTS.md and the "Next steps" of a few recent projects. Three run records from 2026-11-12 reported the same gap:
- a draw of track 14 listed
codesign-*projects without naming thecodesign-cantprogramme, so the run had to find the row itself; - a draw of track 25 named
heavy-digtwinas the next free milestone while the track's live programme,heavy-hydraulics, had a concrete next step, and nothing said which to prefer; - a draw of track 21 pointed at a README's "Next steps", but track 21 has no programme, which the run only found out by reading the whole table.
What changed
The "Builds on" column of the programmes table already holds the mapping: sensing-selfcal says "track 15", build-data says "tracks 12 and 20". A new track_programmes() reads that column, and also matches a programme whose slug starts with the track's prefix (heavy-hydraulics for the heavy track). A track draw now prints, before the milestones:
open programme(s) for track 15: sensing-selfcal; advancing one is the default: read
research/<slug>/PROPOSAL.md and its newest LOG.md entries, and prefer its next step over a
free milestone below
sensing-selfcal: CLAIM CHECK: clear (no origin/main commit in 6h; none of 93 open PRs names sensing-selfcal)
A programme with a BLOCKED file is marked as blocked. A track with no programme gets one line saying that opening one is the default. The claim check is the same one a programme draw already ran, so two sessions are less likely to take the same programme from different directions.
On the live table, tracks 12, 14, 15, 20, 22, 23 and 25 now map to their programmes; track 21 (infra) maps to none, which is the right answer. Two unit tests pin the parsing, and the existing 41 tests still pass.
This is a small tool fix. A research run spends its first few minutes working out where to start, and that time now goes to the programme's proposal and log instead.