The owner of this repository sets directions in plain language: "AI accounting systems", "autonomous water vehicles", "tools for running orgs of agents". Each became a prose section in prompts/goals.md, about 25 of them so far. Agents read them, but no script could, so nothing could say which directions were getting work and which had been forgotten.
Now portfolio/work/directions.yaml indexes them. Each entry records when the owner set it, where their words live, a weight, and the paths whose commits count as work on it. tasks.yaml holds concrete tasks under each, one PR's worth with a "done when". workctl.py reads both:
coveragecounts this week's commits on main per direction and flags the ones with none;nextranks open tasks by the direction's weight divided by its share of recent work, and skips tasks an open PR claims with[work:<id>]in its title.
On its first run, against the 50 commits in this session's shallow clone (30 October onwards), the solo-business direction had the most commits (13), then AI accounting (10). The AI wealth guide, data center supply chain and vertical foundation model directions, all set today from investors' requests for startups, had none. That is expected for directions a few hours old, and it is exactly the gap the picker is meant to show.
The registry starts with 30 directions (20 areas, 10 working principles) and 14 tasks. The prose stays canonical for what the owner said. The YAML is only the index. Pointing the master prompt and CI at it touches owner-reviewed files, so that waits for its own small PR.