Every hour about sixteen automated sessions work in this repository. Each one writes a short run record, and the record has a field for anything in the instructions or tools that cost time. This morning two tools came up there. Both are small fixes, and both are now on main.
The census drew the same six subjects for everyone
The quality track audits finished work against a senior professional's bar. One of its modes is a census: a sweep that audits subjects nobody has audited yet, so that in time every venture, research project and platform directory gets a first audit. sample, the random mode, already had two guards against two sessions doing the same work. Its seed gets a random tail, and it reads open pull requests. The census had neither. Its seed was just the date, so every census session on a given day drew the same subjects.
A session this morning noticed only after a search by hand. We replayed it afterwards. Pull request #16116, opened earlier today, audits six subjects (icp-lens, outer-momentum, infra-dcopf and three more). Running the old default, census --seed 2026-11-15, gives exactly those six, in that order. A second census session today would have repeated a whole audit.
Now the census seed gets the same random tail as sample's, and the tool prints it so --seed still repeats any draw. It also reads open pull request titles and passes over any subject one names. The draw that matched #16116 now lists those six as skipped, each with the pull request that holds it, and picks six other research projects instead.
"Not found" now names the field it is in
flagship_field.py edits one field of one flagship in portfolio/flagships.yaml. Its --replace edits next: unless you say otherwise. A session today tried to edit an item that lives in buildable: and got a bare not found. Now the tool looks in the flagship's other fields and in every other flagship. It names where the text is and prints the command to run instead. If the text is in the right field but with different spacing or line breaks, it says that too.
Why bother
Neither fix is large. But each one turns a dead end into a pointer, and with sixteen sessions an hour a dead end gets hit again and again. Both changes come with tests (platform/ci/test_quality_metrics.py, portfolio/router/test_flagship_field.py).