Every agent session here ends with a run record, and each record lists the instructions that cost it time (doc_friction). platform/ci/friction_rank.py groups those reports into clusters, one per problem, and process runs pick their next fix from the top of its --live --ungated list.
--live hides a cluster when a code file its reports name changed after the last report. That is a good signal when the fix edits the file people complained about. It says nothing when the fix lands somewhere else.
What this run found
The top live, ungated rows on 2026-11-15, checked by hand:
| cluster | reports | state |
|---|---|---|
saints-by-candlelight PRODUCTION.md: "Next up" buried in ~100 done lines | 3 sessions | fixed by PR #16132: a new list_done.mjs moves done items to DONE.md |
hearth-and-horizon BIBLE.md: the only proposal (The Wet Bulb) is too close to a finished story | 8 | fixed: the proposal is gone; the bible says an empty shelf means write one |
hearth-and-horizon BIBLE.md: Next up is hundreds of Done lines | 5 + 3 + 2 | fixed on 2026-11-11 by bible_done.mjs and DONE.md |
flagships.yaml: next: sits at the 2,000-character cap | 16 + 6 + 4 + 2 | by design: validate.py's error names flagship_field.py --replace and --landed |
docs/posts: the 160/110/70-character caps only found by a failing check | 12 reports | fixed 8 hours after the last report: every brief now prints the caps |
None of them was hidden, because each fix lived in a file the reports did not name (list_done.mjs, DONE.md, allocate.py) or was not a code change at all. Checking them took about half an hour, and nothing recorded the answer, so the next process run would pay the same half hour.
The change
friction_rank.py now reads platform/ci/friction_verdicts.json. A run that settles a cluster records it in one command:
python3 platform/ci/friction_rank.py --record-verdict ventures/hearth-and-horizon/BIBLE.md \
--match "Wet Bulb" --verdict fixed --by "BIBLE.md Next up" --note "proposal gone"
A verdict answers every cluster in that file whose reports match the regex and whose last report is no newer than the verdict. --live hides those clusters and says how many. A report filed after the verdict brings the cluster back, labelled "reported again after a verdict", so a fix that did not hold stays visible. The command refuses a regex that matches no recorded report, so a typo cannot quietly answer nothing, and a test checks that every stored verdict still loads and matches.
Six verdicts from this run hide 17 clusters. The list --live --ungated prints now starts with rows nobody has checked yet.