Every run in this repository ends with a record, and each record lists the instructions or tools that cost it time (doc_friction). A script, platform/ci/friction_rank.py, groups those complaints into clusters and ranks them by how many sessions repeated them. Process runs pick their next fix from the top of that list.
The ranker also checks whether a file a cluster names has changed since the cluster's last report. If it has, the complaint was probably fixed, and --live hides the row. That check reads git log.
The problem
Cloud sessions start from a shallow clone. Today's clone held 50 commits of main, the last few hours of work. Nearly every report in the 14-day window was older than that history, so the ranker printed "touched-since unknown" under most rows and --live hid none of them.
Process runs then checked the rows by hand. Three run records from today say so: the top live clusters "were already fixed on main, so ~5 min went to" confirming it, twice over.
What changed
When the history starts after the oldest report, the ranker now fetches main back to a day before that report (git fetch --shallow-since, capped at 21 days so a --days 0 run never asks for the full ~5 GB history). It also reads origin/main alongside HEAD, so a fix that lands while a session is working still counts. --no-fetch skips the fetch.
The measurement
Same command, same records, this session:
| before the fetch | after | |
|---|---|---|
| history edge | 2026-11-15 (50 commits) | 2026-10-31 |
clusters hidden by --live --ungated | 0 | 68 |
| rows marked "touched-since unknown" | most of the top 20 | none |
| time | about 5 s | about 40 s for the fetch, once |
The 68 are hints, not proof: a hidden cluster's file changed after its last report, and the commit subject says whether the change was the fix. What is left at the top is now mostly data files (flagships.yaml, story shelves) whose changes the ranker does not count as fixes. Those are the rows that really are still open.