The agents that work in this repository may not merge changes to their own instructions. Prompts, CLAUDE.md files, hooks and CI workflows are listed in .github/CODEOWNERS, and a pull request that touches them waits for the owner. That rule is there to stop a rogue agent rewriting its own guardrails. The cost is that a wrong or stale line in a prompt stays wrong until the owner gets to it. Every run that reads the line meanwhile reports it again in its doc_friction notes.
There are 63 such pull requests open today. owner_held_prs.py already listed them oldest first. Oldest first is not the order that helps most, though, so the tool now has a second view:
python3 platform/ci/owner_held_prs.py --friction --days 7
It counts, for each owner-gated file, how many distinct run sessions reported friction on it in the last seven days. Next to each file it lists the held pull requests that change that file.
What it showed on 2026-11-17
- 52 owner-gated files drew at least one friction report in the week. 35 of them have no held pull request. No agent has proposed a fix for those, so nothing is waiting for the owner to approve.
- The root
CLAUDE.mdleads with 441 sessions. Most of those reports are about the cloud container (which Python has numpy) and are not really about the file. - The creative track prompt (
prompts/creative/prompt-0.md) is fourth, with 125 sessions and no held PR. It tells a run to pick the story shelf "that has gone longest without a new piece" by readinggit log. That fails twice over. A cloud clone is shallow and shows three or four commits, and every shelf now gets a new piece every day: For Other Minds got 38 in the last week, Little Lamps 41, Saints by Candlelight 50 and Hearth & Horizon 67. A tool that answers the question from the stories' ownpublished:dates,platform/ci/creative_due.py, has existed since 15 November. The prompt doesn't name it, so the runs that follow the prompt don't use it. prompt.mdhas six held PRs. Three of them make the same step 0 fix, so one review would close several.
The mistake that led here
I nearly built a second copy of creative_due.py. The friction pointed at a missing tool, I wrote one with tests, and only then found the original mentioned in the Hearth & Horizon README. The friction didn't come from a missing tool. It came from a gated prompt that never mentions the tool, and no one had proposed that edit. The new view separates the two cases: a file with reports and a held PR needs review, and a file with reports and no held PR needs someone to propose the fix.
The counts come from the run records in prompts/runs/. They are distinct sessions, not distinct problems, and some reports describe behaviour that is working as designed. Treat the table as a reading order for the owner's review, not as 35 confirmed bugs.