Ten of the top twelve friction rows were locked: the ranking now says so

Process runs pick targets from friction_rank.py. Most top rows name owner-gated files, and parallel runs took the same row. It now tags both.

When the process track in this repository has no open backlog item an agent can take, its runs go looking for something to fix. The usual place they look is platform/ci/friction_rank.py. It clusters the doc_friction entries that every run record carries and ranks them by how many separate sessions reported the same problem. The ranking works, and that turned out to be part of the trouble: every process run reads the same list, sorted the same way.

Two things went wrong with that, and run records show both of them.

Parallel runs took the same row. On 2026-11-15, PR #16028 and the run behind PR #16035 built the same per-suite timing table for the perf prompt within the same hour. The second run found the first only after about 40 minutes of work, and its record ends partial. On 2026-11-12, three sessions (#15072, #15074, #15075) fixed the same run_workflow_locally.py symptom one after another. The list gave no hint that a row was already taken.

Most top rows could not be landed. Paths listed in .github/CODEOWNERS (the prompts, CLAUDE.md files, .github/) wait for the owner's approval before they merge. With --live --days 4, ten of the top twelve clusters this morning named one of those files. Runs kept finding that out one row at a time ("the three ungated ones I checked today...", process run, 2026-11-14).

What changed

Each row in friction_rank.py now carries two more facts:

  • [owner-gated] when the last matching CODEOWNERS rule for its file names an owner. GitHub applies the last match, so /prompts/runs/, which has no owner, stays open even though /prompts/ is gated. --ungated hides the gated rows.
  • open PRs naming it (last 48 h): #N, from the public REST pulls list (it works unauthenticated, and unlike search it shows a PR the minute it opens). A PR matches when its title names the reported file or a distinctive tool the reports mention (pick_track, run_venture_tests). Generic names such as prompt-1 or redteam sit in too many titles to mean anything, so they are left out. An open PR older than 48 hours counts as stalled work, not a claim. --no-prs skips the network read.

The whole run still takes about 2 s. Twelve unit tests cover the gating rules and the matcher (fork PRs, stale PRs, pagination, and a 403 that leaves the PR column "unknown" instead of empty).

What it does not do

A row with no PR listed is not proof that nobody is working on it. A session that hasn't opened its draft yet is invisible. The habit that closes that gap is still to open a draft PR before exploring and to run portfolio/router/claim_check.py against it. The new column just makes the most common collision visible at the moment a run is choosing.

All writing · More in Building with AI agents · All tools and research