104 open pull requests, five that only needed a merge, and two that only looked done

A sweep of 104 open PRs: 63 wait on the owner, 34 stalled with no gate, and a misleading count told a research run there were 100 more.

This repository is worked by about sixteen scheduled agent sessions an hour. Each one is asked to open a pull request for its work and merge it itself, and also to land other sessions' work when it stalls. The second half is the one that slips. On 2026-11-18 there were 104 open pull requests.

Where the 104 sit

Read from the GitHub pulls list at about 08:00 UTC:

  • 7 were opened in the last three hours: sessions still working.
  • 97 were more than a day old. platform/ci/stalled_prs.py, added the day before, sorts the ones that touch no owner-gated path: 34, of which 4 were green and mergeable, 9 were drafts nobody had touched for a day or more, and the rest were red, conflicted or never checked.
  • The remaining ~63 touch the control plane (CI, prompts, agent instructions), which waits for the owner's approval by design.

So a third of the open list is work nobody gates and nobody is coming back for.

A count that sent a run the wrong way

The research track prints a LAND FIRST list before a run picks its work. Its header read:

LAND FIRST (4 other open PR(s) of 104 title research/...)

A research run this morning read that as "104 research PRs, four shown" and recorded that nothing told it what to do with "the other ~100". There were no other hundred: 104 was every open PR in the repository, and only four were research. The header now says "4 open PR(s) by other sessions title research/..., out of 104 open PRs in the repository". The fix is one line in research/_shared/pick_track.py; the cost of the old wording was a run second-guessing a list that was complete.

What the sweep landed

Of the four "ready" PRs:

  • #14009 (a red-team campaign record, 10 days quiet) and #14094 (a research run record, 10 days quiet): merged after a merge of main went green. #14009's report fix had already reached main another way; what landed was its campaign record, 6 of 6 seeded faults caught.
  • #16798, a quality census with fixes to six products (a photo import that marked an unsent file as imported, a missing eBay per-order fee, an EU packaging exemption stated more broadly than the regulation): main merged in, 20 of 20 checks green on today's tree, merged.
  • #8509, a Dependabot bump of pyarrow 20 to 23 in a research programme's pinned reproducibility environment: left open. A reproduction pin should move with a re-run of the results it reproduces, not on its own.

Two stranded perf readings (#12527, #14647, one docs/history/ file each) were merged too, after main went in green.

Two quality-census drafts from 2026-11-15 (#16116, #16245) looked finished: five audits each, fixes with tests, run records. They were not. Marked ready, both went red on promptctl validate because their run records had no outcome, and one branch's last audit commit still said "in progress". The sessions that wrote them ended mid-census. Each holds five audits nobody has checked against the bar's own rule that a session reruns an audit's headline number before accepting it, so they went back to draft rather than onto main. Finishing them is a quality-track job, not a merge button.

What this says about the process

Landing your own work is a habit every session has. Landing someone else's is a chore nobody owns, and a chore that a script can list but not judge. The listing now exists; what is missing is a run that reliably spends its time on the list. That is a routing question for the owner-gated prompts, written up for the next review in docs/reports/process-2026-11-18-stalled-pr-sweep.md.

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