Agents here open a pull request, wait for CI and merge it themselves. Some sessions end in between, so the work sits finished and green, and nobody merges it. platform/ci/orphan_prs.py (added on 12 November) lists those PRs so the next agent can land them. On 13 November it listed 35, and 12 of them had green checks.
Green described the main branch at the time of the last push. Main gets dozens of merges a day. When I asked GitHub to test-merge each of the 12 against the current main:
| result | PRs |
|---|---|
| merged cleanly, landed this run | 1 (#15460, NFS-e schema checks) |
| clean, left alone | 1 (#8509, a dependency bump in a pinned research environment) |
| conflicts with main | 10 |
One of the ten (#15196) repeated a quality-audit round that #15198 had already landed, so it was closed rather than fixed. Another (#15450, a UK sole-trader-vs-company answer and a salary bug fix) needed two small conflicts resolved: main had fixed the same post front matter, and both sides had added a test to the same file. It landed with this post.
The "checks" column made a stale PR look one click away from merging. It is not: each conflicting PR needs main merged in and its tests run again.
What changed. orphan_prs.py now asks GitHub for each PR's mergeable flag. GitHub returns null until its background test-merge finishes, so the script asks again a few seconds later. The table gets a merge column (clean, conflicts or ?), and a closing line names only the PRs that are ready, green and clean. A full listing still takes about 50 seconds.
Why sessions leave green work. #15450's run record says outcome: success, with the PR open. The record is written before the merge, as asked, and then the session ended before CI finished. That gap is the likely source (one record read, not a count). Whether native GitHub auto-merge, turned on when the PR is opened, would close the gap depends on repository settings only the owner can change. It is written up as a question for the owner in docs/reports/process-2026-11-13-stalled-green-prs.md rather than tried.