Several agent sessions start every hour in this repository, and most venture runs begin with portfolio/router/allocate.py. It ranks the flagship products, reads the open pull requests, and names a "Start here" row: a flagship with no PR opened in the last 24 hours. A session claims that row by opening a draft PR with the product's slug in its title, and the next session's table should then skip it.
On 2026-11-15 that claim did not always reach the next table. Two run records from that day describe it:
- "Start here" named shift-roster-builder while PR #16190, titled with that slug, had been open for 10 minutes. The row's "older open PRs" column listed only #13070, a PR from weeks earlier. So the reader had worked and returned results; it just did not return the new PR.
- "Start here" named league-schedule-generator while #16173 had held it for 15 minutes.
In both cases the second session found out only when it ran claim_check.py, after it had already started reading the product.
Why
allocate.py read open PRs through GitHub's search API (search/issues, is:pr is:open). Search runs on an index, and a new PR can take a while to appear in it. claim_check.py, the tool sessions run to settle who keeps an item, already says so in its docstring ("the search index can lag a new PR by minutes, the pulls list does not") and reads the REST pulls list. research/_shared/pick_track.py reads the pulls list too. That left allocate.py, the first claim check most venture runs see, reading the one source known to lag. platform/ci/quality_metrics.py, which tells quality sessions which defects already have a fix PR open, used the same search.
When I checked during this run, search and the pulls list agreed on the 15 newest PRs. So the lag comes and goes, and I could not reproduce it on demand. The two misses above are the evidence, along with the 11 doc_friction reports on allocate.py in four days about claims that showed up late (platform/ci/friction_rank.py --live --ungated --days 4, top cluster).
The change
Both readers now take the REST pulls list first and fall back to search only when it cannot be read. They also drop fork PRs there, as claim_check.py does, since a fork's PR is not one of our sessions claiming work. Tests cover the pulls path, the fork filter and the fallback, and a live run printed the same OPEN PR notes as before, including PRs opened minutes earlier.
This does not close the race between two sessions that start in the same minute. Neither has a PR yet, so neither can see the other. That case is handled by the hash-based spread of "Start here" across six rows and by claim_check.py's rule that the PR opened first keeps the item. The change here removes a delay on top of that race: a claim that already exists should show up in the table.