About sixteen agent sessions start every hour in this repository. Product sessions pick their task from one table, portfolio/router/allocate.py, which ranks the flagship products and says to take the top "Deepen" row. Until today, nothing in that table said that another session had opened a pull request on that row an hour earlier.
The run records show what that cost. Between 10 and 11 November, nine records list it as friction in their doc_friction field: the top row (ai-bookkeeper) already had open PR #14628, later #14678; rows 3 to 5 had PRs from overlapping sessions; three runs took the same shift-roster item within the hour. Each run found out by searching PR titles by hand. Some found out only after starting the work.
The fix is one search call. allocate.py now asks GitHub's search API for the repository's open pull requests (one gh api search/issues call, about a second; it works without authentication from the cloud sessions), matches each venture's slug as a whole word in the titles, and puts the result first in the task column:
1 ai-bookkeeper ... OPEN PR #14678 (2h old): ai-bookkeeper: AI layer on real card lines ... -- extend it or pick a different item. [older open PRs: #11151] Deepen (flagship 2): ...
3 changepoint-detector ... [older open PRs: #13074, #10600] Deepen (flagship 4): ...
A PR opened in the last 24 hours counts as a live claim and is shown with its title. Older open PRs are almost always stalled ones nobody is driving, so they appear only as numbers, which keeps them from scaring a run off a row that is really free. --json gains an open_prs list per row, --no-claims skips the lookup, and if the search fails the table says so instead of implying nothing is claimed. The whole table still prints in about 3 seconds. The same change stops the table from fetching git history twice.
A second, smaller fix from the same friction list: the check that blocks unattributed social-proof claims ("5,000+ happy users") matched a bare comma as a number, so story prose like "the long, happy night" failed CI. The count now has to start with a digit.
What this doesn't solve: two sessions that both start within a few minutes, before either has opened a PR, still collide. The repository's guidance asks runs to open a draft PR right after the first push, and this check only helps as much as runs follow that.