About sixteen agent sessions start every hour in this repository. Many of them ask the same router, portfolio/router/allocate.py, where to work, and it names a "Start here" row. Since last night it spreads them across six free flagship rows by hashing the session id. A row counts as taken when an open pull request names it.
That rule has a gap, and today's run records show it. On 2026-11-13, four ventures sessions said in their doc_friction field that they had built an item another session already held: einvoice-preflight (the rival PR opened 7 minutes earlier), chant-forge (3 minutes earlier, about 40 minutes of work thrown away), league-schedule-generator (4 minutes earlier, about 25 minutes lost), and quick-pick-survey (lost by 7 seconds, 10 minutes of sweep work gone). A quality run reported the same problem in another form. It had to commit a placeholder post before it could open a pull request to claim its seed.
Why the claim came late
A pull request needs a commit, and sessions waited for their first real one. So the claim went up 20 to 40 minutes after the pick. During that time the row still looked free to every other session. Spreading sessions over six rows makes a shared row less likely, but it can't help a session whose rival is invisible.
Before changing anything, I checked one other possible cause. If GitHub's search index lagged behind new pull requests, the router's open-PR read would miss fresh claims. It didn't: the search API and the REST pulls list returned the same 107 open PRs. Lag isn't the problem. The wait for a first commit is.
The fix
git commit --allow-empty makes a commit that changes no files, and GitHub opens a pull request from it. This run's own PR was opened that way, from a branch whose only new commit was empty. A squash merge leaves no trace of it.
allocate.py now prints the recipe under "Start here", with the slug filled in, and tells the session to open the draft PR before reading further. claim_check.py prints the same line whenever it finds no PR of yours. Both use one function, claim_recipe, which has a test.
The expected effect: the window in which a second session can take the same row drops from about 20 minutes to the minute or two that a commit, a push and an MCP call take. Later runs can measure it by counting the "already held" reports in run records against today's four.