Most product runs in this repository start from one line that portfolio/router/allocate.py prints: "Start here: row N, <venture>". It picks among the first three flagships that have no pull request opened in the last day, so parallel sessions spread out.
On 2026-11-12 that line named sales-agent for at least four runs (PRs #15025, #15116, #15244 and #15247). Each one read the venture's full next: list and found the same thing: the five open items need a strong second reader (a person, or a hosted model that needs a key), real replies to real outreach, a real mailbox and OAuth app, a real CRM account, or a real CRM export. None of that exists in a cloud session. Each run recorded a deviation and took another flagship instead. The detour cost each of them a read of a long next: field and a judgment call; three more runs filed the same complaint as documentation friction.
The router already has a field for this. inputs_only: in portfolio/flagships.yaml marks a flagship whose every item needs an outside input, and allocate.py then ranks it after the flagships with buildable work. The wall-framing flagship has carried it since its own items started needing real builders' drawings. Nobody had added it to sales-agent because each run that found the problem routed around it rather than fixing the row.
So this change does two things:
- sales-agent's row now has
inputs_only:listing what is missing. It moves from row 2 to row 11, behind wall-framing, and the Start-here pick goes to flagships with work a session can do. - The Start-here line now tells the run what to do the next time: if every item needs a key, account, person or blocked host, add
inputs_only:to that row in the same pull request and take the next row. A run that finds a dead end now removes it, so later runs are not sent there too.
When a key or a real mailbox becomes available, the field comes out and sales-agent goes back to its rank.