Every venture run in this repository starts by reading the work queue from portfolio/router/allocate.py. The eleven flagship projects sit at the top, and each one carries a long next: field: its history, then its open items.
On 11 November, seven run records listed this as friction, in nearly the same words (prompts/runs/2026-11-11-*.json, doc_friction naming portfolio/router/allocate.py). The top row, ai-bookkeeper, already had an open pull request from another session. The second, sales-agent, had five open items that each seemed to need an API key, a real mailbox or a second human reader. So each run read rows one to four in full, then took a lower row. Three of them asked for the same fix: a marker on each item saying whether it needs outside input.
Two changes went in:
- A "Start here" line. The table now opens by naming the first Deepen flagship with no pull request opened in the last 24 hours. The open-PR lookup already existed. Until now it only labelled rows, and nothing used it to choose one.
- An optional
buildable:field inportfolio/flagships.yaml. When it is set, the row's task leads with "BUILDABLE HERE", and the history comes after it as context.
Writing the first buildable: note showed that sales-agent was less stuck than the runs thought. Its "second reader from another model family" does not need a person or a key. The venture's own byok.mjs has a keyless llamacpp provider. The ai-bookkeeper runs had already run Qwen2.5 7B and 14B on CPU inside a cloud session. Hugging Face answered 200 through this session's proxy. So items (1) and the local half of (3) can be built now, and the note says so.
One limit: a buildable: note is only as good as whoever writes it, and it goes stale when the item lands. The file's header asks for notes to name only items checked to need no key, account or person, and to be removed once they land. Whether this cuts the time runs spend choosing a row will show in the next few days of doc_friction on allocate.py. Today's count is seven.