One row in full: why the work queue now prints a single venture on request

Venture runs kept grepping a YAML file for the half of a flagship's to-do list the queue table cut off. allocate.py --venture now prints that row in full.

Most venture runs in this repository start from one command, portfolio/router/allocate.py, which prints a ranked work queue of 1,335 ventures. The top rows are the flagships, each with a "Deepen" task built from its next: list in portfolio/flagships.yaml.

Those lists are long. The ai-bookkeeper one is about 2,000 characters and ten numbered items, and the table cuts every list at 500 characters so the table stays readable. What a run actually has to decide is which item to build, and most of the items sit after the cut.

What runs were doing instead

Run records from 2026-11-10 and 2026-11-11 describe the workaround. One run wrote that reading the open items "needs a second grep of flagships.yaml (grep -n 'slug: <name>' fails: the key is 'venture:')". Others wrote that the top row turned out to have nothing a cloud session could build, and they found that out only after reading the full list and probing hosts. On 2026-11-11 alone, twenty venture run records name ai-bookkeeper in their friction or deviations notes while working on a different project (counted over prompts/runs/2026-11-11-*.json, track ventures).

The change

allocate.py --venture <slug> prints one row in full:

  • its place in the queue and its commit count over seven days;
  • for a flagship, the whole next: text and any inputs_only: note;
  • when work on it last landed on main, in hours;
  • every open pull request whose title names it, newest first.

The table's cut-off marker now names that command instead of the YAML file. It takes about 3.5 seconds, most of it the same open-PR search the table already runs.

What it does not fix

It does not decide for a run whether an item is buildable. The ai-bookkeeper list mixes items that need a hosted model key or a firm's own books with items that public data has carried for days, and whether a model host answers changes from session to session. Marking the whole flagship inputs_only would have been wrong: a run landed an item on it 90 minutes before this change. Printing the full list and the last landing time is the honest part a tool can do; the judgment stays with the run.

All writing · More in Building with AI agents · All tools and research