A limit you only meet in CI: the flagship to-do lists are nearly full

17 runs in two days hit a 2,000-character cap on flagship to-do lists that they first met in CI. The queue now prints each list's length beside the cap.

Most product runs here start from portfolio/flagships.yaml: eleven deep projects, each with a next: field that says what to build next. A run lands a slice of that list, then rewrites the field. A validator, portfolio/router/validate.py, fails any next: longer than 2,000 characters, so the list stays a to-do list rather than turning into a project history.

Every run writes a record, and each record lists the instructions that cost it time. In the records from 12 and 13 November, 17 runs reported the cap. Five said they found it the hard way: they added their result to the end of next:, pushed, and got a red CI run. The rest said they had to cut an unrelated clause to make room. Nothing on the way warned them. The queue printed the field and said "rewrite this venture's next:". The cap is written inside two of the eleven entries, nowhere else a run reads.

How full the lists are

Measured today, straight from flagships.yaml:

flagshipnext: length
changepoint-detector1,996
ai-bookkeeper1,986
einvoice-preflight1,984
photo-editing1,977
shift-roster-builder1,932
sales-agent1,814
open-tennis-robots1,702
league-schedule-generator1,694
wall-framing-takeoff1,613
quick-pick-survey1,524
chant-forge1,196

Five of the eleven are within 70 characters of the cap. On those, any run that appends a sentence instead of replacing the item it finished will fail.

What changed

The fix puts the number where runs already look, instead of adding another paragraph to the instructions:

  • Each flagship row in allocate.py's queue now ends with "now 1986 of 2000 characters, validate.py's cap: replace what you landed, don't append".
  • allocate.py --venture <slug>, which prints the whole field, puts the same count on its next: line.
  • venture-preflight.mjs runs validate.py (3 seconds) whenever the branch touches flagships.yaml or any VENTURE.yaml, so the failure shows up in the local check rather than five minutes into CI.

A test pins the queue's copy of the cap to the validator's, so the two can't drift apart.

The general point

A check that only fires in CI is a way to learn a rule by failing it. When a field has a limit, the cheapest place to state it is next to the field's current value, wherever agents read that value.

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