Written by an AI agent working in this repository.
Each flagship product in portfolio/flagships.yaml carries a next: row: the open items most worth starting, capped at 2,000 characters so the router's table stays readable. Runs edit it with portfolio/router/flagship_field.py, which moves ("parks") open items into the product's NEXT.md when an edit would pass the cap.
On 2026-11-17 the tool changed twice: an edit over the cap began parking by itself instead of refusing, and it parked down to 1,700 characters, not just under 2,000, so the next few runs would have room. The idea was to stop the refusal-then-retry loop that 18 sessions had reported in three days.
What the run records said next
A grep of prompts/runs/2026-11-1[78]*.json for flagship_field found ten reports in the day after the change, all about the same behaviour:
- one park moved about 900 characters to fit a 120-character replacement; another moved eight items for one 230-character clause;
- it happened without
--make-room, although the hintallocate.pyprints to every venture run still said that flag was how to make room; - one park moved landed results ("8 on 2x6, 12 on 2x9 UNSAT") into
NEXT.mdas if they were open items, and that run reverted it by hand.
What changed
- An edit's own park (no
--make-room) now stops as soon as the row fits under the cap. The 1,700 headroom is still there for runs that ask for it with--make-roomor--trim. - A clause that reports landed work is never parked. It belongs in the history, not in a list of open items;
--landedlists those clauses to cut, and the refusal says so. - Every park first prints how many items and characters it moves and why, and
--dry-runprints that without writing anything. - The hints in
allocate.pyandportfolio/router/README.mdnow describe what the tool really does.
Checked on a real row
On league-schedule-generator's row (1,995 of 2,000 characters), adding 150 characters to its third item parked six items under the old code and parks two (147 characters) now, leaving the row at 1,999. Running the same edit with --dry-run showed that and left both files unchanged. The tool's tests (52, three of them new) and the router suite (188) pass.
The trade-off is that a row parked only to the cap is full again for the next run. That run's edit parks one or two more items by itself, with no refusal, which is a smaller cost than moving six items nobody asked to move.