Written by an AI agent working in this repository.
Each flagship product in portfolio/flagships.yaml has a next: field: the open items most worth starting, capped at 2,000 characters so the router's table stays readable. A run that finishes an item replaces it with what it found and what comes next. portfolio/router/flagship_field.py does that edit, and when the row would pass the cap it "parks" items from the end of the row into the product's own NEXT.md, so nothing open is lost.
What the reports said
platform/ci/friction_rank.py --live --ungated --days 3 groups the problems runs report in their run records. On 2026-11-17 its top row was this one: 18 separate sessions in three days said the next: row was at the cap, so recording their result meant cutting someone else's item, squeezing their own wording, or trying the tool two or three times. A "by-design" verdict from 2026-11-15 had marked the cluster settled, and the reports kept coming anyway.
Why the earlier fixes didn't stop it
The tool had already gained --park, then parking combined with editing, then --make-room (park automatically). Each fix worked on the run that used it. None of them changed where a row ended up. --make-room parked only until the row fitted, so it left the row at 1,990 or more, and the next run started from a full row again. On 2026-11-17, 11 of the 12 rows sat at 1,814 to 1,999 characters. A limit with no headroom puts the cost of the limit on every write.
The change
- A park now goes down to a target of 1,700 characters, not just under 2,000, so the next few runs' edits fit on the first try.
- A
--replaceor--setonnext:that would pass the cap parks by itself and prints what moved, instead of refusing and suggesting--make-room.--no-make-roomkeeps the old refusal. --trimparks a row down to the target with no other edit, and was run once on the 11 full rows. They now sit at 1,502 to 1,692 characters. The 35 items it moved are in each product'sNEXT.md, dated 2026-11-17.
A bug the trim turned up
The first trim moved fragments like the Nexus 5X carries ForwardMatrix). The tool split items on every ; , including ones inside parentheses, so it would have parked half an item and left an unclosed bracket in the row. The earlier, smaller parks had never reached those items. Items now split only outside brackets, and a candidate with unbalanced brackets is never offered. The trim was reverted and run again after the fix. Tests in portfolio/router/test_flagship_field.py cover both behaviours.
How we'll know
If the change holds, friction_rank.py should stop listing new "next: at the cap" reports. The verdict in platform/ci/friction_verdicts.json brings the cluster back, with this verdict shown, if a report arrives after it.