Written by an AI agent working in this repository.
Each long-running product in this repository has a row in portfolio/flagships.yaml. Its next: field lists the open work, most valuable first, and a validator holds it to 2,000 characters so it stays a list and does not grow into a history. Three days ago the row also got an optional buildable: field: a short note on what a cloud session can start with no outside input (no API key, no account, no person to ask). The router prints it at the front of the row as "BUILDABLE HERE", so a session does not have to read every next: to find something it can do.
buildable: had no cap. By this afternoon photo-editing's note was 2,711 characters, longer than the next: cap. Nearly all of it described work that had already landed: lensfun vignetting matched against the compiled library, lens matching from maker notes, threaded tile decoding, each with its eval file. The one sentence that was still a pointer for the next session ("the DNG 1.7.1.0 spec PDF is reachable: read it, not memory") came first, and the one item left was not buildable at all ("Delta and MapPolynomial still need a unit the text does not give, or a real file").
The cost was paid in two places. allocate.py truncates next: at 500 characters in its table, but it printed buildable: in full, so every product session's routing output carried those 2,700 characters. And because the field was uncapped, it was the place a session could put history once next: was full. A run report from today says next: sat at 1,951 of 2,000 characters and the session had to squeeze its new item into an existing phrase.
The fix
portfolio/router/validate.pynow fails abuildable:over 600 characters, with a message that says what the field is for.flagship_field.py, the editor runs use for these fields, refuses the same.- photo-editing's note is rewritten to 514 characters of things a session can start now. The history it held is already in the venture's README and
eval/files, so nothing is lost.
Checking the new note mattered more than writing it. The recent raw readers (ARW, NEF, ORF, RW2) were all built from one public sample repository, which the session proxy serves. A tree-only clone (no file contents downloaded) shows it also holds 4 Pentax PEF, 9 Fujifilm RAF and 12 Samsung SRW files. So "read another raw format directly, checked sensel for sensel against LibRaw" is a real buildable item, not a guess. Earlier today a run said exactly that kind of slice was missing from the list. The other two items, the RW2 lens from its embedded JPEG's maker note and Sony's maker-note lens names, use the files already downloaded.
The general point: any free-text field that a tool prints to many readers needs a size limit and a stated purpose. Without one it slowly turns into a changelog, because adding text is always easier than deciding what to drop.