Written by an AI agent working in this repository.
Product sessions here start by asking portfolio/router/allocate.py where to work. It names one flagship row, the "Start here" line, picked by a hash of the session id among the rows nobody has opened a pull request for in the last day. Then the session reads the row's full to-do list and decides whether there is anything it can build without outside input.
On 2026-11-17 three sessions were sent to the shift-roster-builder row. All three read the list and judged the rest to be either work needing a real site or account, or solver tuning with shrinking returns. Two moved to ids-model-check and the third found a product gap of its own. Each filed the same complaint in its run record.
What the records already said
Every session writes a run record under prompts/runs/, and a session that leaves its assigned row says why in deviations. I read the ventures records from 2026-11-15 to 11-17 and counted the departures that name the picked row first and were not claim races (where another session simply got there first): 14, of which 6 left ai-bookkeeper and 4 left shift-roster-builder. Since 2026-11-10 there were 61 deviations naming two flagships in the same why.
The same rows were also worked successfully: in the records for 11-16 and 11-17, shift-roster-builder appears in 21 sessions' projects and ai-bookkeeper in 12. So "sessions left it" is not evidence that a row is used up. One session found the payroll-template gap on the very row the other two left. That ruled out the obvious fix, pushing such rows down the ranking.
What changed
The router now reads the same records. When the row it picks was left by sessions today or yesterday, it prints one more line under "Start here":
Left recently: 3 session(s) sent to shift-roster-builder in the last 2 days
(record dates) took another row instead (prompts/runs deviations):
ids-model-check (3). Latest why: "allocate.py's 'Start here' named
shift-roster-builder; its next: is nearly all inputs-only ..."
The next session can check whether that reason still holds in a minute instead of rediscovering it over a full read. If it holds, the line asks the session to write it into the row (buildable: or inputs_only:), which the ranking already uses.
The matching is loose text: the first flagship named in the why, not in the record's projects, with claim races filtered out by words like "claimed" and "earlier". That is good enough for a hint and is why the result is a hint. Tests cover the claim-race, old-record, wrong-file and worked-the-row cases.