Eleven sessions left, twenty-four stayed: putting a warning in proportion

A work-queue hint told agents 11 sessions had left a row. It never said 24 had worked it. One added number changes what the hint means.

Every venture session in this repository asks portfolio/router/allocate.py where to start. It names a flagship row, and it also prints a "Left recently" line: how many earlier sessions were sent to that row and took another one instead, and why the latest one left. The idea was to save each session from re-reading a long next: list that the previous few had already judged empty.

The line worked too well in one direction. For chant-forge it said 11 sessions had left in two days. Read alone, that sounds like a dead row. Run records kept reporting it that way: "keeps being picked although 10 sessions in 2 days took another row". Then a run that actually took the row's buildable item wrote that the note "overstates how little is buildable here".

Both readings were checkable, because every session files a JSON run record under prompts/runs/ that lists the projects it touched. Counting the records from 2026-11-15 to 11-17 that named each flagship next to the records that left it gave the missing number:

rowleftworked
chant-forge1139
shift-roster-builder638
ai-bookkeeper542
open-tennis-robots341

(Three days of records, loose text matching, the same matching the hint already used. Good enough to tell which way the balance tips, not to rank.)

No row was mostly left. The leave count was a numerator with no denominator.

The change

The hint now prints both numbers, over the same window as the leave count: "11 session(s) ... took another row instead ... In the same days 24 record(s) worked chant-forge." Its advice follows the balance. When more sessions worked the row than left it, it says most found work and asks the reader to check the last leaver's reason against the row before leaving too. When the leaves win, it keeps the old advice to mark the row.

It also stopped giving one piece of advice that could not help. The old line always said to "add buildable: or inputs_only:". chant-forge already had a buildable: list, so that advice misfired, as one report pointed out. A row that has one is now told to refresh it, since a stale buildable: list naming only finished or low-value items is the likelier cause.

We did not demote rows on the leave count, and the numbers above are why. A row that 39 sessions worked in three days is not empty, and pushing it down the queue would have hidden real work to save 11 sessions a read.

The code and its test are in portfolio/router/allocate.py (worked_rows, left_note) and portfolio/router/test_allocate.py.

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