Repository navigation
Question: should the Model Pool's judge go free, or be skipped, while the OpenRouter balance is low? #1493
Description
A question for a ruling, not a defect. Nothing here behaves differently from how it was written. #1440 made a new situation possible, and nobody has decided how the Model Pool's judge should behave in it.
What is true today
- chat: know the OpenRouter balance, start a near-zero account on free models, keep the whole 402 sentence #1440 (merged into
devasf101ad2e7) reads the OpenRouter balance. When the account is known to be low (credits.LowCreditsUSD, $0.50 or less remaining), every crew tier the person has not written a model for goes to a free model, so a near-zero account keeps working on free models. - After a task lands, the Model Pool records it and asks a judge model to score the seats that ran it (
cmd/codeaf/poolrecord.go, the loop that callsjudge.Judge). It asks up tojudgeTriescandidates in turn. - The judge is never free.
judge.Candidatesininternal/pool/judge/judge.goleaves out every row whose prompt and completion prices sum to zero, and every id ending in:free. Its comment gives the reason: a free row is rate-limited and answers with a 429 when a judge most needs an answer.
Together: on an account codeaf itself knows is almost out of credit, the task's own work runs on free models, but the judge pass after it still picks the cheapest paid model that qualifies. That call spends from the balance #1440 is trying to protect. When the balance is nearly empty, the call is also likely to fail with a 402 that nobody sees, and the pool then records a failed judge attempt instead of scores.
The options
- A free judge while the balance is low. Drop the free-row exclusion when the reading says
Low, and accept 429s. A judge that fails costs nothing, and the pool already records a failed judge as such. - Skip the judge pass while the balance is low. The landing is recorded unjudged, with a reason such as
balance low. Nothing is spent, and the pool loses those scores. - Keep it paid. The scores are worth the cents. If so, the manual should say that a judge call is paid even when the crew has gone free, and the low-credit notice should not imply that everything is now free.
Whichever is chosen, the words a person reads about low credit (internal/manual/chat/openrouter-credits.md and models-and-cost.md) should say what the judge does.
Where
internal/pool/judge/judge.go:Candidates. Search forA free row is a rate-limited.cmd/codeaf/poolrecord.go: the judge loop. Search forjudge.Candidates(models(), seats, judge.DefaultFloor).internal/credits/credits.go:LowCreditsUSDandReading.Low.
Activity
santos/dev2 lands on dev: the worker harness as default, runs as work…