mirror of
https://github.com/alibaba/open-code-review.git
synced 2026-10-02 09:24:52 +08:00
Micah Goulart
8d57bc9e83
* fix(agent): enforce --max-tokens-budget inside a running group The budget was only checked when the next group was dispatched, so a review with fewer groups than --concurrency could never hit it. Check it before every LLM round, give the over-budget group one final round to submit findings, and report it as failed(budget). * fix(agent): record the budget when the round gate skips a round The round gate in executeGroupSubtask breaks out of the round loop when the aggregate budget is already spent, but it was the only budget stop that did not set budgetExceeded. A group whose round finished normally — task_done, findings submitted — and only then tripped the gate left the run reporting the budget as intact, because no other gate ever looked again. It now records the stop the same way the in-conversation check does, with a warning naming the skipped round, the group, and the usage against the cap. The grace round stays on StopTokenBudget, in answer to the review on this PR. It is one bounded call whose whole job is to let the model submit findings it already has; skipping it spends the tokens the run has already paid for and throws the findings away. What was missing was not the skip but the record, plus docs that say the cap is checked before every round rather than only at dispatch — the cli-reference and CI input tables in all five locales now say so. Test covers the gate specifically: a group whose round 1 ends in task_done already over budget must not start round 2, must not get a grace round (it has nothing unreported), must stay classified completed, and must still report BudgetExceeded.