+- **Reviewer-revise treadmill** — failing sessions could not converge: a task that failed at max iterations or was rejected was re-picked with zero memory of the reviewer's feedback (measured: 4 successes vs 51 max-iteration exhaustions in a recent window; issues re-picked up to 59×). Every failure/rejection now persists its last reviewer feedback (task-state `lastFailures`, cleared on success) and injects it into the worker's first iteration on re-attempt — on both the parallel and serial heartbeat paths. Two consecutive near-identical revise feedbacks also end the session early instead of burning the remaining iterations (the reflection stagnation brake only counted objective sources, so a repeating reviewer never tripped it). (INT-2474, #230)
0 commit comments