CodeRescue: Budget-Calibrated Recovery Routing for Coding Agents
The paper argues failed coding-agent runs should be routed between cheap retry and costly escalation, not treated as automatic handoffs.
CodeRescue trains a supervised router from execution rollouts to decide what recovery action is worth the budget after a failure. A Conformal Risk Control layer lets the same router adjust to different cost limits at deployment time without retraining. On held-out failures from five coding benchmarks, cheap recovery and escalation succeed on different cases. In the reported GPT-5.4-nano/GPT-5.4 setup, one calibrated point beats always escalating while using 35% of its mean recovery cost. ArXiv · AI/CL/LG's note
CodeRescue trains a supervised router from execution rollouts to decide what recovery action is worth the budget after a failure. A Conformal Risk Control layer lets the same router adjust to different cost limits at deployment time without retraining. On held-out failures from five coding benchmarks, cheap recovery and escalation succeed on different cases. In the reported GPT-5.4-nano/GPT-5.4 setup, one calibrated point beats always escalating while using 35% of its mean recovery cost. ArXiv · AI/CL/LG's note
score 4