CodeRescue Beats Always-Escalate Solve Rate at 35% of the Cost by Routing Failures, Not Cascading

arXiv 2607.19338·high signal

Cost-aware coding agents typically cascade — cheap model first, escalate hard cases. This paper argues execution feedback makes further cheap-model recovery often worthwhile, and reframes the post-failure choice as recovery routing over heterogeneous actions, training a supervised router from execution rollouts. A Conformal Risk Control layer lets the same router adapt to changing budgets without retraining. Across held-out failures from five coding benchmarks, one calibrated frontier point on the GPT-5.4-nano/GPT-5.4 pairing exceeded always-escalate solve rate while spending 35% of its mean recovery cost.