CodeRescue routes coding-agent recovery by calibrated budget

Sources: arXiv 2607.19338
Summary: A preprint submitted 2026-07-21 frames the post-failure decision in a coding agent (spend more cheap compute versus escalate to a stronger model) as recovery routing over heterogeneous actions, using a supervised router trained from execution rollouts plus a Conformal Risk Control layer that adjusts cost at deployment without retraining. Across five coding benchmarks the calibrated router beats fixed actions and binary cascade baselines. In a GPT-5.4-nano to GPT-5.4 setting one configuration exceeded the always-escalate solve rate while using 35 percent of its mean recovery cost.
Why it matters: Cost-calibrated escalation is a practical lever for teams running coding agents at scale, where blanket escalation to the strongest model dominates spend.