CodeRescue: Budget-Calibrated Recovery Routing for Coding Agents

Forum topic · 小凯 · 2026-07-23

Summary

CodeRescue (arXiv:2507.17084) addresses a key question in deploying coding agents under budget constraints: after a failed attempt, should the agent spend more cheap compute on recovery, or escalate to a stronger, more expensive model? The authors, Qijia He, Jiayi Cheng, and Chenqian Le, formulate this post-failure decision as recovery routing over heterogeneous actions and train a supervised router from execution rollouts. To keep the same router usable across varying budgets, they add a Conformal Risk Control (CRC) layer that selects a deployment-time cost penalty without retraining, while providing marginal expected cost guarantees under exchangeability. On held-out failure cases across five coding benchmarks, cheap recovery and escalation show complementary success patterns. The calibrated frontier outperforms fixed-action policies, prompt-only routers, and binary cascade baselines. In the main GPT-5.4-nano/GPT-5.4 setup, one CRC-calibrated frontier point exceeds the solve rate of always-escalating while using only 35% of its average recovery cost. The code is open-sourced.

Overview

Field: Machine Learning Authors: Qijia He, Jiayi Cheng, Chenqian Le Published: 2026-07-22 arXiv: 2507.17084

Abstract

Coding agents increasingly operate in executable environments where a failed attempt produces actionable feedback rather than merely an incorrect answer. Existing cost-aware systems typically treat such failures as cascade decisions: try a cheap model first, then escalate hard cases to a stronger and more expensive model. In coding, however, execution feedback can also make further cheap-model recovery worthwhile, raising a budgeted deployment question: when should an agent spend more cheap compute, and when should it escalate?

The authors formulate this post-failure decision as recovery routing over heterogeneous actions and train a supervised router from execution rollouts. To make the same router usable under changing budgets, they add a Conformal Risk Control (CRC) layer that selects a deployment-time cost penalty without retraining, while providing marginal expected cost guarantees under exchangeability.

Key Results

Evaluated on held-out failure cases across five coding benchmarks.
Cheap recovery and escalation exhibit complementary success patterns.
The calibrated frontier outperforms fixed-action policies, prompt-only routers, and binary cascade baselines.
In the main GPT-5.4-nano/GPT-5.4 setup, one CRC-calibrated frontier point exceeds the solve rate of always-escalating while using only 35% of its average recovery cost.

The code has been open-sourced.

--- *Auto-collected on 2026-07-23.*