coderescue-9f5b6704·1 events·first seen Aliases: CodeRescue
CodeRescue introduces a post-failure decision framework for coding agents that routes between cheap-model retry and escalation to a stronger model based on execution feedback. A supervised router trained on execution rollouts is combined with a Conformal Risk Control (CRC) layer that adjusts cost penalties at deployment time without retraining. Evaluated across five coding benchmarks, one CRC-calibrated configuration exceeds always-escalate solve rates while using only 35% of the mean recovery cost in a GPT-5.4-nano/GPT-5.4 setting. The work addresses a practical inference-economics problem in agentic coding deployments.