TraceGate: Policy-Based Disclosure of Structured Crash Evidence for LLM-Assisted Debugging

Large Language Models (LLMs) are increasingly used in automated debugging and repair. However, most current systems treat failure-time evidence in one of two ways: as a fixed prompt artifact or as the result of an open-ended, model-driven investigation. This misses a fundamental opportunity. In LLM-assisted debugging, the ability to inspect and control what information is made available when a program fails should be treated as a first-class control problem. We present TraceGate, an adaptation layer that sits between failure execution and patch generation. TraceGate captures structured crash snapshots and adaptively decides which evidence to disclose, whether to invoke diagnostic helpers, and when to escalate budget for repair. These decisions are made by lightweight online policies conditioned on failure characteristics, repair history, model identity, and remaining budget, without changing the underlying repair backend. We evaluate TraceGate by augmenting representative LLM-based repair approaches across multiple benchmarks and model sizes. Our results demonstrate that policy-controlled evidence disclosure recovers additional correct repairs in previously unsolved cases while improving the success-cost tradeoff in most evaluated settings. ... mehrOverall, TraceGate shows that rethinking debugging through controlled observability, rather than relying solely on stronger models or larger prompts, can make LLM-assisted repair more effective, efficient and controllable.