Skill Profile: Scope-Aware Experience Reuse for Agent Skill Adaptation

Agent skills package procedural know-how as reusable artifacts for runtime loading on demand. Once a skill runs across many tasks, its trajectories hold further procedural experience that the skill did not capture, such as missed checks, model workarounds, and handoff rules, much of it found through failure and correction. This experience should outlive the run that produced it, yet it is often valid only under the execution conditions in which it was learned. Reusing it without preserving those conditions leads to wrong-scope contamination, where locally correct adaptations become invalid guidance in new contexts. We present Skill Profile, a runtime layer that decides which retained experience may act on the current run. Offline, traces become Scoped Adaptation Records, each binding a behavioral change to a target skill, a validity scope, a trigger condition, an execution hook, and its source evidence. At runtime, a compiler admits only records whose scope, trigger, and output contract match the invocation, then selects on marginal gain and stops once coverage is met, leaving the underlying skill package unchanged. Experimental results across three transfer benchmarks demonstrate that Skill Profile achieves significant gains and holds contamination near its floor as more experience is added.