Go is widely used in cloud-native infrastructure

while the frequent evolution of third-party libraries poses persistent challenges for code maintenance. Although large language models (LLMs) have shown strong performance in code generation

their effectiveness in Go library migration scenarios still lacks fine-grained evaluation. Based on the non-equivalent migration pair from gorilla/websocket to coder/websocket in the Go-MigBench dataset

this study constructs a four-stage automated evaluation workflow and systematically evaluates DeepSeek-Chat on 20 cases. The results show that the build success rate reaches 95.0%

whereas the end-to-end test pass rate is only 35.0%

yielding a gap of 60 percentage points. Failure attribution identifies three error patterns: API hallucination

message/return-value/close-semantics errors

and improper use of context lifecycles. The simplified CodeBLEU score shows only a weak correlation with test results (r = 0.26

p = 0.28)

indicating that static similarity alone is insufficient to predict migration correctness.