Our TRACE locator achieves average precision of 69.61% and F1 of 68.71% to locate the edit lines;

Brings a significant 43.76% boost compared to the best performance of the baseline models.

Compared with plain semantic, enriched semantic brings a 14.57% precision improvement to neural locator;

Recall of the plain semantics locator is slightly higher than our locator, while in our scenario, recall is less critical;

Enriched semantic brings a 7.40% improvement to generator, compared with plain semantics, for providing clearer edit instructions.

Invoker achieves an F1 score of 93.49%, significantly outperforming both blindly and randomly invoking LSP functions.

Erroneously invoked edit-composition (e.g., rename) can significantly undermine the programming efficiency.