Rider 2026.2.1 lets AI agents invoke its refactoring engine directly, cutting C# task time 83%

The core change is that Rider's refactoring engine, powered by ReSharper, is now exposed to AI agents as a callable skill rather than remaining an IDE-internal feature. The agent can invoke operations like rename_refactoring, extract_method, extract_interface, extract_base_class, change_api_signature, move_type_to_namespace, reorganize_namespaces, and safe_delete. These operations work from the IDE's resolved syntax tree, which already knows which declaration every identifier binds to, which overload each call resolves to, and where every reference lives across the solution. This is knowledge the agent previously had to reconstruct one build at a time.

The evaluation methodology was a paired comparison: the same model (gpt-5.5) ran through Codex CLI on the same fifteen C# refactoring tasks, roughly ten times each, with the only difference being whether refactoring-code was available. The tasks covered eight refactoring operations in straightforward and harder variants with more call sites or entangled dependencies. Results were judged by a paired permutation test, and the eight operations were chosen specifically because they have clean contracts: a defined target, a defined result, and a refusal when the change is unsafe.

The before picture reveals how agents cope without structural tools. Across 2,513 tool calls, the agent piped text into interactive commands 468 times, called git 422 times, sed 392 times, and dotnet build 163 times, while performing a structural refactoring operation exactly zero times. The 163 builds were not for checking finished work but for discovering what a prior edit had done, because a correct rename requires overload resolution, partial class awareness, and distinguishing a type name from the same word in a comment. None of that is recoverable from a regular expression, so the agent guesses in text and lets the build score the guess.

After the skill was available, total tool calls fell from 2,513 to 926, and the eight refactoring operations accounted for only 167 of those 926 calls. The agent did not stop editing text: sed remained its most-used tool, and ordinary edits stayed in the editor's medium. What changed was the division of labour, where structural changes whose consequences ripple beyond what the agent can see go to the engine. The 95th percentile task time fell from 346.4 seconds to 56.9 seconds, because the slowest runs were the ones trapped in the edit-build-read-error cycle and those runs effectively stopped existing.

The skill is limited to C# refactoring and was evaluated only on eight operations with clean contracts. The article does not claim coverage of all of Rider's dozens of refactorings, and the evaluation used a single model through a single CLI. Only one feed carried this story, so there is no independent corroboration of the measurements. The cost figures are medians from a controlled comparison, not a guarantee of production outcomes, and the skill's value depends on the agent being asked to perform the specific structural operations the engine supports.