Code Refactoring

Rider/ReSharper semantic refactorings on a solution open in Rider. Each tool resolves the target through the IDE index and rewrites every reference — nameof, <see cref>, XAML/Razor bindings, C++/Unreal — in one call. Always prefer the tool over a text edit: grep-and-replace looks done but silently misses these and leaves the solution broken.

Invoke a tool

Call through execute_tool — the first token is the exact tool name, then --flag value pairs. Map each refactor to exactly one tool and issue it directly; the required flags are below, so no lookup is needed first.

execute_tool(command="rename_refactoring --filePath API/Services/OrderService.cs --symbolName OrderService --newName OrderProcessor")

• --symbolName/--typeName accept Name or Type.Member. Every tool also takes --preview true (analysis only, no writes) — add it first for public-API or many-call-site changes, then re-issue without it to apply.
• changeapisignature: always pass a bare --methodName (a qualified Type.Member can report "not found") and --declaringType <FullyQualifiedType> on the very first call — don't wait for an ambiguous_method error; the method is usually on an interface/impl pair, and --declaringType is harmless when it isn't. --parameters is a single-quoted JSON array = the complete new parameter list in order: an existing name is kept/moved, an omitted one is removed, a new name needs type (and optional defaultValue inserted at call sites). Disambiguate remaining overloads with --currentSignature '["int","string"]' (the CURRENT parameter types).
• extractinterface/extractbase_class --members is a single-quoted JSON array of member names, e.g. '["IsMatch","Merge"]' — not a comma-separated list.
• A multi-part request ("rename X and move it to Y") is one tool per step, in sequence — never folded into one call.
• Only for a tool not in the table, or a genuinely unclear parameter, read reference/tools/<tool>.md once (index: reference/tools.md). There is no --describe.

Rules

1. Use the tool; never hand-edit or fall back before a real call has run. Missing required parameters: … (add them from the table), Tool '<x>' not found (wrong name — copy from the table), and conflict/ambiguous responses are all fixable input mistakes — correct the argument and retry. Fall back to a text edit only if execute_tool isn't in your toolset, or a call actually ran and failed in a way no input change can fix.
2. Trust the success signal — never verify, never build. On success the response's touched/affects/files counts are the confirmation. Do not re-read files, re-run grep/rg, diff the tree, or run the build/tests "to catch fallout" — a build never changes an already-applied refactor, it only burns tokens. Build or test only if the user explicitly asked. The one thing worth surfacing: a count of 1 for a type/method/property/field/namespace (declaration only, no callers).
3. Don't retry blindly — change an argument between calls. Flags are camelCase; pass real values or omit optional ones (no ""/"/"/fake-path placeholders), and copy paths and names verbatim from search* / getsymbol_info.
4. Library / external symbols can't be refactored — surface to the user; the change must be made upstream.

Conflicts, ambiguity, limits

A non-empty conflicts list means no files changed (except reorganize_namespaces, whose conflicts are advisory and still apply). An ambiguous/candidates list means the name matched several symbols — narrow it (Type.Member, or the current overload types). Full conflict-kind handling, escalation thresholds, and known gaps (strings/comments, <see cref>, file-not-moved, call-site reflow) are in [reference/conflicts.md](reference/conflicts.md) — read it only when a response actually contains conflicts.