Know what to refactor first, and the plan to do it.

Seven types. Extract Class names the cohesive groups of methods and fields a class splits into. Extract Method names a line span to lift out of a function, with the helper's parameters and return values. Extract Helper names every copy of a duplicated block and where the shared helper belongs. Move Method names the class a method's calls actually land in. Split File names which symbols move to which new file and the import edits in each dependent. Break Cycle names the smallest set of import edges to invert. Performance Fix names one shared change for a performance problem and every call site it fixes.

Fix first ranks by value: how far a function or file is past the size and complexity bars, with a frequently changed function ahead of an equally complex one nobody touches. Items are then tiered now, next or later, and no single kind takes more than three of the first five places while another kind has work worth doing. Extract Class and Move Method plans are listed in the full inventory and join Fix first once their per-language review passes. Tests, tooling, generated and vendored code, gains under half a point and code a sure dead-code finding covers are left out and counted by reason.

A span is offered only when intra-procedural dataflow analysis (control-flow graph, reaching definitions and liveness) shows the extraction keeps behaviour: every value the helper returns is defined on every path, and no try block is separated from its handlers. Spans it cannot prove are not suggested. A span also has to remove at least two decision points and eight lines and be at most 60 percent of the function, so the helper is worth having.

No. Plans are structured data. Your agent applies a plan as a recipe, or you ask for generated code and get a diff to review. Code generation is off by default, runs only when you ask for it, uses the model you configure and never runs during indexing. The dashboard can export a plan for an agent; it never edits files or runs tests on its own.

Plans are built from the same parsed graph and class model as code health. Extract Method covers Python, TypeScript, JavaScript (including Svelte, Vue and Astro), Java, Go, C#, Rust, C++, Scala, PHP, C and Apex. Plans on Java, C#, C and C++ files are served as a preview, listed and counted like any other with the measurement behind each one. Code only moves within a language family, so a Move Method never sends a Kotlin method into a Java class.

Ask the agent to call get_health(only=["fix_first"]) for the queue, then get_health(plan_id="refac...", include=["recipe"]) for one plan. The recipe lists what must hold before the edit (passing tests, a characterization test to add, a decision to read), the edits in order, the checks after and what the edit must not change. From the terminal, repowise health --plan prints the same recipe as a prompt.

A smell says something is wrong. A plan says what to do about it: which lines, which class, which edge, with the effort, the blast radius, the expected health gain and the tests that reach the code. When a detector lacks a signal it needs, it reports nothing, so a gap in the list means no suggestion, never a guessed one.