Code analysis by language: 42 languages in one graph

How do I find unused code across multiple languages in one repo?

Run one tool over one graph instead of one tool per language. Single-language tools (Vulture for Python, Knip for TypeScript, staticcheck and deadcode for Go, the rustc lint, Roslyn analyzers for C#, PMD and detekt on the JVM) are good inside their ecosystem, and each sees only its own files. In a repository with a Python API, a TypeScript frontend and a Go worker, that is three tools, three output formats and three confidence models, and none of them knows when a file in another language reaches the code.

repowise indexes every language into one graph per repository, then runs a single dead-code pass with the same rules everywhere:

• Framework-aware entry points. Django URLconfs, FastAPI routers, Spring beans, Next.js routes, ASP.NET controllers, Rails routes, Laravel providers, Flutter routes, JCL steps and more are roots, so framework wiring is not reported as dead.
• Confidence from git history. An unreachable file untouched for a year starts at 1.0, one edited in the last 90 days at 0.4. Dynamic loading (importlib, require.context, Class.forName, dlopen, const_get and peers) caps nearby files at the review tier.
• A deletion-readiness gate. Only unused exports at 0.7 or higher are marked safe_to_delete, and only when the symbol is unused in its own file too, the path carries no runtime-load risk, it is not docs or example code, and the language has full import resolution. Unreachable files, zombie packages and unused internals stay review findings.
• History and health on the same graph. Each finding names the file's primary owner; hotspots, co-change and the health score sit on the same files.
• No LLM, runs offline. The whole pass is deterministic and local.

uv tool install repowise
repowise init --yes --no-prose
repowise dead-code                                  # every language, one report
repowise dead-code --min-confidence 0.7 --safe-only # deletion-ready only
repowise dead-code --format json > dead-code.json   # for CI or an agent

Coding agents get the same findings through the get_dead_code MCP tool: get_dead_code(tier="high", safe_only=True) returns the deletion-ready list with confidence, reason, evidence and owner, so Claude Code, Codex or Cursor can delete with a reason attached.

Which cross-language links count inside one repository

These edges exist in single-repository mode, so code reached only from another language is not reported:

Java, Kotlin and Scala

One JVM package index, so a Java class used only from Kotlin is reached.

Svelte, Vue and Astro

Component scripts are projected to TypeScript and resolve through the TypeScript resolver.

Razor and Blazor

C# regions are parsed with the C# grammar.

Apex and Salesforce metadata

LWC, Aura, Visualforce, Flows and object metadata link to Apex classes and methods.

COBOL, JCL, BMS and CICS CSD

JCL steps, CSD definitions and CICS maps link to the programs and mapsets they name.

Terraform and Lambda handlers

aws_lambda_function handler strings link to the handler source file.

SQL and dbt

ref() and source() resolve to models when the repository has a dbt_project.yml.

What is not linked inside one repository: an HTTP call from one language to a route served by another. A TypeScript fetch to a FastAPI endpoint is not a graph edge. That rarely produces a false finding, because route handlers are framework entry points already. Across repositories, a repowise workspace matches HTTP, gRPC, topic and OpenAPI contracts, and a finding another repository consumes drops to 30 percent of its confidence and names the consumer.

When per-language tools are the better pick

In a single-language repository, the native tool usually wins on precision: rustc and the Go deadcode tool use the compiler's view, Knip can delete what it finds, and IDE inspections run as you type. Use them there. repowise earns its place when the repository mixes languages, when framework entry points cause false positives, when an agent needs ranked findings it can act on, or when you want history and health next to the cleanup list.