Architecture

A check runs in five stages. Each has its own folder under src.

Load the config

config/ finds edgefit.config.*, bundles it with esbuild so TypeScript works on any Node version, imports it, and validates it. CLI flags are merged on top in cli/project.ts.

Create the targets

targets/ builds one Target per requested runtime. A target knows its export conditions, the esbuild plugins it needs (Deno's specifiers and import map, for example), its globals, and how to look up an API in its data. Runtime settings are read here: wrangler config and compatibility gates for workerd, the pinned version for Bun, deno.json for Deno.

Resolve the graph

resolve/ runs esbuild from the entry with the target's conditions, without writing output, and turns the metafile into a module graph. Node built-ins and runtime-provided modules are kept external. Every module records who imported it, which is where the via chain comes from.

Extract usages

extract/ parses each module with oxc-parser and walks the AST with scope tracking. It records imports of built-ins, require calls, member access on those bindings, and globals that are not shadowed. Anything it cannot resolve statically, such as require(name) or fs[method], becomes a dynamic usage.

For --built, built/ then maps each usage through the sourcemaps to its original file and package, and marks bundled unenv code as mocked.

Classify and report

core/ looks up each usage in the target's data (data/), turns problems into findings, applies ignore rules and levels, and groups repeated uses. report/ renders the result as text, JSON or GitHub annotations, and computes compare tables and diff results.

Data

data/ loads the vendored files from data:

• providers/matrix.ts reads the workers-nodejs-compat-matrix dumps and compares each runtime with the Node baseline.
• providers/overrides.ts applies the curated override file for the target.
• providers/runtime-compat-data.ts answers for Web APIs the matrix does not describe.
• compat-index.ts merges them into one lookup, where the more precise source wins.

Design rules

• A false error is worse than a missed warning. Resolution follows the real bundler, and every heuristic is measured against real projects before it ships.
• Never claim safety. When the analysis cannot tell, the finding is unknown.
• Results are reproducible. No network access and no data that changes between runs.
• Each finding is traceable to a data source and, for curated entries, to a line of runtime source.