I went looking for why my linter felt slow to start. I found a TypeScript compiler in node_modules — once per plugin, never asked for.

My first theory was that I'd inherited it from @typescript-eslint/utils. My second was that it was my own fault. Both were half right, which is why the first fix didn't work.

Devkit at 1.0.0:

"peerDependencies": { "typescript": ">=4.0.0", "get-tsconfig": "^4.13.0", "enhanced-resolve": "^5.18.3", "@typescript-eslint/utils": "^8.46.2" }, "peerDependenciesMeta": { "get-tsconfig": { "optional": true }, "enhanced-resolve": { "optional": true } } 

Four peers, two marked optional. npm 7+ auto-installs the ones that aren't (npm/cli#4828). So 24MB of tsc came down, for code that used the compiler for one thing: a handful of ts.TypeFlags integers.

One caveat, and it has moved since I first wrote this. 24MB is TypeScript 6.x — measured 2026-09-04, 6.0.3 installs 24MB on the nose. I expected TypeScript 7's Go port to make that argument obsolete by shrinking to a couple of megabytes. It does the opposite: typescript 7.0.2 is now npm's latest and lands 30.7MB — a 3.6MB JavaScript shim plus a 26.5MB platform-native Go binary the shim will not run without. Force-installing a compiler nobody asked for was wrong at 24MB and is more wrong at 30MB.

I knew about the optional flag. I used it twice in that object and missed the two that mattered. Nobody reviews a manifest the way they review a function.

But marking it optional didn't empty node_modules. There was a second door:

npm view @typescript-eslint/utils@8.0.0 peerDependencies # { eslint: '^8.57.0 || ^9.0.0' } npm view @typescript-eslint/utils@8.46.2 peerDependencies # { eslint: '^8.57.0 || ^9.0.0', typescript: '>=4.8.4 <6.0.0' } 

Somewhere in the 8.x line utils picked up its own non-optional typescript peer — on utils itself, not a transitive dependency — and 8.46.2 is what I had pinned. Two independent paths to the same 24MB, either sufficient alone. The fix was not a flag. utils had to leave.

Two numbers, and they are not the same number

What stops landing in node_modules. The cuts below keep ~30MB off every consumer of one plugin — typescript 24MB, utils 4.5MB, oxc-resolver 1.5MB.

What ships inside the tarballs. From the registry at both ends, same instrument, 20 packages: 5,432 KB → 3,037 KB, −44.1%. None of the three cuts caused it — a package's own unpackedSize excludes its dependencies. That drop is dead bytes: source maps, AGENTS.md, JSDoc in emitted .js.

Two gains, two axes. Conflating them is the mistake I nearly published.

The three cuts

typescript, 24MB. Used for ts.TypeFlags, integer constants. Integers do not need a compiler. Inlined.

@typescript-eslint/utils, 4.5MB. The second door. Used for one factory function, ESLintUtils.RuleCreator. Ported in-tree.

oxc-resolver, 1.5MB native binary. One of the 19 plugins used it. Now a lazy optional peer; the other 18 never see it.

Before, both were hard dependencies. Now:

"peerDependencies": { "@typescript-eslint/utils": "^7.0.0 || ^8.0.0", "eslint": "^8.40.0 || ^9.0.0 || ^10.0.0", "typescript": ">=4.8.4", "oxc-resolver": "^11.24.2" }, "peerDependenciesMeta": { "@typescript-eslint/utils": { "optional": true }, "typescript": { "optional": true }, "oxc-resolver": { "optional": true } } 

Zero runtime dependencies. Everything real is optional and lazy.

The load path moved too, timed at the pinned commit: devkit cold require 242ms → 13.6ms, 433 modules down to 29. ESLint end-to-end 288 → 216ms; oxlint 320 → 145ms against a 68ms Rust floor.

The part I got wrong

I assumed tree-shaking would help. It cannot: ESLint plugins are CommonJS and nothing bundles them — no build step sits between my dist/ and your node_modules. The lever is not bundle size, it is what evaluates at require time.

The devkit's own unpacked size even went up, 339KB → 377KB, while what you install collapsed. Optimising the npm-page number would have optimised the wrong one — any proxy metric.

One trap: removeComments strips .d.ts docs too, silently killing editor hover. Emit to a scratch directory and copy back only the .js.

Still on the floor: about 49% of rule modules load unused under recommended, because every plugin eagerly requires every rule and the config picks a subset. Lazy loading measured −70ms on one plugin, unshipped.

Check your own tree. npm ls typescript in a project that only installs a linter is an uncomfortable command.

npm install --save-dev @interlace/eslint-devkit 

Its eslint peer is ^8.40.0 || ^9.0.0 || ^10.0.0, so 8, 9 and 10 all work. Every other peer — typescript, @typescript-eslint/utils, oxc-resolver — is optional, which is the whole point: nothing installs a compiler on your behalf.

The devkit is on npm; the 19 plugins on it share one repo.

Why a slow import is a slow lint: where the 45 seconds goes, and what replacing it bought.