Browser and edge deployment

Why AOT matters in the browser

Runtime JIT needs globalThis.Function, which strict Content Security Policy often blocks. It also requires shipping schema builders and compiler code. AOT performs that work during the build and emits only ordinary JavaScript.

// jit/user.jit.ts
const selected = JIT.validator(UserSchema).get("is", "safeParse");

export const User = JIT.compile(UserSchema, {
  is: selected.is,
  safeParse: selected.safeParse,
});

import { User } from "@jit/generated";

Tree-shaking checklist

1. Export only operations used by the application.
2. Prefer standalone exports for route-level granularity.
3. Keep generated package sideEffects: false.
4. Import from generated modules, not @jit-compiler/jit/runtime, in production UI code.
5. Inspect the bundler output, not just source package size.
6. Run jit inspect <export> --stage source when a helper is unexpectedly retained.

A grouped object is convenient but importing one property retains the object and its declared siblings. Standalone functions provide the narrowest bundler unit.

Callbacks and bindings

Regexes and primitive constants are serializable. Arbitrary closures are not. When a query/mapper/refinement captures a runtime callback, generation reports why it was skipped instead of embedding an unsafe or environment-dependent value.

Keep browser AOT declarations declarative. Move application side effects to the consumer of a visitor/iterator rather than capturing them in a generated plan.

Cold start

AOT removes runtime schema traversal and Function compilation. Module parse still has a cost: one giant generated barrel can be slower than route-specific subpath modules. Enable subpaths and import only the feature used by a route.

Memory

• Prefer is for a boolean gate; safeParse allocates issues on failure.
• Use projection before storing UI models.
• Use iterator/visitor when the full result does not need to stay live.
• Avoid binary rowsets for small UI lists; conversion can cost more than the scan.
• Use exact rowsets when multiple batches stay alive; dynamic scratch buffers are for sequential reuse.

Edge runtimes

Generated code has zero imports from JIT and is suitable for locked-down edge workers. Confirm support for APIs your schema uses: Temporal may need a polyfill, File differs across runtimes, and Node-specific constructors should not appear in browser schemas.

Verification

pnpm jit doctor
pnpm jit generate
pnpm build

The repository's AOT tests import generated ESM/CJS modules, execute typed functions and bundle selected exports with esbuild to prove unused operations are removed.