advanced35 minby DevFox

Interop with Untyped Code

Migrating incrementally, containing any at a boundary, typing a legacy module from the outside, and measuring progress so a migration actually finishes.

TypeScript

The migration has been running for fourteen months. Roughly half the files are TypeScript, the strict flags are off because too much fails, and nobody can say whether it is closer to finished than it was in March. Meanwhile every new file imports something untyped, so the checking that exists is partly illusory.

Most real adoption happens like this — inside a codebase that already works, alongside people shipping features. By the end of this lesson you will know how to contain untyped code rather than let it spread, and how to measure a migration so it converges.

Two ways to be untyped

They need different responses.

JavaScript files in your own project. You control them, they can be converted, and the goal is to convert them.

Dependencies with no types. You do not control them, they will not be converted, and the goal is to describe the parts you use — the practice course covered declare module for exactly this.

The first is a migration. The second is a boundary, permanently.

Letting the checker see JavaScript

json

{
  "compilerOptions": {
    "allowJs": true,
    "checkJs": false
  }
}

allowJs lets TypeScript files import JavaScript ones, so the checker knows something about them from inference. checkJs turns on checking of the JavaScript itself, which on a large codebase is thousands of errors at once.

Enable it per file instead:

javascript

// @ts-check
export function makeCaption(filename) { ... }

That one comment gives a file real checking, and JSDoc supplies the types it cannot infer:

javascript

/**
 * @param {string} filename
 * @param {{ separator?: string }} [options]
 * @returns {string}
 */
export function makeCaption(filename, options = {}) { ... }

JSDoc types are checked exactly like annotations — the same type system, different syntax. For a file that must stay JavaScript, this is a genuine option rather than a consolation.

Containing any at the boundary

Bad — importing untyped code directly.

typescript

import { processOrder } from "./legacy/orders.js";

const result = processOrder(order);
const total = result.total * 1.2;
saveInvoice({ amount: total, customer: result.customer.id });

Good — one typed module in front of it.

typescript

// src/legacy/orders.ts
import { processOrder as raw } from "./orders.js";

export type ProcessedOrder = {
  total: number;
  customer: { id: string };
};

export function processOrder(order: Order): ProcessedOrder {
  return OrderResultSchema.parse(raw(order));
}

In the first version result is any, so result.total, everything computed from it, and everything passed onward are all unchecked — a typo in customer.id compiles, and the contamination reaches modules that never touched the legacy code. That is the spreading any from the practice course, and an untyped dependency is its most common source.

The wrapper is the only file that touches the untyped module. The rest of the codebase imports a typed function, and — because this is a runtime boundary — the schema actually verifies the shape rather than asserting it.

The general rule: an untyped module gets exactly one typed neighbour, and nothing else imports it. A lint rule can enforce that, using the layering technique from the practice course.

Migrating a file

An order that keeps the build green throughout.

Convert leaves first
A file with no local imports has nothing to fight. Working inward means every file you convert already has typed dependencies.
Rename, then fix
.js to .ts produces a list of errors, and that list is the actual work. Resist adding types while renaming — two tasks, two very different review sizes.
Reach for unknown before any
Where a type is genuinely not known yet, unknown forces a check at the point of use instead of spreading.
Leave a marked suppression
@ts-expect-error with a ticket number, never a silent one.

Every step here is chosen to keep the pull request small enough that somebody will review it

That last one, in full:

@ts-expect-error fails when there is nothing to suppress, so it cannot outlive the problem — unlike @ts-ignore, which stays forever.

Never @ts-nocheck. It turns off checking for a whole file, and the file then looks converted while being entirely unchecked. Every migration that stalls has some of these.

Measuring convergence

A migration without a number does not finish, because nobody can tell whether a given week helped.

Print those in CI on every build. Four numbers, three of which should go down, and a trend that is visible without anyone running an audit.

Two rules that make the trend hold:

No new JavaScript files. A lint rule or a CI check, so the denominator stops growing.

Convert what you touch. A feature in an untyped file converts it first. That distributes the work across the people already reading that code, which is far more effective than a migration team working through a list.

type-coverage gives a single percentage, which is useful for a target — and the raw counts are better for seeing which category is stuck.

Turning strictness on as you go

The strict-mode adoption from the practice course applies, and combines usefully here:

Both in CI. The strict list only grows, a converted module joins it, and a new module starts on it. That gives you a second number — how much of the codebase is strictly checked — which is the one that actually corresponds to safety.

A cheat sheet to keep

Where to go next

That closes Advanced TypeScript, and the catalog with it.

You started this track unable to write the language and now know its type system as a programmable thing: conditional and mapped types, template literals, inference and variance, declaration files, library authorship, the compiler API, and what all of it costs.

The question running through this course is the one worth keeping. What does this cost, and who pays it? A recursive conditional type costs everyone who reads the error. A (...args: any[]) => any wrapper costs every call site its checking. A metaclass-equivalent — a decorator hiding behaviour — costs whoever has to find it. A brand without a checked constructor costs you false confidence. None of those are reasons not to; they are the second half of a decision usually made with only the first half in view.

The other thread is narrower and more practical: the checker finishes before your program starts. Everything about validation at boundaries, branded constructors, schemas over assertions, and unknown over any follows from that one fact.

Before you go, pick the most elaborate type in a codebase you work on and try to state what it costs and who pays. If the answer comes easily, it was probably a good decision. If it takes a while, that is worth knowing too — and it is the most useful thing this course can leave you with.