Freerange: Static fit checks for ordinary TypeScript layout code

Freerange shows you the range of every number

in your TypeScript codebase, letting you find potential NaN

, Infinity

, division by zero, out-of-bounds array indexes, and more.

Uses the official TypeScript API. Not a new language, not a fork. No annotations, no library functions.** Static**. Freerange works at compile (build) time, like TS. No need to start your app. AI agents can now guarantee UI layouts without ever touching the browser!Fast. Uses a negligible fraction of TypeScript's analysis time.** Robust**. Adversarially tested by agents against thousands of edge cases.

Freerange is deliberately designed to cater to a useful (and growing) subset of TypeScript, and gives concrete guidance for moving important calculations into that subset, so that your code and math can meet in the middle to unlock the most proof power without much ergonomics drawbacks. AI agents are especially well-suited to refactor such code, and we highly recommend you asking them to do so. However, if you/they do find an unsupported TS feature truly valuable, please file an issue!

bun install --dev @chenglou/freerange

There's no API =). Your TypeScript code provides enough information for Freerange's analysis. We recommend that your agents shape code in the analysis-friendly ways described below.

fr

: print project errors and warningsfr --audit

: print every function's contracts, plus refactor suggestions to help Freerange analyze better. Great for agents

Pass a file path to either command to filter down to just that file's report.

fr

directly uses TypeScript under the hood, so it naturally respects your tsconfig

. We output TS errors before our analysis, so technically, you can swap out your explicit tsc --noEmit

command for fr

and nothing changes!

function gridColumnCount(containerWidth: number) {
  return Math.floor(containerWidth / 240)
}

function gridItemWidth(containerWidth: number) {
  return containerWidth / gridColumnCount(containerWidth)
}

gridItemWidth(200)

bun fr

outputs:

index.ts:9:1 - error [inferred-requirement]: call to gridItemWidth violates its nonzero divisor requirement (division at index.ts:6:10)

How it works: Freerange follows 200

into gridItemWidth

, then through the call to gridColumnCount

. It works out that Math.floor(200 / 240)

is 0

, then catches the later division by that result. TypeScript only knows that these values are numbers; Freerange follows their ranges through both functions.

Did you know that console.log

has a lesser-known sibling: console.assert

? When the assertion is true, it stays silent. When the assertion is false, it reports a failure.

By itself, console.assert

isn't as universally useful as console.log

. Freerange changes that by analyzing console.assert

statically:

export function itemColumn(itemIndex: number, columnCount: number): number {
  console.assert(Number.isInteger(columnCount))
  console.assert(columnCount >= 1)

  const index = Math.max(0, Math.floor(itemIndex))
  const column = index % columnCount

  console.assert(column < columnCount)
  return column
}

In the example above, calling itemColumn(0, 2.2)

produces an error (columnCount

should be an integer) at compile time, not at runtime! No need to start a browser to know that the code is wrong here.

console.assert

calls at the very beginning of a function, before any other statement, are caller requirements. Like parameter types, every caller must satisfy them. Any console.assert

later in the function will be proven by Freerange for the function itself. Otherwise, Freerange reports an error.

For simplicity and predictability, console.assert

currently works only in named top-level functions and accepts simple numeric checks:

Number.isInteger

,Number.isFinite

,Number.isNaN

• Strict comparisons ( ===

,!==

,<

,>

,<=

,>=

) using number literals, object paths, andarray.length

• References to module constants. For caller requirements, the constant must resolve to a numeric literal

We also don't support aliasing console.assert

, e.g. const assert = console.assert

.

For more complex assertions, like inline calculations, extract them into variables:

const availableWidth = frame.right - frame.left
console.assert(availableWidth >= 0)

(You can strip console.assert

in production with bundler or Bun's drop feature, as you may already be doing.)

There are infinitely many assertable things. Here are some good, non-noisy ones:

• Guarantee that two UI items don't overlap:

console.assert(input.bottom < content.top)

• Guarantee that a virtualized list never renders more items than intended:

const visibleItemCount = endIndex - startIndex
console.assert(visibleItemCount <= MAX_VISIBLE_ITEM_COUNT)

• Ensure that two separately calculated values are equal:

const frame = {
  input: {bottom: inputBottom},
  inputTray: {bottom: inputBottom},
}
console.assert(frame.inputTray.bottom === frame.input.bottom)

Every plain number

parameter, including a number field in a fixed-shape object parameter, already requires a finite, non-NaN

value. Freerange also checks whether a divisor may be 0

and the other conditions shown by fr --audit

. You don't need to assert the same information explicitly.

Freerange supports a subset of TS:

• Named, synchronous top-level functions in a file; Freerange follows calls between functions in the same file
• Numbers, booleans, strings, nullable values, plain objects, tagged unions, dense arrays, and fixed tuples if

/else

, ternaries, non-fallthroughswitch

,&&

,||

,!

,??

,for

,while

, andfor...of

loops- Arithmetic, comparisons, object field and array reads, selected Math

operations, andNumber.isInteger

,Number.isFinite

, andNumber.isNaN

Freerange could theoretically support a much larger subset of TS, and did before its public release. Those patterns often made numeric inference and proofs much harder and slower, however, and some questions are undecidable in general. Now that AI agents write code, we strongly recommend asking agents to refactor important calculations into shapes that Freerange analyzes well, guided by fr --audit

. Code that is easy to analyze tends to resemble functional programming: immutable data, explicit inputs and outputs, and clean, direct control flow.

Put important calculations in small named functions. A React component, callback, or async function can call a plain synchronous helper. Keep any helper functions that Freerange needs to inspect in the same file.

export function fittedImageHeight(frameWidth: number, imageWidth: number, imageHeight: number): number {
  const width = Math.max(1, imageWidth)
  const height = Math.max(1, imageHeight)
  return (frameWidth * height) / width
}

function ImageCard(props: {frameWidth: number; imageWidth: number; imageHeight: number}) {
  const height = fittedImageHeight(props.frameWidth, props.imageWidth, props.imageHeight)
  return <img style={{height}} />
}

Name a calculation before checking it. If the divisor isoldMax - oldMin

, writeconst oldSpan = oldMax - oldMin

, checkoldSpan === 0

, and divide byoldSpan

. CheckingoldMin === oldMax

does not tell Freerange about the separately calculatedoldSpan

. Audit code:[guard-derived-value]

. - Decide how invalid inputs should be handled before using them. IfcolumnCount

must be a positive integer, either require that with leadingconsole.assert

calls or normalize it withMath.max(1, Math.floor(columnCount))

. Only normalize when the application actually wants that runtime behavior. Audit code:[encode-input-rule]

. - Choose the order of arithmetic deliberately. Formulas that are equivalent on paper can round, overflow, or underflow differently with JavaScript numbers. For example,frameWidth / (imageWidth / imageHeight)

introduces a ratio that can round to zero;(frameWidth * imageHeight) / imageWidth

avoids that particular problem. The two expressions can still round differently. Audit code:[use-direct-operands]

. - Decide what a missing array element means. Usevalues[index] ?? fallback

only when the application really wants a fallback. Otherwise, prove thatindex

is an integer from zero throughvalues.length - 1

before usingvalues[index]!

. A bounds check cannot detect a hole in a sparse array, so Freerange expects arrays to be dense. Audit codes:[handle-missing-element]

,[guard-array-index]

. - Write the condition and loop directly. Preferwidth === 0

over using a number as a condition, such aswidth || 1

. Use a regular loop for a simple dense-array calculation when callback arguments, callback effects, and a newly allocated result array do not matter. Audit codes:[write-explicit-condition]

,[use-loop-for-aggregation]

. - Write object copies explicitly. Use{width: layout.width, height: layout.height}

instead of{...layout}

. Use dense arrays and fixed-length tuples, and do not modify objects or arrays after creating them. Rebuilding an object is not equivalent to mutation when other code observes its identity or the mutation. - Give each union case a tag and switch on it. Use a tagged union when different cases carry different fields. Make theswitch

exhaustive and do not use fallthrough. - Check and use the same local value. When a value comes from module, class, or reactive state, store it in a local first. For example, writeconst currentScale = scale; if (currentScale !== null) return currentScale

instead of checking one read ofscale

and returning another. - Use precise TypeScript types. Avoidany

, casts, and suppression comments. Parse external data before passing it to a numeric helper, give the helper typed parameters, and pass only the fields it uses. A file containing@ts-ignore

,@ts-expect-error

,@ts-nocheck

, oreval

is rejected because its declared types cannot be trusted.

The only TypeScript compiler option that Freerange mandates is strictNullChecks

(otherwise the analysis is too unsafe), which is enabled when strict

is on. We generally recommend enabling the options below as well. They aren't necessary for Freerange's analysis, but they help AI agents and humans write safer code that is more likely to be analyzable:

{
  "compilerOptions": {
    "strict": true,
    "noImplicitAny": true,
    "noUncheckedIndexedAccess": true,
    "exactOptionalPropertyTypes": true,
    "noImplicitReturns": true,
    "noFallthroughCasesInSwitch": true
  }
}

Freerange uses a few terms consistently:

requires

: a condition the caller must satisfy. The function's guarantees assume the condition is true.ensures

: a guarantee about the returned value whenever the function returns.assumes

: an input condition Freerange accepts without proving, such as an array being dense or every element of anumber[]

being finite.proves

: a successful staticconsole.assert

check.unsupported

: Freerange cannot analyze the function because it uses code outside the analyzed subset. Freerange shows the first blocker you can potentially refactor.partially supported

: Freerange can analyze some, but not all, of the function.skipped

: some top-level statements in the modules weren't analyzed.

A caller requirement is not automatically a bug. For example, requires: columns >= 1

means the function is safe under that condition; it does not mean Freerange found a caller passing zero. Freerange checks supported same-file calls, but it is not a repository-wide call-site verifier. Imported calls and unsupported callers may remain unchecked.

An ensures

line assumes its requires

and assumes

. A requirement may be a real API rule, or it may expose a relationship Freerange cannot currently prove. An assumption may identify a real input boundary or an analysis limitation. Decide what the program should do before changing code to remove either one.

Always read the coverage line. No findings does not mean an unsupported file is safe. A derived guarantee becoming weaker, for example at least 54

becoming at least 0

, appears in the audit rather than the shorter findings output.

If your production code actually needs support beyond these limits, please file an issue! We're open to relaxing the limits.

Freerange's numeric analysis is designed around arithmetic used in layouts and other everyday application code. For each TypeScript number

, Freerange tracks one continuous range, whether the value is an integer, whether it may be NaN

or infinite, and at most one exact number that a branch proved impossible. For example, after value !== 0

, Freerange can remember that value

cannot be 0

.

Freerange does not keep separate ranges or arbitrary sets of possible numbers. If one branch produces 1..2

and another produces 10..11

, Freerange keeps the combined range 1..11

. A later check that rules out a different exact number may replace the number remembered from an earlier check.

Freerange recognizes repeated uses of one already-computed value, including aliases, stable property and array reads, lengths, and duplicate arguments passed to a supported function in the same file. For example, const span = right - left; span - span

is exactly 0

unless span

is infinite or NaN

. Two separate evaluations are not assumed equal, even when their source code looks the same; store the result in a local when that identity matters.

Earlier versions included an exact rational linear prover based on Farkas' lemma and other relational machinery. The current analyzer does not use that machinery or an SMT solver. In practice, those approaches made analysis unpredictable without proving much more real-world code. We also decided against analyzing numbers as real numbers, which would have been sweet, because doing so produces false proofs: floating-point arithmetic is not associative, rounds results, and can overflow or underflow.

Freerange follows calls to supported functions in the same file using the ranges known at each call. Imported functions are not followed. Imported constants are followed only when they resolve to a numeric literal such as export const GAP = 24

. Literal default parameters work in supported calls; object and calculated defaults do not. Passing more arguments than the implementation declares is also unsupported. Unknown calls, when callbacks run, caught exceptions, changes made through another reference to the same object, and most framework behavior remain outside the subset.

Inside a console.assert

, Freerange can follow common UI calculations through named values: Math.min

and Math.max

, addition or subtraction by a nonnegative value, the same multiplication by a nonnegative value on both sides of a comparison, the fact that index % columnCount < columnCount

when columnCount

is positive, and fields read from a freshly constructed record. Freerange does not chain arbitrary comparisons: left <= middle

and middle <= right

do not by themselves prove left <= right

.

Freerange analyzes a loop again until the ranges known at the start of an iteration stop changing. Freerange does not simulate every runtime iteration or try to produce a formula for the final value. Ordinary counting loops usually settle after two or three analysis passes. If the ranges still change after 16 passes, analysis stops for that path.

Freerange follows records, tuples, arrays, and tagged unions declared in your project through at most eight nested levels. A deeper property becomes unknown. A function is unsupported when its top-level input type cannot be represented.

Property reads are assumed to be stable and side-effect-free during one analyzed function call. A getter or Proxy that changes its answer or performs work is outside the model. Property writes, including assignments that invoke setters, are unsupported. Object spread is also unsupported because JavaScript copies only an object's own enumerable properties, which may not match the fields declared by its TypeScript type.

Every plain number

parameter must be finite and not NaN

. The same rule applies to number fields in fixed-shape object parameters, even when the function does not read them. Numeric literal types such as 1 | 2

already satisfy the rule. Nullable numbers, arrays, tuples, and tagged unions keep their more specific assumes

lines instead. A supported literal default is used when an argument is omitted, so the default can satisfy the requirement automatically.

Calls to supported functions in the same file either prove these requirements, pass them outward to their own callers, or report a definitely invalid argument. A successful call also proves that the same stored argument is finite afterward. Writing console.assert(Number.isFinite(value))

at the start of the function is allowed but normally redundant; Number.isInteger(value)

is stronger and replaces the finite requirement.

Division and array reads can create additional requirements. Freerange tries to express each one using the function's parameters so callers can be checked. A later operation on the same stored value and path may rely on that requirement; assigning a new value or repeating the calculation starts over. Freerange follows each intermediate calculation at most once. If one pass cannot reach the parameters, Freerange prints a local assumes

condition instead.

These limits may make a result less precise or stop analysis, but they cannot make a guarantee stronger than the code supports.

bun install
bun run check

source & further reading

github.com — original article