| name | ocaml2moonbit-migration |
| description | Guide for migrating OCaml projects, libraries, modules, and test suites to idiomatic MoonBit. Use when translating OCaml code to MoonBit, planning a large OCaml-to-MoonBit port, preserving byte/string-heavy behavior, replacing OCaml variants/records/exceptions/refs/arrays, mapping OCaml APIs to MoonBit packages, or building verification and test strategy for a migration. |
OCaml to MoonBit Migration
Port behavior, data invariants, and public contracts first. Translate syntax only after the source semantics are classified. The most common porting bug is silently coercing OCaml string (a byte sequence) into MoonBit String (UTF-16 text); classify every field by meaning before choosing a type.
When in doubt, probe with moon run -c '...'. Probes in this guide were verified on moon 0.1.20260512-class toolchains; rerun the relevant probe if the local toolchain is newer and the behavior is load-bearing.
Migration Workflow
- Inventory the OCaml module boundary: public types, functions, exceptions, optional arguments, mutable state, lazy/deferred state, C/Unix/filesystem dependencies, and existing tests or golden fixtures.
- Classify every OCaml
string by meaning before choosing a MoonBit type. Do this field by field, even inside the same OCaml record.
- Choose MoonBit package boundaries and imports before coding. Add imports to
moon.pkg; MoonBit source files do not use OCaml-style open.
- Port one behavioral slice at a time with tests. Prefer a thin public API skeleton, then fill parser/serializer/algorithm internals behind it.
- Probe uncertain language or library behavior with
moon run -c and, when needed, a small OCaml toplevel probe. Keep probes minimal.
- Finish each slice with
moon check, moon test, moon info, and moon fmt. You can add more warnings moon check --warn-list +... to be more strict.
Type-Mapping Cheatsheet
| OCaml use | MoonBit default |
|---|
| binary payload, file contents, compressed/encrypted/checksummed data, parser input | Bytes or BytesView |
| human-readable text, diagnostics, labels that are truly Unicode | String |
| single byte with known 0..255 range | Byte |
| indexes, counts, small identifiers, deliberate signed 32-bit wrapping | Int |
| file offsets, serialized positions, large object numbers | Int64 or UInt64 |
OCaml array with fixed length | FixedArray[T] |
| mutable growable builder | Array[T] |
| read-only sequence parameter | ArrayView[T] or BytesView |
| compile-time lookup table (literal or comprehension) | ReadOnlyArray[T] |
| keyed lookup with deterministic iteration | Map[K, V] |
OCaml Buffer.t byte builder | @buffer.Buffer |
OCaml ref | Ref[T] |
| OCaml variant | enum, often priv enum for internal states |
| OCaml record | struct, with { ..old, field: value } for immutable update |
| OCaml exception flow | suberror plus checked raise |
| lazy/deferred state | Lazy[T] |
OCaml int32 needing wrapping | Int |
OCaml int32 needing wide arithmetic | Int64 or UInt64 |
OCaml float | Double |
There is no Int32 type in the current toolchain.
Bytes vs String
OCaml String.length counts bytes; MoonBit String::length() counts UTF-16 code units. The same source character takes different positions in the two languages, and that difference is silent.
ocaml -noprompt -noinit <<'EOF'
let s = "𝄞";;
Printf.printf "%d\n" (String.length s);;
Printf.printf "%d\n" (Char.code s.[0]);;
EOF
moon run -c 'fn main { let s = "𝄞"; println(s.length()); println(s.char_length()) }'
String::length() is UTF-16 code units, String::char_length() is Unicode scalars, and neither is bytes. For byte-oriented formats use Bytes/BytesView. Convert between Bytes and String only through a named encoding helper (@ascii.encode, @utf8.encode, etc.) that documents the encoding assumption.
moon run -c 'fn main { let raw : Array[Byte] = [65, 0, 255]; let b = Bytes::from_array(raw); println(b.length()); println(b[2].to_int()) }'
moon run -c 'fn main { let source : Array[Byte] = [1, 2]; let bytes = Bytes::from_array(source); source[0] = 9; println(bytes[0].to_int()); println(source[0].to_int()) }'
Bytes::from_array produces an immutable owned Bytes. Mutating the source array afterwards does not affect the frozen value. Port OCaml Bytes mutators either as BytesView -> Bytes transforms, or keep state in Array[Byte]/FixedArray[Byte] until the final freeze.
moon run -c 'fn main { let empty = Bytes::new(0); let zeros = Bytes::new(2); println(empty.length()); println(zeros.length()); println(zeros[0].to_int()) }'
Bytes::new(length) always zero-fills and requires an explicit length. Use it where OCaml Bytes.empty or Bytes.make len '\000' would have appeared.
Bytes::copy is deprecated because Bytes is immutable. When a port must materialize a fresh physical copy (e.g. to preserve an OCaml promise that no buffer is shared), use Bytes::makei(len, i => src[i]).
moon run -c $'let cached_filter = b"/Filter"\nfn main { let b = b"PDF"; println(b.length()); println(b[0].to_int()); println(cached_filter.length()); println(cached_filter[0].to_int()) }'
b"..." is a compile-time byte-string literal of type Bytes. Use it for ASCII format syntax (b"/Filter", b"PDF", etc.) — no @ascii import, no runtime call, no type annotation needed at top level. Reserve @ascii.encode(text) for String → Bytes conversion when the source is a dynamic String. Do not build binary file formats by String concatenation.
moon run -c $'fn main { let buf = @buffer.new(); buf.write_bytes(b"PDF"); buf.write_byte(10); let bytes = buf.contents(); println(bytes.length()); println(bytes[0].to_int()); println(bytes[3].to_int()) }'
The canonical MoonBit byte-builder is @buffer.Buffer (from moonbitlang/core/buffer). Construct with @buffer.new() (optionally size_hint=N), append static ASCII via buf.write_bytes(b"..."), single bytes via buf.write_byte(n), dynamic text via buf.write_bytes(@ascii.encode(text)), binary payloads via buf.write_bytes(view), then freeze with buf.contents(). Reserve Array[Byte] plus Bytes::from_array for cases that need random-access mutation of in-flight bytes; Buffer is the OCaml Buffer.t analogue.
moon run -c 'fn main { let s = "/UniJIS-UCS2-H"; println(s.has_prefix("/Uni")); println(s.contains("-UCS2-")); println(s.has_suffix("-H")); println(s[1:]); println(s[1:].to_owned()) }'
String::has_prefix, has_suffix, and contains cover predicate work. Slice with s[start:end] — like BytesView, this returns a borrowed StringView, cheap and good for inspection/pattern matching. When the callee needs an owned String (e.g. for storage or a String parameter), use s[start:end].to_owned(). Offsets are UTF-16 code-unit offsets, not bytes; use these only for ASCII-validated tokens or genuinely textual parsing.
Bytes Views
moon run -c 'fn has_ab(view : BytesView) -> Bool { match view { [65, 66, ..] => true; _ => false } }
fn main { let bytes : Bytes = [65, 66, 67]; let view = bytes[:2]; println(view.length()); println(view[0].to_int()); println(has_ab(bytes[:])); println(has_ab(view)) }'
BytesView is the byte-sequence counterpart to ArrayView. Views are cheap slices, expose read-only byte operations, and support pattern matching with rest patterns. Prefer BytesView for read-only byte APIs; call .to_owned() only at explicit ownership boundaries. BytesView::to_owned() returns the original bytes for a whole view but allocates and copies for a partial slice.
Bytes is assignable to a BytesView parameter, so a single API can accept owned or borrowed input. A returned BytesView keeps its backing Bytes alive across function boundaries — useful for exposing decode/decrypt results without forcing a copy.
The reverse direction is not automatic: a BytesView does not type-check where owned Bytes is required, including in native extern declarations. Treat that as the ownership boundary and call .to_owned() deliberately. Equality works when the view is the left operand; if Bytes is on the left, slice it (bytes[:]) before comparing.
moon run -c 'fn main { let data : Bytes = [60, 65, 62, 0, 12, 60, 66, 62]; println(data.length()); println(data[3].to_int()); println(data[4].to_int()); let view = data[5:]; match view { [60, 66, 62] => println("match"); _ => println("miss") } }'
NUL (0) and form-feed (12) are ordinary bytes in Bytes/BytesView. Port OCaml byte predicates literally; do not narrow a format whitespace predicate to a string-oriented space/tab check.
Integers
moon run -c 'fn main { let (b, u, u16, i64, u64) : (Byte, UInt, UInt16, Int64, UInt64) = (255, 7, 65535, 42, 42); println(b.to_int()); println(u.to_string()); println(u16.to_int()); println(i64.to_string()); println(u64.to_string()) }'
Scalar types: Byte, Int16, UInt16, Int, UInt, Int64, UInt64, Float, Double. There is no Int32; use Int for deliberate signed 32-bit wrapping, or Int64/UInt64 when an OCaml int32 value must be represented without truncation. When calling a method on an integer literal, parenthesize: (65).to_byte(), not 65.to_byte() — 65. parses as the start of a float.
moon run -c 'fn main { let max = 2147483647; println(max + 1); println(1 << 31); println(0xF0 & 0x0F); println(0x80 >> 7); println(0xAA ^ 0xFF) }'
Int arithmetic wraps modulo 2^32 with signed interpretation. Symbolic operators map directly from OCaml: & (land), | (lor), ^ (lxor), << (lsl), >> (right shift, arithmetic for signed).
moon run -c 'fn main { println((-8) >> 1); println(1 << 32); println(1 >> 32); let logical = ((-8).reinterpret_as_uint() >> 1).reinterpret_as_int(); println(logical) }'
Int right shift is arithmetic, and shift counts are masked to 5 bits, so 1 << 32 == 1. For OCaml Int32.shift_right_logical, reinterpret signed→unsigned, shift, reinterpret back. Numeric conversion (.to_uint()) is not the same as bit reinterpretation (.reinterpret_as_uint()).
moon run -c $'fn main { try @string.parse_int("2147483648") catch { _ => println(true) } noraise { _ => println(false) }; let mut value = 0; for digit in [50,49,52,55,52,56,51,54,52,56] { value = value * 10 + digit - 48 }; println(value); let mut wide = 0L; for digit in [50,49,52,55,52,56,51,54,52,56] { wide = wide * 10L + (digit - 48).to_int64() }; println(wide.to_string()) }'
@string.parse_int rejects out-of-range decimals, but handwritten digit accumulation in Int silently wraps. Accumulate offsets, lengths, object numbers, or serialized counters in Int64/UInt64; bounds-check before narrowing to Int.
moon run -c 'fn main { println(0x8EA2A1A1 < 0); println(0x8EA2A1A1); println(0x7FFFFFFF < 0x8EA2A1A1); println((-1) % 256) }'
Int literals above 0x7FFFFFFF are negative. A table sorted by unsigned byte order is not sorted for signed Int comparison — compare through UInt (reinterpret_as_uint) or store as Int64/UInt64. % preserves the sign of the left operand; normalize (a - b) mod 256 style expressions into 0..255 before converting to Byte. For byte-codec arithmetic that multiplies by large radices, promote Byte to UInt64; UInt64::to_int() truncates to signed 32 bits for large values.
moon run -c 'fn rotr64(value : UInt64, bits : Int) -> UInt64 { (value >> bits) | (value << (64 - bits)) }
fn main { let value : UInt64 = 0x0123456789abcdef; let full : UInt64 = UInt64::lnot(0); println(rotr64(value, 8).to_string()); println((full + (1 : UInt64)).to_string()); println(((0x80 : UInt64) >> 7).to_string()) }'
UInt64 arithmetic wraps modulo 2^64. Unsigned right shifts are logical. Provide explicit UInt64 context (x : UInt64) when the literal exceeds signed Int range.
moon run -c 'fn main { println(Int::lnot(0)); println((0x0F).lnot()); println((!true).to_string()) }'
Bitwise complement: Int::lnot(value) or value.lnot(). Boolean negation: !expr. Do not port OCaml not x literally.
Floats
ocaml -noprompt -noinit <<'EOF'
print_endline (string_of_float 2.);;
print_endline (string_of_float infinity);;
EOF
moon run -c 'fn main raise { println((2.0).to_string()); println((1.0 / 0.0).to_string()); let tiny : Double = @string.from_str("1e-10"); println(tiny.to_string()) }'
Double::to_string() is not a drop-in for OCaml string_of_float or Printf "%.12g". Whole floats render as 2 (not 2.); infinities render as Infinity/-Infinity (not inf/-inf); small values use scientific notation. When a digest, file format, or snapshot depends on the exact spelling, port the OCaml formatter as an explicit boundary function. Overflow produces Infinity; division by infinity produces 0. Cover those edges with a probe when porting defensive float branches.
moon run -c 'fn main { println(@math.pow(0.5, 2.0)); println((4.0).sqrt()); println((-3.0).abs()); println((5.5).mod(2.0)) }'
There is no ** operator — use @math.pow(base, exponent). Common helpers are Double methods: value.sqrt(), value.abs(), value.mod(other). Free functions live in @math: sin, cos, atan2, ln, log10, floor, ceil, round, trunc, exp. Note @math.sqrt/@math.abs are not valid — those are method-only. Probe with moon ide doc @math before assuming a path.
Data Structures
Map OCaml variants to MoonBit enum. Derive Debug, Eq for types that will be inspected in tests.
moon run -c 'enum E { A(Int) } derive(Debug)
fn E::value(self : E) -> Int { match self { A(n) => n } }
fn main { let value = A(7); println(value.value()) }'
moon run -c 'enum E { A(Int); B(Int) } derive(Debug)
fn main { let xs : Array[E] = [E::A(1), E::B(2)]; println(xs.length()) }'
When code crosses package boundaries or builds heterogeneous-looking enum arrays, prefer explicit Type::Constructor forms or a concrete element-type annotation. This keeps moon check --warn-list +73 useful — strip annotations only when the compiler reports them as truly unnecessary.
moon run -c 'priv enum Section { NoneSection; ActiveSection }
fn choose(flag : Bool) -> Section { if flag { ActiveSection } else { NoneSection } }
fn main { match choose(true) { ActiveSection => println("active"); NoneSection => println("none") } }'
moon run -c 'priv struct S { value : Int }
fn make() -> S { { value: 1 } }
fn main { let s = make(); println(s.value) }'
Use priv enum/priv struct for internal helpers that should not appear in the package interface. Avoid deriving traits on private helpers unless a debug path actually uses the implementation.
moon run -c 'struct S { a : Int; b : Int? } derive(Debug, Eq)
fn main { let s = { a: 1, b: None }; let t = { ..s, b: Some(2) }; match t.b { Some(value) => println(value); None => println(0) } }'
Record update is { ..old, field: value }, the direct replacement for OCaml { old with field = value }.
moon run -c 'struct S { values : Int } derive(Debug)
fn make(values : Int) -> S? { Some({ values, }) }
fn main { println(make(7).unwrap().values) }'
For single-field struct literals using field shorthand inside another expression, write a trailing comma ({ values, }) to silence the ambiguous-block warning.
When an OCaml function returns a broad variant and later code relies on an informal invariant, introduce a narrow private MoonBit enum for the post-checked state. This makes impossible fallback branches explicit at the type boundary rather than carrying broad Object-like values through the rest of the port.
Lookup Tables
moon run -c 'fn range_lookup(ranges : ArrayView[(Int, Int, Int)], key : Int) -> Int? { for range in ranges { if key >= range.0 && key <= range.1 { return Some(range.2 + key - range.0) } }; None }
fn main { let ranges = [(0x21, 0x23, 100), (0x30, 0x30, 200)]; println(range_lookup(ranges, 0x22).unwrap()); println(range_lookup(ranges, 0x24) is None) }'
For large OCaml mapping tables, prefer compact ArrayView-accepted range tables ((first, last, base) plus small exception/sequence tables) over mechanically expanding every entry. Generated MoonBit sources stay smaller and native-target validation runs faster, with the same deterministic lookup behavior.
moon run -c $'let table : ReadOnlyArray[Int?] = [for i in 0..<8 => if i == 5 { Some(10) } else { None }]\nfn lookup(i : Int) -> Int? { if i >= 0 && i < table.length() { table[i] } else { None } }\nfn main { println(lookup(5).unwrap()); println(lookup(0) == None); println(lookup(99) == None); println(table.length()) }'
For dense lookup domains, define a private top-level let table : ReadOnlyArray[T?] = [...]. ReadOnlyArray[T] is the read-only counterpart of Array[T]; an array literal coerces directly, and list comprehensions also produce ReadOnlyArray[T] — so [for i in 0..<N => entry(i)] is the idiomatic way to populate a compile-time table without makei ceremony. Keep the bounds check at the lookup boundary.
Mutation, Refs, Arrays
moon run -c $'fn first_sorted(xs : ArrayView[Int]) -> Int { let ys = xs.to_owned(); ys.sort(); ys[0] }\nfn main { let r : Ref[Int] = Ref::{ val: 0 }; r.val += 1; println(r.val); let fixed : FixedArray[Int] = [3, 1, 2]; let grow = [4, 2, 3]; println(first_sorted(fixed)); println(first_sorted(grow)); println(fixed[0]); println(grow[0]) }'
Ref[T] for primitive mutability; mut fields inside structs for larger state. Array is a growable mutable vector; FixedArray matches OCaml array more directly. ArrayView[T] is the read-only sequence parameter type — callers can pass Array, FixedArray, or another view without copying. .to_owned() materializes a mutable Array at explicit ownership boundaries (e.g. before sort()/sort_by()); avoid it in hot paths. Represent lazy/deferred states as explicit enum variants instead of nested mutable containers.
moon run -c 'fn main { let xs = [(2, "b"), (1, "a")]; xs.sort_by(fn(a, b) { a.0.compare(b.0) }); println(xs[0].0); let ys = [1, 2, 3]; let zs = ys.rev(); println(zs[0]); println(ys[0]) }'
Array::sort_by sorts in place with an Int-returning comparator. Array::rev() returns a reversed copy and leaves the original unchanged — useful when porting OCaml list-building code that conses in reverse and conditionally applies List.rev.
moon run -c $'fn main { let m : Map[Int, String] = Map([]); m[3] = "three"; m[1] = "one"; m[2] = "two"; for k, v in m { println("\\{k}=\\{v}") } }'
For keyed lookup, prefer the built-in Map[K, V] — it is a linked hashmap that preserves insertion order, so iteration is deterministic and matches the order keys were written. Construct with Map([]) (or Map([], capacity=N)); Map::new() is deprecated. Reach for @hashmap.HashMap[K, V] (from moonbitlang/core/hashmap) only when insertion order is irrelevant and you want a marginally cheaper structure. Use ordered arrays of pairs only when you need duplicate keys or sequence-style processing.
Errors
OCaml exceptions (Not_found, End_of_file, Invalid_argument, domain errors) should not be copied as unchecked control flow.
moon run -c $'suberror E\nfn make(flag : Bool) -> ((Int) -> Int?) raise E { if flag { raise E } else { fn(x) { Some(x) } } }\nfn main raise { println(make(false)(3).unwrap()) }'
moon run -c $'fn parse(s : String) -> Int raise { @string.parse_int(s) }\nfn helper() -> Int raise { parse("7") }\ntest "raising test body can propagate" { let value = parse("7"); @test.assert_eq(value, 7) }\nfn main raise { println(helper().to_string()) }'
Define a project-level suberror once the first fallible functions are ported. Fallible functions declare raise ProjectError or plain raise. main itself can be declared fn main raise { ... }, which is the cleanest way to write probes and small entry points that call fallible APIs — no try! boilerplate. Inside a raising function, raising test helper, or test body, do not add explicit try! around every fallible call; checked errors propagate from those contexts automatically. Test helpers that can call fail must declare raise, even when only called from test blocks.
When a raising function returns another function, parenthesize the function type before raise: -> ((A) -> B) raise E, not -> (A) -> B raise E (the latter binds raise to the returned function type).
moon run -c $'fn may_fail(flag : Bool) -> Unit raise { if flag { fail("boom") } }\nfn main { try may_fail(true) catch { _ => println(true) } noraise { _ => println(false) } }'
moon run -c $'fn parse(s : String) -> Int raise { @string.parse_int(s) }\nfn main {\n try parse("7") catch {\n _ => println(0)\n } noraise {\n value => println(value)\n }\n}'
try expr catch { ... } noraise { value => ... } is the OCaml try ... with ... analogue with explicit success and error branches. When only propagating an error, call the fallible function directly from a raise context.
To catch one constructor and propagate the rest while keeping the project error type, use expr catch { SpecificError => fallback; error => raise error }. catch arms are checked for exhaustiveness; do not omit the final propagating arm. When porting OCaml exception tests, use a wildcard catch arm for "this should raise"; match a specific variant only when that variant is itself the contract under test.
Do not convert checked errors into generic success/failure values for control flow. Prefer try f() catch { ... } noraise { ... } when both branches matter, or a plain call from a raise context when the error should propagate.
Functions and Callbacks
moon run -c 'fn helper(x : Int) -> Int { x + 1 }
fn main { println(helper(1)) }'
Top-level helper functions (anything other than main) need explicit return-type annotations. This applies to moon run -c probes too.
moon run -c 'fn[A, B, C] apply_pair(f : (A, B) -> C, a : A, b : B) -> (C, A) { (f(a, b), a) }
fn main { let result = apply_pair(fn(x, y) { x + y }, 2, 3); println(result.0); println(result.1) }'
Polymorphic functions write type parameters before the name: fn[A, B, C] name(...). The older fn name[A, B, C](...) spelling is deprecated.
moon run -c 'fn apply_twice(f : (Int) -> Int, value : Int) -> Int { f(f(value)) }
fn main { println(apply_twice(fn(x) { x + 1 }, 40)) }'
Callback parameters use function types like (Int) -> Int; callback literals are fn(x) { ... }. Capturing mutable Ref state in a closure is a direct port of OCaml closures over refs.
moon run -c 'fn apply(f : (Int) -> Int raise, x : Int) -> Int raise { f(x) }
fn main raise { println(apply(fn(x) raise { if x == 0 { fail("zero") }; x + 1 }, 1)) }'
Raising callbacks must include raise in the parameter type and in the literal: (Int) -> Int raise and fn(x) raise { ... }. Inferred raising effects on fn literals are deprecated. For a narrower project error type, annotate: fn(x) raise ProjectError { ... }.
moon run -c $'fn apply(f : (Int) -> Int raise Error) -> Int raise Error { f(1) }\nfn main { try apply(fn(x) raise Error { if x > 0 { fail("x") } else { x } }) catch { _ => println(true) } noraise { _ => println(false) } }'
Raising and non-raising callback types are not interchangeable. A raising closure cannot satisfy a non-raising slot, and vice versa. When porting OCaml callbacks whose implementation may later touch fallible APIs (filesystem, async, random), declare raise in the type from the start; do not hide failure inside the callback.
Labelled and Default Arguments
moon run -c $'fn greet(name? : String = "pdf") -> String { name }\nfn main { println(greet()); println(greet(name="moon")) }'
Defaults attach to labelled parameters: name? : T = default. Call sites pass name=value. Do not place a default on an unlabelled positional parameter.
moon run -c 'fn inner(flag? : Bool = false) -> Bool { flag }
fn outer(flag? : Bool = false) -> Bool { inner(flag~) }
fn main { println(outer()); println(outer(flag=true)) }'
label~ forwards the current value of a same-named labelled argument through wrapper layers — the MoonBit counterpart to preserving OCaml optional-argument semantics across helpers.
moon run -c 'fn takes_view(xs : ArrayView[Int]) -> Int { xs.length() }
fn sample(xs? : ArrayView[Int] = []) -> Int { takes_view(xs) }
fn main { println(sample()); println(sample(xs=[1, 2, 3])) }'
Optional view parameters can default to []. Optional-only arguments must be passed by label (sample(xs=[1, 2, 3]), not sample([1, 2, 3])).
moon run -c 'fn f(a : Int, flag? : Bool = true, xs? : ArrayView[Int] = []) -> Int { if flag { a + xs.length() } else { a - xs.length() } }
fn main { println(f(3, xs=[1, 2], flag=false)); println(f(3, flag=true)); println(f(3)) }'
After required positional arguments, multiple optionals can be supplied by label in any order.
Default values can call local functions, useful for structured-record or array defaults.
For an OCaml parameter that is genuinely "absent vs explicitly None vs explicitly Some(x)" — three distinct states — an Option-typed labelled argument (x? : T? = None) is acceptable, and label~ forwarding still works through it. But this is rare. The common case is just x? : T = default_value; do not reach for Option-typed labels when a plain typed default would express the same thing.
Pattern Matching
moon run -c 'enum E { A(Int); B }
fn f(e : E) -> Int { match e { A(n) if n > 0 => n; _ => 0 } }
fn main { println(f(A(3))); println(f(A(-1))); println(f(B)) }'
Pattern guards belong on the same arm: Pattern if condition => .... Do not split the if onto its own line.
Match arms run top to bottom. Put specific tuple or variant cases before broad wildcard or guarded catch-all arms — a guarded broad arm can make a later specific case behaviorally dead, and the compiler does not always catch that.
moon run -c 'fn main { let xs = [1, 2, 3]; match xs { [_, .. rest] => { let owned = [ for x in rest => x ]; println(rest.length()); println(owned.length()); println(owned[0]) }; _ => () } }'
Array rest patterns (.. rest) bind read-only views. If an enum payload or API needs an owned Array[T], copy with a comprehension: [ for x in rest => x ].
Loops and Comprehensions
moon run -c 'fn main { let fixed : FixedArray[Int] = [1, 2, 3]; let doubled = [ for x in fixed => x * 2 ]; println(doubled.length()); println(doubled[2]) }'
moon run -c 'fn main { let pairs = [(1, 2), (3, 4)]; let firsts = [ for pair in pairs => pair.0 ]; println(firsts[0]); println(firsts[1]) }'
Comprehension syntax: [ for x in xs => expr ]. Use a simple identifier as the binder and destructure tuples or access fields inside the body.
moon run -c $'fn parse(s : String) -> Int raise { @string.parse_int(s) }\nfn main { let _ = [ for s in ["1"] => parse(s) ]; println("done") }'
Comprehension bodies cannot call error-raising functions. Port OCaml List.map/Array.map with raising mappers as an explicit loop in a raising function: push each result into an output array and let the error propagate.
moon run -c 'fn main { let m : Map[Int, Int] = Map([]); m[1] = 10; m[2] = 20; let mut total = 0; for key in m.keys() { total += key }; println(total) }'
for x in xs iterates any iterable, not just arrays/views. Map, @hashmap.HashMap, ranges, and views all support it directly. Do not convert iterables to Array just to loop; reserve .to_array()/.to_owned() for cases that need an owned snapshot, sorting, indexing, or mutation.
moon run -c 'fn first_positive(xs : Array[Int]) -> Int? { let mut i = 0; while i < xs.length() { if xs[i] > 0 { break Some(xs[i]) }; i += 1 } nobreak { None } }
fn main { println(first_positive([-2, 0, 7]).unwrap()); println(first_positive([-2, 0]) is None) }'
while loops may produce a value with break value; the branch when the condition becomes false is nobreak { ... }. The older else spelling is deprecated. Avoid the older functional loop ... { ... } form in new ports — MoonBit warns on it.
Surface Hazards
alias is a reserved keyword; avoid it as a local name, loop binder, or helper name to keep warning-enabled checks clean.- On the JavaScript backend, also avoid sentinel names like
undefined for local test results — cross-target package tests can expose generated-JS name collisions.
- Source files do not contain OCaml-style
open. Add package imports in moon.pkg, then call imported packages with their @alias.
- For
moon run -c snippets that need extra packages, use an MBTX import block at the start: import { "moonbitlang/x/crypto" } (comma-separated entries for multiple packages). The source-file form import "moonbitlang/core/encoding/ascii" is rejected in command snippets.
Process and Random
@env (in moonbitlang/core/env) provides args() -> Array[String], current_dir() -> String?, get_env_var(String) -> String?, get_env_vars() -> Map[String, String], now() -> UInt64, set_env_var, unset_env_var. current_dir() is optional — some platforms cannot always determine it. Prefer x is Some(_) over the deprecated .is_some().
moon run -c 'fn main { let r = @random.Rand::chacha8(seed=@ascii.encode("01234567890123456789012345678901")); println(r.uint(limit=256).to_string()); println(r.uint(limit=256).to_string()) }'
@random provides deterministic PRNGs. Rand::new() is reproducible, not OS entropy. Do not substitute @random or @env.now() for a cryptographic random source when porting security-sensitive OCaml code (AES IVs, salts, file keys); keep an explicit random provider or use a target-specific secure API.
Targets and Async I/O
moon run --target native -c $'#cfg(target="native")\nfn target_name() -> String { "native" }\n#cfg(not(target="native"))\nfn target_name() -> String { "other" }\nfn main { println(target_name()) }'
#cfg(...) on its own line immediately before a declaration provides target-specific implementations. Combining #cfg and the declaration on the same line leaves the attribute unused.
MoonBit async has no await keyword — async functions call other async functions directly.
- Keep CPU-bound pure transformations over
Bytes synchronous.
- Make filesystem/network entry points async when they touch
moonbitlang/async.
- Prefer async wrappers that load file contents into
Bytes, then call the synchronous core. This avoids making every recursive helper async.
- Async entry points and async tests need
moonbitlang/async in the relevant moon.pkg; syntax alone is not enough.
- Async functions can raise by default. Do not add
raise to async functions just to propagate errors; add noraise only when the async function must reject unhandled errors.
moon run --target native -c 'async fn main { println("async ok") }'
moon run --target native -c $'import {\n "moonbitlang/async",\n "moonbitlang/async/fs" @fs,\n}\nasync fn main {\n try @fs.read_file("/tmp/definitely-missing-fixture") catch {\n _ => println(true)\n } noraise {\n _ => println(false)\n }\n}'
If a port should keep its pure core available on non-native targets, put async/file-system wrappers in a native-only package rather than importing async into the core. Imports are package-level; supported_targets = "+native" controls backend inclusion. A single test file can be target-gated with options(targets: { "file_test.mbt": [ "native" ] }), but imports remain package-level and may report unused on non-native checks.
Testing
Test file roles:
*_test.mbt: black-box tests. Call only public APIs through the package alias.*_wbtest.mbt: white-box tests. Run inside the package; may test private helpers.*.mbt.md: documentation with checked code blocks. mbt check for code that must compile and test; mbt nocheck for illustrative-only snippets.
Assertion style:
@test.assert_eq for stable scalar or structural results (Debug bound; avoids the deprecated Show-based path).- Boolean assertions:
@test.assert_eq(condition, true) — there is no @test.assert_true helper.
@test.assert_eq accepts a named msg= argument for shared assertion helpers.- Shared test helpers that call
@test.assert_eq, fail, or fallible APIs should declare -> Unit raise Error unless they use a narrower project error type.
- Pattern checks:
assert_true(value is Pattern(...)) or guard ... else { fail(...) }.
- Snapshots:
inspect(value, content="...") for small values; debug_inspect(value, content=...) for complex values with Debug. If the expected snapshot is unknown, write inspect(value), run moon test --update, then review the diff.
- Success paths through raising functions: call the fallible API directly and let the test fail on any error. Expected failures:
try f() catch { err => inspect(err) } noraise { _ => fail("expected to fail") }.
- Use a wildcard
catch arm for "this should raise"; match a specific error variant only when that variant is itself part of the compatibility contract.
moon run -c 'fn helper() -> Unit raise Error { @test.assert_eq(true, true); @test.assert_eq([1, 2], [1, 2]) }
fn main raise { helper() }'
moon run -c 'import { "moonbitlang/core/test" }
fn helper() -> Unit raise { @test.assert_eq(true, true, msg="named assertion") }
fn main raise { helper(); println("ok") }'
For parser/serializer, encoder/decoder, or loader/writer pairs, pair focused edge tests with at least one public-API round-trip test. Round-trips catch ownership, byte/text, and object-boundary mistakes that isolated unit tests miss.
Cover the byte/text edges that OCaml callers depended on: non-ASCII, NUL, form-feed, high-bit bytes, integer overflow, empty input, and boundary offsets.
Command Cheatsheet
moon run -c 'fn main { ... }'
moon check --warn-list +73
moon test --outline
moon test
moon test path/to/file_test.mbt
moon test package/dir --filter 'glob'
moon test --update
moon test --target native
moon coverage analyze > uncovered.log
moon info && moon fmt
Validation rule for each migration patch:
- Add or update tests before or with the ported code.
- Run targeted
moon test while developing.
- Finish with
moon check --warn-list +73, moon test, moon info, and moon fmt.
Update Discipline
When a migration teaches a reusable rule that contradicts or refines this guide, update this file with:
- The OCaml behavior being replaced.
- The MoonBit API or idiom chosen.
- A minimal
moon run -c probe and its observed output.
- Any known incompatibility, target limitation, or deferred behavior.
