Ocaml

OCaml

OCaml is a statically-typed, mostly-functional language with a strong type inference engine (Hindley-Milner). Coming from JS/TS/PHP, the biggest mental shift is not syntax — it's that the compiler is a design partner, not a linter: exhaustive pattern matching and a sound type system catch entire classes of bugs (null checks, missed cases) at compile time that you're used to catching at runtime or not at all.

Default to explaining why something is idiomatic, not just that it is — the user is learning the language, not just translating syntax.

Type inference: don't annotate unless asked

OCaml infers types from usage, almost always correctly and often more precisely than a human would bother to write. Unless the user explicitly asks for type annotations (or a .mli signature requires them), write bindings and function definitions without type annotations:

(* Prefer *)
let total_ttc price_ht tax_rate = price_ht *. (1. +. tax_rate)

(* Not, unless asked *)
let total_ttc (price_ht : float) (tax_rate : float) : float = price_ht *. (1. +. tax_rate)

If the user is confused about what type something is, the answer is to ask the tools, not to preemptively annotate everything: ocamlc -i file.ml prints the full inferred signature of a file without touching it, dune utop <dir> / dune build @check do the same interactively, and the ocaml toplevel (or utop) reports the inferred type after every expression you enter.

When reading an inference error, trace the expression the way the compiler does:

• Start from the most generic type ('a), add one constraint per construct: function application pins an arrow type, if pins its branches to the same type, arithmetic pins its operands to int or float.
• The reported error is the last constraint that couldn't be satisfied, not the actual mistake site — locate the real one by walking constraints in order.
• A message saying a variable "occurs inside" a type it's supposed to equal means an infinite/recursive type — walk the constraints back in order rather than staring at the flagged line.

Core language cheat sheet

• let bindings are immutable by default. let x = 5 cannot be reassigned. Mutation needs an explicit ref: let x = ref 5 in x := !x + 1. This is the single biggest habit-break coming from JS — reach for immutable rebinding / recursion before reaching for ref.
• ;; is a REPL-only separator, not a statement terminator. In .ml files, let/in chains and top-level let definitions don't need it.
• ; sequences unit-typed expressions: print_string "a"; print_string "b" — only the left side must be unit, and the whole sequence takes the right side's type. A common beginner compile error is forgetting a value is non-unit and needs ignore or a let _ = ... in. When an imperative sequence sits inside if ... then/else or a match arm, parenthesize it — or use begin ... end for the same effect, more readably.
• Pattern matching (match ... with) is exhaustive-checked by the compiler. Adding a case to a variant type and forgetting to handle it somewhere produces a compiler warning/error, not a silent runtime bug — this is the direct analogue of TypeScript's discriminated-union exhaustiveness checks, except OCaml doesn't need a never-typed default branch to get it.
• Variant types replace TS union types / enums, and are usually the right tool where a JS dev would reach for a string literal union or a class hierarchy:

  type shape =
    | Circle of float           (* radius *)
    | Rectangle of float * float (* width, height *)
  
  let area = function
    | Circle r -> Float.pi *. r *. r
    | Rectangle (w, h) -> w *. h
  

• Records replace plain-object interfaces: type point = { x : float; y : float }. Fields are immutable unless declared mutable. Update via functional update syntax: { p with x = 0. }.
• option replaces null/undefined, result replaces exceptions-as-control-flow: type 'a option = None | Some of 'a, type ('a, 'e) result = Ok of 'a | Error of 'e. There is no null in OCaml — an option has to be pattern-matched or explicitly unwrapped, so "forgot to check for null" is not a category of bug that exists.
• Exceptions exist but the type system doesn't track them. raise Not_found, try ... with Not_found -> ..., and custom exception NotPrime of int all work, but a function's type says nothing about what it raises — that invisible side-channel is exactly why the style default below is option/result for expected failure. Under the hood exn is an extensible variant: you can grow it with new exception declarations, and failwith s is literally raise (Failure s). Exceptions also pattern-match in a plain match, not just try ... with: | exception (Failure s) -> s. And a try arm that raises an unmatched exception silently re-propagates the original one — try can catch but not rewrap. Built-ins worth knowing: Not_found (lookup misses), Invalid_argument (bad arguments), End_of_file (channel reads), Sys_error (file errors), Division_by_zero.
• |> is the pipe operator, directly analogous to a JS .then/method chain but for plain functions: xs |> List.filter is_even |> List.map square.
• Folds are where list iteration lands. List.fold_left threads an accumulator left-to-right, List.fold_right right-to-left, and the function's argument order flips: f acc h for fold_left vs f h acc for fold_right. fold_left is tail-recursive (safe on huge lists), fold_right is not. The order flip bites on non-commutative operators: List.fold_left ( - ) 0 [1;2;3] = ((0-3)-2)-1 = -6, but List.fold_right ( - ) [1;2;3] 0 = 3-(2-(1-0)) = 2. ListLabels variants take labeled args (~f ~init:) to remove the ambiguity. And short-circuit: List.for_all/List.exists stop at the first non-passing element, while a fold_left over &&/|| visits everything.
• Evaluation order is unspecified for tuple and function arguments, and the current implementation evaluates right-to-left — so (raise A, raise B) raises B, not A. Don't rely on it; if order matters, force it with let a = ... in let b = ... in.
• Every function is curried. let add x y = x + y is shorthand for fun x -> fun y -> ... — add 6 is itself a value (int -> int), and partial application composes everywhere: List.map (add 6) xs. Operators become functions with parentheses — ( * ) has type int -> int -> int, so List.map (( * ) 2) doubles. But sections keep argument order, which bites on non-commutative operators: ( / ) 2 is fun x -> 2 / x, so List.map (( / ) 2) does not halve a list — write fun x -> x / 2 for that.
• rec is required for recursive functions: let rec fact n = if n = 0 then 1 else n * fact (n - 1). Without rec, the function name inside the body refers to a shadowed outer binding (or doesn't exist), which is a common source of "unbound value" errors for beginners.
• Non-tail recursion grows the stack; tail recursion doesn't. 1 + f (n - 1) builds a pending expression of the whole call chain before any + runs — that's what "Stack overflow during evaluation (looping recursion?)" means. Fix it with an accumulator and wrap it so callers can't pass a bad initial value:

  let length l =
    let rec go acc = function [] -> acc | _ :: t -> go (acc + 1) t in
    go 0 l
  

  List costs matter here too: :: is O(1), but @ walks its whole left argument — acc @ [x] inside a fold/loop makes it O(n²). Cons onto the front and List.rev at the end instead.
• Structural vs physical equality: =/<> compare structurally (like a deep-equal), ==/!= compare physical identity (like JS === on objects, i.e. same box/reference). This is the reverse of what a JS/PHP dev's fingers expect — = is almost always the one you want. The asymmetry matters beyond style: == always finishes in constant time, while = walks the whole structure and can hang forever on a cyclic value (e.g. a value built with let rec) — if a structural comparison seems to hang, cyclic data is a real suspect, not just a slow machine.
• int is not arbitrary width. It's a native machine word minus one tag bit — 63-bit on a 64-bit machine, 31-bit on a 32-bit machine — because OCaml represents an unboxed int as an odd machine integer (2n+1) so the runtime can tell integers and pointers apart at a glance. This is narrower than a JS Number's safe integer range in the 32-bit case, and it means overflow wraps silently rather than promoting to a bigger representation (there's no automatic bignum). For arbitrary-precision arithmetic, reach for Int64/Int32 (fixed-width, boxed) or the zarith library, not the default int.
• int and float never mix, and their operators differ. Floats use +. -. *. /. (plus ** for powers); using + on a float is a type error, not a coercion — convert explicitly with float_of_int/int_of_float/Float.of_int. Literal floats always need a decimal point (1., not 1). Unlike integer division, float division never raises: 1. /. 0. is infinity (with neg_infinity and nan as the other special values), so "divide by zero" shows up as a wrong answer rather than an exception.
• Modules (module Foo = struct ... end) and signatures (.mli files or module type) are the OCaml analogue of TS interfaces + implementation, but enforced at compile time with no structural escape hatch. A .mli file next to a .ml file restricts what's visible outside the module — the closest OCaml equivalent to export/private.
• No classes-by-default culture. OCaml has an object system (class, object ... end) but idiomatic OCaml reaches for modules + records + functions first, objects/classes rarely. Don't default to translating a JS class into an OCaml class — translate it into a module exposing functions over a record type instead.

Runtime model (use this when explaining equality or mutation)

Ground any answer about = vs ==, mutation, or "why is it like that" in this model, and state it explicitly — it is not optional background:

• An OCaml value is either an unboxed integer or a pointer to a heap block (a header word + N value words).
• Tuples, arrays, records, and non-constant variant constructors are all just blocks — { x = 1.; y = 2. } is a pointer to a 2-word block, so two records with identical fields are still two separately allocated blocks.
• A ref is just a one-field mutable record, a list is a chain of 2-word :: cells, a closure is a block with a code pointer plus its captured environment — which is why ref/:=/! compose with everything else instead of being a distinct language feature.
• Consequence for equality: == compares only the two words (constant time, even on cyclic values); = walks into the blocks and can loop forever on a cycle.

Full walkthrough: references/runtime-model.md.

Project structure: dune, opam, Alcotest

Initializing a new project

Don't hand-write dune-project/dune files from scratch when scaffolding a new project — use dune's own generator, then adjust:

dune init proj <name>

(prefix with opam exec -- if the user hasn't run eval $(opam env) in the current shell). This creates a <name>/ directory with bin/, lib/, and test/ subdirectories (each with their own dune file), a top-level dune-project, and a generated <name>.opam. bin/main.ml is the entry point and is runnable immediately with dune exec <name>, dune build compiles, dune runtest/dune test runs the Alcotest suite already wired up in test/. This is the right default when the user asks to "set up a new dune project" — it gives them the conventional three-directory layout (below) for free, rather than the minimal single-file setup.

For a throwaway single-file project, or to show what dune actually requires under the hood, the bare minimum is two files: a dune-project with just (lang dune 3.6) (or later), and a dune file with an executable stanza, e.g. (executable (name foo)), sitting next to foo.ml. No bin/lib/test split is required — dune only needs a dune file in each directory containing something to build.

A typical dune project the user will hand you:

dune-project        ; project-wide config: dune lang version, package metadata
<lib>.opam           ; generated from dune-project when (generate_opam_files true) — don't hand-edit, edit dune-project instead
lib/
  dune               ; library stanza: (library (name foo) (libraries ...))
  foo.ml
test/
  dune               ; test stanza: (test (name test_foo) (libraries alcotest foo))
  test_foo.ml

• dune build compiles. dune runtest (or dune test) runs the test suite. dune exec <target>.exe runs an executable stanza directly — dune always names the built executable with a .exe suffix, even for native binaries on Linux/macOS, so dune exec ./foo.exe is correct even when there's no foo.exe file to be found by hand (it lives under _build/default/).
• After scaffolding or editing a dune-project file (or any dune stanza), validate it before moving on. Dune has no standalone dune validate subcommand — the practical check is dune build @check, the built-in alias that builds just enough (.cmi/.cmt/.cmti files) to catch stanza and syntax errors without paying for full linking, making it the fast, cheap way to confirm a fresh dune-project/dune setup is well-formed. Follow up with a plain dune build (and dune runtest if tests exist), since @check alone won't surface every downstream issue (e.g. link-time errors). See the dune manual for the full alias reference.
• An executable stanza spanning multiple modules needs a (modules ...) field listing them explicitly, e.g. (executable (name main) (modules util main) (libraries ...)) — dune compiles any .mli for a module before its .ml, mirroring the manual ocamlc -c order a .mli-then-.ml pair requires.
• The .opam file at the project root is generated output when dune-project has (generate_opam_files true) — if the user asks you to change the synopsis, authors, or dependencies, edit dune-project, then regenerate with dune build, never hand-edit the .opam file.
• If the user is working outside a dune project (a one-off script, or debugging in isolation), the underlying tools are still worth knowing: ocamlc file.ml compiles to a portable bytecode a.out (or -o name for a chosen name), ocamlopt does the same to native code, and ocaml (or utop) drops into a REPL — the toplevel conventions (the ;; terminator, #quit;;/Ctrl-D to exit) apply there, per the cheat sheet above.
• Alcotest (mirage/alcotest) is the standard test framework in this kind of setup. Shape:

  let test_area_of_circle () =
    Alcotest.(check (float 0.001)) "area" 12.566 (Foo.area (Foo.Circle 2.))
  
  let () =
    Alcotest.run "foo"
      [ ("area", [ Alcotest.test_case "circle" `Quick test_area_of_circle ]) ]
  

  Alcotest.check takes a testable (Alcotest.float epsilon, Alcotest.int, Alcotest.string, Alcotest.(list int), etc.) — for custom types you either derive one with Alcotest.testable pp equal or compare projected primitive fields. Alcotest test names show up in dune runtest output, so name them for what they assert, not generic ("area of a zero-radius circle is zero", not "test1"). Asserting that a function raises needs a thunk (a unit -> unit, never an already-called value): Alcotest.check_raises "division by zero" Division_by_zero (fun () -> 1 / 0). Pick test cases two ways: black-box (cases from the spec, no implementation knowledge) and glass-box/white-box (cases from each implementation branch) — cover both; black-box catches spec mismatches, glass-box catches missed branches.

The Alcotest shape above is unit testing — fixed examples you hand-pick. Real dune projects layer two more test types on top of it: property-based testing (qcheck, random inputs + shrinking to prove invariants) and mutation testing (mutaml, inject faults to prove your tests would catch a bug). Two adjacent quality tools — style linting (camelot) and dead-code analysis (dead_code_analyzer) — fill gaps the compiler doesn't. When to reach for each and the one-line idiom: references/testing-and-quality.md.

Debugging & compiler errors

• Treat dune build output as ground truth, not a nuisance to silence — the type errors it produces are usually pointing at a real logic gap (a missed pattern match arm, a type mismatch that reveals a wrong assumption), not a syntax nit.
• Printf.printf "%d\n" x / Printf.printf "%s\n" x for quick print-debugging (%d int, %s string, %f float, %b bool — printf format specifiers are type-checked at compile time, so a mismatched specifier is itself a compile error, not a runtime surprise).
• Canonical messages (verify against these before guessing):
  • This expression has type X but an expression was expected of type Y — read the inferred type first, then find where your usage disagrees with it; the error location is often the second use site, not the actual mistake. Apply the left-to-right constraint walk from the type inference section above.
  • The function ... is applied to too many arguments — usually a missing ; between two side-effecting calls in a sequence, so the second call is parsed as an extra argument to the first.
  • Warning 8 [partial-match]: this pattern-matching is not exhaustive — compiles, then raises Match_failure at runtime on the missed case. The warning prints an example unmatched value; handle it. Promote warnings to errors in the dune stanza for real projects, and never assume "it compiled" means every case is handled.
  • Warning 10: this expression should have type unit — you discarded a non-unit value in a sequence; make it explicit with ignore e or let _ = e in ....
  • Warning 11 [redundant-case]: this match case is unused — an earlier arm (often _) already covers this case, so this one is dead code; reorder or delete.
  • Stack overflow during evaluation (looping recursion?) — either unbounded recursion (missing/wrong base case) or a non-tail-recursive function on a huge argument; fix the base case or add an accumulator (see the tail-recursion bullet above).
• Core typing rules (fixed — don't re-derive):
  • if e1 then e2 else e3 requires e1 : bool and e2/e3 the same type; if e1 then e2 (no else) requires e2 : unit.
  • while/for loop bodies must be unit.
  • Sequence e1; e2 requires e1 : unit — if it isn't, write ignore e1; e2 or let _ = e1 in e2.
  • match/try-with: every branch's right-hand side shares one type; every match pattern shares the scrutinee's type.

Style defaults

• Favor small, composable functions over long imperative bodies — List.map/List.filter/List.fold_left over hand-rolled loops with ref accumulators, unless performance or clarity genuinely calls for the latter.
• Favor pattern matching over if/else chains once there are more than two cases, especially over a variant type — it's both more idiomatic and gets exhaustiveness checking for free.
• Keep functions total where reasonable (return option/result for partiality) rather than raising exceptions for expected failure cases; reserve exceptions for genuinely exceptional/programmer-error conditions.

For katas and exercises, or when the user wants a structured path from beginner to writing real programs, OCaml from the Very Beginning (John Whitington, free online: johnwhitington.net/ocamlfromtheverybeginning) teaches the language no-prerequisites with worked questions per chapter — a good source of practice tasks and a canonical explanation order to mirror. A heavier, more rigorous companion is the Cornell CS 3110 textbook OCaml Programming: Correct + Efficient + Beautiful (Fall 2026 edition, free online: cs3110.github.io/textbook) — an entire university course with 200+ embedded lecture videos and per-chapter exercises.

Benchmark

Scenario: .benchmarks/scenarios/ocaml-001-equality-semantics.md · Run: 2026-08-31 · Log: .benchmarks/runs/2026-08-31/ocaml-001-equality-semantics.json

NEG (run 2026-08-31). Opus −17 (100→83): runtime-model grounding missed — the Runtime Model section is a one-liner punting to references. Sonnet/haiku unchanged. Follow-up: expand the one-liner (cap reached this cycle). Gate per .agents/skills/skill-optimizer/rules/release-gates.md.

Scenario: .benchmarks/scenarios/ocaml-001-equality-semantics.md · Run: 2026-09-01 (salience re-run wf_aae2e0cb) · Log: .benchmarks/runs/2026-09-01/ocaml-001-equality-semantics.triage-rerun.json

PASS (run 2026-09-01). Salience re-run (runtime-model one-liner promoted to an imperative inline section, wf_aae2e0cb): the 2026-08-31 opus regression is gone (−17 → +0) and the universal c5 grounding criterion cleared with-skill on haiku (83→100). Opus/sonnet still miss c5 in both conditions; opus without-skill also dropped 100→83 this run, so some baseline noise is in play. Follow-up: next cycle, check whether opus/sonnet need the grounding tied directly into the equality cheat-sheet bullet rather than its own section. Gate per .agents/skills/skill-optimizer/rules/release-gates.md.

PENDING (2026-09-02). Content update sourced from OCaml from the Very Beginning (float operators, currying, exceptions, tail recursion, canonical error messages, ;-sequence fix) and the CS 3110 textbook (folds, exception patterns / exn extensibility, evaluation order, Alcotest raise-testing, black-box vs glass-box test selection, resource link). Equality/runtime-model content unchanged, so ocaml-001 should be unaffected — rerun ocaml-001 before merging per release-gates.