Raven Language Skill

Raven Language (v2)

Raven v2 is a statically-typed, ahead-of-time compiled language implemented in Rust. Source lowers through a resolver, type checker, HIR, and a monomorphizing MIR, then a Cranelift backend emits a native binary. Memory is managed by a tracing garbage collector. The syntax reads like a blend of Rust and Swift, with local type inference, no semicolons, traits, generics, sum types, and pattern matching. The single most useful thing this skill does is steer you away from the non-obvious pitfalls so the first version of your code parses, type-checks, and links.

v2 is a clean break from v1. If you find old material mentioning elseif, C-style for loops, int/string lowercase types, format("{}", x), import math;, or main(); at the bottom of a file, that is v1 and does not apply here.

When to read what

• This file: read every time. The pitfalls and templates cover most tasks.
• references/builtins.md: the always-available functions and the methods on String, List, Map, Set.
• references/stdlib.md: per-module API for std/io, std/string, std/collections, std/math, std/iter, std/fs, std/time, std/json, and more.
• references/rvpm.md: starting a project, rv.toml, GitHub-direct packages, the raven and rvpm CLIs.
• references/grammar.md: exact lexical and grammar rules when something parses surprisingly.

The stdlib source lives in stdlib/std/*.rv in the Raven repo and is written in Raven itself, so reading it is the authoritative answer for any signature.

Pitfalls to internalise

These are the mistakes that ruin first-try compilation. Burn them in.

1. No semicolons. Statements end at the newline. Do not terminate lines with ;.
2. Types are PascalCase. Int, Float, Bool, String, Char, Unit. There is no int/string. Unit is the no-value type (the implicit return).
3. Type annotations are optional. Local inference works: let x = 5 infers Int. Annotate when you want to be explicit or when inference can't tell (let xs: List<Int> = []). Function parameters and return types are still written out.
4. let is mutable; there is no let mut. Reassign freely. Compound assignment works: += -= *= /= %=. const is for immutable compile-time constants and works both as a local (const X = 5) and at module level (const MAX: Int = 100).
5. Logical operators are &&, ||, !. The words and/or/not are NOT operators.
6. else if is two words (v1's elseif is gone). if is also an expression: let s = if x > 0 { "pos" } else { "neg" }.
7. for is range/iterator based. for i in 0..10 { } (exclusive), for i in 1..=10 { } (inclusive), for item in list { }. while cond { } and loop { } exist. break and continue work.
8. String interpolation is "${expr}". The expression can be arbitrary: a nested string literal ("${a.concat("y")}"), a macro call ("${square!(5)}"), and a struct value that derives ToString ("${p}") all interpolate directly.
9. String length is .length(), not .len(). .len() type-checks but fails in codegen. List/Map/Set use .len(). String methods (.to_upper(), .trim(), .split(), .concat(), …) require import std/string.
10. No null. Absence is Option<T> (sugar T?) with Some(x)/None. Fallible results are Result<T, E> with Ok(x)/Err(e) and the ? operator. Some, None, Ok, Err are in scope without imports.
11. Enum variants are constructed qualified: Shape.Circle(2.0), Color.Red. In match, the patterns are bare: Circle(r) -> .... match is exhaustive.
12. No visibility modifiers. Every top-level fun/struct/enum/trait is importable. There is no export and no pub.
13. fun main() is the entry point. Define it; do NOT call main() yourself. There is no top-level statement execution.
14. Prefer selector imports for free functions. import std/fs { write } then write(path, data) is the form that always works. Module-qualified access (import std/fs then fs.exists(...)) works for some runtime-backed modules but not all (import std/math then math.sqrt(...) fails), so reach for the selector form. Types (Map) and methods (.to_upper) come in through import std/collections { Map } / import std/string.
15. match on String literal patterns works. match s { "a" -> ..., _ -> ... } matches as expected, and strings support </> ordering.
16. Reading module-level bindings works; mutating them does not. A top-level let/const is readable from any function (let greeting: String = "hi"), but reassigning a module-level binding from a function mis-compiles (binop lhs used a Unit value). For mutable state, keep it inside functions, thread it through parameters, or hold it in a struct you pass around.

Anatomy of a Raven file

// imports: std modules by name, locals and packages by quoted path
import std/io { println }
import std/collections { Map }
import "./board"                              // local ./board.rv
import "github.com/user/raven-json" { parse } // GitHub-direct package

// a trait
trait Display {
    fun show(self) -> String
}

// a struct + an impl of the trait + inherent methods
struct Point {
    x: Float,
    y: Float,
}

impl Display for Point {
    fun show(self) -> String = "(${self.x}, ${self.y})"
}

impl Point {
    fun magnitude(self) -> Float = sqrt(self.x * self.x + self.y * self.y)
}

import std/math { sqrt }

// a free function (single-expression body)
fun midpoint(a: Point, b: Point) -> Point =
    Point { x: (a.x + b.x) / 2.0, y: (a.y + b.y) / 2.0 }

// entry point: defined, never called by you
fun main() {
    let p = Point { x: 3.0, y: 4.0 }
    println("mag = ${p.magnitude()}")
    println(p.show())
}

Implicit rules:

• Imports: import std/<module> { names } for the bundled stdlib; import "./rel" for a sibling file; import "github.com/user/repo" { names } for a package fetched by rvpm. Selectors bring the names into scope unqualified.
• No entry call: the runtime calls main for you.
• Single-expression bodies: fun f(...) -> T = <expr> is shorthand for { return <expr> }.
• Trailing commas are allowed in struct definitions/literals and collection literals.

Templates

Working code you can adapt. These compile.

Variables, control flow

fun main() {
    let total = 0
    for i in 1..=10 {
        if i % 2 == 0 {
            total += i
        }
    }
    let label = if total > 20 { "big" } else { "small" }
    print("${total} is ${label}")
}

Collections and iterators

import std/io { println }
import std/collections { Map }
import std/iter { collect, fold }

fun main() {
    let nums = [1, 2, 3, 4, 5, 6]
    println("len = ${nums.len()}, first = ${nums[0]}")

    // lazy pipeline: consume with collect / fold / count
    let doubledEvens: List<Int> =
        collect(nums.iter().filter(fun(x: Int) -> Bool = x % 2 == 0).map(fun(x: Int) -> Int = x * 2))
    println("count = ${doubledEvens.len()}")

    let sum = fold(nums.iter(), 0, fun(acc: Int, v: Int) -> Int = acc + v)
    println("sum = ${sum}")

    let tally: Map<String, Int> = Map.new()
    tally.set("a", 1)
    match tally.get("a") {
        Some(n) -> println("a = ${n}"),
        None -> println("missing"),
    }
}

Struct, trait, dynamic dispatch

trait Shape {
    fun area(self) -> Float
}

struct Circle { r: Float }
struct Square { side: Float }

impl Shape for Circle {
    fun area(self) -> Float = 3.14159 * self.r * self.r
}
impl Shape for Square {
    fun area(self) -> Float = self.side * self.side
}

fun describe(s: dyn Shape) {
    print("area = ${s.area()}")
}

fun main() {
    describe(Circle { r: 2.0 })
    describe(Square { side: 3.0 })
}

Note: List<dyn Trait> (a heterogeneous list of trait objects) is not supported in this release. Pass trait objects to dyn-typed parameters or assign to a dyn-typed local.

Enums, match, Option, Result, ?

enum Shape {
    Circle(Float)
    Rectangle(Float, Float)
}

fun area(s: Shape) -> Float =
    match s {
        Circle(r) -> 3.14159 * r * r,
        Rectangle(w, h) -> w * h,
    }

fun checked_div(a: Int, b: Int) -> Result<Int, String> {
    if b == 0 {
        return Err("divide by zero")
    }
    return Ok(a / b)
}

fun halve_then_div(a: Int, b: Int) -> Result<Int, String> {
    let q = checked_div(a, b)?    // ? propagates Err, unwraps Ok
    return Ok(q / 2)
}

fun main() {
    print("${area(Shape.Circle(2.0))}")
    match halve_then_div(20, 5) {
        Ok(v) -> print("ok ${v}"),
        Err(e) -> print("err ${e}"),
    }
}

Generics with a trait bound

trait Describe {
    fun describe(self) -> String
}

struct Dog {}
impl Describe for Dog {
    fun describe(self) -> String = "a dog"
}

fun announce<T: Describe>(x: T) {
    print("this is ${x.describe()}")
}

fun main() {
    announce(Dog {})
}

defer (LIFO, runs at function exit)

fun main() {
    let log = [1]
    defer log.push(3)    // runs second
    defer log.push(2)    // runs first
    print("len now ${log.len()}")
}

Concurrency: spawn + channels

import std/sync { channel, channel_buffered, yield_now }

fun main() {
    let ch = channel()
    spawn(fun() -> Unit {
        let i = 1
        while i <= 5 {
            ch.send(i)
            i = i + 1
        }
    })
    let sum = 0
    let n = 0
    while n < 5 {
        sum += ch.recv()
        n += 1
    }
    print("sum = ${sum}")   // 15
}

Goroutines run in parallel on a pool of worker threads (one per core), so spawn is true parallelism, not cooperative time-slicing. A goroutine only ever suspends at a blocking point (a full/empty channel, yield_now, sleep_millis), and the scheduler may resume it on a different worker than it ran on before. Channels carry Int and block when full (send) or empty (recv).

std/sync also has a Mutex, a WaitGroup, and select:

import std/sync { mutex, wait_group, channel, select_recv }

fun main() {
    // Mutex: lock() blocks until free, unlock() releases (call only while held).
    let m = mutex()
    m.lock()
    m.unlock()

    // WaitGroup: add() before spawning, done() as each finishes, wait() blocks to zero.
    let wg = wait_group()
    wg.add(1)
    spawn(fun() -> Unit {
        wg.done()
    })
    wg.wait()

    // select_recv: block on several channels, lowest ready index wins. The
    // SelectResult is { index, value } (index -1 if the list was empty).
    let a = channel()
    let b = channel()
    spawn(fun() -> Unit { a.send(7) })
    let r = select_recv([a, b])
    print("chan ${r.index} -> ${r.value}")
}

Channels, wait groups, and select sets hold runtime registry entries with no destructor; call .free() (or select_recv frees its own set) when you are done with one to avoid leaking the entry.

Metaprogramming: derive + macros + reflection

@derive(Eq, Hash, ToString, Debug)
struct Point { x: Int, y: Int }

macro square { ($x:expr) => { ($x) * ($x) } }

fun main() {
    let p = Point { x: 1, y: 2 }
    print(p.to_string())          // Point { x: 1, y: 2 }
    print("${square!(5)} fields=${field_names<Point>().len()}")
}

Derivable traits: Eq, Hash, ToString, Debug, and Ord. @derive(Ord) adds compare(self, other) -> Int (negative / zero / positive), comparing structs field-by-field in declaration order and enums by variant order then payload. Pair it with import std/cmp { sort } to sort a List:

import std/cmp { sort }

@derive(Ord)
struct Version { major: Int, minor: Int }

fun main() {
    let vs = sort([Version { major: 1, minor: 2 }, Version { major: 1, minor: 0 }])
    print("${vs[0].minor}")   // 0
}

C FFI

import std/ffi { alloc, free, load, store }

extern "C" {
    fun abs(x: CInt) -> CInt
    fun strlen(s: CStr) -> CSize
}

fun main() {
    print("${abs(-7)}")
    let len = strlen(c"hello")     // c"..." is a C string literal
    print("${len}")
    let buf = alloc<CInt>(2)
    store<CInt>(buf, 42)
    print("${load<CInt>(buf)}")
    free<CInt>(buf)
}

Built-in cheatsheet

Always available, no import. See references/builtins.md for details.

Stdlib at a glance

Bundled into the compiler; import with import std/<module> { names }. See references/stdlib.md.

If a name isn't here, read stdlib/std/<module>.rv in the Raven repo. Note: a dependency package can also import std free functions (fixed in 2.0.2).

Running and project layout

my_project/
├── rv.toml          # [package] / [dependencies] / optional [ffi], [fmt]
└── src/
    └── main.rv      # entry point — defines fun main(), not called manually

rvpm init my_project        # scaffold in the current dir (--lib for a library, lib.rv)
rvpm new my_project         # scaffold in a fresh my_project/ dir (--lib)
rvpm run                    # build src/main.rv and run it (forwards args: rvpm run -- a b)
rvpm build                  # compile to target/raven-out/<name>, or type-check a lib.rv
rvpm test                   # run fun test_*() tests in *_test.rv files (assert with import std/test)
rvpm doc                    # generate Markdown API docs into target/doc
rvpm fmt                    # format .rv files in place (pass paths for a library: rvpm fmt lib.rv)
rvpm fmt --check            # CI: non-zero if anything would change

rvpm add github.com/user/repo@v1.0.0   # add a GitHub-direct dependency, write rv.lock
rvpm install                # resolve rv.toml against rv.lock and fill the cache
rvpm update [pkg]           # re-resolve and rewrite rv.lock for one package or all
rvpm fetch github.com/user/repo@v1.0.0 # fetch a package into the shared cache
rvpm lock                   # generate or validate rv.lock
rvpm cache list             # inspect the shared cache (also dir / clean)
rvpm version                # print the rvpm version (also --version, -V)

raven build file.rv -o out  # compile a single file to a native binary
./out                       # run it

There is no raven file.rv direct-run, no -c type-check flag, and no REPL in v2; compiling is the check. A C linker must be on PATH (link.exe on Windows, cc/clang on Unix) so the compiler can link the runtime into your program.

Recommended workflow

1. Sketch types first: structs, enums, traits.
2. Write small free functions; group methods in impl blocks.
3. Write fun main() (do not call it).
4. raven build src/main.rv -o app (or rvpm run). Compile early and often; the type checker catches the bulk of mistakes.
5. If a parse error appears, the usual cause is a pitfall above: a stray ;, lowercase type name, .len() on a String, or an unqualified enum constructor.
6. rvpm fmt before committing.

Style

• 4-space indent, no semicolons.
• Order: imports → traits → structs/enums → impls → free functions → fun main.
• snake_case for functions and variables, PascalCase for types and enum variants.
• Prefer single-expression bodies (fun f() -> T = expr) for one-liners.
• Use string interpolation ("${x}") instead of manual concatenation.
• Keep match arms exhaustive; add a _ -> ... arm only when a catch-all is truly intended.