Instructions¶
Library¶
From a git checkout:
Mojo is pinned to 1.1.0. The Modular channel may require PREFIX_API_KEY in a local .env file. scripts/ci-setup.sh reads that file. Do not commit it.
Encode and decode¶
A struct needs Defaultable so decode can build it, and Movable so the value can be returned. Declare the fields in the order they should appear on the wire. Kiln writes those values and leaves the names in the source, so two Int32 fields occupy eight bytes.
from std.collections import Span
from kiln import Writer, decode, encode
@fieldwise_init
struct Point(Copyable, Movable, Defaultable):
var x: Int32
var y: Int32
def __init__(out self):
self.x = 0
self.y = 0
def main() raises:
var w = Writer(64)
var p = Point(Int32(1), Int32(2))
encode(p, w)
var raw = w.copy_bytes()
var back = decode[Point](Span(raw))
_ = back
Writer keeps its allocation. The next encode clears the length and writes into the same storage when the value fits. encode_append does not clear the length.
decode checks that each length stays inside the buffer and that strings are UTF-8. Pass False as the second argument to skip the UTF-8 check for bytes you already trust.
Frames and tensors¶
encode_framed writes the 16-byte KLN0 header and one Frozen value. decode_framed checks the magic and the schema hash.
TensorDir is the tensor directory. f64_directory builds one unnamed f64 tensor from a List[Float64]. encode and decode read and write that value with the same rules as any other struct field.
JSON Schema¶
gld-kilngen-mojo --schema testdata/schema/point.json --out tests/generated writes structs whose encode_to and decode_from call the same reflection codec. The subset is described in the spec: objects, arrays, booleans, integers, numbers, strings, binary strings, and #/$defs/ references.
schema_from_json builds the same description for encode_value and decode_value, which walk a Value tree instead of a Mojo struct. That path is for data whose type arrives at run time. It is not the fast path.