remove is the inverse of add: the require line leaves package.ludic; the lock keeps exactly what the
remaining requires still reach, read from the store's copy of each locked package.ludic (no
network), so a package another still requires stays locked and what only the removed one brought in
leaves with it; each leaving package loses the ludic_modules/ symlink add made, never the shared
store entry. Refused (exit 1) when package.ludic does not require the module; source still
importing a removed package is a warning. With no store copy to read, only the named module leaves.
get --json diffs the lock before and after in memory and prints {added, removed, changed,
unchanged} on stdout (an entry as the lock records it: name, version, hash, kind, provides; a change
as name, from, to, from_hash, to_hash), the resolver's lines on stderr. Cases added to test-pkg,
not run.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
16 KiB
Ludic packages
Ludic has a package manager built into the task runner (bin/ludic). It fetches,
resolves, stores and links third-party packages with no new infrastructure to
run — it drives plain git and rides on the Forgejo host and its release tags.
This is the v1 implementation of the direction decided in issue #63.
The four decisions
| Axis | Ludic's choice |
|---|---|
| Where source comes from | URL-as-identity, no registry. The import path is the git location — git.workshopsoft.io/user/pkg. A git tag vX.Y.Z publishes a version. No account, no publish step, no index to run. |
| Version resolution | Minimum Version Selection (MVS), Go-style. A require is a minimum; the resolver picks, per module, the greatest of every required minimum, then keeps the reachable closure at those versions. Deterministic, no SAT solver. |
| Where dependencies live | A content-addressed global store + per-project links (pnpm-style). One immutable store keyed by a file-content hash (~/.ludic/store/<sha256>); each project gets a light linked view under ludic_modules/ instead of a full copy. |
| Manifest / lock | package.ludic (declared deps + provided namespaces + kind + targets) and package.lock.ludic (resolved versions + content hashes). |
Commands
ludic add <module>[@version] add a dependency to package.ludic, then resolve + fetch + link
ludic remove <module> the inverse of add: drop the require, what only it locked, its links
ludic get [--json] resolve every dependency in package.ludic, link them, write the lock
ludic update [module] bump a dependency (or all) to its latest published version, then relock
ludic verify check every locked package against the store by content hash
ludic vendor copy the resolved packages into ./vendor for hermetic/offline builds
ludic add with no @version picks the latest published tag and records it as the
minimum. All the install commands print the resolved build list and write
package.lock.ludic.
ludic get --json prints what the run changed in the lock as one JSON object on stdout (the
resolver's lines go to stderr): {"added": [...], "removed": [...], "changed": [...], "unchanged": n},
an added or removed entry being the lock's line - {"name", "version", "hash", "kind", "provides"} -
and a changed one {"name", "from", "to", "from_hash", "to_hash"} (the versions, and the content
hashes beside them). The lock is compared before and after, in memory.
ludic remove <module> takes the require line out of package.ludic and re-reads the dependency
graph from the store's own copy of each locked package's manifest - no network - so the lock keeps
exactly what the remaining requires still reach: a package another one still requires stays locked
(at the version it had; ludic get settles versions), and what only the removed one brought in
leaves the lock with it. Each package leaving the lock loses its ludic_modules/ link - the link
add made, never the store entry it points at, which other projects share; anything there that is
not such a link is left and reported, as is a copy under vendor/. A module package.ludic does
not require is refused (exit 1). Source that still imports a removed package is a warning, not a
failure - the caller may be about to delete it. If the store has no copy of a locked package, only
the named module leaves the lock, and it says so.
The manifest — package.ludic
A line-oriented manifest. # starts a comment; strings are double-quoted.
package "git.workshopsoft.io/orkun/greeter" # this package's import path
version "1.2.0" # the version this checkout publishes
kind source # source | prebuilt
provides "Greet" # the Foo.* namespace(s) it registers (repeatable)
require "git.workshopsoft.io/orkun/util" "1.0.0" # a dependency and its minimum version
# a prebuilt lib also declares the targets it ships:
# kind prebuilt
# targets "native-arm64" "wasm32"
A consumer project's manifest is the same file, usually with only require
lines (the package/version fields describe a publishable package and are
optional for a leaf application).
The entry program
entry "src/game.ludic" names the program ludic run, build and bundle compile
when no file is given. Without it the CLI tries src/main.ludic, main.ludic, then
the one file under src/ that declares a program.
Scripts and hooks
script "dev" "ludic run --headless" # ludic dev (or: ludic script dev)
script "shots" "tools/shots.sh" # ludic shots a b -> tools/shots.sh 'a' 'b'
hook before build "tools/gen_atlas.sh" # a failing before hook stops the build
hook after bundle "tools/notarize.sh" # after hooks run only when the command succeeded
hook before dev "ludic fmt --check" # a script is an event too
A script is a named shell command, run from the package root. ludic <name> runs
it when no built-in command has that name; ludic script <name> always does, and
ludic scripts lists them. Arguments after the name are passed on, quoted.
A hook runs a command before or after a built-in command (build, run,
test, bundle, pack, clean, fmt, get, add, update, verify, vendor,
assets, build-lib) or a script. Hooks of one event run in the order written. A
before hook that exits non-zero stops the command with that exit code; after
hooks run only when the command succeeded.
Both run through the shell with the toolchain's bin/ first on PATH and
LUDIC_PACKAGE_ROOT, LUDIC_EVENT (the command or script) and LUDIC_PHASE
(before, after or run) set. Inside a quoted value, \" is a quote and \\ a
backslash.
The lockfile — package.lock.ludic
Generated by ludic get; do not edit by hand. One line per resolved module, pinning
its selected version, content hash, kind and provided namespaces:
# package.lock.ludic — generated by `ludic get`. Do not edit by hand.
lock 1
module "git.workshopsoft.io/orkun/greeter" version "1.2.0" hash "sha256:…" kind "source" provides "Greet"
module "git.workshopsoft.io/orkun/util" version "1.0.0" hash "sha256:…" kind "source" provides "Util"
ludic verify rehashes each store entry and confirms the project links to it, so a
tampered or missing dependency is caught before it reaches a build.
The store and the project view
Fetched packages live once in a global, immutable, content-addressed store:
~/.ludic/store/<sha256>/… the package tree at a version (no .git)
~/.ludic/store/cache/<module>/ a git clone cache used during resolution
Each project gets a lightweight view — ludic_modules/<import-path> is a symlink
into the store — so many projects share one copy and nothing is duplicated
per-project. Override the store location with $LUDIC_STORE.
Consuming a package — namespace registration
A source package ships plain Ludic. The consumer imports the package files by their import path:
program App {
import "git.workshopsoft.io/orkun/greeter/greet.ludic"
entry { print(greet_hello()) }
}
The compiler resolves an import first relative to the importing file, then — for
a non-absolute path that is not found — under the package module root
($LUDIC_MODULES, default ludic_modules/). So a fetched package's code is
spliced into the build and its namespace becomes available exactly the way the
built-in stdlib namespaces (Regex.*, Grid.*, …) are.
The engine runtime is the exception (issue #75): the compiler auto-splices
runtime/native/* for any ECS game, and that runtime ships with the toolchain,
not the project. A runtime/... import that is not found relative to the build is
resolved from the install root $LUDIC_HOME (default: the compiler binary's
directory — the same place the platform .ll files come from), before the
package module root. So an external game does not have to copy or symlink the
engine runtime into its ludic_modules/; that directory holds only third-party
packages. In-repo builds are unaffected — the runtime resolves locally there. Because Ludic compiles
ahead-of-time, a source package is compiled into the consumer's binary — no
ABI seam, and the whole-program guarantees (determinism, replay, world_save)
still hold.
Two packages may not register the same Foo.* namespace — a collision is a hard
error naming both modules.
Package-declarable namespaces and engine systems (issue #62)
A package's namespace can also be an alias block (L6): namespace Weapon { alias fire(slot) = wp_fire_slot } names each method and the function it calls, with its own labels, so the public
names need not be the implementation's. The engine's own Http.*, Json.*, Screen.* and the
rest are declared exactly this way, in runtime/native/namespaces.ludic.
A package can register two things that used to be compiler-hardcoded — a Foo.*
namespace and an engine-owned system — with no compiler edit, via two
keyword-free annotations. This is what lets gameplay-controller libraries ship as
ordinary packages instead of living in the compiler's stdlib.
-
@Namespace(Name)on a function opens aName.method(…)namespace that dispatches to the barename_method(…)(the same generic path the built-in namespaces use, applied only after them so it never shadows a core one):@Namespace(Coach) function coach_bonus() -> int { return 99 } # a consumer then writes Coach.bonus() -
@EngineSystem(Component, Phase)registers an engine-owned system: the frame loop calls it everyPhase(after the game's own handlers) whenever the namedComponentis present — the package-declarable form of the built-inSpriteAnim/Motion/Light2Dsystems. It reads and writes components by name through the reflection ABI, so an unused registration is byte-identical:@EngineSystem(Score, Update) function esys_score() -> void { … }
The core stdlib namespaces and engine systems keep their own optimized codegen;
packages flow through the generic registry alongside them. (For a prebuilt
binary module, the runtime counterpart of @EngineSystem is @System(Phase) —
see below.)
Prebuilt binary packages (issue #64)
A kind prebuilt package ships a compiled artifact (a native dylib per
target it lists in targets) instead of source. A consumer uses its exported
functions, systems and components without ever seeing the source. This is
the escape hatch for closed-source or other-language code; source packages stay
the default because they keep cross-compilation (including wasm) and the
compile-time ECS first-class. Prebuilt packages are native-only and ride the
stable reflection C-ABI — second-class ECS (dynamic, by-name components; one
indirect call per registered system), not part of the deterministic/replay core.
How a binary module works. The module is compiled with --emit-module: no
main, no world table (the consumer owns the single world). It carries a
load-time constructor that, when the dylib loads, registers its pieces against
the host through the C-ABI:
- Components —
world_register_prop(name, nfields)in amodule_initfunction; the host owns storage, the module reads/writes by name withworld_get/world_set/world_has/world_attach_dyn. - Systems — a function marked
@System(Phase)is registered withludic_register_system; the host's frame loop calls it every frame in that phase, after its own handlers. Phases:Input,FixedUpdate,Update,LateUpdate,Render,Start,OnQuit. - Functions — plain functions become dylib symbols; a consumer binds them
with
extern function name(...) -> T = "fn_name".
A module registers its ludic_* calls as undefined and binds them back to the
host image at load (-undefined dynamic_lookup); the host exports its ABI
(-export_dynamic). Every Ludic game is a capable host — the reflection ABI is
always emitted (unused parts dead-strip).
Publishing. In the package repo:
ludic build-lib module.ludic # -> lib/<target>/lib<name>.dylib
# add `kind prebuilt` and `targets "<target>"` to package.ludic, commit lib/, git tag
Consuming. In the game project:
ludic add git.host/user/module # kind prebuilt is resolved + the dylib linked into the view
ludic get # links the artifact for the build target (hard error if the target is missing)
then link the module dylibs into the game build. ludic link-flags prints the exact
clang flags (the dylib, an rpath to the store, -export_dynamic) for any build
system to splice into its link step:
clang -O2 game.ll $(ludic link-flags) -o game
(ludic build links them automatically when building in-repo.) If the package does
not ship the build target, ludic get fails — build from source instead where the
package offers it.
Native libraries - a source package that carries a C/C++ library (phase 15)
A source package may link a native library of its own - a physics engine, a navmesh builder, an
audio renderer - by naming it per target in its package.ludic:
native "macos-arm64" "lib/macos-arm64/libjoltc.dylib"
native "windows-x64" "lib/windows-x64/joltc.dll"
The targets are macos-arm64, macos-x64, windows-x64 and linux-x64
($LUDIC_NATIVE_TARGET names another). The package's Ludic reaches the library through
extern function declarations, which only a package (or the runtime) may call (L7).
Linked at build time, not opened with dlopen. When the compiler loads a file from a
package, it reads that package's manifest once and keeps the libraries for the build's target;
they are written into the IR as ; ludic-native: <path> lines, and every link reads that one
list - ludicc -o, ludic build, ludic test and ludic bundle. A program that imports no
such package links nothing new. Linking was chosen over a Vulkan-style loader because a missing
or mismatched library then fails at start-up, with the loader's own message naming it, rather
than at the first call deep in a frame; because an extern function is all the binding needs (no
generated thunk table per library); and because nothing here is optional the way Vulkan is - a
game that imports the physics package cannot run without it.
- macOS: the dylib's install name is
@rpath/lib<name>.dylib. A build gets an rpath to the package'slib/<target>/(soludic runandludic testwork in place) and one to@executable_path/../Frameworks.ludic bundlecopies each library intoContents/Frameworks- neverContents/MacOS, which holds mach-o executables only (a Velopack update strips a detached signature) - removes the build machine's rpath from the executable, and signs every library before the app. - Windows: the
.dllis linked through its import library,<name>.libbeside it;ludicccopies the.dllbeside the executable it writes, andludic bundlebeside the game's.exe.
Built here from a pinned source. A package's native/build.sh fetches the upstream tag,
checks its SHA-256, compiles the library and the package's shim with clang, and writes
lib/<target>/ (committed through Git LFS). tools/native/lib.sh is what every such script
shares - native_fetch, native_link, the target and compiler - and it runs unchanged in Git
Bash on Windows with the LLVM installer's clang. The shim's rules (what may cross, how a
callback comes back) are in packages/README.md;
ludic.nativeecho is the worked example.
Offline / hermetic builds
ludic vendor copies the resolved packages out of the store into ./vendor. Build
against the copy with LUDIC_MODULES=vendor, so the build needs neither the
network nor the global store.