lingua-engine/docs/kernel-contract.md

Kernel Contract v0

A draft, not a final decision. The microkernel, the plugin contract, and the hot-reload model for Lingua Engine — a modular engine on C#/.NET, with a constraint that shapes half the decisions below: most of the code, kernel and plugins alike, will be written by an LLM agent rather than a human.

• Stack — .NET 9, C# 13
• Platforms — Linux, Windows
• Kernel — BCL only, no dependencies
• Object model — GameObject / Component
• Physics — Box3D (C, MIT), bound via P/Invoke — not our own
• Primary author — an agent
• Goal — Play-in-editor with no domain reload

Built for a small team's own use, at indie scale — not AAA. That scope licenses several calls made below: accepting GC pauses instead of chasing a zero-allocation hot path, picking GameObject/Component over a faster struct-of-arrays ECS, buying rendering and windowing off the shelf instead of writing them. None of those are free choices at a bigger scale; at this one, dev velocity outweighs the performance left on the table.

---

1. Principle: the kernel is a shared language, not "the engine minus plugins"

The tempting version of "everything is a plugin" makes even the object model a plugin. That's a trap: if every other plugin depends on the GameObject/Component plugin, that plugin is the kernel already — just with an extra layer of indirection and none of the stability guarantees a kernel should provide.

What becomes a plugin is behavior, not the shared data model.

The kernel is a lingua franca: the minimal set of types and mechanisms plugins need in order to understand each other at all. Anything two independent plugins are required to agree on lives in the kernel. Everything else lives outside it.

This is how every plugin architecture that survived contact with reality is built — Eclipse, VS Code, OSGi, Bevy: a small, stable, extensible kernel plus everything else on top. Trying to make the shared language itself swappable produces either indirection overhead or a kernel so empty it guarantees nothing.

2. Scope: what's in, what's out

Kernel — roughly 4,000 lines, BCL only

#	Piece	Role
01	World	GameObject hierarchy (parent/children, name, tags) plus typed Component instances, with a type index so Query<T>() costs O(matches), not O(all). Plain classes, zero unsafe — see §7.
02	Scheduler	Frame stages, topological system ordering, parallel execution of systems with disjoint declared access, and debug-mode enforcement of that access — see §7. Structural changes (adding/removing a GameObject or Component) are queued and applied at the stage boundary, so a running system never sees a collection mutate under it.
03	Plugin Host	Manifest parsing, dependency resolution, ALC loading, unloading, reload.
04	Service Registry	Publishing and discovering interfaces between plugins. Control path, not the hot path.
05	Event Bus	Publish/Subscribe, ownership-tracked and leak-safe the same way as the Scheduler's systems — see §7. PluginHost publishes PluginLoaded/PluginUnloaded; engine.assets publishes TextureReloaded when a watched file changes on disk (M2). Still nothing publishes GameObject created/destroyed facts — GameWorld doesn't touch the bus at all, deliberately: a publish on every structural change would tax the hot path for listeners that usually don't exist.
06	Time & Log	Frame clock (DeltaTime, ElapsedTime, FrameCount, FixedDeltaTime) and logging, both on IPluginContext. The fixed-step accumulator (Time.ConsumeFixedSteps) shipped with M4, driving Stage.FixedUpdate — engine.physics is its first real consumer.

GameObject.Transform is the one field embedded directly rather than modeled as a Component subclass — it's a plain struct holding local position, rotation, and scale, because nearly every system in the engine touches it every frame, and routing that through the same virtual-dispatch path as every other component would tax the one thing everything depends on. GameObject.WorldMatrix composes it with the parent chain on every read rather than caching a value — a cache here would need invalidating on every reparent and every ancestor's change, which is more bookkeeping than a handful of matrix multiplies costs at indie scale. Everything else — MeshRenderer, Rigidbody, AudioSource, game-specific components — is a plain class, heap-allocated, no special treatment.

Plugins — everything else, no exceptions

windowing · render · physics · audio · input · assets · scene-format · animation · ui · scripting · editor-shell · inspector · gizmos · profiler · introspect · build-pipeline · the game itself.

The editor is also just a set of plugins over the same kernel. This is the architecture's real test: if the editor can't be assembled as plugins, the extensibility claim is decorative. A game build is the same kernel minus the editor plugins.

Per-project configuration

A plugin's manifest declares what it needs; a project's manifest declares which plugins it loads, at which versions, and where to find its own. This is the piece that actually makes modularity a per-project property rather than a claim about the engine in the abstract — a new project doesn't fork the engine to swap an implementation, it points its manifest at a different plugin satisfying the same contracts, or adds project-local plugins that never leave its own tree.

// MyGame/project.json
{
  "engineVersion": "^0.3",
  "plugins": [
    { "id": "engine.windowing" },
    { "id": "engine.render",  "version": "^0.3" },
    { "id": "engine.physics", "version": "^0.2" },
    { "id": "mygame.enemies" }
  ],
  "pluginPaths": ["./plugins"]
}

engine.render here could just as well point at a project-local fork with the same contracts and a bumped id — the Plugin Host resolves a project's manifest through the exact same dependency graph it already builds for plugin-to-plugin dependsOn, so nothing new has to be built to support it.

Two channels, two costs

Plugins talk to the kernel — and to each other — through two paths with deliberately different prices:

Channel	For	Cost	Frequency
World (GameObjects & Components)	Anything per-entity: transforms, meshes, colliders, health. Render reads what physics wrote without knowing physics exists.	Direct field access on a cached component reference; Query<T>() is a type-index lookup, not a scan	10⁴–10⁶ / frame
Services (interfaces)	Commands and resources: load an asset, open a window, compile a shader, open an editor panel.	Virtual call, negligible	a handful / scene
Events (bus)	Facts with no fixed consumer at design time: asset reloaded, plugin unloaded, entity destroyed.	Allocation + fan-out to subscribers	tens / frame

The line that must never be crossed. Never write IPhysicsService.GetPosition(GameObject go). A single call is cheap, but that shape of API invites calling it in a loop over entities — and now the plugin boundary sits in the hot path. Position is on GameObject.Transform, not a service method. Services hand out capabilities; World hands out data.

The same rule applies one level down, inside World itself: don't call otherGameObject.GetComponent<T>() for a different entity from inside a per-entity loop — that's a type-indexed lookup multiplied by iteration count, the exact perf trap Unity code is famous for. Resolve the components you need once, before the loop starts, and index into that.

3. The plugin contract

A plugin is two entry points and a manifest next to them. The manifest is a separate file, not assembly attributes — the host has to build the dependency graph before loading anything, or plugin load order becomes a chicken-and-egg problem with ALC loading itself.

// Engine.Kernel / IPlugin.cs

// A plugin holds no game state. None.
// State lives in World; the plugin is code that operates on it.
public interface IPlugin
{
    // Registration: services, systems, component types.
    void Configure(IPluginContext ctx);

    // Full undo of Configure. Whether this method is honest
    // determines whether the ALC unloads at all — see §4.
    void Shutdown(IPluginContext ctx);
}

public interface IPluginContext
{
    IWorld           World    { get; }  // data
    IServiceRegistry Services { get; }  // Provide<T> / Require<T>
    ISchedule        Schedule { get; }  // systems and ordering
    IEventBus        Events   { get; }  // Publish / Subscribe
    ILogger          Log      { get; }
    ITime            Time     { get; }  // DeltaTime, ElapsedTime, FrameCount
}

// plugins/engine.render/plugin.json
{
  "id":        "engine.render",
  "version":   "0.3.1",
  "contracts": "Engine.Render.Contracts.dll",  // Default ALC
  "assembly":  "Engine.Render.dll",             // Collectible ALC
  "dependsOn": {
    "engine.windowing": "^0.3",
    "engine.assets":    "^0.2"
  },
  "reloadable": true
}

// plugins/engine.render/Contracts/MeshRenderer.cs

// Plain data, no methods. Lives in the Contracts assembly — see §4
// for why that split is what makes reload safe.
public sealed class MeshRenderer : Component
{
    public MeshHandle Handle;
}

// plugins/engine.render/RenderPlugin.cs

public sealed class RenderPlugin : IPlugin
{
    public void Configure(IPluginContext ctx)
    {
        // control plane: hand out an interface, take one in
        var window = ctx.Services.Require<IEngineWindow>();
        ctx.Services.Provide<IRenderer>(new VulkanRenderer(window));

        // data plane: the system queries GameObjects by component type.
        // Reads/Writes are declared explicitly — the scheduler uses
        // them to run systems with disjoint access in parallel, and
        // in debug builds enforces that a system only touches what
        // it declared — see §7.
        ctx.Schedule.Add(Stage.Render, SubmitDrawCalls)
           .After("engine.transform:propagate")
           .Reads<MeshRenderer>();
    }

    public void Shutdown(IPluginContext ctx)
    {
        // undo everything: systems, services, subscriptions, GPU resources
        ctx.Services.Revoke<IRenderer>();
        ctx.Schedule.RemoveAllFrom("engine.render");
    }

    static void SubmitDrawCalls(in Frame f, IWorld world)
    {
        // type-indexed lookup, not a scan — see the World row in §2
        foreach (var go in world.Query<MeshRenderer>())
            f.Draw(go.GetComponent<MeshRenderer>().Handle, go.WorldMatrix);
    }
}

4. Hot reload: why every plugin is two assemblies

A collectible AssemblyLoadContext only unloads once nothing references its contents. One forgotten event subscription, one live Task, one cached Type — and the unload silently fails to happen, leaking a little more memory on every reload.

The most treacherous reference isn't a subscription — it's the component classes themselves. If a plugin declares class MeshRenderer : Component and a GameObject holds one in its component list, the kernel holds a reference to a type from the context you're trying to unload. That plugin will never unload.

This is why every plugin splits into two assemblies:

• Contracts (*.Contracts.dll) — component classes, service interfaces. Loaded into the Default ALC, which lives for the process lifetime and never unloads. World owning references into it is fine, because it isn't supposed to unload.
• Implementation (*.dll) — systems, service implementations. Loaded into a collectible ALC, recreated on every reload. No static state — only code that operates on objects it doesn't own.

References only point from implementation to contracts, never the reverse, which is what lets Unload() actually succeed. Because component instances live in World, owned by the kernel, an implementation-only reload never touches game data at all — it isn't snapshotted and restored, it's simply never in the collectible ALC to begin with.

Reload sequence

1. A file watcher sees a freshly built Engine.Render.dll. The build happens externally, via plain dotnet build — the editor doesn't need its own compiler.
2. The scheduler finishes the current frame and pauses. Reload never happens mid-stage.
3. Shutdown() runs: systems, services, subscriptions, and native resources are torn down. Anything Configure registered has to be undone here, or step 4 fails. World's component instances aren't touched — the implementation assembly never held them.
4. alc.Unload() + GC.Collect(), then a WeakReference check. If the context doesn't collect, that's a loud error naming the pinning reference — not a silent leak.
5. A new ALC, the new assembly loads, Configure() runs. The plugin doesn't know it was reloaded.
6. The scheduler rebuilds its ordering graph and resumes. Typical budget: 200–400 ms, almost all of it spent waiting on the build.

In practice this covers ~95% of iteration, because most changes are to system logic, not component shape.

Changing a component's own fields is a different, rarer case — and it doesn't hot-reload at all. A component's fields live in the Contracts assembly, and the Default ALC hosting it never unloads by design. There is no in-process path to swap it. This isn't a gap to fill later; it's a deliberate seam. Field changes are rare enough that paying for an editor restart there — reloading the scene from its serialized file rather than migrating live objects — is a better trade than writing and maintaining live-migration code for the 95% case that doesn't need it.

Leak testing belongs in CI from day one. Load and unload a test plugin 200 times in a row; after each cycle, verify the ALC's WeakReference is dead and working-set memory hasn't grown. This is the one thing that keeps the architecture from slowly degrading — ALC leaks accumulate invisibly and surface months later, by which point the cause is indistinguishable from noise.

5. Play mode without domain reload

Unity's Play-mode wait isn't about compilation — it's about serializing all script state, tearing the domain down, and recreating it. That step doesn't exist here: state never lived in plugin code to begin with. It lives in World, owned by the kernel, untouched by reload and untouched by entering Play. Built and shipped in M3 — this is what actually runs, not the design sketch that preceded it:

// IWorld (Engine.Kernel)
string Snapshot();          // a scene-format dump of the whole graph
void Restore(string snapshot);

// Engine.Editor.Contracts / IPlayModeController — engine.editor's real
// implementation wraps exactly these two calls, nothing else:
void EnterPlay() => _snapshot = world.Snapshot();
void ExitPlay()  { world.Restore(_snapshot!); _snapshot = null; }

Snapshot/Restore reuse SceneFormat — the same JSON serialization a scene file on disk already uses — rather than a bespoke clone mechanism. A scene file and a Play-mode snapshot are the same problem (capture every GameObject's state faithfully enough to reconstruct it) at two different moments; reusing already-proven serialization beat maintaining a second way to walk the same graph. WorldSnapshotTests times a 300-GameObject snapshot+restore at under 100ms — M3's literal "done when" — and the real editor path confirms it: PlayModeController.EnterPlay logs its own wall-clock cost, and a real run against samples/WindowDemo's scene measured 13ms.

There's no SystemGroup.Play/SystemGroup.Edit split in the scheduler — that would mean every system declaring which group it belongs to, for a distinction the host loop alone can already make. Engine.Host checks IPlayModeController.IsPlaying once per frame and only runs Stage.Update while it's true; Stage.Render/Stage.Present run every frame regardless, so Edit mode still shows a live, responsive scene view — just one that never ticks. A project with no engine.editor loaded sees no behavior change at all: Stage.Update runs unconditionally, same as before Play mode existed.

A side effect of Play being nothing but a snapshot and a stage-gate: system code can be edited during Play without restarting — hot-reloading a plugin mid-Play still works, because nothing about Play mode touches the ALC or plugin loading at all. That's the feedback loop the whole engine exists to enable.

6. Where this breaks

Risk	The problem	Mitigation
ALC leaks — the main killer	Unload silently fails from one forgotten reference. Symptom: memory growth after N reloads; cause takes days to find.	200-cycle test in CI. Diagnose pinning references in the host itself, not via an external profiler.
Scope	The kernel is 3–5k lines and a couple of months. The renderer, asset pipeline, and editor are years, and they decide whether the engine ships.	Don't write your own RHI or physics engine. Silk.NET or Veldrid underneath the renderer, Box3D underneath physics; originality goes into the architecture on top.
GC pressure from Components	Every component is a heap object; churn from creating/destroying GameObjects at runtime (bullets, particles, pickups) means allocation and collection, against a 16.6 ms frame budget.	Pool GameObjects and components for anything spawned/destroyed at high frequency. Server GC. Query<T>() iterators must not allocate.
Creeping abstraction	The temptation to hide World behind a "cleaner" interface. Kills performance invisibly and irreversibly.	The rule in §2 is law. Review rejects any service method that takes a GameObject.
Plausible-but-wrong code — agent-specific	The agent produces code that compiles, passes a smoke test, and breaks on someone else's GPU — sync, barriers, resource lifetime.	Minimize new subsystems; Silk.NET/Veldrid is risk management, not time-saving. A conformance harness gates every plugin merge.
Contract drift	A contract change requires updating every dependent plugin, and a stale implementation keeps compiling while silently diverging from spec.	Versions in the manifest, plus running every plugin's harness on every build, not just the changed one.

7. Written by an agent, not a human

This isn't an afterthought — it's an input condition. It's part of why §2 picked the most conventional possible object model instead of a performance-first one, and it adds a surface no classic editor needs at all.

What works in our favor:

• A plugin's boundary matches a context window's boundary. Writing engine.physics only requires the kernel API, physics' own contracts, and its own code — nothing else. Modularity chosen for team reasons turns out to also be how you fit a task in an agent's head.
• Blast radius is bounded by the plugin. Plausible-but-wrong code is inevitable; the question is what it can break. The kernel is written once, tested, and frozen — the agent never touches it again after that. A bug in a plugin stays a bug in that plugin.
• GameObject/Component is the most over-represented pattern in an LLM's training data of any game architecture. That's also a reason it won over a hand-rolled ECS: components are plain classes with plain fields, no stride arithmetic, no manual layout, nothing that compiles cleanly and corrupts memory at runtime. The one performance-motivated exception, Transform as an inline struct, is confined to the kernel and never written by the agent at all.

What has to change:

• Explicit over clever. Naming conventions, code generators, reflection magic save a human keystrokes but hide behavior from something that reasons over text. Verbose, explicit system and service registration is a deliberate cost. Reflection stays where it's safe: the editor inspector.

Verification instead of trust. Reads<> / Writes<> declarations must be enforced in debug builds: a system touching an undeclared component fails immediately, with a message naming the violation. For a human this is hygiene; for an agent it's structural — otherwise a wrong access declaration becomes a race that reproduces once in a hundred runs and is otherwise undiagnosable. Same principle for the plugin conformance harness: load, reload 200 times, verify Shutdown fully undoes Configure. The agent needs to be able to tell, on its own, that it's actually done.

Introspection surface

The agent doesn't look at a screen. Whatever the editor shows a human's eyes has to be available as data, or the feedback loop closes on a human and the whole point of fast reload is lost.

# run a scene headless and check an assertion about world state
engine run --headless --frames 60 \
           --scene tests/physics_stack.scene \
           --dump out/world.json \
           --assert "count(Rigidbody where sleeping) == 12"

# why a plugin won't unload — instead of guessing from a profiler
engine diag why-pinned engine.render

The loop this enables: agent edits a system → dotnet build for one plugin → ALC reload (world state untouched) → 60 headless frames → machine-readable dump + assertions → back to the agent, no human in the loop. Fast reload saves a human time on its own; paired with headless runs and a state dump, it becomes a loop the agent can close by itself — and the minutes-to-seconds iteration speedup multiplies by however many iterations the agent can now run.

The kernel is written once and frozen. It's the one place where a mistake is expensive and spreads everywhere. A small kernel isn't only an architectural preference — it bounds how much code has to be correct.

8. Build order

Each milestone ends in a working demo, not a "finished subsystem." The order is chosen so the riskiest bet — ALC unloading — gets tested first, while the cost of changing course is still zero.

	Milestone	Done when
M0	Kernel only. World as a GameObject/Component hierarchy with type-indexed queries, staged scheduler with access enforcement, plugin host with ALC, service registry, JSON world dump. No window, no graphics.	An agent runs the full loop from §7 unassisted: edits a headless plugin, rebuilds, reads the changed dump — and the 200-cycle leak test is green.
M1	Window, input, a triangle. Three separate plugins over Silk.NET. First real-load test of the data channel.	The triangle's color changes by editing system code, with no app restart.
M2	Assets and scenes. Hot-reloading asset plugin, scene format, World serialization.	Swapping a texture on disk changes the picture with nothing stopped; a scene loads and saves.
M3	Editor as plugins — done. Shell (engine.editor, ImGui over the live scene, see §5 and the new Stage.Present), reflection-based inspector and hierarchy, Play/Stop on snapshots, a real 3D translate gizmo driven by the actual camera.	Entering Play takes under 100 ms — the original complaint about Unity is closed. Proven twice: WorldSnapshotTests (kernel, 300 GameObjects) and a real editor run against samples/WindowDemo (13ms, logged by PlayModeController).
M4	One small game, end to end — done when met, game itself in progress. engine.physics (Box3D) and engine.audio (miniaudio) done, both over a narrow, scalars-only P/Invoke shim; samples/PhysicsDemo ties physics/audio/input/render together with engine.editor deliberately excluded from its project.json. .github/workflows/build.yml builds both native shims from source, tests, and headless-verifies on real ubuntu-latest/windows-latest runners — both green on the first real run, which itself caught two real cross-platform bugs no local testing could have (a locally-built .so too new for ubuntu-latest's glibc; --headless not stopping a loaded engine.render from demanding a real GL context on a GPU-less Windows runner — see README's Status section). Not yet: the demo is a physics sandbox, not yet a scored 20-minute game.	The build runs on both platforms with no editor plugins in the shipped binary.

---

The kernel's open questions are resolved. What M0 shipped without deciding, in order:

• Time and Log: both in the kernel, both on IPluginContext. Every plugin needs logging and a frame clock; there's no realistic case for a project wanting to swap either out per-project the way it would swap physics or rendering. Log was already built this way by the time the question got asked explicitly — Time (DeltaTime, ElapsedTime, FrameCount) shipped alongside closing the question, not before it. Explicitly still deferred: the fixed-step accumulator the original kernel scope named alongside the frame clock. Building it now, with no physics system to test it against, would be untested speculative machinery — exactly what this project has avoided everywhere else. It arrives with M4, alongside the Stage.FixedUpdate it would drive.
• Event Bus: real, not event-components in World. A disposable event-as-GameObject fits a pure ECS's cheap-entity model better than ours, where GameObject carries persistent identity and hierarchy. A conventional Publish/Subscribe bus is the better fit for this object model specifically. PluginHost publishing PluginLoaded / PluginUnloaded is the concrete proof it's real infrastructure, not an API nobody calls — and Subscribe/RemoveAllFrom follow the exact leak-safety shape Schedule already established (ownership tracked by the subscribing delegate's declaring assembly), proven the same way: sandbox.echo subscribes for real, and the 200-cycle leak test in AlcUnloadTests now exercises that cleanup path, not just Schedule's.
• Frame stages: fixed, kernel-defined — not plugin-extensible. A stage is part of the shared language every plugin and the host loop rely on; letting plugins register arbitrary custom stages would mean the host's frame loop can no longer just call a known, closed set of RunStages. No stage gets added without something real to run in it — which is exactly what happened once: M3 added Stage.Present alongside {Update, Render} when engine.editor's ImGui overlay needed to draw after the scene but before the buffers swap, and engine.render's old single Clear+Draw+ SwapBuffers system had nowhere else to put the swap that wouldn't race it. Stage.FixedUpdate is still the only stage left on the M4 list.
• Data-oriented fast path: still open, on purpose, with a trigger condition instead of a deadline. Not "undecided" the way the other three were — deliberately not worth deciding before there's a concrete system to decide it against. Revisit when a specific system (particles is the standing example) needs tens of thousands of GameObjects updated per frame and profiling — not intuition — shows GameObject/Component overhead is the actual bottleneck. Until then, an early decision here would be optimizing against a guess.

Assembly names in the examples are placeholders.