reznok/gmcabilitysystem/gmas-testing

Behavior below was verified against GMC 2.3.x and GMAS 1.4. Confirm signatures in this project's own headers (paths are relative to each plugin's root, e.g. Source/GMCCore/Public/Components/GMCReplicationComponent.h); GMC's source is not part of this plugin.

GMAS gameplay is tested at three depths: the pure structs, the ability component driven through a stub movement component without a world, and the real move loop in a networked play-in-editor (PIE) session. The authoring skills (gmas:gmas-ability, gmas:gmas-effect, gmas:gmas-attribute, gmas:gmas-task) say which spec to write for each class; this skill is the mechanics: the spec harness and its limits, the headless run, and a PIE harness that starts a dedicated server plus clients, waits until every player has a pawn, asserts inside confirm windows and tears the session down. The harness code is in references/net-test-skeleton.md. What a healthy session logs and how to read a bad one: gmas:gmas-debug. GMAS paths below are relative to Source/GMCAbilitySystem/; engine paths to Engine/Source/.

Three layers

Which layer for which claim: "the value is X after N ticks" → 1 or 2; "the gate refuses when the tag is set" → 2; "the server accepts the client's spend without a correction", "the shot exists once on every machine", "the effect reaches the other client", "the ability survives 1 % loss" → 3. Every checklist line in the authoring skills that says verified under networked PIE is a layer-3 claim; a green layer-2 spec with a red PIE session puts the fault in the network path (gmas:gmas-debug). Standalone PIE proves nothing about prediction: its pawn is a locally controlled server pawn, never predicted, replayed or smoothed (gmas:gmc-prediction).

GMAS specs

The bundled specs are the model: on 1.4.1+ the test module Source/GMCAbilitySystemTests/ (UncookedOnly: built for the editor and uncooked targets, never into a cooked or Shipping game; 533 cases under GMAS.*, all passing), on 1.4.0 Source/GMCAbilitySystem/Private/Tests/ inside the runtime module (449 cases, about 150 failing). A spec is a BEGIN_DEFINE_SPEC block whose members are the harness, a Define() with BeforeEach / AfterEach / Describe / It, guarded by #if WITH_DEV_AUTOMATION_TESTS (Runtime/Core/Public/Misc/AutomationTest.h; GMAS 1.4.0 used WITH_AUTOMATION_WORKER). Put yours in a test module of the same kind that depends on GMCAbilitySystem, GMCCore, GameplayTags and GameplayTasks, as GMCAbilitySystemTests.Build.cs does: test-only UCLASSes then never ship.

BEGIN_DEFINE_SPEC(FMyAbilitySpec, "MyGame.Unit.Ability",
    EAutomationTestFlags::EditorContext | EAutomationTestFlags::ProductFilter)
    UMyTestMovementCmp*          MoveCmp     = nullptr;   // UCLASS() class UMyTestMovementCmp : public UGMC_MovementUtilityCmp {}: empty
    UGMC_AbilitySystemComponent* AbilityComp = nullptr;
    UGMCAttributesData*          Attributes  = nullptr;   // your real rows: same clamp, same bGMCBound
END_DEFINE_SPEC(FMyAbilitySpec)
 
void FMyAbilitySpec::Define()
{
    BeforeEach([this]()
    {
        MoveCmp     = NewObject<UMyTestMovementCmp>(GetTransientPackage());          MoveCmp->AddToRoot();
        AbilityComp = NewObject<UGMC_AbilitySystemComponent>(GetTransientPackage()); AbilityComp->AddToRoot();
        Attributes  = NewObject<UGMCAttributesData>(GetTransientPackage());          Attributes->AddToRoot();
        // rows: Attribute.MaxStamina 100 / Clamp.Max 1000; Attribute.Stamina bStartFull, MaxAttributeTag = Attribute.MaxStamina
        AbilityComp->GMCMovementComponent = MoveCmp;
        AbilityComp->AttributeDataAssets.Add(Attributes);
        AbilityComp->BindReplicationData();                 // instantiates and binds the attributes; nothing consumes the binds
        AbilityComp->ActionTimer = -1.0;                    // non-zero: 0 fails every apply (1.4.1+: SetActionTimerForTest)
        GetMutableDefault<UMyAbility_Dash>()->CooldownTime = 5.f;   // the ASC instantiates from the CDO: configure it here...
    });
    AfterEach([this]()
    {
        GetMutableDefault<UMyAbility_Dash>()->CooldownTime = 0.f;   // ...and restore it, or the next spec inherits it
        AbilityComp->RemoveFromRoot(); MoveCmp->RemoveFromRoot(); Attributes->RemoveFromRoot();
        AbilityComp = nullptr; MoveCmp = nullptr; Attributes = nullptr;
    });
    Describe("Dash", [this]() { It("spends stamina once", [this]() { /* activate, drive, assert */ }); });
}

The stub. An ownerless component reports NM_Standalone from GetNetMode() and false from HasAuthority(), which is exactly what makes the harness cheap and what bounds it: no net role, no moves, no world. UGMAS_TestMovementCmp is that empty subclass (Source/GMCAbilitySystemTests/Private/ on 1.4.1+, Private/Tests/ on 1.4.0); write your own in your module rather than including a plugin-private header (1.4.1 no longer exports the test folder). On a negative clock (-1.0) each effect-id generator wraps into its range and reports it once ([EffectID], an Error, 1.4.1+): call SilenceEffectIDWrapReportForTest() in a harness that is not about ids.

Test classes. The bundled stubs (UGMAS_TestAbility, UGMAS_TestAbilityB, UGMAS_TestCostEffect, same folder) are empty subclasses whose configuration is written to the class default object per test (an instance copies every property of its class default object, 1.4.1+). 1.4.1 adds GMAS_TestHelpers.h (GMASTest::MakeAttr / MakeClampedAttr for bare attributes with an explicit clamp, and the clock seeds StableActionTimer = -1.0 and ClientAuthActionTimer = 1.0), UGMAS_TestEventRecorder (binds the component's dynamic delegates and records effect ids, ability activated / ended / cancelled counts and custom events; it can remove an effect or cancel an ability from inside a listener) and UGMAS_TestCountingEffect (records Start / Tick / Period / End in order). Configuration per test: GetMutableDefault<UMyAbility_Dash>()->AbilityTag = …, GetMutableDefault<UMyEffect_Burn>()->EffectData.Modifiers = …, with every field restored in AfterEach. Testing your real classes the same way works as long as they read their configuration from class defaults and not from assets that exist only in a cooked map.

Driving the component. Pick the call by what you need to advance:

Seams (1.4+, `WITH_AUTOMATION_WORKER` only, bottom of `Public/Components/GMCAbilityComponent.h`). They exist because GMC's role and replay queries are not virtual and several dispatch paths are private. GMAS_BugFixSpec.cpp, GMAS_ClientAuthSpec.cpp, GMAS_TaskSpec.cpp and GMAS_AbilityEndSpec.cpp show each in use. They are inspection hatches, not a substitute for layer 3: a forced flag proves the branch runs, not that the network takes it.