Layer 5: The Gate That Works
Layer 5 is done. TacticsTask now activates only when scouting has found a threat
and strategy has set a course of action. In the early game — no enemy contact — tactics
never runs.
The mechanism is entryCriteria() on each plugin. The CaseEngine checks which keys are
present in the CaseFile before scheduling a task; any missing key means the task is
skipped. So TacticsTask declaring {READY, STRATEGY, NEAREST_THREAT} as its criteria
means it sits out until scouting writes NEAREST_THREAT. Scouting only writes that when
enemies are visible. The gate is real.
StrategyTask got {READY, ENEMY_ARMY_SIZE} — not a genuine gate (scouting always
writes ENEMY_ARMY_SIZE, even as 0), but an ordering commitment. Strategy runs after
scouting every tick. When C2 lands and strategy migrates to scouting-derived intel
(ENEMY_POSTURE, ENEMY_BUILD_ORDER) instead of raw ENEMY_UNITS, the ordering
becomes a proper data dependency. The stub read is in both strategy implementations,
tracked as #169.
Two things the foundation didn’t do that we had to work around.
First: TaskDefinition.canActivate() in the installed snapshot returns true
unconditionally. The interface has the method; the default implementation ignores
entryCriteria() entirely. Bytecode confirmed: iconst_1; ireturn. Updating
entryCriteria() alone had no runtime effect. We overrode canActivate() explicitly
on all four affected plugins — the correct implementation the default should have had.
When casehub-core fixes the default, the overrides are removable. Garden entry filed
(GE-20260603-e98633). Protocol written.
Second: createAndSolve() returns the CaseFile immediately — before plugins have run.
The duration parameter goes to a PropagationContext timeout, not to a blocking await.
So the CaseFile we get back from the tick contains only translator-written keys
(MINERALS, ARMY, ENEMY_UNITS). Plugin-written keys (NEAREST_THREAT, STRATEGY)
are absent. Garden entry (GE-20260603-a944d1). Tests use canActivate() directly on
the injected CDI beans to verify gate semantics — cleaner than waiting on async solve
completion, and it tests the right thing.
Also extracted GameTickExecutor from AgentOrchestrator.gameTick(). The original
discarded the CaseFile returned by createAndSolve(). Now AgentOrchestrator exposes
getLastTickResult() — the test surface the integration tests needed.
The distinction between ordering dependencies and conditional gates turned out to be
important enough to document explicitly in ARC42STORIES.MD and the LAYER-LOG. They
look identical in the code (entryCriteria() returning a set of keys) but serve
different architectural purposes. ENEMY_ARMY_SIZE on strategy is a sequencing
constraint that will become a data dependency at C2. NEAREST_THREAT on tactics is
a genuine adaptive gate today.