Six Issues, One Branch
The last session landed the universal routing strategy convention (#634) — NamedStrategy, StrategyResolver, five SPIs retrofitted. That work left a trail of follow-on issues: evaluation happening in the wrong place, non-deterministic defaults, a concurrency race nobody had noticed, and three methods that still threw UnsupportedOperationException.
I wanted to clear the deck. Six issues, all on one branch, ranging from a one-line annotation swap to a full case-lifecycle tracker.
The Race That Was Always There
The PlanItem concurrency fix turned out to be more interesting than the issue title suggested. The SequentialPlanningStrategy integration test was flagged as “bindings fire concurrently” — which sounds like a test timing problem. It’s not. It’s a production race condition.
PlanItem.status was volatile with check-then-set transitions. Two concurrent CONTEXT_CHANGED evaluations — perfectly normal on the Vert.x worker pool — could both read PENDING, both call markRunning(), both dispatch. The fix was AtomicReference with a tryMarkRunning() CAS, plus merging the two-step filter-then-index into a single atomic filterAndIndexForDispatch. Only the thread that wins the CAS dispatches; the loser’s binding quietly drops from the returned list.
The harder question was whether to also serialize all CONTEXT_CHANGED processing per case. We filed that separately — the CAS guard is correct for CapabilityTarget bindings, but non-CapabilityTarget bindings (HumanTask, SubCase) still have a TOCTOU gap because their handlers own the state transition. That gap needs its own analysis of re-entrant paths and deadlock surface.
Spawning Cases From Workers
The WorkerRuntime.spawnCase()/awaitCase() implementation was the largest piece. The design needed three things: a way to find a case definition by name, a way to track when a child case completes, and careful handling of the race where the child finishes before the parent registers interest.
CaseDefinitionRegistry.findByName() scans by name and throws on ambiguity — in Tier 1 (same JVM), case names are almost always unique. CaseCompletionTracker is a ConcurrentHashMap<UUID, CompletableFuture<CaseContext>> with a deliberate asymmetry: register() creates entries, complete() only acts on existing entries. No orphans — if nobody is awaiting a case, the terminal notification is a no-op.
The race guard in awaitCase() handles the out-of-order case by checking CaseInstanceCache after registration. The cache has no eviction, so terminal instances are always present. The design review caught several things I’d have missed: CaseContext mutability (fixed with snapshot()), the need for CaseTerminatedException to propagate through ExecutionException unwrapped, and a tracker leak where complete() originally used remove() instead of get().
The Default That Wasn’t Deterministic
EngineStrategyResolver used putIfAbsent to select the default strategy per type — which meant whichever CDI bean iterated first won. CDI iteration order is not deterministic. The fix uses Quarkus ARC’s InjectableBean.isDefaultBean() to detect @DefaultBean strategies explicitly. @DefaultBean wins unconditionally; multiple @DefaultBean for the same type throws at startup. This is Quarkus-specific, but the resolver is already @Alternative @Priority(1) @ApplicationScoped — it’s not pretending to be portable.
Moving the Gate Evaluation
The gate evaluation fix was structurally clean. ActionGateWorkItemHandler was evaluating CandidateSetStrategy with null context — it lives in work-adapter and has no case context. Moving the evaluation to WorkflowExecutionCompletedHandler.handleGate(), where CaseInstance is in scope, and carrying the resolved Set<String> in the event, was the obvious fix. The design review added failure recovery — if the strategy evaluation fails, the gate still fires with empty candidate groups and a warning, rather than silently dropping.
The routing strategy convention protocol and PLATFORM.md updates landed as cross-repo commits to casehubio/parent and the garden. The convention now has a formal home: per-case selectable strategies extend NamedStrategy, declare id(), ship @DefaultBean, resolve via StrategyResolver.
Six issues, fourteen commits, two follow-on issues filed. The deck is clear.