diff --git a/core/src/JobSystem.cpp b/core/src/JobSystem.cpp
index f072136..bf341db 100644
--- a/core/src/JobSystem.cpp
+++ b/core/src/JobSystem.cpp
@@ -158,3 +158,207 @@ JobSystem::~JobSystem() {
     for (std::thread& worker : workers) worker.join();
 }
 
+void JobSystem::completeState(const std::shared_ptr<JobHandle::State>& state,
+                              std::exception_ptr error) {
+    {
+        std::lock_guard<std::mutex> lock(state->mutex);
+        state->error = error;
+        state->finished = true;
+    }
+    state->cv.notify_all();
+}
+
+std::size_t JobSystem::queuedLocked() const {
+    std::size_t total = 0;
+    for (int i = 0; i < jobPriorityCount; ++i) total += queues[i].size();
+    return total;
+}
+
+JobSystem::QueuedJob JobSystem::popLocked() {
+    for (int i = 0; i < jobPriorityCount; ++i) {
+        if (queues[i].empty()) continue;
+        QueuedJob job = std::move(queues[i].front());
+        queues[i].pop_front();
+        return job;
+    }
+    return QueuedJob{};
+}
+
+void JobSystem::enqueue(std::function<void()> work, std::shared_ptr<JobHandle::State> state,
+                        JobPriority priority) {
+    {
+        std::lock_guard<std::mutex> lock(mutex);
+        ++counters.submitted;
+        queues[static_cast<int>(priority)].push_back(QueuedJob{std::move(work), std::move(state)});
+        std::uint32_t depth = static_cast<std::uint32_t>(queuedLocked());
+        if (depth > counters.peakQueueDepth) counters.peakQueueDepth = depth;
+    }
+    workCv.notify_one();
+}
+
+void JobSystem::noteInlineStart() {
+    std::lock_guard<std::mutex> lock(mutex);
+    ++counters.submitted;
+    if (counters.peakBusyWorkers < 1u) counters.peakBusyWorkers = 1u;
+}
+
+void JobSystem::noteInlineFinish(bool failed) {
+    std::lock_guard<std::mutex> lock(mutex);
+    ++counters.executed;
+    if (failed) ++counters.failed;
+}
+
+void JobSystem::noteFailure() {
+    std::lock_guard<std::mutex> lock(mutex);
+    ++counters.failed;
+}
+
+void JobSystem::runInline(std::function<void()> work,
+                          const std::shared_ptr<JobHandle::State>& state) {
+    noteInlineStart();
+    std::exception_ptr captured;
+    try {
+        work();
+    } catch (...) {
+        captured = std::current_exception();
+    }
+    completeState(state, captured);
+    noteInlineFinish(captured != nullptr);
+}
+
+JobHandle JobSystem::submit(std::function<void()> work, JobPriority priority) {
+    auto state = std::make_shared<JobHandle::State>();
+    if (workers.empty()) {
+        runInline(std::move(work), state);
+        return JobHandle(state);
+    }
+    enqueue(std::move(work), state, priority);
+    return JobHandle(state);
+}
+
+JobBatch JobSystem::submitBatch(std::vector<std::function<void()>> work, JobPriority priority) {
+    JobBatch batch;
+    for (std::function<void()>& job : work) batch.add(submit(std::move(job), priority));
+    return batch;
+}
+
+void JobSystem::workerLoop() {
+    for (;;) {
+        std::unique_lock<std::mutex> lock(mutex);
+        workCv.wait(lock, [this] { return stopping || queuedLocked() > 0; });
+        if (queuedLocked() == 0) {
+            if (stopping) return;
+            continue;
+        }
+
+        QueuedJob job = popLocked();
+        ++counters.queuePops;
+        ++busy;
+        if (busy > counters.peakBusyWorkers) counters.peakBusyWorkers = busy;
+        lock.unlock();
+
+        std::exception_ptr captured;
+        try {
+            job.work();
+        } catch (...) {
+            captured = std::current_exception();
+        }
+        if (job.state) completeState(job.state, captured);
+
+        lock.lock();
+        --busy;
+        ++counters.executed;
+        if (captured) ++counters.failed;
+        bool idle = busy == 0 && queuedLocked() == 0;
+        lock.unlock();
+        if (idle) idleCv.notify_all();
+    }
+}
+
+void JobSystem::waitAll() {
+    if (workers.empty()) return;
+    std::unique_lock<std::mutex> lock(mutex);
+    idleCv.wait(lock, [this] { return busy == 0 && queuedLocked() == 0; });
+}
+
+bool JobSystem::executeGraphNode(const std::shared_ptr<GraphRun>& run, std::size_t index) {
+    std::exception_ptr captured;
+    if (run->depFailed[index].load(std::memory_order_acquire)) {
+        captured = std::make_exception_ptr(
+            std::runtime_error("dependency failed, skipped job " + run->names[index]));
+    } else {
+        try {
+            run->work[index]();
+        } catch (...) {
+            captured = std::current_exception();
+        }
+    }
+    completeState(run->states[index], captured);
+
+    for (std::size_t dependent : run->dependents[index]) {
+        if (captured) run->depFailed[dependent].store(true, std::memory_order_release);
+        if (run->remaining[dependent].fetch_sub(1, std::memory_order_acq_rel) == 1 && !run->serial)
+            scheduleGraphNode(run, dependent);
+    }
+    return captured != nullptr;
+}
+
+void JobSystem::scheduleGraphNode(const std::shared_ptr<GraphRun>& run, std::size_t index) {
+    JobPriority priority = run->priorities[index];
+    enqueue([this, run, index] {
+                if (executeGraphNode(run, index)) noteFailure();
+            },
+            nullptr, priority);
+}
+
+JobBatch JobSystem::runGraph(const JobGraph& graph) {
+    JobBatch batch;
+    std::size_t count = graph.nodeCount();
+    std::vector<std::size_t> order = graph.topologicalOrder();
+    if (order.size() != count) {
+        batch.runnable = false;
+        return batch;
+    }
+
+    auto run = std::make_shared<GraphRun>(count);
+    run->serial = workers.empty();
+    for (std::size_t i = 0; i < count; ++i) {
+        const JobNode& node = graph.nodeAt(i);
+        run->work[i] = node.work;
+        run->names[i] = node.name;
+        run->priorities[i] = node.priority;
+        run->remaining[i].store(static_cast<int>(node.dependencies.size()),
+                                std::memory_order_relaxed);
+        run->states[i] = std::make_shared<JobHandle::State>();
+        batch.add(JobHandle(run->states[i]));
+    }
+    for (std::size_t i = 0; i < count; ++i)
+        for (std::size_t dep : graph.nodeAt(i).dependencies) run->dependents[dep].push_back(i);
+
+    if (run->serial) {
+        for (std::size_t index : order) {
+            noteInlineStart();
+            bool failed = executeGraphNode(run, index);
+            noteInlineFinish(failed);
+        }
+        return batch;
+    }
+
+    std::vector<std::size_t> roots;
+    for (std::size_t i = 0; i < count; ++i)
+        if (run->remaining[i].load(std::memory_order_relaxed) == 0) roots.push_back(i);
+    for (std::size_t index : roots) scheduleGraphNode(run, index);
+    return batch;
+}
+
+JobStats JobSystem::stats() const {
+    std::lock_guard<std::mutex> lock(mutex);
+    return counters;
+}
+
+void JobSystem::resetStats() {
+    std::lock_guard<std::mutex> lock(mutex);
+    counters = JobStats{};
+}
+
+} // namespace vf