diff --git a/.changeset/cache-step-port-discovery.md b/.changeset/cache-step-port-discovery.md
new file mode 100644
index 0000000000..28903e56df
--- /dev/null
+++ b/.changeset/cache-step-port-discovery.md
@@ -0,0 +1,5 @@
+---
+'@workflow/core': patch
+---
+
+Cache self-hosted step port discovery per process.
diff --git a/packages/core/src/runtime/get-port-lazy.ts b/packages/core/src/runtime/get-port-lazy.ts
index e38e908bb3..eab677ee43 100644
--- a/packages/core/src/runtime/get-port-lazy.ts
+++ b/packages/core/src/runtime/get-port-lazy.ts
@@ -1,11 +1,11 @@
 import { getPort } from '@workflow/utils/get-port';
 
-// Per-process cache of the resolved port. The workflow dev server listens on a
+// Per-process cache of the resolved port. The workflow server listens on a
 // stable port for the lifetime of the process, but `getPort()` rediscovers it
 // on every call by querying the OS for the process's listening sockets — on
 // macOS that shells out to `lsof` (~60ms), which the runtime pays on EVERY
-// workflow replay (the non-Vercel branch of `runWorkflow`). Since the port does
-// not change within a process, resolve it once and reuse it. `_inFlight`
+// workflow replay or step invocation. Since the port does not change within a
+// process, resolve it once and reuse it. `_inFlight`
 // dedupes concurrent first calls so discovery never runs more than once.
 //
 // The first concrete port is pinned for the lifetime of the process — there is
diff --git a/packages/core/src/runtime/step-handler.test.ts b/packages/core/src/runtime/step-handler.test.ts
index aefac8aee3..13476ab0be 100644
--- a/packages/core/src/runtime/step-handler.test.ts
+++ b/packages/core/src/runtime/step-handler.test.ts
@@ -134,23 +134,33 @@ vi.mock('../private.js', () => ({
   getStepFunction: vi.fn().mockReturnValue(mockStepFn),
 }));
 
-// Mock get-port
-vi.mock('@workflow/utils/get-port', () => ({
-  getPort: vi.fn().mockResolvedValue(3000),
-}));
-
 // Import the module AFTER all mocks are set up - this triggers createQueueHandler
 // which populates capturedHandlerRef
 import './step-handler.js';
-import { MAX_QUEUE_DELIVERIES } from './constants.js';
 import { getStepFunction } from '../private.js';
 import {
   getErrorName,
   getErrorStack,
   normalizeUnknownError,
 } from '../types.js';
+import { MAX_QUEUE_DELIVERIES } from './constants.js';
+import {
+  resetPortCacheForTesting,
+  setPortResolverForTesting,
+} from './get-port-lazy.js';
 import { getWorld } from './world.js';
 
+const mockPortResolver = vi.fn(async () => 3000);
+
+beforeEach(() => {
+  mockPortResolver.mockReset().mockResolvedValue(3000);
+  setPortResolverForTesting(mockPortResolver);
+});
+
+afterEach(() => {
+  resetPortCacheForTesting();
+});
+
 function capturedHandler(
   message: unknown,
   metadata: { queueName: string; messageId: string; attempt: number }
@@ -228,6 +238,13 @@ describe('step-handler 409 handling', () => {
     vi.restoreAllMocks();
   });
 
+  it('reuses cached port discovery across step invocations', async () => {
+    await capturedHandler(createMessage(), createMetadata('myStep'));
+    await capturedHandler(createMessage(), createMetadata('myStep'));
+
+    expect(mockPortResolver).toHaveBeenCalledOnce();
+  });
+
   describe('step_completed 409', () => {
     it('should warn and return when step_completed gets a 409', async () => {
       // step_started succeeds, step function succeeds, step_completed returns 409
diff --git a/packages/core/src/runtime/step-handler.ts b/packages/core/src/runtime/step-handler.ts
index 440612826d..c0efefd524 100644
--- a/packages/core/src/runtime/step-handler.ts
+++ b/packages/core/src/runtime/step-handler.ts
@@ -10,7 +10,6 @@ import {
   WorkflowWorldError,
 } from '@workflow/errors';
 import { createWorkflowBaseUrl, pluralize } from '@workflow/utils';
-import { getPort } from '@workflow/utils/get-port';
 import {
   getQueueTopicPrefix,
   resolveQueueNamespace,
@@ -41,6 +40,7 @@ import {
 } from '../types.js';
 
 import { MAX_QUEUE_DELIVERIES } from './constants.js';
+import { getPortLazy } from './get-port-lazy.js';
 import {
   getQueueOverhead,
   getWorkflowQueueName,
@@ -158,7 +158,7 @@ const stepHandler = createQueueHandler(
 
       // Resolve local async values concurrently before entering the trace span
       const [port, spanKind] = await Promise.all([
-        isVercel ? undefined : getPort(),
+        isVercel ? undefined : getPortLazy(),
         getSpanKind('CONSUMER'),
       ]);
 
diff --git a/packages/world-postgres/src/queue.test.ts b/packages/world-postgres/src/queue.test.ts
index 9a48b0d37b..2c2079d5d1 100644
--- a/packages/world-postgres/src/queue.test.ts
+++ b/packages/world-postgres/src/queue.test.ts
@@ -29,6 +29,9 @@ vi.mock('graphile-worker', () => ({
 
 vi.mock('@workflow/utils/get-port', () => ({
   getWorkflowPort: vi.fn(),
+  // stepEntrypoint (imported from the core dist bundle below) calls
+  // getPortLazy(), which imports getPort from this same module.
+  getPort: vi.fn().mockResolvedValue(undefined),
 }));
 
 vi.mock('@workflow/world-local', async (importOriginal) => {