diff --git a/src/lib/gateway-chat-pool.test.ts b/src/lib/gateway-chat-pool.test.ts
index 493776b..273ffb1 100644
--- a/src/lib/gateway-chat-pool.test.ts
+++ b/src/lib/gateway-chat-pool.test.ts
@@ -187,4 +187,23 @@ describe("getGatewayClientForRuntime", () => {
     expect(mockState.clients[0].close).toHaveBeenCalledTimes(1);
     expect(getGatewayPoolDiagnostics()).toMatchObject({ poolSize: 1 });
   });
+
+  it("replaces disconnected held clients instead of returning them", async () => {
+    addRuntime("runtime-a");
+
+    const disconnectedClient = await getGatewayClientForRuntime("runtime-a");
+    holdClient(disconnectedClient);
+    mockState.clients[0].isConnected = false;
+
+    const freshClient = await getGatewayClientForRuntime("runtime-a");
+
+    expect(freshClient).not.toBe(disconnectedClient);
+    expect(mockState.clients).toHaveLength(2);
+    expect(mockState.clients[0].close).toHaveBeenCalledTimes(1);
+    expect(getGatewayPoolDiagnostics()).toMatchObject({
+      activeClients: 0,
+      activeHolds: 0,
+      poolSize: 1,
+    });
+  });
 });
diff --git a/src/lib/gateway-chat-pool.ts b/src/lib/gateway-chat-pool.ts
index 2263ec2..cf469c4 100644
--- a/src/lib/gateway-chat-pool.ts
+++ b/src/lib/gateway-chat-pool.ts
@@ -505,7 +505,7 @@ async function getClientForRuntime(runtime: GatewayRuntimeConnection | null | un
 
   // Close stale connection — but only if no active request is using it
   if (existing) {
-    if (activeClientHolds.has(existing.client)) {
+    if (activeClientHolds.has(existing.client) && existing.client.isConnected) {
       // Client is mid-request, return it anyway (don't recycle under its feet)
       publishAgentModeDiagnostic({
         scope: "gateway-pool",
@@ -519,6 +519,7 @@ async function getClientForRuntime(runtime: GatewayRuntimeConnection | null | un
       });
       return existing.client;
     }
+    activeClientHolds.delete(existing.client);
     publishAgentModeDiagnostic({
       scope: "gateway-pool",
       event: "client.recycle.close-stale",