diff --git a/packages/opencode/src/bus/global.ts b/packages/opencode/src/bus/global.ts
index 43386dd6b20..dc23020d27c 100644
--- a/packages/opencode/src/bus/global.ts
+++ b/packages/opencode/src/bus/global.ts
@@ -8,3 +8,4 @@ export const GlobalBus = new EventEmitter<{
     },
   ]
 }>()
+GlobalBus.setMaxListeners(50) // kilocode_change — surface warning if SSE listeners accumulate
diff --git a/packages/opencode/src/server/routes/event.ts b/packages/opencode/src/server/routes/event.ts
index 989b857710d..565486d3c3a 100644
--- a/packages/opencode/src/server/routes/event.ts
+++ b/packages/opencode/src/server/routes/event.ts
@@ -70,14 +70,27 @@ export const EventRoutes = () =>
 
         stream.onAbort(stop)
 
+        // kilocode_change start
+        // On Windows, stream.onAbort() may never fire after a client disconnects
+        // (delayed TCP RST detection via IOCP). Without this try/catch, the
+        // GlobalBus listener, heartbeat interval, and AsyncQueue stay alive
+        // indefinitely for each dead connection — leaking memory on every
+        // SSE reconnect. Catching write errors lets us clean up eagerly.
         try {
           for await (const data of q) {
             if (data === null) return
-            await stream.writeSSE({ data })
+            try {
+              await stream.writeSSE({ data })
+            } catch {
+              log.info("event write failed, cleaning up dead stream")
+              stop()
+              return
+            }
           }
         } finally {
           stop()
         }
+        // kilocode_change end
       })
     },
   )
diff --git a/packages/opencode/src/server/routes/global.ts b/packages/opencode/src/server/routes/global.ts
index 733a0ef6b27..ddfa4c9acad 100644
--- a/packages/opencode/src/server/routes/global.ts
+++ b/packages/opencode/src/server/routes/global.ts
@@ -56,14 +56,27 @@ async function streamEvents(c: Context, subscribe: (q: AsyncQueue<string | null>
 
     stream.onAbort(stop)
 
+    // kilocode_change start
+    // On Windows, stream.onAbort() may never fire after a client disconnects
+    // (delayed TCP RST detection via IOCP). Without this try/catch, the
+    // GlobalBus listener, heartbeat interval, and AsyncQueue stay alive
+    // indefinitely for each dead connection — leaking memory on every
+    // SSE reconnect. Catching write errors lets us clean up eagerly.
     try {
       for await (const data of q) {
         if (data === null) return
-        await stream.writeSSE({ data })
+        try {
+          await stream.writeSSE({ data })
+        } catch {
+          log.info("global event write failed, cleaning up dead stream")
+          stop()
+          return
+        }
       }
     } finally {
       stop()
     }
+    // kilocode_change end
   })
 }