diff --git a/pycodeloop/cli/serve.py b/pycodeloop/cli/serve.py
index 23bc3c9..82efbd4 100644
--- a/pycodeloop/cli/serve.py
+++ b/pycodeloop/cli/serve.py
@@ -72,6 +72,12 @@ def _respond_error(self, request_id, code: int, message: str) -> None:
     def _wire_callbacks(self) -> None:
         agent = self.flow.agent
 
+        def on_request(message_count: int, tool_count: int) -> None:
+            self._notify(
+                "chat/request",
+                {"messageCount": message_count, "toolCount": tool_count},
+            )
+
         def on_text_delta(delta: str) -> None:
             self._notify("chat/textDelta", {"delta": delta})
 
@@ -135,6 +141,7 @@ def confirm(name: str, preview: str) -> bool | str:
             finally:
                 self._confirm_waiters.pop(request_id, None)
 
+        agent.on_request = on_request
         agent.on_text_delta = on_text_delta
         agent.on_turn_end = on_turn_end
         agent.on_tool_call = on_tool_call
diff --git a/tests/functional/test_serve_end_to_end.py b/tests/functional/test_serve_end_to_end.py
index 2cd83c5..988dfa7 100644
--- a/tests/functional/test_serve_end_to_end.py
+++ b/tests/functional/test_serve_end_to_end.py
@@ -143,6 +143,37 @@ def test_chat_send_streams_tool_call_and_returns_final_text(self):
             response["result"]["text"], "Here's the directory listing."
         )
 
+    def test_chat_send_emits_on_request_notification(self):
+        self.server = FakeLLMServer([chat_completion(text="hi")])
+        self.client = ServeSubprocess(self.server.url, self.tmp_path)
+        self.client.next_message()  # ready
+
+        self.client.send(
+            {
+                "jsonrpc": "2.0",
+                "id": "1",
+                "method": "chat/send",
+                "params": {"prompt": "hello"},
+            }
+        )
+
+        messages = []
+        response = None
+        while response is None:
+            message = self.client.next_message()
+            if message.get("id") == "1" and "result" in message:
+                response = message
+            else:
+                messages.append(message)
+
+        request_notification = next(
+            m for m in messages if m.get("method") == "chat/request"
+        )
+        self.assertIsInstance(
+            request_notification["params"]["messageCount"], int
+        )
+        self.assertIsInstance(request_notification["params"]["toolCount"], int)
+
     def test_dangerous_tool_blocks_on_confirm_round_trip(self):
         self.server = FakeLLMServer(
             [