diff --git a/src/kimi_cli/wire/server.py b/src/kimi_cli/wire/server.py
index fecc48c83..17275c955 100644
--- a/src/kimi_cli/wire/server.py
+++ b/src/kimi_cli/wire/server.py
@@ -79,6 +79,8 @@ def __init__(self, soul: Soul):
         # soul running stuffs
         self._soul = soul
         self._cancel_event: asyncio.Event | None = None
+        self._turn_active: bool = False
+        """True only while the soul is actively running a turn (between run_soul start and return)."""
         self._pending_requests: dict[str, Request] = {}
         """Maps JSON RPC message IDs to pending `Request`s."""
         self._client_supports_question: bool = False
@@ -540,6 +542,7 @@ async def _handle_prompt(
             )
 
         self._cancel_event = asyncio.Event()
+        self._turn_active = True
         try:
             runtime = self._soul.runtime if isinstance(self._soul, KimiSoul) else None
             await run_soul(
@@ -580,6 +583,9 @@ async def _handle_prompt(
                 result={"status": Statuses.CANCELLED},
             )
         finally:
+            # Mark the turn as inactive immediately so concurrent steer/cancel
+            # handlers see the correct state before _cancel_event is cleared.
+            self._turn_active = False
             # Clean up any remaining pending requests from this turn.
             # After run_soul() returns, the soul and all subagents are done,
             # so any unresolved requests are stale.
@@ -609,7 +615,7 @@ async def _handle_prompt(
     async def _handle_steer(
         self, msg: JSONRPCSteerMessage
     ) -> JSONRPCSuccessResponse | JSONRPCErrorResponse:
-        if not isinstance(self._soul, KimiSoul) or not self._is_streaming:
+        if not isinstance(self._soul, KimiSoul) or not self._turn_active:
             return JSONRPCErrorResponse(
                 id=msg.id,
                 error=JSONRPCErrorObject(
diff --git a/tests/core/test_wire_server_steer.py b/tests/core/test_wire_server_steer.py
index 15a35edce..02ee1497d 100644
--- a/tests/core/test_wire_server_steer.py
+++ b/tests/core/test_wire_server_steer.py
@@ -66,6 +66,7 @@ async def test_handle_steer_queues_input_when_streaming(
 
     monkeypatch.setattr(soul, "steer", lambda user_input: queued.append(user_input))
     server._cancel_event = asyncio.Event()
+    server._turn_active = True
 
     response = await server._handle_steer(
         JSONRPCSteerMessage(
@@ -79,6 +80,31 @@ async def test_handle_steer_queues_input_when_streaming(
     assert queued == [[TextPart(text="follow-up")]]
 
 
+@pytest.mark.asyncio
+async def test_handle_steer_rejects_after_turn_ends(
+    runtime: Runtime,
+    tmp_path: Path,
+) -> None:
+    """Steer sent after run_soul() returns but before _cancel_event is cleared should be rejected."""
+    soul = _make_soul(runtime, tmp_path)
+    server = WireServer(soul)
+
+    # Simulate the state after run_soul() returns but before finally cleanup:
+    # _cancel_event is still set (not yet None), but _turn_active is False.
+    server._cancel_event = asyncio.Event()
+    server._turn_active = False
+
+    response = await server._handle_steer(
+        JSONRPCSteerMessage(
+            id="1",
+            params=JSONRPCSteerMessage.Params(user_input=[TextPart(text="too late")]),
+        )
+    )
+
+    assert isinstance(response, JSONRPCErrorResponse)
+    assert response.error.code == ErrorCodes.INVALID_STATE
+
+
 @pytest.mark.asyncio
 async def test_shutdown_rejects_foreground_approval_in_runtime(
     runtime: Runtime,