diff --git a/tests/tools/test_wake_word.py b/tests/tools/test_wake_word.py
index 0ac450f5c5..469f823531 100644
--- a/tests/tools/test_wake_word.py
+++ b/tests/tools/test_wake_word.py
@@ -641,7 +641,11 @@ def test_resolve_capture_mode_auto_and_prefer_client(monkeypatch):
     assert ww.resolve_capture_mode({"capture": "auto"}, force_local=True) == "local"
     monkeypatch.setattr(ww, "_local_input_device_ready", lambda: True)
     assert ww.resolve_capture_mode({"capture": "auto"}) == "local"
-    assert ww.resolve_capture_mode({"capture": "auto"}, prefer_client=True) == "client"
+    # A working backend mic wins under auto even for a preferring surface, so
+    # local desktops keep PortAudio + wake_word.input_device selection.
+    assert ww.resolve_capture_mode({"capture": "auto"}, prefer_client=True) == "local"
+    # Explicit client still forces streaming (backend mic exists but is wrong).
+    assert ww.resolve_capture_mode({"capture": "client"}, prefer_client=True) == "client"
 
 
 def test_requirements_client_capture_without_local_mic(monkeypatch):
@@ -674,7 +678,7 @@ def test_requirements_client_capture_without_local_mic(monkeypatch):
 
 
 def test_client_capture_feed_fires(monkeypatch, tmp_path):
-    import numpy as np
+    np = pytest.importorskip("numpy")
 
     monkeypatch.setattr(ww, "_build_engine", lambda cfg: _FakeEngine(fire=True))
     monkeypatch.setattr(ww, "_lock_path", lambda: tmp_path / "wake.lock")
diff --git a/tools/wake_word.py b/tools/wake_word.py
index b71c47cb0b..960bed50e2 100644
--- a/tools/wake_word.py
+++ b/tools/wake_word.py
@@ -269,15 +269,17 @@ def resolve_capture_mode(
         return "client"
     if raw == "local":
         return "local"
-    # auto
-    if prefer_client:
-        return "client"
-    # Prefer local when a PortAudio input exists. Without an explicit client
-    # preference (desktop remote), stay on local so CLI/TUI/status still
-    # require a server mic instead of advertising a capture path only the
-    # desktop can feed.
+    # auto: a working backend input always wins so local desktops keep
+    # PortAudio and the configured ``input_device`` selection. Client capture
+    # is the fallback for a preferring surface (desktop remote) on a backend
+    # with no usable mic — the headless VPS/Cloud case.
     if _local_input_device_ready():
         return "local"
+    if prefer_client:
+        return "client"
+    # No local mic and no client preference (CLI/TUI): stay local so status
+    # reports the real requirement instead of advertising a capture path
+    # nothing will feed.
     return "local"