diff --git a/cecli/coders/base_coder.py b/cecli/coders/base_coder.py
index 126781384fd..57c7ea15784 100755
--- a/cecli/coders/base_coder.py
+++ b/cecli/coders/base_coder.py
@@ -3984,7 +3984,7 @@ async def show_send_output(self, completion):
         if (
             not len(self.partial_response_content)
             and not len(self.partial_response_tool_calls)
-            and not len(self.partial_response_reasoning_content)
+            and not _is_meaningful_reasoning(self.partial_response_reasoning_content)
         ):
             self.empty_response = True
             return
@@ -4114,7 +4114,8 @@ async def show_send_output_stream(self, completion):
 
                             text += reasoning_content
                             self.got_reasoning_content = True
-                            received_content = True
+                            if _is_meaningful_reasoning(reasoning_content):
+                                received_content = True
 
                         self.token_profiler.on_token()
                         self.io.update_spinner_suffix(reasoning_content)
@@ -4190,7 +4191,14 @@ async def show_send_output_stream(self, completion):
             self.io.tool_warning("Execution stopped by on message hook")
             return
 
-        if not received_content and len(self.partial_response_tool_calls) == 0:
+        # Treat the response as empty when nothing was received, or when the
+        # only thing received was reasoning made entirely of non-alphanumeric
+        # characters (e.g. moonshotai/kimi-k3 returning "!!!!").
+        if (
+            not received_content
+            and len(self.partial_response_tool_calls) == 0
+            and not _is_meaningful_reasoning(self.partial_response_reasoning_content)
+        ):
             self.empty_response = True
             return
 
@@ -5379,3 +5387,15 @@ def _first_usage_tokens(usage: object, paths: list[str], default: int = 0) -> in
         if value is not None:
             return value
     return default
+
+
+def _is_meaningful_reasoning(text):
+    """Return True if reasoning text contains at least one alphanumeric character.
+
+    Some providers (e.g. moonshotai/kimi-k3) occasionally return completions
+    with empty ``content`` and a ``reasoning_content`` made entirely of
+    punctuation (e.g. ``"!!!!"``). Those responses are effectively empty, so
+    the empty-response detector only lets reasoning count as response
+    content when it holds at least one alphanumeric character.
+    """
+    return bool(text) and any(ch.isalnum() for ch in text)
diff --git a/cecli/website/docs/config/retries.md b/cecli/website/docs/config/retries.md
index 3140c2de6ed..0aa01c39f85 100644
--- a/cecli/website/docs/config/retries.md
+++ b/cecli/website/docs/config/retries.md
@@ -11,6 +11,7 @@ Cecli can be configured to retry failed API calls. This is useful for handling i
 - `retry-timeout`: The timeout in seconds for each retry.
 - `retry-backoff-factor`: The backoff factor to use between retries.
 - `retry-on-unavailable`: Whether to retry on 503 Service Unavailable errors.
+- `retry-on-empty`: Whether to retry when the model returns an empty response. A response is considered empty when it has no content, no tool calls, and no *meaningful* reasoning. Reasoning that contains no alphanumeric characters (for example a `reasoning_content` of `"!!!!"`) does not count as a response, so it is retried like any other empty response.
 - `retry-on-unauthorized`: Whether to retry on 401 Unauthorized (and 403 Forbidden) errors. Default: false.
 
 Example usage in `.cecli.conf.yml`:
diff --git a/tests/basic/test_empty_response.py b/tests/basic/test_empty_response.py
new file mode 100644
index 00000000000..4b562fc7222
--- /dev/null
+++ b/tests/basic/test_empty_response.py
@@ -0,0 +1,223 @@
+"""Tests for empty-response detection with non-meaningful reasoning content.
+
+Regression coverage for the ``retry-on-empty`` fix: a completion with empty
+``content``, no tool calls, and a ``reasoning_content`` made entirely of
+non-alphanumeric characters (e.g. moonshotai/kimi-k3 returning ``"!!!!"``) must
+be classified as an empty response so the existing retry loop engages.
+"""
+
+from types import SimpleNamespace
+
+import pytest
+
+from cecli.coders.base_coder import Coder, _is_meaningful_reasoning
+from cecli.helpers.threading import ThreadSafeEvent
+from cecli.llm import litellm
+
+
+# --------------------------------------------------------------------------- #
+# Test doubles
+# --------------------------------------------------------------------------- #
+class _AlwaysSetEvent:
+    def is_set(self):
+        return True
+
+
+class _FakeIO:
+    def __init__(self):
+        self.confirmation_in_progress_event = _AlwaysSetEvent()
+        self.assistant_outputs = []
+
+    def tool_error(self, *a, **k):
+        pass
+
+    def tool_warning(self, *a, **k):
+        pass
+
+    def update_spinner_suffix(self, *a, **k):
+        pass
+
+    def reset_streaming_response(self):
+        pass
+
+    def stream_output(self, *a, **k):
+        pass
+
+    def ai_output(self, *a, **k):
+        pass
+
+    def tool_output(self, *a, **k):
+        pass
+
+    def assistant_output(self, *a, **k):
+        self.assistant_outputs.append(a)
+
+
+class _FakeTokenProfiler:
+    def start(self):
+        pass
+
+    def on_token(self):
+        pass
+
+    def on_error(self):
+        pass
+
+    def add_to_usage_report(self, *a, **k):
+        return a[0] if a else ""
+
+
+def _make_coder(stream=False):
+    coder = Coder.__new__(Coder)
+    coder.stream = stream
+    coder.verbose = False
+    coder.args = SimpleNamespace(debug=False, show_thinking=False)
+    coder.io = _FakeIO()
+    coder.interrupt_event = ThreadSafeEvent()
+    coder.pretty = False
+    coder.reasoning_tag_name = "THINKING"
+    coder.got_reasoning_content = False
+    coder.ended_reasoning_content = False
+    coder.empty_response = False
+    coder.tool_reflection = False
+    coder.partial_response_content = ""
+    coder.partial_response_reasoning_content = ""
+    coder.partial_response_chunks = []
+    coder.partial_response_tool_calls = []
+    coder.partial_response_function_call = dict()
+    coder.partial_response_consolidated = None
+    coder.multi_response_content = ""
+    coder.chat_completion_response_hashes = []
+    coder._streaming_buffer_length = 0
+    coder.token_profiler = _FakeTokenProfiler()
+    coder._output_loop_detected = False
+    coder._output_loop_message = ""
+    coder._has_empty_reflected = False
+    coder.edit_format = "code"
+    return coder
+
+
+def _tc(index, call_id, name, arguments):
+    return litellm.ChatCompletionMessageToolCall(
+        id=call_id,
+        function=litellm.Function(arguments=arguments or "", name=name),
+        type="function",
+        index=index,
+    )
+
+
+def _non_streaming_completion(content="", reasoning=None, tool_calls=None):
+    message = {"role": "assistant", "content": content}
+    if reasoning is not None:
+        message["reasoning_content"] = reasoning
+    if tool_calls is not None:
+        message["tool_calls"] = tool_calls
+    return litellm.ModelResponse(
+        id="test-completion",
+        created=0,
+        model="gpt-test",
+        object="chat.completion",
+        choices=[{"finish_reason": "stop", "index": 0, "message": message}],
+        usage={"completion_tokens": 1, "prompt_tokens": 1, "total_tokens": 2},
+    )
+
+
+def _reasoning_chunk(text):
+    delta = litellm.Delta(role="assistant", content=None, reasoning_content=text)
+    choice = litellm.StreamChoice(finish_reason=None, index=0, delta=delta)
+    return litellm.StreamChunk(
+        id="cmpl-test", created=1000, model="gpt-test", choices=[choice], usage=None
+    )
+
+
+async def _agen(chunks):
+    for chunk in chunks:
+        yield chunk
+
+
+# --------------------------------------------------------------------------- #
+# _is_meaningful_reasoning unit tests
+# --------------------------------------------------------------------------- #
+def test_is_meaningful_reasoning_empty_and_none():
+    assert _is_meaningful_reasoning("") is False
+    assert _is_meaningful_reasoning(None) is False
+
+
+def test_is_meaningful_reasoning_punctuation_only():
+    assert _is_meaningful_reasoning("!!!!") is False
+    assert _is_meaningful_reasoning("...?!,;:") is False
+
+
+def test_is_meaningful_reasoning_whitespace_only():
+    assert _is_meaningful_reasoning("   \n\t  ") is False
+
+
+def test_is_meaningful_reasoning_normal_text():
+    assert _is_meaningful_reasoning("Let me think about this") is True
+
+
+def test_is_meaningful_reasoning_mixed_punctuation_and_alnum():
+    assert _is_meaningful_reasoning("42?") is True
+    assert _is_meaningful_reasoning("!!!a!!!") is True
+
+
+# --------------------------------------------------------------------------- #
+# Non-streaming path (show_send_output)
+# --------------------------------------------------------------------------- #
+@pytest.mark.asyncio
+async def test_junk_reasoning_is_empty_non_streaming():
+    coder = _make_coder(stream=False)
+    await coder.show_send_output(_non_streaming_completion(content="", reasoning="!" * 40))
+    assert coder.empty_response is True
+
+
+@pytest.mark.asyncio
+async def test_meaningful_reasoning_is_not_empty_non_streaming():
+    coder = _make_coder(stream=False)
+    await coder.show_send_output(
+        _non_streaming_completion(content="", reasoning="Let me think about this")
+    )
+    assert coder.empty_response is False
+
+
+@pytest.mark.asyncio
+async def test_tool_calls_are_not_empty_non_streaming():
+    coder = _make_coder(stream=False)
+    await coder.show_send_output(
+        _non_streaming_completion(
+            content="",
+            reasoning="",
+            tool_calls=[_tc(0, "call_1", "Local--ls", "{}")],
+        )
+    )
+    assert coder.empty_response is False
+
+
+@pytest.mark.asyncio
+async def test_mixed_reasoning_is_not_empty_non_streaming():
+    coder = _make_coder(stream=False)
+    await coder.show_send_output(_non_streaming_completion(content="", reasoning="42?"))
+    assert coder.empty_response is False
+
+
+# --------------------------------------------------------------------------- #
+# Streaming path (show_send_output_stream)
+# --------------------------------------------------------------------------- #
+@pytest.mark.asyncio
+async def test_junk_reasoning_is_empty_streaming():
+    coder = _make_coder(stream=True)
+    coder.args.show_thinking = True
+    async for _ in coder.show_send_output_stream(_agen([_reasoning_chunk("!" * 40)])):
+        pass
+    assert coder.empty_response is True
+
+
+@pytest.mark.asyncio
+async def test_meaningful_reasoning_is_not_empty_streaming():
+    coder = _make_coder(stream=True)
+    coder.args.show_thinking = True
+    async for _ in coder.show_send_output_stream(
+        _agen([_reasoning_chunk("Let me think about this")])
+    ):
+        pass
+    assert coder.empty_response is False