fix(compression): use extract_content_or_reasoning for reasoning model summaries

Reasoning models (DeepSeek-R1, Qwen-QwQ, glm-5-turbo) sometimes put
all output inside think/reasoning blocks with an empty content field.
The compressor was reading raw response.choices[0].message.content
directly, getting an empty string, and silently dropping middle turns
without a meaningful summary.

Use the existing extract_content_or_reasoning() helper (from
auxiliary_client) which already handles:
- Empty content + structured reasoning field fallback
- XML-style think/thinking/reasoning tag stripping

Also normalize dict content (llama.cpp) before extraction to prevent
type errors.

Tests: 3 new cases covering reasoning-only, think-tag, and normal
content extraction paths.

agent/context_compressor.py

@@ -352,8 +352,16 @@ Write only the summary body. Do not include any preamble or prefix."""
             if self.summary_model:
                 call_kwargs["model"] = self.summary_model
             response = call_llm(**call_kwargs)
-            content = response.choices[0].message.content
-            # Handle cases where content is not a string (e.g., dict from llama.cpp)
+            # Normalize dict content (e.g. llama.cpp tool calls) before
+            # extracting, then use extract_content_or_reasoning to handle
+            # models that put all output inside think/reasoning blocks with
+            # empty content field (e.g. DeepSeek-R1, Qwen-QwQ, glm-5-turbo).
+            raw_content = response.choices[0].message.content
+            if isinstance(raw_content, dict):
+                raw_content = str(raw_content) if raw_content else ""
+                response.choices[0].message.content = raw_content
+            from agent.auxiliary_client import extract_content_or_reasoning
+            content = extract_content_or_reasoning(response)
             if not isinstance(content, str):
                 content = str(content) if content else ""
             summary = content.strip()

tests/agent/test_context_compressor.py

@@ -198,6 +198,68 @@ class TestNonStringContent:
         assert summary == SUMMARY_PREFIX
 
 
+class TestReasoningOnlyExtraction:
+    """Regression: reasoning models (DeepSeek-R1, QwQ, glm-5-turbo) that put
+    all output in reasoning/think blocks with empty content must still produce
+    a valid summary via extract_content_or_reasoning."""
+
+    @pytest.fixture
+    def _compressor(self):
+        with patch("agent.context_compressor.get_model_context_length", return_value=100000):
+            return ContextCompressor(model="test", quiet_mode=True)
+
+    def test_reasoning_content_extracted_as_summary(self, _compressor):
+        mock_response = MagicMock()
+        mock_response.choices = [MagicMock()]
+        mock_response.choices[0].message.content = ""
+        mock_response.choices[0].message.reasoning = "The user was working on feature X and completed steps 1-3."
+
+        messages = [
+            {"role": "user", "content": "do something"},
+            {"role": "assistant", "content": "ok"},
+        ]
+
+        with patch("agent.context_compressor.call_llm", return_value=mock_response):
+            summary = _compressor._generate_summary(messages)
+        assert isinstance(summary, str)
+        assert "feature X" in summary
+
+    def test_think_blocks_stripped_and_content_used(self, _compressor):
+        mock_response = MagicMock()
+        mock_response.choices = [MagicMock()]
+        mock_response.choices[0].message.content = "Let me analyze this.\n\n## Summary\nWorked on the API module."
+        mock_response.choices[0].message.reasoning = None
+
+        messages = [
+            {"role": "user", "content": "summarize"},
+            {"role": "assistant", "content": "done"},
+        ]
+
+        with patch("agent.context_compressor.call_llm", return_value=mock_response):
+            summary = _compressor._generate_summary(messages)
+        assert isinstance(summary, str)
+        assert "API module" in summary
+
+    def test_inline_think_tags_stripped(self, _compressor):
+        mock_response = MagicMock()
+        mock_response.choices = [MagicMock()]
+        # extract_content_or_reasoning strips <think/> XML tags but not emoji blocks
+        mock_response.choices[0].message.content = "<thinking>Internal reasoning here.</thinking>\n\nThe actual summary content."
+        mock_response.choices[0].message.reasoning = None
+
+        messages = [
+            {"role": "user", "content": "summarize"},
+            {"role": "assistant", "content": "done"},
+        ]
+
+        with patch("agent.context_compressor.call_llm", return_value=mock_response):
+            summary = _compressor._generate_summary(messages)
+        assert isinstance(summary, str)
+        assert "actual summary" in summary
+        # Think block content should not leak into summary
+        assert "Internal reasoning" not in summary
+
+
 class TestSummaryPrefixNormalization:
     def test_legacy_prefix_is_replaced(self):
         summary = ContextCompressor._with_summary_prefix("[CONTEXT SUMMARY]: did work")