Merge ce805b771e into 4fade39c90

run_agent.py

@@ -5854,6 +5854,26 @@ class AIAgent:
             except Exception:
                 pass
 
+    def _fire_reasoning_available(self, assistant_message) -> None:
+        """Fire structured reasoning progress if registered.
+
+        This reuses the existing reasoning extractor so the callback only
+        receives tagged/structured reasoning, never ordinary assistant reply
+        text.
+        """
+        cb = self.tool_progress_callback
+        if cb is None:
+            return
+
+        reasoning_text = self._extract_reasoning(assistant_message)
+        if not reasoning_text:
+            return
+
+        try:
+            cb("reasoning.available", "_thinking", reasoning_text[:500], None)
+        except Exception:
+            pass
+
     def _fire_tool_gen_started(self, tool_name: str) -> None:
         """Notify display layer that the model is generating tool call arguments.
 
@@ -11675,11 +11695,8 @@ class AIAgent:
                             self.tool_progress_callback("_thinking", first_line)
                         except Exception:
                             pass
-                    elif _think_text:
-                        try:
-                            self.tool_progress_callback("reasoning.available", "_thinking", _think_text[:500], None)
-                        except Exception:
-                            pass
+                    else:
+                        self._fire_reasoning_available(assistant_message)
                 
                 # Check for incomplete <REASONING_SCRATCHPAD> (opened but never closed)
                 # This means the model ran out of output tokens mid-reasoning — retry up to 2 times

tests/agent/test_subagent_progress.py

@@ -322,6 +322,18 @@ class TestThinkingCallback:
         )
         assert len(calls) == 0
 
+    def test_thinking_callback_keeps_stripped_preview(self):
+        """Subagent thinking relay should still expose the stripped preview."""
+        calls = []
+        self._simulate_thinking_callback(
+            "<think>internal reasoning</think> visible summary",
+            lambda name, preview=None: calls.append((name, preview))
+        )
+        assert len(calls) == 1
+        assert calls[0][0] == "_thinking"
+        assert "internal reasoning" in calls[0][1]
+        assert "visible summary" in calls[0][1]
+
 
 # =========================================================================
 # Gateway batch flush tests
@@ -386,4 +398,3 @@ class TestBatchFlush:
 
 if __name__ == "__main__":
     pytest.main([__file__, "-v"])
-

tests/run_agent/test_run_agent.py

@@ -1453,6 +1453,30 @@ class TestBuildAssistantMessage:
         assert result["content"] == ""
 
 
+class TestReasoningAvailableProgress:
+    def test_emits_only_extracted_reasoning(self, agent):
+        calls = []
+        agent.tool_progress_callback = lambda *args, **kwargs: calls.append(args)
+
+        msg = _mock_assistant_msg(
+            content="<think>internal analysis</think>The actual answer."
+        )
+
+        agent._fire_reasoning_available(msg)
+
+        assert calls == [("reasoning.available", "_thinking", "internal analysis", None)]
+
+    def test_skips_plain_answer_text(self, agent):
+        calls = []
+        agent.tool_progress_callback = lambda *args, **kwargs: calls.append(args)
+
+        msg = _mock_assistant_msg(content="The actual answer only.")
+
+        agent._fire_reasoning_available(msg)
+
+        assert calls == []
+
+
 class TestFormatToolsForSystemMessage:
     def test_no_tools_returns_empty_array(self, agent):
         agent.tools = []