record_search: three-view GIF recorder (map | first person | chat with model reasoning); capture reasoning_content in TurnResult; search_mission.gif artifact

testbed/llm_agent.py

@@ -314,6 +314,7 @@ class TurnResult:
     calls: list[ToolCallResult] = field(default_factory=list)
     interacted: bool = False
     message: str | None = None
+    thinking: str = ""
 
 
 def detect_multimodal(base_url: str, model: str) -> bool:
@@ -502,6 +503,15 @@ class LLMController:
         choice = resp.choices[0]
         text = choice.message.content or ""
         result = TurnResult(text=text)
+        # Some providers return reasoning tokens separately (e.g. deepseek-r1
+        # style models); surface them when present.
+        msg = choice.message
+        thinking = (
+            getattr(msg, "reasoning_content", None)
+            or getattr(msg, "reasoning", None)
+            or ""
+        )
+        result.thinking = thinking if isinstance(thinking, str) else ""
         calls = choice.message.tool_calls
         if not calls:
             self.messages.append({"role": "assistant", "content": text})
@@ -741,6 +751,8 @@ async def run_llm_agent_loop(
             turn = await controller.invoke()
         if turn.text:
             log("agent", turn.text[:400])
+        if turn.thinking:
+            log("think", turn.thinking[:400])
         summary["tool_calls"] += len(turn.calls)
         for c in turn.calls:
             if (