Trajectory record (messages, raw tool results, metadata; reasoning apart) and trajectory runner

Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>

harness/agents.py

@@ -85,6 +85,7 @@ class LlmAgent:
         self.empty_retries = 2
         self.loop_guard = loop_guard  # end the run after this many identical pushes in a row (None = off)
         self.end_reason = None
+        self.messages, self.tools, self.reasoning, self.turn_usage = [], [], [], []  # for the trajectory record
         self.chat_template_kwargs = chat_template_kwargs  # local server only, e.g. {"enable_thinking": False}
 
     def _chat(self, messages, tools):
@@ -118,6 +119,7 @@ class LlmAgent:
                  for t in proxy.schemas()]
         messages = [{"role": "system", "content": SYSTEM_PROMPT},
                     {"role": "user", "content": task.spec}]
+        self.messages, self.tools, self.reasoning, self.turn_usage = messages, tools, [], []
         if ":cloud" in self.model:
             check_budget()
         final = ""
@@ -132,6 +134,7 @@ class LlmAgent:
             try:
                 msg, usage = self._chat(messages, tools)
                 add_usage(self.model, usage, kind="run", ref=proxy.prefix)
+                self.turn_usage.append(usage)
             except RuntimeError as e:
                 final = f"Stopped: {e}"
                 self.end_reason = "model_error"
@@ -139,6 +142,9 @@ class LlmAgent:
             proxy.note("assistant", {"content": msg.get("content"),
                                      "tool_calls": msg.get("tool_calls"), "usage": usage})
             messages.append({k: v for k, v in msg.items() if k in ("role", "content", "tool_calls")})
+            think = msg.get("reasoning") or msg.get("reasoning_content") or msg.get("thinking")
+            if think:  # kept apart from the messages: not training data
+                self.reasoning.append({"message_index": len(messages) - 1, "reasoning": think})
             calls = msg.get("tool_calls") or []
             if not calls and not (msg.get("content") or "").strip() and empty < self.empty_retries:
                 empty += 1  # empty turn (output limit reached in reasoning): ask the same question again