Loop guard (3 identical pushes), end_reason and activation error records per run; thinking off for the stage 1 baseline

Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_014aUaQeLnwbb1zTpN7kHeat

harness/runner.py

@@ -61,7 +61,9 @@ def _norm(src):
 def trajectory_stats(path):
     """Rebuild agent statistics from a trajectory (same rules as ToolProxy)."""
     from .proxy import WRITE_TOOLS
-    calls = activations = streak = max_streak = 0
+    from .proxy import activation_messages
+    calls = activations = streak = max_streak = fails = 0
+    errors = []
     final, t_first, t_last = "", None, None
     for line in open(path):
         e = json.loads(line)
@@ -79,7 +81,13 @@ def trajectory_stats(path):
             failed = e["is_error"] or '"success":false' in e["result"].replace(" ", "")
             streak = streak + 1 if failed else 0
             max_streak = max(max_streak, streak)
+            if failed:
+                fails += 1
+                for m in activation_messages(e["result"], e["is_error"]):
+                    if m not in errors and len(errors) < 30:
+                        errors.append(m)
     return {"tool_calls": calls, "activations": activations, "max_fail_streak": max_streak,
+            "activation_failures": fails, "activation_error_messages": errors,
             "final_report": final, "agent_seconds": round((t_last or 0) - (t_first or 0), 1)}
 
 
@@ -207,6 +215,9 @@ class Runner:
                 rep["agent_seconds"] = round(time.time() - t1, 1)
                 rep["tool_calls"], rep["activations"] = proxy.calls, proxy.activations
                 rep["max_fail_streak"] = proxy.max_fail_streak
+                rep["activation_failures"] = proxy.activation_failures
+                rep["activation_error_messages"] = proxy.activation_errors
+                rep["end_reason"] = getattr(agent, "end_reason", None)
                 if proxy.fallbacks:
                     rep["adt_fallbacks"] = proxy.fallbacks