fix(simulator): preserve traditional mode in dry-run preparation

prepare_dry_run now respects agent_mode_max_iterations=0 (traditional
mode) instead of unconditionally forcing agent mode. Only overrides
to 1 when the user configured agent mode (non-zero).

autogpt_platform/backend/backend/executor/simulator.py

@@ -287,10 +287,16 @@ def prepare_dry_run(block: Any, input_data: dict[str, Any]) -> dict[str, Any] |
     block should be LLM-simulated instead.
     """
     if isinstance(block, OrchestratorBlock):
+        # Preserve the user's configured mode: 0 means traditional (single
+        # LLM call, no agent loop). Only override to 1 when the user chose
+        # agent mode (non-zero) so the dry run still exercises the loop once
+        # without running away.
+        original = input_data.get("agent_mode_max_iterations", 0)
+        max_iters = 1 if original != 0 else 0
         return {
             **input_data,
             "model": DRY_RUN_MODEL,
-            "agent_mode_max_iterations": 1,
+            "agent_mode_max_iterations": max_iters,
         }
     return None