Add local sandbox session manager support and config

- Add agent_sdk_use_local_sandbox setting and LocalSandboxSessionManager
  branch in AgentSessionMixin._ensure_agent_session
- Update predicator_v3 config to use docker sandbox by default

Author: yichao-liang

predicators/approaches/agent_session_mixin.py

@@ -109,6 +109,17 @@ def _ensure_agent_session(self) -> None:
                 image=CFG.agent_sdk_docker_image,
                 extra_reference_files=self._get_sandbox_reference_files(),
             )
+        elif CFG.agent_sdk_use_local_sandbox:
+            from predicators.agent_sdk.local_sandbox import \
+                LocalSandboxSessionManager
+            self._agent_session = LocalSandboxSessionManager(
+                system_prompt=self._get_agent_system_prompt(),
+                log_dir=self._get_log_dir(),
+                model_name=self._get_agent_model_name(),
+                tool_context=self._tool_context,
+                tool_names=tool_names,
+                extra_reference_files=self._get_sandbox_reference_files(),
+            )
         else:
             from claude_agent_sdk import create_sdk_mcp_server
 

predicators/settings.py

@@ -968,9 +968,10 @@ class GlobalSettings:
     agent_sdk_max_trajectories_in_context = 3
     agent_sdk_log_agent_responses = True
 
-    # Docker sandbox settings for agent SDK
+    # Sandbox settings for agent SDK
     agent_sdk_use_docker_sandbox = False  # run agent inside Docker container
     agent_sdk_docker_image = "predicators-sandbox"  # Docker image name
+    agent_sdk_use_local_sandbox = False  # sandbox dir with built-in tools, no Docker
 
     # Agent explorer settings
     agent_explorer_max_turns = 5  # max agent turns per exploration query

scripts/configs/predicatorv3/predicator_v3.yaml

@@ -34,6 +34,7 @@ APPROACHES:
       terminate_on_goal_reached_and_option_terminated: True
       agent_planner_isolate_test_session: True
       agent_sdk_use_docker_sandbox: True
+      # agent_sdk_use_local_sandbox: True
       agent_sdk_max_agent_turns_per_iteration: 40
   # mb_agent:
   #   NAME: "agent_abstraction_learning"