feat: Initial library extraction from PIN LLM benchmark

DockerSandbox + LangChain file/shell tools extracted into a standalone package.

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

docker_agent_sandbox/tools/bash.py

@@ -0,0 +1,38 @@
+"""bash.py – tool for executing shell commands inside the sandbox."""
+
+from __future__ import annotations
+
+from typing import TYPE_CHECKING
+
+from langchain_core.tools import BaseTool, tool
+from loguru import logger
+
+from docker_agent_sandbox.tools._utils import truncate_output
+
+if TYPE_CHECKING:
+    from docker_agent_sandbox.sandbox import DockerSandbox
+
+
+def make_bash_tool(sandbox: "DockerSandbox") -> BaseTool:
+    """
+    Return a bash tool that executes commands inside the Docker sandbox container.
+
+    The model's working directory is the sandbox root; all paths it uses are
+    identical on the host (via the bind mount) and inside the container.
+    """
+
+    @tool
+    def bash(command: str, timeout: int = 120) -> str:
+        """
+        Execute a shell command in the sandbox container.
+
+        Returns EXIT:<code> followed by combined stdout+stderr.
+        Large outputs are truncated to stay within token limits.
+        Use for: running the target binary, processing PIN output,
+        compiling plugins, or any other shell operation during analysis.
+        """
+        logger.debug("Running inside sandbox: {}", command)
+        exit_code, output = sandbox.exec(command, timeout=timeout)
+        return f"EXIT:{exit_code}\n{truncate_output(output)}"
+
+    return bash