feat: add end to end integration test

pyproject.toml

@@ -143,6 +143,15 @@ test = [
     "hatch test --cover --cov-report term --cov-report html --cov-report xml {args}"
 ]
 
+test-integ = [
+    "hatch test tests_integ {args}"
+]
+
+[tool.pytest.ini_options]
+testpaths = [
+    "tests"
+]
+
 [tool.mypy]
 python_version = "3.10"
 warn_return_any = true
@@ -161,7 +170,7 @@ ignore_missing_imports = false
 
 [tool.ruff]
 line-length = 120
-include = ["src/**/*.py", "tests/**/*.py", "tools/**/*.py"]
+include = ["src/**/*.py", "tests/**/*.py", "tools/**/*.py","tests_integ/**/*.py"]
 
 [tool.ruff.lint]
 select = [

tests_integ/conftest.py

@@ -0,0 +1,13 @@
+"""Common fixtures for integration tests."""
+
+from pathlib import Path
+
+import pytest
+
+
+@pytest.fixture
+def tmp_file_structure(tmp_path: Path):
+    """Creates a temporary directory structure for tool creation and returns paths."""
+    tools_dir = tmp_path / ".strands" / "tools"
+    tools_dir.mkdir(parents=True, exist_ok=True)
+    return {"tools_dir": tools_dir, "root_dir": tmp_path}

tests_integ/test_build_agent_from_specifications.py

@@ -0,0 +1,38 @@
+import sys
+from unittest import mock
+
+from strands import Agent
+
+from strands_agents_builder import strands
+
+
+def test_build_agent_from_specification_file():
+    """Test building agent from specification using semantic creation."""
+
+    # Agent specification content
+    spec_content = """Agent Specification:
+- Role: Math Tutor Agent
+- Purpose: Help students with basic arithmetic and algebra
+- Tools needed: calculator, file_read, current_time
+- Create specialized tools for math tutoring
+"""
+
+    # Simulate: cat agent-spec.txt | strands "Build a specialized agent based on these specifications"
+    query = f"Build a specialized agent based on these specifications:\n\n{spec_content}"
+
+    with mock.patch.object(sys, "argv", ["strands", query]):
+        strands.main()
+
+    # Validate agent
+    agent = Agent(
+        load_tools_from_directory=True,
+        callback_handler=None,
+    )
+
+    # Validate agent was created successfully
+    assert agent is not None
+
+    # Validate agent has exactly 3 specified tools
+    required_tools = ["calculator", "file_read", "current_time"]
+    for tool_name in required_tools:
+        assert hasattr(agent.tool, tool_name), f"Agent missing required tool: {tool_name}"

tests_integ/test_semantic_create_tool.py

@@ -0,0 +1,42 @@
+import sys
+from unittest import mock
+
+from strands.agent import Agent
+
+from strands_agents_builder import strands
+
+
+@mock.patch("strands_tools.utils.user_input.get_user_input", return_value="y")
+@mock.patch.dict("os.environ", {"STRANDS_TOOL_CONSOLE_MODE": "enabled"})
+def test_interactive_model_create_tool_then_validate(mock_get_user_input, capsys, tmp_file_structure):
+    """
+    Test creating a calculator tool via CLI and validating its functionality.
+    """
+    with mock.patch.dict("os.environ", {"STRANDS_TOOLS_DIR": str(tmp_file_structure["tools_dir"])}):
+        test_query = "create a tool that can only calculate sum of two number called calculator"
+
+        with mock.patch.object(sys, "argv", ["strands", test_query]):
+            strands.main()
+
+        agent = Agent(
+            load_tools_from_directory=True,
+        )
+
+        # assert agent has a tool called calculator
+        assert hasattr(agent.tool, "calculator"), "Agent should have a 'calculator' tool after creation."
+
+        test_cases = [("what is 1 plus 2?", "3"), ("calculate 5 plus 7", "12")]
+
+        for question, expected in test_cases:
+            response = agent(question)
+            response_str = str(response).lower()
+            assert expected in response_str, f"Expected '{expected}' in response for '{question}', got '{response_str}'"
+
+            # Verify the calculator tool was actually used by checking for toolResults
+            calculator_used = any(
+                content.get("toolUse", {}).get("name") == "calculator"
+                for message in agent.messages
+                for content in message.get("content", [])
+                if content.get("toolUse")
+            )
+            assert calculator_used, f"Calculator tool should have been used for question: '{question}'"

tests_integ/test_shell_then_exit.py

@@ -0,0 +1,18 @@
+from unittest import mock
+
+from strands_agents_builder import strands
+
+
+@mock.patch.object(strands, "get_user_input")
+@mock.patch.dict("os.environ", {"STRANDS_TOOL_CONSOLE_MODE": "enabled"})
+def test_cli_shell_and_interactive_mode(mock_user_input, capsys):
+    """Test interactive mode with a shell command."""
+
+    user_inputs = ["!echo hello", "exit"]
+    mock_user_input.side_effect = user_inputs
+
+    strands.main()
+    out = capsys.readouterr().out
+
+    assert "hello" in out, "Expected 'hello' (from echo) in output"
+    assert "thank you for using strands" in out.lower(), "Expected exit message"