|
| 1 | +from calendar import c |
| 2 | +from unittest.mock import AsyncMock, MagicMock, patch |
| 3 | + |
| 4 | +import pytest |
| 5 | +from pydantic import Field |
| 6 | + |
| 7 | +from app.agent.base import ROLE_TYPE, AgentState, BaseAgent, Message |
| 8 | +from app.llm import LLM |
| 9 | +from app.sandbox.client import SANDBOX_CLIENT |
| 10 | +from app.schema import Memory, Role |
| 11 | + |
| 12 | + |
| 13 | +class MockAgent(BaseAgent): |
| 14 | + |
| 15 | + name: str |
| 16 | + description: str | None = None |
| 17 | + |
| 18 | + system_prompt: str | None = None |
| 19 | + next_step_prompt: str | None = None |
| 20 | + |
| 21 | + llm: LLM = Field(default_factory=LLM) |
| 22 | + memory: Memory = Field(default_factory=Memory) |
| 23 | + state: AgentState = AgentState.IDLE |
| 24 | + |
| 25 | + max_steps: int = 10 |
| 26 | + current_step: int = 0 |
| 27 | + |
| 28 | + async def step(self) -> str: |
| 29 | + return "Mock step result" |
| 30 | + |
| 31 | + |
| 32 | +class MockFeaturedAgent(BaseAgent): |
| 33 | + |
| 34 | + name: str |
| 35 | + description: str | None = None |
| 36 | + |
| 37 | + system_prompt: str | None = None |
| 38 | + next_step_prompt: str | None = None |
| 39 | + |
| 40 | + llm: LLM = Field(default_factory=LLM) |
| 41 | + memory: Memory = Field(default_factory=Memory) |
| 42 | + state: AgentState = AgentState.IDLE |
| 43 | + |
| 44 | + max_steps: int = 10 |
| 45 | + current_step: int = 0 |
| 46 | + |
| 47 | + __cnt = 0 |
| 48 | + |
| 49 | + async def step(self) -> str: |
| 50 | + self.__cnt += 1 |
| 51 | + if self.__cnt > 4: |
| 52 | + self.state = AgentState.FINISHED |
| 53 | + return "Mock step result {}".format((self.__cnt - 1) % 2) |
| 54 | + |
| 55 | + |
| 56 | +@pytest.fixture |
| 57 | +def mock_agent(): |
| 58 | + return MockAgent( |
| 59 | + name="TestAgent", |
| 60 | + system_prompt="System prompt", |
| 61 | + next_step_prompt="Next step prompt", |
| 62 | + ) |
| 63 | + |
| 64 | + |
| 65 | +@pytest.fixture |
| 66 | +def mock_featured_agent(): |
| 67 | + return MockFeaturedAgent( |
| 68 | + name="TestAgent", |
| 69 | + system_prompt="System prompt", |
| 70 | + next_step_prompt="Next step prompt", |
| 71 | + ) |
| 72 | + |
| 73 | + |
| 74 | +@pytest.mark.asyncio |
| 75 | +async def test_run_happy_path(mock_agent): |
| 76 | + mock_agent.step = AsyncMock(return_value="Mock step result") |
| 77 | + SANDBOX_CLIENT.cleanup = AsyncMock() |
| 78 | + result = await mock_agent.run() |
| 79 | + assert ( |
| 80 | + result |
| 81 | + == """Step 1: Mock step result |
| 82 | +Step 2: Mock step result |
| 83 | +Step 3: Mock step result |
| 84 | +Step 4: Mock step result |
| 85 | +Step 5: Mock step result |
| 86 | +Step 6: Mock step result |
| 87 | +Step 7: Mock step result |
| 88 | +Step 8: Mock step result |
| 89 | +Step 9: Mock step result |
| 90 | +Step 10: Mock step result |
| 91 | +Terminated: Reached max steps (10)""" |
| 92 | + ) |
| 93 | + assert mock_agent.current_step == 0 |
| 94 | + assert mock_agent.state == AgentState.IDLE |
| 95 | + SANDBOX_CLIENT.cleanup.assert_called_once() |
| 96 | + |
| 97 | + |
| 98 | +@pytest.mark.asyncio |
| 99 | +async def test_run_with_request(mock_agent): |
| 100 | + mock_agent.step = AsyncMock(return_value="Mock step result") |
| 101 | + SANDBOX_CLIENT.cleanup = AsyncMock() |
| 102 | + result = await mock_agent.run(request="User request") |
| 103 | + assert ( |
| 104 | + result |
| 105 | + == """Step 1: Mock step result |
| 106 | +Step 2: Mock step result |
| 107 | +Step 3: Mock step result |
| 108 | +Step 4: Mock step result |
| 109 | +Step 5: Mock step result |
| 110 | +Step 6: Mock step result |
| 111 | +Step 7: Mock step result |
| 112 | +Step 8: Mock step result |
| 113 | +Step 9: Mock step result |
| 114 | +Step 10: Mock step result |
| 115 | +Terminated: Reached max steps (10)""" |
| 116 | + ) |
| 117 | + assert mock_agent.current_step == 0 # reset at max_steps |
| 118 | + assert mock_agent.state == AgentState.IDLE |
| 119 | + assert mock_agent.messages[0].role == Role.USER |
| 120 | + assert mock_agent.messages[0].content == "User request" |
| 121 | + SANDBOX_CLIENT.cleanup.assert_called_once() |
| 122 | + |
| 123 | + |
| 124 | +@pytest.mark.asyncio |
| 125 | +async def test_run_max_steps(mock_agent): |
| 126 | + mock_agent.step = AsyncMock(return_value="Mock step result") |
| 127 | + SANDBOX_CLIENT.cleanup = AsyncMock() |
| 128 | + mock_agent.max_steps = 2 |
| 129 | + result = await mock_agent.run() |
| 130 | + assert ( |
| 131 | + result |
| 132 | + == "Step 1: Mock step result\nStep 2: Mock step result\nTerminated: Reached max steps (2)" |
| 133 | + ) |
| 134 | + assert mock_agent.current_step == 0 |
| 135 | + assert mock_agent.state == AgentState.IDLE |
| 136 | + SANDBOX_CLIENT.cleanup.assert_called_once() |
| 137 | + |
| 138 | + |
| 139 | +@pytest.mark.asyncio |
| 140 | +async def test_run_agent_already_running(mock_agent): |
| 141 | + mock_agent.state = AgentState.RUNNING |
| 142 | + with pytest.raises(RuntimeError) as e: |
| 143 | + await mock_agent.run() |
| 144 | + assert str(e.value) == "Cannot run agent from state: AgentState.RUNNING" |
| 145 | + |
| 146 | + |
| 147 | +@pytest.mark.asyncio |
| 148 | +async def test_run_agent_stuck(mock_agent): |
| 149 | + mock_agent.step = AsyncMock(return_value="Stuck result") |
| 150 | + SANDBOX_CLIENT.cleanup = AsyncMock() |
| 151 | + mock_agent.duplicate_threshold = 1 |
| 152 | + mock_agent.update_memory(Role.ASSISTANT, "Stuck result") |
| 153 | + mock_agent.update_memory(Role.ASSISTANT, "Stuck result") |
| 154 | + result = await mock_agent.run() |
| 155 | + assert "Observed duplicate responses" in mock_agent.next_step_prompt |
| 156 | + assert "Step 1: Stuck result" in result |
| 157 | + assert ( |
| 158 | + mock_agent.current_step == 0 |
| 159 | + ) # it will run 10 times and just reset the counter |
| 160 | + assert mock_agent.state == AgentState.IDLE |
| 161 | + SANDBOX_CLIENT.cleanup.assert_called_once() |
| 162 | + |
| 163 | + |
| 164 | +@pytest.mark.asyncio |
| 165 | +async def test_update_memory_happy_path(mock_agent): |
| 166 | + mock_agent.update_memory(Role.USER, "User message") |
| 167 | + assert mock_agent.messages[0].role == Role.USER |
| 168 | + assert mock_agent.messages[0].content == "User message" |
| 169 | + |
| 170 | + |
| 171 | +@pytest.mark.asyncio |
| 172 | +async def test_update_memory_unsupported_role(mock_agent): |
| 173 | + with pytest.raises(ValueError) as e: |
| 174 | + mock_agent.update_memory("unsupported", "Content") |
| 175 | + assert str(e.value) == "Unsupported message role: unsupported" |
| 176 | + |
| 177 | + |
| 178 | +@pytest.mark.asyncio |
| 179 | +async def test_is_stuck_not_enough_messages(mock_agent): |
| 180 | + assert not mock_agent.is_stuck() |
| 181 | + |
| 182 | + |
| 183 | +@pytest.mark.asyncio |
| 184 | +async def test_is_stuck_no_duplicates(mock_agent): |
| 185 | + mock_agent.update_memory(Role.ASSISTANT, "Unique result 1") |
| 186 | + mock_agent.update_memory(Role.ASSISTANT, "Unique result 2") |
| 187 | + assert not mock_agent.is_stuck() |
| 188 | + |
| 189 | + |
| 190 | +@pytest.mark.asyncio |
| 191 | +async def test_is_stuck_with_duplicates(mock_agent): |
| 192 | + mock_agent.update_memory(Role.ASSISTANT, "Duplicate result") |
| 193 | + mock_agent.update_memory(Role.ASSISTANT, "Duplicate result") |
| 194 | + mock_agent.update_memory(Role.ASSISTANT, "Duplicate result") |
| 195 | + assert mock_agent.is_stuck() |
| 196 | + |
| 197 | + |
| 198 | +@pytest.mark.asyncio |
| 199 | +async def test_is_stuck_different_roles(mock_agent): |
| 200 | + mock_agent.update_memory(Role.ASSISTANT, "First Message") |
| 201 | + mock_agent.update_memory(Role.ASSISTANT, "Duplicate result") |
| 202 | + mock_agent.update_memory(Role.USER, "Duplicate result") |
| 203 | + mock_agent.update_memory(Role.ASSISTANT, "Duplicate result") |
| 204 | + assert not mock_agent.is_stuck() |
| 205 | + |
| 206 | + |
| 207 | +@pytest.mark.asyncio |
| 208 | +async def test_auto_fix_weird_llm(mock_agent): |
| 209 | + class not_a_llm_object: |
| 210 | + a = "hello" |
| 211 | + b = "handsome" |
| 212 | + c = "manna" and "poem" |
| 213 | + |
| 214 | + mock_agent.llm = not_a_llm_object_instance = not_a_llm_object() |
| 215 | + |
| 216 | + mock_agent.initialize_agent() |
| 217 | + assert ( |
| 218 | + isinstance(mock_agent.llm, LLM) |
| 219 | + and not_a_llm_object_instance is not mock_agent.llm |
| 220 | + ) |
| 221 | + |
| 222 | + |
| 223 | +@pytest.mark.asyncio |
| 224 | +async def test_auto_fix_weird_memory(mock_agent): |
| 225 | + class not_a_memory_object: |
| 226 | + a = "not" |
| 227 | + b = "memory" |
| 228 | + c = "object" |
| 229 | + d = "absolutely" |
| 230 | + |
| 231 | + mock_agent.memory = not_a_memory_object_instance = not_a_memory_object() |
| 232 | + |
| 233 | + mock_agent.initialize_agent() |
| 234 | + |
| 235 | + assert isinstance(mock_agent.memory, Memory) |
| 236 | + assert not_a_memory_object_instance is not mock_agent.memory |
| 237 | + |
| 238 | + |
| 239 | +@pytest.mark.asyncio |
| 240 | +async def test_state_context_invalid_new_state(mock_agent): |
| 241 | + class not_a_state_object: |
| 242 | + a = "not" |
| 243 | + b = "state" |
| 244 | + c = "object" |
| 245 | + |
| 246 | + new_state = not_a_state_object() |
| 247 | + |
| 248 | + with pytest.raises(ValueError) as e: |
| 249 | + async with mock_agent.state_context(new_state): |
| 250 | + pass |
| 251 | + assert "Invalid state:" in str(e.value) |
| 252 | + |
| 253 | + |
| 254 | +@pytest.mark.asyncio |
| 255 | +async def test_state_context_raise_exception(mock_agent): |
| 256 | + prev_state = mock_agent.state |
| 257 | + with pytest.raises(RuntimeError) as e: |
| 258 | + async with mock_agent.state_context(AgentState.RUNNING): |
| 259 | + raise RuntimeError("Test exception 123123") |
| 260 | + # assert mock_agent.state == AgentState.ERROR |
| 261 | + assert str(e.value) == "Test exception 123123" |
| 262 | + assert mock_agent.state == prev_state |
| 263 | + |
| 264 | + |
| 265 | +@pytest.mark.asyncio |
| 266 | +async def test_run_with_request_featured_agent(mock_featured_agent): |
| 267 | + SANDBOX_CLIENT.cleanup = AsyncMock() |
| 268 | + result = await mock_featured_agent.run(request="User request") |
| 269 | + assert ( |
| 270 | + result |
| 271 | + == """Step 1: Mock step result 0 |
| 272 | +Step 2: Mock step result 1 |
| 273 | +Step 3: Mock step result 0 |
| 274 | +Step 4: Mock step result 1 |
| 275 | +Step 5: Mock step result 0""" |
| 276 | + ) |
| 277 | + assert mock_featured_agent.current_step == 5 |
| 278 | + assert mock_featured_agent.state == AgentState.IDLE |
| 279 | + assert mock_featured_agent.messages[0].role == Role.USER |
| 280 | + assert mock_featured_agent.messages[0].content == "User request" |
| 281 | + SANDBOX_CLIENT.cleanup.assert_called_once() |
| 282 | + |
| 283 | + |
| 284 | +@pytest.mark.asyncio |
| 285 | +async def test_is_stuck_invalid_last_message(mock_agent): |
| 286 | + mock_agent.update_memory(Role.ASSISTANT, "First Message") |
| 287 | + mock_agent.update_memory(Role.ASSISTANT, "Duplicate result") |
| 288 | + mock_agent.update_memory(Role.USER, "Duplicate result") |
| 289 | + mock_agent.update_memory(Role.ASSISTANT, None) |
| 290 | + assert not mock_agent.is_stuck() |
| 291 | + |
| 292 | + |
| 293 | +@pytest.mark.asyncio |
| 294 | +async def test_agent_message_assignment(mock_agent): |
| 295 | + class test_messages_object: |
| 296 | + manna = "poem" |
| 297 | + |
| 298 | + a = test_messages_object() |
| 299 | + |
| 300 | + assert mock_agent.messages is not a |
| 301 | + |
| 302 | + mock_agent.messages = a |
| 303 | + |
| 304 | + assert mock_agent.messages is a |
0 commit comments