| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | |
| 4 | import pytest |
| 5 | from openai.types.responses import FunctionTool |
| 6 | from openai_harmony import DeveloperContent, Message, Role |
| 7 | |
| 8 | from tests.entrypoints.openai.utils import verify_harmony_messages |
| 9 | from vllm.entrypoints.openai.chat_completion.protocol import ChatCompletionToolsParam |
| 10 | from vllm.entrypoints.openai.parser.harmony_utils import ( |
| 11 | auto_drop_analysis_messages, |
| 12 | create_tool_definition, |
| 13 | extract_function_from_recipient, |
| 14 | get_system_message, |
| 15 | has_custom_tools, |
| 16 | is_function_recipient, |
| 17 | parse_chat_input_to_harmony_message, |
| 18 | ) |
| 19 | from vllm.entrypoints.openai.responses.harmony import ( |
| 20 | response_input_to_harmony, |
| 21 | response_previous_input_to_harmony, |
| 22 | ) |
| 23 | from vllm.exceptions import VLLMValidationError |
| 24 | |
| 25 | _TOOL_PARAMETERS = { |
| 26 | "type": "object", |
| 27 | "properties": {"status": {"type": "string"}}, |
| 28 | "required": ["status"], |
| 29 | "additionalProperties": False, |
| 30 | } |
| 31 | |
| 32 | |
| 33 | class TestCreateToolDefinition: |
| 34 | def test_chat_completion_omitted_description_defaults_to_empty_string(self): |
| 35 | tool = ChatCompletionToolsParam( |
| 36 | function={ |
| 37 | "name": "report_status", |
| 38 | "parameters": _TOOL_PARAMETERS, |
| 39 | } |
| 40 | ) |
| 41 | |
| 42 | tool_definition = create_tool_definition(tool) |
| 43 | |
| 44 | assert tool_definition.name == "report_status" |
| 45 | assert tool_definition.description == "" |
| 46 | assert tool_definition.parameters == _TOOL_PARAMETERS |
| 47 | |
| 48 | def test_chat_completion_none_description_defaults_to_empty_string(self): |
| 49 | tool = ChatCompletionToolsParam( |
| 50 | function={ |
| 51 | "name": "report_status", |
| 52 | "description": None, |
| 53 | "parameters": _TOOL_PARAMETERS, |
| 54 | } |
| 55 | ) |
| 56 | |
| 57 | tool_definition = create_tool_definition(tool) |
| 58 | |
| 59 | assert tool_definition.name == "report_status" |
| 60 | assert tool_definition.description == "" |
| 61 | assert tool_definition.parameters == _TOOL_PARAMETERS |
| 62 | |
| 63 | def test_response_tool_none_description_defaults_to_empty_string(self): |
| 64 | tool = FunctionTool( |
| 65 | name="report_status", |
| 66 | description=None, |
| 67 | parameters=_TOOL_PARAMETERS, |
| 68 | type="function", |
| 69 | ) |
| 70 | |
| 71 | tool_definition = create_tool_definition(tool) |
| 72 | |
| 73 | assert tool_definition.name == "report_status" |
| 74 | assert tool_definition.description == "" |
| 75 | assert tool_definition.parameters == _TOOL_PARAMETERS |
| 76 | |
| 77 | |
| 78 | class TestIsFunctionRecipient: |
| 79 | @pytest.mark.parametrize( |
| 80 | "recipient", |
| 81 | [ |
| 82 | "functions.get_weather", |
| 83 | "functions.search_web", |
| 84 | "functions.math.sum", |
| 85 | ], |
| 86 | ) |
| 87 | def test_functions_prefix_accepted(self, recipient): |
| 88 | assert is_function_recipient(recipient) is True |
| 89 | |
| 90 | @pytest.mark.parametrize( |
| 91 | "recipient", |
| 92 | [ |
| 93 | "get_weather", |
| 94 | "search_web", |
| 95 | "calculator", |
| 96 | "my-tool", |
| 97 | ], |
| 98 | ) |
| 99 | def test_bare_function_name_accepted(self, recipient): |
| 100 | assert is_function_recipient(recipient) is True |
| 101 | |
| 102 | @pytest.mark.parametrize( |
| 103 | "recipient", |
| 104 | [ |
| 105 | "assistant", |
| 106 | ], |
| 107 | ) |
| 108 | def test_assistant_rejected(self, recipient): |
| 109 | assert is_function_recipient(recipient) is False |
| 110 | |
| 111 | @pytest.mark.parametrize( |
| 112 | "recipient", |
| 113 | [ |
| 114 | "math.sum", |
| 115 | "code.run", |
| 116 | "namespace.tool_name", |
| 117 | "my.deeply.nested.tool", |
| 118 | ], |
| 119 | ) |
| 120 | def test_dotted_function_names_accepted(self, recipient): |
| 121 | assert is_function_recipient(recipient) is True |
| 122 | |
| 123 | @pytest.mark.parametrize( |
| 124 | "recipient", |
| 125 | [ |
| 126 | "python", |
| 127 | "browser", |
| 128 | "container", |
| 129 | ], |
| 130 | ) |
| 131 | def test_builtin_tool_names_rejected(self, recipient): |
| 132 | assert is_function_recipient(recipient) is False |
| 133 | |
| 134 | @pytest.mark.parametrize( |
| 135 | "recipient", |
| 136 | [ |
| 137 | "python.run", |
| 138 | "python.execute", |
| 139 | "browser.search", |
| 140 | "browser.open", |
| 141 | "container.exec", |
| 142 | ], |
| 143 | ) |
| 144 | def test_builtin_dotted_variants_rejected(self, recipient): |
| 145 | assert is_function_recipient(recipient) is False |
| 146 | |
| 147 | @pytest.mark.parametrize( |
| 148 | "recipient", |
| 149 | [ |
| 150 | "", |
| 151 | "functions.", |
| 152 | ], |
| 153 | ) |
| 154 | def test_empty_recipients_rejected(self, recipient): |
| 155 | assert is_function_recipient(recipient) is False |
| 156 | |
| 157 | @pytest.mark.parametrize( |
| 158 | "recipient", |
| 159 | [ |
| 160 | "<|start|>", |
| 161 | "<|end|>", |
| 162 | "<|channel|>", |
| 163 | ], |
| 164 | ) |
| 165 | def test_harmony_tokens_rejected(self, recipient): |
| 166 | assert is_function_recipient(recipient) is False |
| 167 | |
| 168 | |
| 169 | class TestIsFunctionRecipientWithAllowedNames: |
| 170 | """Tests for is_function_recipient with allowed_function_tool_names.""" |
| 171 | |
| 172 | def test_prefixed_always_accepted(self): |
| 173 | """Functions. prefix is always accepted regardless of allowed names.""" |
| 174 | fn_names = frozenset({"other_tool"}) |
| 175 | assert is_function_recipient("functions.get_weather", fn_names) is True |
| 176 | |
| 177 | def test_bare_name_accepted_when_in_allowed_names(self): |
| 178 | fn_names = frozenset({"get_weather", "search_web"}) |
| 179 | assert is_function_recipient("get_weather", fn_names) is True |
| 180 | assert is_function_recipient("search_web", fn_names) is True |
| 181 | |
| 182 | def test_bare_name_rejected_when_not_in_allowed_names(self): |
| 183 | fn_names = frozenset({"get_weather"}) |
| 184 | assert is_function_recipient("unknown_tool", fn_names) is False |
| 185 | |
| 186 | def test_dotted_name_accepted_when_in_allowed_names(self): |
| 187 | fn_names = frozenset({"math.sum", "namespace.tool_name"}) |
| 188 | assert is_function_recipient("math.sum", fn_names) is True |
| 189 | assert is_function_recipient("namespace.tool_name", fn_names) is True |
| 190 | |
| 191 | def test_dotted_name_rejected_when_not_in_allowed_names(self): |
| 192 | fn_names = frozenset({"get_weather"}) |
| 193 | assert is_function_recipient("custom_server.search", fn_names) is False |
| 194 | |
| 195 | def test_empty_allowed_names_rejects_bare_names(self): |
| 196 | """Empty frozenset means no function tools — bare names are not functions.""" |
| 197 | fn_names: frozenset[str] = frozenset() |
| 198 | assert is_function_recipient("get_weather", fn_names) is False |
| 199 | assert is_function_recipient("math.sum", fn_names) is False |
| 200 | |
| 201 | def test_builtin_tools_always_rejected(self): |
| 202 | fn_names = frozenset({"python", "browser", "container"}) |
| 203 | assert is_function_recipient("python", fn_names) is False |
| 204 | assert is_function_recipient("browser", fn_names) is False |
| 205 | assert is_function_recipient("container", fn_names) is False |
| 206 | |
| 207 | def test_builtin_dotted_always_rejected(self): |
| 208 | fn_names = frozenset({"python.run", "browser.search"}) |
| 209 | assert is_function_recipient("python.run", fn_names) is False |
| 210 | assert is_function_recipient("browser.search", fn_names) is False |
| 211 | |
| 212 | def test_none_allowed_names_uses_heuristic(self): |
| 213 | """When allowed names is None (Chat Completions), use heuristic.""" |
| 214 | assert is_function_recipient("get_weather", None) is True |
| 215 | assert is_function_recipient("math.sum", None) is True |
| 216 | assert is_function_recipient("python", None) is False |
| 217 | |
| 218 | |
| 219 | class TestExtractFunctionFromRecipient: |
| 220 | @pytest.mark.parametrize( |
| 221 | "recipient,expected", |
| 222 | [ |
| 223 | ("functions.get_weather", "get_weather"), |
| 224 | ("functions.search_web", "search_web"), |
| 225 | ("functions.", ""), |
| 226 | ], |
| 227 | ) |
| 228 | def test_strips_functions_prefix(self, recipient, expected): |
| 229 | assert extract_function_from_recipient(recipient) == expected |
| 230 | |
| 231 | @pytest.mark.parametrize( |
| 232 | "recipient", |
| 233 | [ |
| 234 | "get_weather", |
| 235 | "calculator", |
| 236 | "my-tool", |
| 237 | ], |
| 238 | ) |
| 239 | def test_bare_name_returned_as_is(self, recipient): |
| 240 | assert extract_function_from_recipient(recipient) == recipient |
| 241 | |
| 242 | @pytest.mark.parametrize( |
| 243 | "recipient,expected", |
| 244 | [ |
| 245 | ("functions.math.sum", "math.sum"), |
| 246 | ("math.sum", "math.sum"), |
| 247 | ("namespace.tool_name", "namespace.tool_name"), |
| 248 | ], |
| 249 | ) |
| 250 | def test_dotted_function_name_extraction(self, recipient, expected): |
| 251 | assert extract_function_from_recipient(recipient) == expected |
| 252 | |
| 253 | |
| 254 | class TestCommonParseInputToHarmonyMessage: |
| 255 | """Tests for scenarios that are common to both Chat Completion |
| 256 | parse_chat_input_to_harmony_message and Responses API |
| 257 | response_previous_input_to_harmony functions. |
| 258 | """ |
| 259 | |
| 260 | @pytest.fixture( |
| 261 | params=[parse_chat_input_to_harmony_message, response_previous_input_to_harmony] |
| 262 | ) |
| 263 | def parse_function(self, request): |
| 264 | return request.param |
| 265 | |
| 266 | def test_assistant_message_with_tool_calls(self, parse_function): |
| 267 | """Test parsing assistant message with tool calls.""" |
| 268 | chat_msg = { |
| 269 | "role": "assistant", |
| 270 | "tool_calls": [ |
| 271 | { |
| 272 | "function": { |
| 273 | "name": "get_weather", |
| 274 | "arguments": '{"location": "San Francisco"}', |
| 275 | } |
| 276 | }, |
| 277 | { |
| 278 | "function": { |
| 279 | "name": "search_web", |
| 280 | "arguments": '{"query": "latest news"}', |
| 281 | } |
| 282 | }, |
| 283 | ], |
| 284 | } |
| 285 | |
| 286 | messages = parse_function(chat_msg) |
| 287 | |
| 288 | assert len(messages) == 2 |
| 289 | |
| 290 | # First tool call |
| 291 | assert messages[0].author.role == Role.ASSISTANT |
| 292 | assert messages[0].content[0].text == '{"location": "San Francisco"}' |
| 293 | assert messages[0].channel == "commentary" |
| 294 | assert messages[0].recipient == "functions.get_weather" |
| 295 | assert messages[0].content_type == "json" |
| 296 | |
| 297 | # Second tool call |
| 298 | assert messages[1].author.role == Role.ASSISTANT |
| 299 | assert messages[1].content[0].text == '{"query": "latest news"}' |
| 300 | assert messages[1].channel == "commentary" |
| 301 | assert messages[1].recipient == "functions.search_web" |
| 302 | assert messages[1].content_type == "json" |
| 303 | |
| 304 | def test_assistant_message_with_empty_tool_call_arguments(self, parse_function): |
| 305 | """Test parsing assistant message with tool call having None arguments.""" |
| 306 | chat_msg = { |
| 307 | "role": "assistant", |
| 308 | "tool_calls": [ |
| 309 | { |
| 310 | "function": { |
| 311 | "name": "get_current_time", |
| 312 | "arguments": None, |
| 313 | } |
| 314 | } |
| 315 | ], |
| 316 | } |
| 317 | |
| 318 | messages = parse_function(chat_msg) |
| 319 | |
| 320 | assert len(messages) == 1 |
| 321 | assert messages[0].content[0].text == "" |
| 322 | assert messages[0].recipient == "functions.get_current_time" |
| 323 | |
| 324 | def test_system_message(self, parse_function): |
| 325 | """Test parsing system messages, which are parsed into developer messages |
| 326 | with DeveloperContent.""" |
| 327 | chat_msg = { |
| 328 | "role": "system", |
| 329 | "content": "You are a helpful assistant", |
| 330 | } |
| 331 | |
| 332 | messages = parse_function(chat_msg) |
| 333 | |
| 334 | assert len(messages) == 1 |
| 335 | assert messages[0].author.role == Role.DEVELOPER |
| 336 | assert isinstance(messages[0].content[0], DeveloperContent) |
| 337 | assert messages[0].content[0].instructions == "You are a helpful assistant" |
| 338 | |
| 339 | def test_developer_message(self, parse_function): |
| 340 | """Test parsing developer message.""" |
| 341 | chat_msg = { |
| 342 | "role": "developer", |
| 343 | "content": "Use concise language", |
| 344 | } |
| 345 | |
| 346 | messages = parse_function(chat_msg) |
| 347 | |
| 348 | assert len(messages) == 1 |
| 349 | assert messages[0].author.role == Role.DEVELOPER |
| 350 | assert isinstance(messages[0].content[0], DeveloperContent) |
| 351 | assert messages[0].content[0].instructions == "Use concise language" |
| 352 | |
| 353 | def test_user_message_with_string_content(self, parse_function): |
| 354 | """Test parsing user message with string content.""" |
| 355 | chat_msg = { |
| 356 | "role": "user", |
| 357 | "content": "What's the weather in San Francisco?", |
| 358 | } |
| 359 | |
| 360 | messages = parse_function(chat_msg) |
| 361 | |
| 362 | assert len(messages) == 1 |
| 363 | assert messages[0].author.role == Role.USER |
| 364 | assert messages[0].content[0].text == "What's the weather in San Francisco?" |
| 365 | |
| 366 | def test_user_message_with_array_content(self, parse_function): |
| 367 | """Test parsing user message with array content.""" |
| 368 | chat_msg = { |
| 369 | "role": "user", |
| 370 | "content": [ |
| 371 | {"text": "What's in this image? "}, |
| 372 | {"text": "Please describe it."}, |
| 373 | ], |
| 374 | } |
| 375 | |
| 376 | messages = parse_function(chat_msg) |
| 377 | |
| 378 | assert len(messages) == 1 |
| 379 | assert messages[0].author.role == Role.USER |
| 380 | assert len(messages[0].content) == 2 |
| 381 | assert messages[0].content[0].text == "What's in this image? " |
| 382 | assert messages[0].content[1].text == "Please describe it." |
| 383 | |
| 384 | def test_assistant_message_with_string_content(self, parse_function): |
| 385 | """Test parsing assistant message with string content (no tool calls).""" |
| 386 | chat_msg = { |
| 387 | "role": "assistant", |
| 388 | "content": "Hello! How can I help you today?", |
| 389 | } |
| 390 | |
| 391 | messages = parse_function(chat_msg) |
| 392 | |
| 393 | assert len(messages) == 1 |
| 394 | assert messages[0].author.role == Role.ASSISTANT |
| 395 | assert messages[0].content[0].text == "Hello! How can I help you today?" |
| 396 | |
| 397 | def test_pydantic_model_input(self, parse_function): |
| 398 | """Test parsing Pydantic model input (has model_dump method).""" |
| 399 | |
| 400 | class MockPydanticModel: |
| 401 | def model_dump(self, exclude_none=True): |
| 402 | return { |
| 403 | "role": "user", |
| 404 | "content": "Test message", |
| 405 | } |
| 406 | |
| 407 | chat_msg = MockPydanticModel() |
| 408 | messages = parse_function(chat_msg) |
| 409 | |
| 410 | assert len(messages) == 1 |
| 411 | assert messages[0].author.role == Role.USER |
| 412 | assert messages[0].content[0].text == "Test message" |
| 413 | |
| 414 | def test_tool_call_with_missing_function_fields(self, parse_function): |
| 415 | """Test parsing tool call with missing name or arguments.""" |
| 416 | chat_msg = { |
| 417 | "role": "assistant", |
| 418 | "tool_calls": [ |
| 419 | { |
| 420 | "function": {} # Missing both name and arguments |
| 421 | } |
| 422 | ], |
| 423 | } |
| 424 | |
| 425 | messages = parse_function(chat_msg) |
| 426 | |
| 427 | assert len(messages) == 1 |
| 428 | assert messages[0].recipient == "functions." |
| 429 | assert messages[0].content[0].text == "" |
| 430 | |
| 431 | def test_array_content_with_missing_text(self, parse_function): |
| 432 | """Test parsing array content where text field is missing.""" |
| 433 | chat_msg = { |
| 434 | "role": "user", |
| 435 | "content": [ |
| 436 | {}, # Missing text field |
| 437 | {"text": "actual text"}, |
| 438 | ], |
| 439 | } |
| 440 | |
| 441 | messages = parse_function(chat_msg) |
| 442 | |
| 443 | assert len(messages) == 1 |
| 444 | assert len(messages[0].content) == 2 |
| 445 | assert messages[0].content[0].text == "" |
| 446 | assert messages[0].content[1].text == "actual text" |
| 447 | |
| 448 | |
| 449 | class TestParseChatInputToHarmonyMessage: |
| 450 | """Tests for scenarios that are specific to the Chat Completion API |
| 451 | parse_chat_input_to_harmony_message function. |
| 452 | """ |
| 453 | |
| 454 | def test_user_message_with_empty_content(self): |
| 455 | chat_msg = { |
| 456 | "role": "user", |
| 457 | "content": "", |
| 458 | } |
| 459 | |
| 460 | messages = parse_chat_input_to_harmony_message(chat_msg) |
| 461 | |
| 462 | verify_harmony_messages( |
| 463 | messages, |
| 464 | [ |
| 465 | { |
| 466 | "role": "user", |
| 467 | "content": "", |
| 468 | }, |
| 469 | ], |
| 470 | ) |
| 471 | |
| 472 | def test_user_message_with_none_content(self): |
| 473 | chat_msg = { |
| 474 | "role": "user", |
| 475 | "content": None, |
| 476 | } |
| 477 | |
| 478 | messages = parse_chat_input_to_harmony_message(chat_msg) |
| 479 | |
| 480 | verify_harmony_messages( |
| 481 | messages, |
| 482 | [ |
| 483 | { |
| 484 | "role": "user", |
| 485 | "content": "", |
| 486 | }, |
| 487 | ], |
| 488 | ) |
| 489 | |
| 490 | def test_assistant_message_with_empty_content(self): |
| 491 | chat_msg = { |
| 492 | "role": "assistant", |
| 493 | "content": "", |
| 494 | } |
| 495 | |
| 496 | messages = parse_chat_input_to_harmony_message(chat_msg) |
| 497 | |
| 498 | assert len(messages) == 0 |
| 499 | |
| 500 | def test_assistant_message_with_none_content(self): |
| 501 | chat_msg = { |
| 502 | "role": "assistant", |
| 503 | "content": None, |
| 504 | } |
| 505 | |
| 506 | messages = parse_chat_input_to_harmony_message(chat_msg) |
| 507 | |
| 508 | assert len(messages) == 0 |
| 509 | |
| 510 | def test_assistant_message_with_content_but_empty_reasoning(self): |
| 511 | chat_msg = { |
| 512 | "role": "assistant", |
| 513 | "content": "The answer is 4.", |
| 514 | "reasoning": "", |
| 515 | } |
| 516 | |
| 517 | messages = parse_chat_input_to_harmony_message(chat_msg) |
| 518 | |
| 519 | verify_harmony_messages( |
| 520 | messages, |
| 521 | [ |
| 522 | { |
| 523 | "role": "assistant", |
| 524 | "channel": "final", |
| 525 | "content": "The answer is 4.", |
| 526 | }, |
| 527 | ], |
| 528 | ) |
| 529 | |
| 530 | def test_assistant_message_with_reasoning_but_empty_content(self): |
| 531 | chat_msg = { |
| 532 | "role": "assistant", |
| 533 | "reasoning": "I'm thinking about the user's question.", |
| 534 | "content": "", |
| 535 | } |
| 536 | |
| 537 | messages = parse_chat_input_to_harmony_message(chat_msg) |
| 538 | |
| 539 | verify_harmony_messages( |
| 540 | messages, |
| 541 | [ |
| 542 | { |
| 543 | "role": "assistant", |
| 544 | "channel": "analysis", |
| 545 | "content": "I'm thinking about the user's question.", |
| 546 | }, |
| 547 | ], |
| 548 | ) |
| 549 | |
| 550 | def test_assistant_message_with_reasoning_but_none_content(self): |
| 551 | chat_msg = { |
| 552 | "role": "assistant", |
| 553 | "reasoning": "I'm thinking about the user's question.", |
| 554 | "content": None, |
| 555 | } |
| 556 | |
| 557 | messages = parse_chat_input_to_harmony_message(chat_msg) |
| 558 | |
| 559 | verify_harmony_messages( |
| 560 | messages, |
| 561 | [ |
| 562 | { |
| 563 | "role": "assistant", |
| 564 | "channel": "analysis", |
| 565 | "content": "I'm thinking about the user's question.", |
| 566 | }, |
| 567 | ], |
| 568 | ) |
| 569 | |
| 570 | def test_assistant_message_with_tool_calls_but_no_content(self): |
| 571 | chat_msg = { |
| 572 | "role": "assistant", |
| 573 | "tool_calls": [ |
| 574 | { |
| 575 | "function": { |
| 576 | "name": "get_weather", |
| 577 | "arguments": '{"location": "San Francisco"}', |
| 578 | } |
| 579 | } |
| 580 | ], |
| 581 | } |
| 582 | |
| 583 | messages = parse_chat_input_to_harmony_message(chat_msg) |
| 584 | |
| 585 | verify_harmony_messages( |
| 586 | messages, |
| 587 | [ |
| 588 | { |
| 589 | "role": "assistant", |
| 590 | "channel": "commentary", |
| 591 | "recipient": "functions.get_weather", |
| 592 | "content": '{"location": "San Francisco"}', |
| 593 | "content_type": "json", |
| 594 | }, |
| 595 | ], |
| 596 | ) |
| 597 | |
| 598 | def test_assistant_message_with_tool_calls_and_content(self): |
| 599 | chat_msg = { |
| 600 | "role": "assistant", |
| 601 | "tool_calls": [ |
| 602 | { |
| 603 | "function": { |
| 604 | "name": "get_weather", |
| 605 | "arguments": '{"location": "San Francisco"}', |
| 606 | } |
| 607 | } |
| 608 | ], |
| 609 | "content": "I'll call the tool.", |
| 610 | } |
| 611 | |
| 612 | messages = parse_chat_input_to_harmony_message(chat_msg) |
| 613 | |
| 614 | verify_harmony_messages( |
| 615 | messages, |
| 616 | [ |
| 617 | { |
| 618 | "role": "assistant", |
| 619 | "channel": "commentary", |
| 620 | "content": "I'll call the tool.", |
| 621 | }, |
| 622 | { |
| 623 | "role": "assistant", |
| 624 | "channel": "commentary", |
| 625 | "recipient": "functions.get_weather", |
| 626 | "content": '{"location": "San Francisco"}', |
| 627 | "content_type": "json", |
| 628 | }, |
| 629 | ], |
| 630 | ) |
| 631 | |
| 632 | def test_assistant_message_with_tool_calls_and_reasoning(self): |
| 633 | chat_msg = { |
| 634 | "role": "assistant", |
| 635 | "tool_calls": [ |
| 636 | { |
| 637 | "function": { |
| 638 | "name": "get_weather", |
| 639 | "arguments": '{"location": "San Francisco"}', |
| 640 | } |
| 641 | } |
| 642 | ], |
| 643 | "reasoning": "I should use the get_weather tool.", |
| 644 | } |
| 645 | |
| 646 | messages = parse_chat_input_to_harmony_message(chat_msg) |
| 647 | |
| 648 | verify_harmony_messages( |
| 649 | messages, |
| 650 | [ |
| 651 | { |
| 652 | "role": "assistant", |
| 653 | "channel": "analysis", |
| 654 | "content": "I should use the get_weather tool.", |
| 655 | }, |
| 656 | { |
| 657 | "role": "assistant", |
| 658 | "channel": "commentary", |
| 659 | "recipient": "functions.get_weather", |
| 660 | "content": '{"location": "San Francisco"}', |
| 661 | "content_type": "json", |
| 662 | }, |
| 663 | ], |
| 664 | ) |
| 665 | |
| 666 | def test_assistant_message_with_tool_calls_and_reasoning_and_content(self): |
| 667 | chat_msg = { |
| 668 | "role": "assistant", |
| 669 | "tool_calls": [ |
| 670 | { |
| 671 | "function": { |
| 672 | "name": "get_weather", |
| 673 | "arguments": '{"location": "San Francisco"}', |
| 674 | } |
| 675 | } |
| 676 | ], |
| 677 | "reasoning": "I should use the get_weather tool.", |
| 678 | "content": "I'll call the tool.", |
| 679 | } |
| 680 | |
| 681 | messages = parse_chat_input_to_harmony_message(chat_msg) |
| 682 | |
| 683 | verify_harmony_messages( |
| 684 | messages, |
| 685 | [ |
| 686 | { |
| 687 | "role": "assistant", |
| 688 | "channel": "commentary", |
| 689 | "content": "I'll call the tool.", |
| 690 | }, |
| 691 | { |
| 692 | "role": "assistant", |
| 693 | "channel": "analysis", |
| 694 | "content": "I should use the get_weather tool.", |
| 695 | }, |
| 696 | { |
| 697 | "role": "assistant", |
| 698 | "channel": "commentary", |
| 699 | "recipient": "functions.get_weather", |
| 700 | "content": '{"location": "San Francisco"}', |
| 701 | "content_type": "json", |
| 702 | }, |
| 703 | ], |
| 704 | ) |
| 705 | |
| 706 | def test_tool_message_with_string_content(self): |
| 707 | tool_id_names = { |
| 708 | "call_123": "get_weather", |
| 709 | } |
| 710 | chat_msg = { |
| 711 | "role": "tool", |
| 712 | "tool_call_id": "call_123", |
| 713 | "content": "The weather in San Francisco is sunny, 72°F", |
| 714 | } |
| 715 | |
| 716 | messages = parse_chat_input_to_harmony_message( |
| 717 | chat_msg, tool_id_names=tool_id_names |
| 718 | ) |
| 719 | |
| 720 | verify_harmony_messages( |
| 721 | messages, |
| 722 | [ |
| 723 | { |
| 724 | "role": "tool", |
| 725 | "name": "functions.get_weather", |
| 726 | "content": "The weather in San Francisco is sunny, 72°F", |
| 727 | "channel": "commentary", |
| 728 | }, |
| 729 | ], |
| 730 | ) |
| 731 | |
| 732 | def test_tool_message_with_array_content(self): |
| 733 | tool_id_names = { |
| 734 | "call_123": "search_results", |
| 735 | } |
| 736 | chat_msg = { |
| 737 | "role": "tool", |
| 738 | "tool_call_id": "call_123", |
| 739 | "content": [ |
| 740 | {"type": "text", "text": "Result 1: "}, |
| 741 | {"type": "text", "text": "Result 2: "}, |
| 742 | { |
| 743 | "type": "image", |
| 744 | "url": "http://example.com/img.png", |
| 745 | }, # Should be ignored |
| 746 | {"type": "text", "text": "Result 3"}, |
| 747 | ], |
| 748 | } |
| 749 | |
| 750 | messages = parse_chat_input_to_harmony_message( |
| 751 | chat_msg, tool_id_names=tool_id_names |
| 752 | ) |
| 753 | |
| 754 | verify_harmony_messages( |
| 755 | messages, |
| 756 | [ |
| 757 | { |
| 758 | "role": "tool", |
| 759 | "name": "functions.search_results", |
| 760 | "content": "Result 1: Result 2: Result 3", |
| 761 | "channel": "commentary", |
| 762 | }, |
| 763 | ], |
| 764 | ) |
| 765 | |
| 766 | def test_tool_message_with_empty_content(self): |
| 767 | tool_id_names = { |
| 768 | "call_123": "empty_tool", |
| 769 | } |
| 770 | chat_msg = { |
| 771 | "role": "tool", |
| 772 | "tool_call_id": "call_123", |
| 773 | "content": "", |
| 774 | } |
| 775 | |
| 776 | messages = parse_chat_input_to_harmony_message( |
| 777 | chat_msg, tool_id_names=tool_id_names |
| 778 | ) |
| 779 | |
| 780 | verify_harmony_messages( |
| 781 | messages, |
| 782 | [ |
| 783 | { |
| 784 | "role": "tool", |
| 785 | "name": "functions.empty_tool", |
| 786 | "content": "", |
| 787 | "channel": "commentary", |
| 788 | }, |
| 789 | ], |
| 790 | ) |
| 791 | |
| 792 | def test_tool_message_with_none_content(self): |
| 793 | tool_id_names = { |
| 794 | "call_123": "empty_tool", |
| 795 | } |
| 796 | chat_msg = { |
| 797 | "role": "tool", |
| 798 | "tool_call_id": "call_123", |
| 799 | "content": None, |
| 800 | } |
| 801 | |
| 802 | messages = parse_chat_input_to_harmony_message( |
| 803 | chat_msg, tool_id_names=tool_id_names |
| 804 | ) |
| 805 | |
| 806 | verify_harmony_messages( |
| 807 | messages, |
| 808 | [ |
| 809 | { |
| 810 | "role": "tool", |
| 811 | "name": "functions.empty_tool", |
| 812 | "content": "", |
| 813 | "channel": "commentary", |
| 814 | }, |
| 815 | ], |
| 816 | ) |
| 817 | |
| 818 | |
| 819 | class TestAutoDropAnalysisMessages: |
| 820 | def test_no_analysis_messages(self) -> None: |
| 821 | messages = [ |
| 822 | Message.from_role_and_content( |
| 823 | Role.ASSISTANT, "The answer is 4." |
| 824 | ).with_channel("final"), |
| 825 | ] |
| 826 | cleaned_messages = auto_drop_analysis_messages(messages) |
| 827 | assert cleaned_messages == messages |
| 828 | |
| 829 | def test_only_analysis_message(self) -> None: |
| 830 | messages = [ |
| 831 | Message.from_role_and_content( |
| 832 | Role.ASSISTANT, "I'm thinking about the user's question." |
| 833 | ).with_channel("analysis"), |
| 834 | ] |
| 835 | cleaned_messages = auto_drop_analysis_messages(messages) |
| 836 | assert cleaned_messages == messages |
| 837 | |
| 838 | def test_multiple_analysis_messages_without_final_message(self) -> None: |
| 839 | messages = [ |
| 840 | Message.from_role_and_content( |
| 841 | Role.ASSISTANT, "I'm thinking about the user's question." |
| 842 | ).with_channel("analysis"), |
| 843 | Message.from_role_and_content( |
| 844 | Role.ASSISTANT, "I'm thinking more." |
| 845 | ).with_channel("analysis"), |
| 846 | Message.from_role_and_content( |
| 847 | Role.ASSISTANT, "I'm thinking even more." |
| 848 | ).with_channel("analysis"), |
| 849 | ] |
| 850 | cleaned_messages = auto_drop_analysis_messages(messages) |
| 851 | assert cleaned_messages == messages |
| 852 | |
| 853 | def test_only_final_message(self) -> None: |
| 854 | messages = [ |
| 855 | Message.from_role_and_content( |
| 856 | Role.ASSISTANT, "The answer is 4." |
| 857 | ).with_channel("final"), |
| 858 | ] |
| 859 | cleaned_messages = auto_drop_analysis_messages(messages) |
| 860 | assert cleaned_messages == messages |
| 861 | |
| 862 | def test_drops_one_analysis_messages_before_final_message(self) -> None: |
| 863 | messages = [ |
| 864 | Message.from_role_and_content( |
| 865 | Role.ASSISTANT, "I'm thinking about the user's question." |
| 866 | ).with_channel("analysis"), |
| 867 | Message.from_role_and_content( |
| 868 | Role.ASSISTANT, "The answer is 4." |
| 869 | ).with_channel("final"), |
| 870 | Message.from_role_and_content( |
| 871 | Role.ASSISTANT, "I should think harder." |
| 872 | ).with_channel("analysis"), |
| 873 | ] |
| 874 | cleaned_messages = auto_drop_analysis_messages(messages) |
| 875 | # Should have dropped the first analysis message |
| 876 | assert cleaned_messages == messages[1:] |
| 877 | |
| 878 | def test_drops_all_analysis_messages_before_final_message(self) -> None: |
| 879 | messages = [ |
| 880 | Message.from_role_and_content( |
| 881 | Role.ASSISTANT, "I'm thinking about the user's question." |
| 882 | ).with_channel("analysis"), |
| 883 | Message.from_role_and_content( |
| 884 | Role.ASSISTANT, "I'm thinking more." |
| 885 | ).with_channel("analysis"), |
| 886 | Message.from_role_and_content( |
| 887 | Role.ASSISTANT, "I'm thinking even more." |
| 888 | ).with_channel("analysis"), |
| 889 | Message.from_role_and_content( |
| 890 | Role.ASSISTANT, "The answer is 4." |
| 891 | ).with_channel("final"), |
| 892 | Message.from_role_and_content( |
| 893 | Role.ASSISTANT, "I should think harder." |
| 894 | ).with_channel("analysis"), |
| 895 | ] |
| 896 | cleaned_messages = auto_drop_analysis_messages(messages) |
| 897 | # Should have dropped the first 3 analysis messages |
| 898 | assert cleaned_messages == messages[3:] |
| 899 | |
| 900 | def test_multiple_analysis_messages_with_multiple_final_messages(self) -> None: |
| 901 | messages = [ |
| 902 | Message.from_role_and_content( |
| 903 | Role.ASSISTANT, "I'm thinking about the user's question." |
| 904 | ).with_channel("analysis"), |
| 905 | Message.from_role_and_content( |
| 906 | Role.ASSISTANT, "I'm thinking more." |
| 907 | ).with_channel("analysis"), |
| 908 | Message.from_role_and_content( |
| 909 | Role.ASSISTANT, "I'm thinking even more." |
| 910 | ).with_channel("analysis"), |
| 911 | Message.from_role_and_content( |
| 912 | Role.ASSISTANT, "The answer is 4." |
| 913 | ).with_channel("final"), |
| 914 | Message.from_role_and_content( |
| 915 | Role.ASSISTANT, "I should think harder." |
| 916 | ).with_channel("analysis"), |
| 917 | Message.from_role_and_content( |
| 918 | Role.ASSISTANT, "The answer is 5." |
| 919 | ).with_channel("final"), |
| 920 | ] |
| 921 | cleaned_messages = auto_drop_analysis_messages(messages) |
| 922 | # Should have dropped all those analysis messages |
| 923 | assert len(cleaned_messages) == 2 |
| 924 | assert cleaned_messages[0].content[0].text == "The answer is 4." |
| 925 | assert cleaned_messages[1].content[0].text == "The answer is 5." |
| 926 | |
| 927 | def test_drops_non_assistant_analysis_messages(self) -> None: |
| 928 | messages = [ |
| 929 | Message.from_role_and_content( |
| 930 | Role.TOOL, "The tool thinks we should think harder." |
| 931 | ).with_channel("analysis"), |
| 932 | Message.from_role_and_content( |
| 933 | Role.ASSISTANT, "The answer is 4." |
| 934 | ).with_channel("final"), |
| 935 | ] |
| 936 | cleaned_messages = auto_drop_analysis_messages(messages) |
| 937 | # Should have dropped the analysis message |
| 938 | assert cleaned_messages == messages[1:] |
| 939 | |
| 940 | |
| 941 | def test_has_custom_tools() -> None: |
| 942 | assert not has_custom_tools(set()) |
| 943 | assert not has_custom_tools({"web_search_preview", "code_interpreter", "container"}) |
| 944 | assert has_custom_tools({"others"}) |
| 945 | assert has_custom_tools( |
| 946 | {"web_search_preview", "code_interpreter", "container", "others"} |
| 947 | ) |
| 948 | |
| 949 | |
| 950 | class TestGetSystemMessage: |
| 951 | """Tests for get_system_message channel configuration.""" |
| 952 | |
| 953 | def test_commentary_channel_present_without_custom_tools(self) -> None: |
| 954 | """Commentary channel must be valid even without custom tools.""" |
| 955 | sys_msg = get_system_message(with_custom_tools=False) |
| 956 | valid_channels = sys_msg.content[0].channel_config.valid_channels |
| 957 | assert "commentary" in valid_channels |
| 958 | |
| 959 | def test_commentary_channel_present_with_custom_tools(self) -> None: |
| 960 | """Commentary channel present when custom tools are enabled.""" |
| 961 | sys_msg = get_system_message(with_custom_tools=True) |
| 962 | valid_channels = sys_msg.content[0].channel_config.valid_channels |
| 963 | assert "commentary" in valid_channels |
| 964 | |
| 965 | def test_all_standard_channels_present(self) -> None: |
| 966 | """All three standard Harmony channels should always be valid.""" |
| 967 | for with_tools in (True, False): |
| 968 | sys_msg = get_system_message(with_custom_tools=with_tools) |
| 969 | valid_channels = sys_msg.content[0].channel_config.valid_channels |
| 970 | for channel in ("analysis", "commentary", "final"): |
| 971 | assert channel in valid_channels, ( |
| 972 | f"{channel} missing when with_custom_tools={with_tools}" |
| 973 | ) |
| 974 | |
| 975 | def test_unsupported_reasoning_effort_raises_clear_error(self) -> None: |
| 976 | with pytest.raises( |
| 977 | VLLMValidationError, |
| 978 | match="reasoning_effort='max' is not supported by Harmony", |
| 979 | ) as exc_info: |
| 980 | get_system_message(reasoning_effort="max") |
| 981 | assert exc_info.value.parameter == "reasoning_effort" |
| 982 | |
| 983 | |
| 984 | class TestResponseInputToHarmonyReasoningItem: |
| 985 | """Tests for response_input_to_harmony handling of reasoning input items. |
| 986 | |
| 987 | Per the OpenAI spec, ResponseReasoningItem.content is |
| 988 | Optional[List[Content]] = None. Clients like langchain-openai may omit |
| 989 | this field when constructing multi-turn input from previous responses. |
| 990 | |
| 991 | Reasoning items with content are converted to Harmony messages on the |
| 992 | 'analysis' channel. All content items are concatenated. Items without |
| 993 | content return None (skipped by the caller). |
| 994 | """ |
| 995 | |
| 996 | def test_reasoning_with_single_content(self): |
| 997 | """Test reasoning item with a single content entry.""" |
| 998 | item = { |
| 999 | "type": "reasoning", |
| 1000 | "id": "rs_123", |
| 1001 | "content": [{"type": "reasoning_text", "text": "Thinking step by step"}], |
| 1002 | } |
| 1003 | |
| 1004 | msg = response_input_to_harmony(item, prev_responses=[]) |
| 1005 | |
| 1006 | assert msg is not None |
| 1007 | assert msg.author.role == Role.ASSISTANT |
| 1008 | assert msg.content[0].text == "Thinking step by step" |
| 1009 | assert msg.channel == "analysis" |
| 1010 | |
| 1011 | def test_reasoning_with_multiple_content_items(self): |
| 1012 | """Test reasoning item with multiple content entries concatenated.""" |
| 1013 | item = { |
| 1014 | "type": "reasoning", |
| 1015 | "id": "rs_123", |
| 1016 | "content": [ |
| 1017 | {"type": "reasoning_text", "text": "First, let me analyze"}, |
| 1018 | {"type": "reasoning_text", "text": "Second, I should consider"}, |
| 1019 | {"type": "reasoning_text", "text": "Finally, the answer is"}, |
| 1020 | ], |
| 1021 | } |
| 1022 | |
| 1023 | msg = response_input_to_harmony(item, prev_responses=[]) |
| 1024 | |
| 1025 | assert msg is not None |
| 1026 | assert msg.author.role == Role.ASSISTANT |
| 1027 | assert msg.content[0].text == ( |
| 1028 | "First, let me analyze\nSecond, I should consider\nFinally, the answer is" |
| 1029 | ) |
| 1030 | assert msg.channel == "analysis" |
| 1031 | |
| 1032 | def test_reasoning_without_content_returns_none(self): |
| 1033 | """Test reasoning item without content field returns None.""" |
| 1034 | item = { |
| 1035 | "type": "reasoning", |
| 1036 | "id": "rs_123", |
| 1037 | "summary": [{"type": "summary_text", "text": "Thinking about math"}], |
| 1038 | } |
| 1039 | |
| 1040 | msg = response_input_to_harmony(item, prev_responses=[]) |
| 1041 | |
| 1042 | assert msg is None |
| 1043 | |
| 1044 | def test_reasoning_with_none_content_returns_none(self): |
| 1045 | """Test reasoning item with content=None returns None.""" |
| 1046 | item = { |
| 1047 | "type": "reasoning", |
| 1048 | "id": "rs_123", |
| 1049 | "content": None, |
| 1050 | "summary": [{"type": "summary_text", "text": "Thinking about math"}], |
| 1051 | } |
| 1052 | |
| 1053 | msg = response_input_to_harmony(item, prev_responses=[]) |
| 1054 | |
| 1055 | assert msg is None |
| 1056 | |
| 1057 | def test_reasoning_with_empty_content_returns_none(self): |
| 1058 | """Test reasoning item with empty content list returns None.""" |
| 1059 | item = { |
| 1060 | "type": "reasoning", |
| 1061 | "id": "rs_123", |
| 1062 | "content": [], |
| 1063 | } |
| 1064 | |
| 1065 | msg = response_input_to_harmony(item, prev_responses=[]) |
| 1066 | |
| 1067 | assert msg is None |