| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | |
| 4 | from unittest.mock import patch |
| 5 | |
| 6 | import pytest |
| 7 | from openai.types.responses.response_function_tool_call import ResponseFunctionToolCall |
| 8 | from openai.types.responses.response_function_tool_call_output_item import ( |
| 9 | ResponseFunctionToolCallOutputItem, |
| 10 | ) |
| 11 | from openai.types.responses.response_output_message import ResponseOutputMessage |
| 12 | from openai.types.responses.response_output_text import ResponseOutputText |
| 13 | from openai.types.responses.response_reasoning_item import ( |
| 14 | Content, |
| 15 | ResponseReasoningItem, |
| 16 | Summary, |
| 17 | ) |
| 18 | |
| 19 | from vllm.entrypoints.openai.responses.utils import ( |
| 20 | _construct_message_from_response_item, |
| 21 | construct_chat_messages_with_tool_call, |
| 22 | construct_input_messages, |
| 23 | should_continue_final_message, |
| 24 | ) |
| 25 | from vllm.exceptions import VLLMValidationError |
| 26 | |
| 27 | |
| 28 | def _single_chat_message(item): |
| 29 | message = _construct_message_from_response_item(item) |
| 30 | assert message is not None |
| 31 | return message |
| 32 | |
| 33 | |
| 34 | def make_output_message( |
| 35 | text: str, |
| 36 | *, |
| 37 | id: str = "msg_1", |
| 38 | status: str = "completed", |
| 39 | ) -> ResponseOutputMessage: |
| 40 | return ResponseOutputMessage( |
| 41 | id=id, |
| 42 | content=[ |
| 43 | ResponseOutputText( |
| 44 | annotations=[], |
| 45 | text=text, |
| 46 | type="output_text", |
| 47 | logprobs=None, |
| 48 | ) |
| 49 | ], |
| 50 | role="assistant", |
| 51 | status=status, |
| 52 | type="message", |
| 53 | ) |
| 54 | |
| 55 | |
| 56 | def make_reasoning_item( |
| 57 | *, |
| 58 | content_text: str | None = None, |
| 59 | summary_text: str | None = None, |
| 60 | content: list[Content] | None = None, |
| 61 | summary: list[Summary] | None = None, |
| 62 | encrypted_content: str | None = None, |
| 63 | id: str = "reasoning_1", |
| 64 | status: str | None = None, |
| 65 | ) -> ResponseReasoningItem: |
| 66 | if content is None and content_text is not None: |
| 67 | content = [Content(text=content_text, type="reasoning_text")] |
| 68 | if summary is None and summary_text is not None: |
| 69 | summary = [Summary(text=summary_text, type="summary_text")] |
| 70 | |
| 71 | return ResponseReasoningItem( |
| 72 | id=id, |
| 73 | summary=[] if summary is None else summary, |
| 74 | type="reasoning", |
| 75 | content=content, |
| 76 | encrypted_content=encrypted_content, |
| 77 | status=status, |
| 78 | ) |
| 79 | |
| 80 | |
| 81 | def make_function_call( |
| 82 | *, |
| 83 | call_id: str, |
| 84 | name: str = "test_function", |
| 85 | arguments: str = "{}", |
| 86 | id: str = "tool_id", |
| 87 | status: str | None = None, |
| 88 | ) -> ResponseFunctionToolCall: |
| 89 | kwargs = { |
| 90 | "type": "function_call", |
| 91 | "id": id, |
| 92 | "call_id": call_id, |
| 93 | "name": name, |
| 94 | "arguments": arguments, |
| 95 | } |
| 96 | if status is not None: |
| 97 | kwargs["status"] = status |
| 98 | |
| 99 | return ResponseFunctionToolCall(**kwargs) |
| 100 | |
| 101 | |
| 102 | def make_function_call_output( |
| 103 | *, |
| 104 | call_id: str, |
| 105 | output: str = "42", |
| 106 | id: str = "output_1", |
| 107 | status: str = "completed", |
| 108 | ) -> ResponseFunctionToolCallOutputItem: |
| 109 | return ResponseFunctionToolCallOutputItem( |
| 110 | id=id, |
| 111 | type="function_call_output", |
| 112 | call_id=call_id, |
| 113 | output=output, |
| 114 | status=status, |
| 115 | ) |
| 116 | |
| 117 | |
| 118 | class TestResponsesUtils: |
| 119 | """Tests for Responses API utils.""" |
| 120 | |
| 121 | def test_construct_chat_messages_with_tool_call(self): |
| 122 | """Test construction of chat messages with tool calls.""" |
| 123 | reasoning_item = ResponseReasoningItem( |
| 124 | id="lol", |
| 125 | summary=[], |
| 126 | type="reasoning", |
| 127 | content=[ |
| 128 | Content( |
| 129 | text="Leroy Jenkins", |
| 130 | type="reasoning_text", |
| 131 | ) |
| 132 | ], |
| 133 | encrypted_content=None, |
| 134 | status=None, |
| 135 | ) |
| 136 | mcp_tool_item = ResponseFunctionToolCall( |
| 137 | id="mcp_123", |
| 138 | call_id="call_123", |
| 139 | type="function_call", |
| 140 | status="completed", |
| 141 | name="python", |
| 142 | arguments='{"code": "123+456"}', |
| 143 | ) |
| 144 | input_items = [reasoning_item, mcp_tool_item] |
| 145 | messages = construct_chat_messages_with_tool_call(input_items) |
| 146 | |
| 147 | assert len(messages) == 1 |
| 148 | message = messages[0] |
| 149 | assert message["role"] == "assistant" |
| 150 | assert message["reasoning"] == "Leroy Jenkins" |
| 151 | assert message["tool_calls"][0]["id"] == "call_123" |
| 152 | assert message["tool_calls"][0]["function"]["name"] == "python" |
| 153 | assert ( |
| 154 | message["tool_calls"][0]["function"]["arguments"] == '{"code": "123+456"}' |
| 155 | ) |
| 156 | |
| 157 | def test_construct_chat_messages_preserves_single_item_conversions(self): |
| 158 | item = ResponseReasoningItem( |
| 159 | id="lol", |
| 160 | summary=[], |
| 161 | type="reasoning", |
| 162 | content=[ |
| 163 | Content( |
| 164 | text="Leroy Jenkins", |
| 165 | type="reasoning_text", |
| 166 | ) |
| 167 | ], |
| 168 | encrypted_content=None, |
| 169 | status=None, |
| 170 | ) |
| 171 | formatted_item = _single_chat_message(item) |
| 172 | assert formatted_item["role"] == "assistant" |
| 173 | assert formatted_item["reasoning"] == "Leroy Jenkins" |
| 174 | |
| 175 | item = ResponseReasoningItem( |
| 176 | id="lol", |
| 177 | summary=[ |
| 178 | Summary( |
| 179 | text='Hmm, the user has just started with a simple "Hello,"', |
| 180 | type="summary_text", |
| 181 | ) |
| 182 | ], |
| 183 | type="reasoning", |
| 184 | content=None, |
| 185 | encrypted_content=None, |
| 186 | status=None, |
| 187 | ) |
| 188 | |
| 189 | formatted_item = _single_chat_message(item) |
| 190 | assert formatted_item["role"] == "assistant" |
| 191 | assert ( |
| 192 | formatted_item["reasoning"] |
| 193 | == 'Hmm, the user has just started with a simple "Hello,"' |
| 194 | ) |
| 195 | |
| 196 | tool_call_output = ResponseFunctionToolCallOutputItem( |
| 197 | id="temp_id", |
| 198 | type="function_call_output", |
| 199 | call_id="temp", |
| 200 | output="1234", |
| 201 | status="completed", |
| 202 | ) |
| 203 | formatted_item = _single_chat_message(tool_call_output) |
| 204 | assert formatted_item["role"] == "tool" |
| 205 | assert formatted_item["content"] == "1234" |
| 206 | assert formatted_item["tool_call_id"] == "temp" |
| 207 | |
| 208 | formatted_item = _single_chat_message( |
| 209 | { |
| 210 | "type": "function_call_output", |
| 211 | "call_id": "temp_dict", |
| 212 | "output": "5678", |
| 213 | } |
| 214 | ) |
| 215 | assert formatted_item["role"] == "tool" |
| 216 | assert formatted_item["content"] == "5678" |
| 217 | assert formatted_item["tool_call_id"] == "temp_dict" |
| 218 | |
| 219 | item = ResponseReasoningItem( |
| 220 | id="lol", |
| 221 | summary=[], |
| 222 | type="reasoning", |
| 223 | content=None, |
| 224 | encrypted_content="TOP_SECRET_MESSAGE", |
| 225 | status=None, |
| 226 | ) |
| 227 | with pytest.raises(VLLMValidationError) as exc_info: |
| 228 | construct_chat_messages_with_tool_call([item]) |
| 229 | assert exc_info.value.parameter == "input" |
| 230 | |
| 231 | output_item = ResponseOutputMessage( |
| 232 | id="msg_bf585bbbe3d500e0", |
| 233 | content=[ |
| 234 | ResponseOutputText( |
| 235 | annotations=[], |
| 236 | text="dongyi", |
| 237 | type="output_text", |
| 238 | logprobs=None, |
| 239 | ) |
| 240 | ], |
| 241 | role="assistant", |
| 242 | status="completed", |
| 243 | type="message", |
| 244 | ) |
| 245 | |
| 246 | formatted_item = _single_chat_message(output_item) |
| 247 | assert formatted_item["role"] == "assistant" |
| 248 | assert formatted_item["content"] == "dongyi" |
| 249 | |
| 250 | |
| 251 | class TestReasoningItemContentPriority: |
| 252 | """Tests that content is prioritized over summary for reasoning items.""" |
| 253 | |
| 254 | def test_content_preferred_over_summary(self): |
| 255 | """When both content and summary are present, content should win.""" |
| 256 | item = ResponseReasoningItem( |
| 257 | id="reasoning_1", |
| 258 | summary=[ |
| 259 | Summary( |
| 260 | text="This is a summary", |
| 261 | type="summary_text", |
| 262 | ) |
| 263 | ], |
| 264 | type="reasoning", |
| 265 | content=[ |
| 266 | Content( |
| 267 | text="This is the actual content", |
| 268 | type="reasoning_text", |
| 269 | ) |
| 270 | ], |
| 271 | encrypted_content=None, |
| 272 | status=None, |
| 273 | ) |
| 274 | formatted = _single_chat_message(item) |
| 275 | assert formatted["reasoning"] == "This is the actual content" |
| 276 | |
| 277 | def test_content_only(self): |
| 278 | """When only content is present (no summary), content is used.""" |
| 279 | item = ResponseReasoningItem( |
| 280 | id="reasoning_2", |
| 281 | summary=[], |
| 282 | type="reasoning", |
| 283 | content=[ |
| 284 | Content( |
| 285 | text="Content without summary", |
| 286 | type="reasoning_text", |
| 287 | ) |
| 288 | ], |
| 289 | encrypted_content=None, |
| 290 | status=None, |
| 291 | ) |
| 292 | formatted = _single_chat_message(item) |
| 293 | assert formatted["reasoning"] == "Content without summary" |
| 294 | |
| 295 | @patch("vllm.entrypoints.openai.responses.utils.logger") |
| 296 | def test_summary_fallback_when_no_content(self, mock_logger): |
| 297 | """When content is absent, summary is used as fallback with warning.""" |
| 298 | item = ResponseReasoningItem( |
| 299 | id="reasoning_3", |
| 300 | summary=[ |
| 301 | Summary( |
| 302 | text="Fallback summary text", |
| 303 | type="summary_text", |
| 304 | ) |
| 305 | ], |
| 306 | type="reasoning", |
| 307 | content=None, |
| 308 | encrypted_content=None, |
| 309 | status=None, |
| 310 | ) |
| 311 | formatted = _single_chat_message(item) |
| 312 | assert formatted["reasoning"] == "Fallback summary text" |
| 313 | mock_logger.warning.assert_called_once() |
| 314 | assert ( |
| 315 | "summary text as reasoning content" in mock_logger.warning.call_args[0][0] |
| 316 | ) |
| 317 | |
| 318 | @patch("vllm.entrypoints.openai.responses.utils.logger") |
| 319 | def test_summary_fallback_when_content_empty(self, mock_logger): |
| 320 | """When content is an empty list, summary is used as fallback.""" |
| 321 | item = ResponseReasoningItem( |
| 322 | id="reasoning_4", |
| 323 | summary=[ |
| 324 | Summary( |
| 325 | text="Summary when content empty", |
| 326 | type="summary_text", |
| 327 | ) |
| 328 | ], |
| 329 | type="reasoning", |
| 330 | content=[], |
| 331 | encrypted_content=None, |
| 332 | status=None, |
| 333 | ) |
| 334 | formatted = _single_chat_message(item) |
| 335 | assert formatted["reasoning"] == "Summary when content empty" |
| 336 | mock_logger.warning.assert_called_once() |
| 337 | assert ( |
| 338 | "summary text as reasoning content" in mock_logger.warning.call_args[0][0] |
| 339 | ) |
| 340 | |
| 341 | def test_neither_content_nor_summary(self): |
| 342 | """When neither content nor summary is present, reasoning is empty.""" |
| 343 | item = ResponseReasoningItem( |
| 344 | id="reasoning_5", |
| 345 | summary=[], |
| 346 | type="reasoning", |
| 347 | content=None, |
| 348 | encrypted_content=None, |
| 349 | status=None, |
| 350 | ) |
| 351 | formatted = _single_chat_message(item) |
| 352 | assert formatted["reasoning"] == "" |
| 353 | |
| 354 | def test_encrypted_content_raises(self): |
| 355 | """Encrypted content should raise VLLMValidationError.""" |
| 356 | item = ResponseReasoningItem( |
| 357 | id="reasoning_6", |
| 358 | summary=[ |
| 359 | Summary( |
| 360 | text="Some summary", |
| 361 | type="summary_text", |
| 362 | ) |
| 363 | ], |
| 364 | type="reasoning", |
| 365 | content=[ |
| 366 | Content( |
| 367 | text="Some content", |
| 368 | type="reasoning_text", |
| 369 | ) |
| 370 | ], |
| 371 | encrypted_content="ENCRYPTED", |
| 372 | status=None, |
| 373 | ) |
| 374 | with pytest.raises(VLLMValidationError) as exc_info: |
| 375 | construct_chat_messages_with_tool_call([item]) |
| 376 | assert exc_info.value.parameter == "input" |
| 377 | |
| 378 | @patch("vllm.entrypoints.openai.responses.utils.logger") |
| 379 | def test_summary_with_multiple_entries_uses_first(self, mock_logger): |
| 380 | """When multiple summary entries exist, the first one is used.""" |
| 381 | item = ResponseReasoningItem( |
| 382 | id="reasoning_7", |
| 383 | summary=[ |
| 384 | Summary( |
| 385 | text="First summary", |
| 386 | type="summary_text", |
| 387 | ), |
| 388 | Summary( |
| 389 | text="Second summary", |
| 390 | type="summary_text", |
| 391 | ), |
| 392 | ], |
| 393 | type="reasoning", |
| 394 | content=None, |
| 395 | encrypted_content=None, |
| 396 | status=None, |
| 397 | ) |
| 398 | formatted = _single_chat_message(item) |
| 399 | assert formatted["reasoning"] == "First summary" |
| 400 | mock_logger.warning.assert_called_once() |
| 401 | assert ( |
| 402 | "summary text as reasoning content" in mock_logger.warning.call_args[0][0] |
| 403 | ) |
| 404 | |
| 405 | @patch("vllm.entrypoints.openai.responses.utils.logger") |
| 406 | def test_no_warning_when_content_used(self, mock_logger): |
| 407 | """No warning should be emitted when content is available.""" |
| 408 | item = ResponseReasoningItem( |
| 409 | id="reasoning_8", |
| 410 | summary=[ |
| 411 | Summary( |
| 412 | text="Summary text", |
| 413 | type="summary_text", |
| 414 | ) |
| 415 | ], |
| 416 | type="reasoning", |
| 417 | content=[ |
| 418 | Content( |
| 419 | text="Content text", |
| 420 | type="reasoning_text", |
| 421 | ) |
| 422 | ], |
| 423 | encrypted_content=None, |
| 424 | status=None, |
| 425 | ) |
| 426 | construct_chat_messages_with_tool_call([item]) |
| 427 | mock_logger.warning.assert_not_called() |
| 428 | |
| 429 | |
| 430 | class TestShouldContinueFinalMessage: |
| 431 | """Tests for should_continue_final_message function. |
| 432 | |
| 433 | This function enables Anthropic-style partial message completion, where |
| 434 | users can provide an incomplete assistant message and have the model |
| 435 | continue from where it left off. |
| 436 | """ |
| 437 | |
| 438 | def test_string_input_returns_false(self): |
| 439 | """String input is always a user message, so should not continue.""" |
| 440 | assert should_continue_final_message("Hello, world!") is False |
| 441 | |
| 442 | def test_empty_list_returns_false(self): |
| 443 | """Empty list should not continue.""" |
| 444 | assert should_continue_final_message([]) is False |
| 445 | |
| 446 | def test_completed_message_returns_false(self): |
| 447 | """Completed message should not be continued.""" |
| 448 | output_item = ResponseOutputMessage( |
| 449 | id="msg_123", |
| 450 | content=[ |
| 451 | ResponseOutputText( |
| 452 | annotations=[], |
| 453 | text="The answer is 42.", |
| 454 | type="output_text", |
| 455 | logprobs=None, |
| 456 | ) |
| 457 | ], |
| 458 | role="assistant", |
| 459 | status="completed", |
| 460 | type="message", |
| 461 | ) |
| 462 | assert should_continue_final_message([output_item]) is False |
| 463 | |
| 464 | def test_in_progress_message_returns_true(self): |
| 465 | """In-progress message should be continued. |
| 466 | |
| 467 | This is the key use case for partial message completion. |
| 468 | Example: The user provides "The best answer is (" and wants |
| 469 | the model to continue from there. |
| 470 | """ |
| 471 | output_item = ResponseOutputMessage( |
| 472 | id="msg_123", |
| 473 | content=[ |
| 474 | ResponseOutputText( |
| 475 | annotations=[], |
| 476 | text="The best answer is (", |
| 477 | type="output_text", |
| 478 | logprobs=None, |
| 479 | ) |
| 480 | ], |
| 481 | role="assistant", |
| 482 | status="in_progress", |
| 483 | type="message", |
| 484 | ) |
| 485 | assert should_continue_final_message([output_item]) is True |
| 486 | |
| 487 | def test_incomplete_message_returns_true(self): |
| 488 | """Incomplete message should be continued.""" |
| 489 | output_item = ResponseOutputMessage( |
| 490 | id="msg_123", |
| 491 | content=[ |
| 492 | ResponseOutputText( |
| 493 | annotations=[], |
| 494 | text="The answer", |
| 495 | type="output_text", |
| 496 | logprobs=None, |
| 497 | ) |
| 498 | ], |
| 499 | role="assistant", |
| 500 | status="incomplete", |
| 501 | type="message", |
| 502 | ) |
| 503 | assert should_continue_final_message([output_item]) is True |
| 504 | |
| 505 | def test_in_progress_reasoning_returns_true(self): |
| 506 | """In-progress reasoning should be continued.""" |
| 507 | reasoning_item = ResponseReasoningItem( |
| 508 | id="reasoning_123", |
| 509 | summary=[], |
| 510 | type="reasoning", |
| 511 | content=[ |
| 512 | Content( |
| 513 | text="Let me think about this...", |
| 514 | type="reasoning_text", |
| 515 | ) |
| 516 | ], |
| 517 | encrypted_content=None, |
| 518 | status="in_progress", |
| 519 | ) |
| 520 | assert should_continue_final_message([reasoning_item]) is True |
| 521 | |
| 522 | def test_incomplete_reasoning_returns_true(self): |
| 523 | """Incomplete reasoning should be continued.""" |
| 524 | reasoning_item = ResponseReasoningItem( |
| 525 | id="reasoning_123", |
| 526 | summary=[], |
| 527 | type="reasoning", |
| 528 | content=[ |
| 529 | Content( |
| 530 | text="Let me think", |
| 531 | type="reasoning_text", |
| 532 | ) |
| 533 | ], |
| 534 | encrypted_content=None, |
| 535 | status="incomplete", |
| 536 | ) |
| 537 | assert should_continue_final_message([reasoning_item]) is True |
| 538 | |
| 539 | reasoning_item = { |
| 540 | "id": "reasoning_123", |
| 541 | "summary": [], |
| 542 | "type": "reasoning", |
| 543 | "content": [], |
| 544 | "status": "incomplete", |
| 545 | } |
| 546 | assert should_continue_final_message([reasoning_item]) is True |
| 547 | |
| 548 | def test_completed_reasoning_returns_false(self): |
| 549 | """Completed reasoning should not be continued.""" |
| 550 | reasoning_item = ResponseReasoningItem( |
| 551 | id="reasoning_123", |
| 552 | summary=[], |
| 553 | type="reasoning", |
| 554 | content=[ |
| 555 | Content( |
| 556 | text="I have thought about this.", |
| 557 | type="reasoning_text", |
| 558 | ) |
| 559 | ], |
| 560 | encrypted_content=None, |
| 561 | status="completed", |
| 562 | ) |
| 563 | assert should_continue_final_message([reasoning_item]) is False |
| 564 | |
| 565 | def test_reasoning_with_none_status_returns_false(self): |
| 566 | """Reasoning with None status should not be continued.""" |
| 567 | reasoning_item = ResponseReasoningItem( |
| 568 | id="reasoning_123", |
| 569 | summary=[], |
| 570 | type="reasoning", |
| 571 | content=[ |
| 572 | Content( |
| 573 | text="Some reasoning", |
| 574 | type="reasoning_text", |
| 575 | ) |
| 576 | ], |
| 577 | encrypted_content=None, |
| 578 | status=None, |
| 579 | ) |
| 580 | assert should_continue_final_message([reasoning_item]) is False |
| 581 | |
| 582 | def test_only_last_item_matters(self): |
| 583 | """Only the last item in the list determines continuation.""" |
| 584 | completed_item = ResponseOutputMessage( |
| 585 | id="msg_1", |
| 586 | content=[ |
| 587 | ResponseOutputText( |
| 588 | annotations=[], |
| 589 | text="Complete message.", |
| 590 | type="output_text", |
| 591 | logprobs=None, |
| 592 | ) |
| 593 | ], |
| 594 | role="assistant", |
| 595 | status="completed", |
| 596 | type="message", |
| 597 | ) |
| 598 | in_progress_item = ResponseOutputMessage( |
| 599 | id="msg_2", |
| 600 | content=[ |
| 601 | ResponseOutputText( |
| 602 | annotations=[], |
| 603 | text="Partial message...", |
| 604 | type="output_text", |
| 605 | logprobs=None, |
| 606 | ) |
| 607 | ], |
| 608 | role="assistant", |
| 609 | status="in_progress", |
| 610 | type="message", |
| 611 | ) |
| 612 | |
| 613 | # In-progress as last item -> should continue |
| 614 | assert should_continue_final_message([completed_item, in_progress_item]) is True |
| 615 | |
| 616 | # Completed as last item -> should not continue |
| 617 | assert ( |
| 618 | should_continue_final_message([in_progress_item, completed_item]) is False |
| 619 | ) |
| 620 | |
| 621 | def test_tool_call_returns_false(self): |
| 622 | """Tool calls should not trigger continuation.""" |
| 623 | tool_call = ResponseFunctionToolCall( |
| 624 | id="fc_123", |
| 625 | call_id="call_123", |
| 626 | type="function_call", |
| 627 | status="in_progress", |
| 628 | name="get_weather", |
| 629 | arguments='{"location": "NYC"}', |
| 630 | ) |
| 631 | assert should_continue_final_message([tool_call]) is False |
| 632 | |
| 633 | tool_call = { |
| 634 | "id": "msg_123", |
| 635 | "call_id": "call_123", |
| 636 | "type": "function_call", |
| 637 | "status": "in_progress", |
| 638 | "name": "get_weather", |
| 639 | "arguments": '{"location": "NYC"}', |
| 640 | } |
| 641 | assert should_continue_final_message([tool_call]) is False |
| 642 | |
| 643 | # Tests for dict inputs (e.g., from curl requests) |
| 644 | def test_dict_in_progress_message_returns_true(self): |
| 645 | """Dict with in_progress status should be continued (curl input).""" |
| 646 | dict_item = { |
| 647 | "id": "msg_123", |
| 648 | "type": "message", |
| 649 | "role": "assistant", |
| 650 | "status": "in_progress", |
| 651 | "content": [{"type": "output_text", "text": "The answer is ("}], |
| 652 | } |
| 653 | assert should_continue_final_message([dict_item]) is True |
| 654 | |
| 655 | def test_dict_incomplete_message_returns_true(self): |
| 656 | """Dict with incomplete status should be continued (curl input).""" |
| 657 | dict_item = { |
| 658 | "id": "msg_123", |
| 659 | "type": "message", |
| 660 | "role": "assistant", |
| 661 | "status": "incomplete", |
| 662 | "content": [{"type": "output_text", "text": "Partial answer"}], |
| 663 | } |
| 664 | assert should_continue_final_message([dict_item]) is True |
| 665 | |
| 666 | def test_dict_completed_message_returns_false(self): |
| 667 | """Dict with completed status should not be continued (curl input).""" |
| 668 | dict_item = { |
| 669 | "id": "msg_123", |
| 670 | "type": "message", |
| 671 | "role": "assistant", |
| 672 | "status": "completed", |
| 673 | "content": [{"type": "output_text", "text": "Complete answer."}], |
| 674 | } |
| 675 | assert should_continue_final_message([dict_item]) is False |
| 676 | |
| 677 | def test_dict_reasoning_in_progress_returns_true(self): |
| 678 | """Dict reasoning item with in_progress status should be continued.""" |
| 679 | dict_item = { |
| 680 | "id": "reasoning_123", |
| 681 | "type": "reasoning", |
| 682 | "status": "in_progress", |
| 683 | "content": [{"type": "reasoning_text", "text": "Let me think..."}], |
| 684 | } |
| 685 | assert should_continue_final_message([dict_item]) is True |
| 686 | |
| 687 | def test_dict_without_status_returns_false(self): |
| 688 | """Dict without status field should not be continued.""" |
| 689 | dict_item = { |
| 690 | "id": "msg_123", |
| 691 | "type": "message", |
| 692 | "role": "assistant", |
| 693 | "content": [{"type": "output_text", "text": "Some text"}], |
| 694 | } |
| 695 | assert should_continue_final_message([dict_item]) is False |
| 696 | |
| 697 | def test_dict_with_none_status_returns_false(self): |
| 698 | """Dict with None status should not be continued.""" |
| 699 | dict_item = { |
| 700 | "id": "msg_123", |
| 701 | "type": "message", |
| 702 | "role": "assistant", |
| 703 | "status": None, |
| 704 | "content": [{"type": "output_text", "text": "Some text"}], |
| 705 | } |
| 706 | assert should_continue_final_message([dict_item]) is False |
| 707 | |
| 708 | |
| 709 | class TestConstructChatMessagesCombinePolicy: |
| 710 | """Tests for contiguous assistant-side merging.""" |
| 711 | |
| 712 | @pytest.mark.parametrize( |
| 713 | ("items", "expected_content", "expected_reasoning", "expected_tool_call_ids"), |
| 714 | [ |
| 715 | pytest.param( |
| 716 | [ |
| 717 | make_reasoning_item(content_text="Let me think"), |
| 718 | make_output_message("Hello"), |
| 719 | ], |
| 720 | "Hello", |
| 721 | "Let me think", |
| 722 | None, |
| 723 | id="reasoning-output-messages", |
| 724 | ), |
| 725 | pytest.param( |
| 726 | [ |
| 727 | make_function_call(call_id="call_123"), |
| 728 | make_function_call(call_id="call_456"), |
| 729 | ], |
| 730 | None, |
| 731 | None, |
| 732 | ["call_123", "call_456"], |
| 733 | id="consecutive-tool-calls", |
| 734 | ), |
| 735 | pytest.param( |
| 736 | [ |
| 737 | make_reasoning_item(content_text="Let me think"), |
| 738 | make_function_call(call_id="call_123"), |
| 739 | ], |
| 740 | None, |
| 741 | "Let me think", |
| 742 | ["call_123"], |
| 743 | id="reasoning-tool-call", |
| 744 | ), |
| 745 | pytest.param( |
| 746 | [ |
| 747 | make_output_message("Hello"), |
| 748 | make_function_call(call_id="call_123"), |
| 749 | ], |
| 750 | "Hello", |
| 751 | None, |
| 752 | ["call_123"], |
| 753 | id="output-tool-call", |
| 754 | ), |
| 755 | pytest.param( |
| 756 | [ |
| 757 | make_reasoning_item(content_text="Thinking"), |
| 758 | make_output_message("Hello"), |
| 759 | make_function_call(call_id="call_123"), |
| 760 | make_function_call(call_id="call_456"), |
| 761 | ], |
| 762 | "Hello", |
| 763 | "Thinking", |
| 764 | ["call_123", "call_456"], |
| 765 | id="reasoning-output-tool-call", |
| 766 | ), |
| 767 | pytest.param( |
| 768 | [ |
| 769 | make_reasoning_item(content_text="Let me think"), |
| 770 | {"type": "message", "role": "assistant", "content": "Hello"}, |
| 771 | ], |
| 772 | "Hello", |
| 773 | "Let me think", |
| 774 | None, |
| 775 | id="reasoning-easyinput-assistant", |
| 776 | ), |
| 777 | ], |
| 778 | ) |
| 779 | def test_assistant_side_items_merge_until_tool_output( |
| 780 | self, |
| 781 | items, |
| 782 | expected_content, |
| 783 | expected_reasoning, |
| 784 | expected_tool_call_ids, |
| 785 | ): |
| 786 | messages = construct_chat_messages_with_tool_call(items) |
| 787 | |
| 788 | assert len(messages) == 1 |
| 789 | assert messages[0]["role"] == "assistant" |
| 790 | if expected_content is None: |
| 791 | assert "content" not in messages[0] |
| 792 | else: |
| 793 | assert messages[0]["content"] == expected_content |
| 794 | if expected_reasoning is None: |
| 795 | assert "reasoning" not in messages[0] |
| 796 | else: |
| 797 | assert messages[0]["reasoning"] == expected_reasoning |
| 798 | if expected_tool_call_ids is None: |
| 799 | assert "tool_calls" not in messages[0] |
| 800 | else: |
| 801 | assert [tool_call["id"] for tool_call in messages[0]["tool_calls"]] == ( |
| 802 | expected_tool_call_ids |
| 803 | ) |
| 804 | |
| 805 | @pytest.mark.parametrize( |
| 806 | ("items", "num_expected_messages"), |
| 807 | [ |
| 808 | pytest.param( |
| 809 | [ |
| 810 | make_output_message("Hello"), |
| 811 | make_output_message("World"), |
| 812 | ], |
| 813 | 2, |
| 814 | id="consecutive-output-messages", |
| 815 | ), |
| 816 | pytest.param( |
| 817 | [ |
| 818 | make_reasoning_item(content_text="Let me think"), |
| 819 | make_reasoning_item(content_text="Let me think more"), |
| 820 | ], |
| 821 | 2, |
| 822 | id="consecutive-reasoning-messages", |
| 823 | ), |
| 824 | pytest.param( |
| 825 | [ |
| 826 | make_function_call(call_id="call_123"), |
| 827 | make_function_call_output(call_id="call_123", output="42"), |
| 828 | make_function_call(call_id="call_456"), |
| 829 | ], |
| 830 | 3, |
| 831 | id="interrupted-by-non-assistant-item", |
| 832 | ), |
| 833 | ], |
| 834 | ) |
| 835 | def test_merge_chain_breaks(self, items, num_expected_messages): |
| 836 | messages = construct_chat_messages_with_tool_call(items) |
| 837 | assert len(messages) == num_expected_messages |
| 838 | |
| 839 | |
| 840 | class TestConstructInputMessagesInstructionsLeak: |
| 841 | """Regression tests for #37697: instructions from a prior response |
| 842 | should NOT leak through previous_response_id.""" |
| 843 | |
| 844 | def test_old_instructions_stripped_from_prev_msg(self): |
| 845 | """System message in prev_msg must be dropped so the new request's |
| 846 | instructions are the only system message in the conversation.""" |
| 847 | prev = [ |
| 848 | {"role": "system", "content": "old instructions"}, |
| 849 | {"role": "user", "content": "What is 2+2?"}, |
| 850 | {"role": "assistant", "content": "4"}, |
| 851 | ] |
| 852 | msgs = construct_input_messages( |
| 853 | request_instructions="new instructions", |
| 854 | request_input="What is 3+3?", |
| 855 | prev_msg=prev, |
| 856 | ) |
| 857 | system_msgs = [m for m in msgs if m.get("role") == "system"] |
| 858 | assert len(system_msgs) == 1 |
| 859 | assert system_msgs[0]["content"] == "new instructions" |
| 860 | |
| 861 | def test_no_instructions_in_new_request(self): |
| 862 | """If the new request has no instructions, old ones should still |
| 863 | be stripped -- they must not carry over.""" |
| 864 | prev = [ |
| 865 | {"role": "system", "content": "old instructions"}, |
| 866 | {"role": "user", "content": "Hi"}, |
| 867 | {"role": "assistant", "content": "Hello"}, |
| 868 | ] |
| 869 | msgs = construct_input_messages( |
| 870 | request_instructions=None, |
| 871 | request_input="What is 3+3?", |
| 872 | prev_msg=prev, |
| 873 | ) |
| 874 | system_msgs = [m for m in msgs if m.get("role") == "system"] |
| 875 | assert len(system_msgs) == 0 |
| 876 | |
| 877 | def test_non_system_messages_preserved(self): |
| 878 | """User/assistant messages from prev_msg must remain intact.""" |
| 879 | prev = [ |
| 880 | {"role": "system", "content": "old instructions"}, |
| 881 | {"role": "user", "content": "Hi"}, |
| 882 | {"role": "assistant", "content": "Hello"}, |
| 883 | ] |
| 884 | msgs = construct_input_messages( |
| 885 | request_instructions="new instructions", |
| 886 | request_input="Follow up", |
| 887 | prev_msg=prev, |
| 888 | ) |
| 889 | roles = [m["role"] for m in msgs] |
| 890 | assert roles == ["system", "user", "assistant", "user"] |
| 891 | assert msgs[0]["content"] == "new instructions" |
| 892 | assert msgs[1]["content"] == "Hi" |
| 893 | assert msgs[2]["content"] == "Hello" |
| 894 | assert msgs[3]["content"] == "Follow up" |
| 895 | |
| 896 | def test_no_prev_msg(self): |
| 897 | """Baseline: when there's no prev_msg, instructions work normally.""" |
| 898 | msgs = construct_input_messages( |
| 899 | request_instructions="be helpful", |
| 900 | request_input="hello", |
| 901 | prev_msg=None, |
| 902 | ) |
| 903 | assert len(msgs) == 2 |
| 904 | assert msgs[0] == {"role": "system", "content": "be helpful"} |
| 905 | assert msgs[1] == {"role": "user", "content": "hello"} |