| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | """Cross-API render parity tests. |
| 4 | |
| 5 | Verifies that the chat completion input path (parse_chat_input_to_harmony_message) |
| 6 | and the responses API input path (response_input_to_harmony) produce identical |
| 7 | Harmony messages and identical rendered token sequences when given equivalent |
| 8 | conversation representations. |
| 9 | |
| 10 | The chat completion API encodes reasoning and tool calls as fields on a single |
| 11 | assistant message dict; the responses API encodes them as separate typed items |
| 12 | in request.input. Both paths must converge on the same Harmony message list and |
| 13 | therefore the same rendered prompt. |
| 14 | |
| 15 | Each test: |
| 16 | 1. Builds Harmony messages from each path for a single message or sequence. |
| 17 | 2. Asserts message-level properties (role, channel, recipient, content) |
| 18 | using verify_harmony_messages. |
| 19 | 3. Asserts that render_for_completion produces identical token sequences. |
| 20 | """ |
| 21 | |
| 22 | from openai.types.responses import ResponseFunctionToolCall |
| 23 | from openai_harmony import Message, Role |
| 24 | |
| 25 | from tests.entrypoints.openai.utils import verify_harmony_messages |
| 26 | from vllm.entrypoints.openai.parser.harmony_utils import ( |
| 27 | get_encoding, |
| 28 | get_system_message, |
| 29 | parse_chat_input_to_harmony_message, |
| 30 | render_for_completion, |
| 31 | ) |
| 32 | from vllm.entrypoints.openai.responses.harmony import ( |
| 33 | response_input_to_harmony, |
| 34 | response_previous_input_to_harmony, |
| 35 | ) |
| 36 | |
| 37 | |
| 38 | def test_render_uses_legacy_tool_call_header(): |
| 39 | tool_call = ( |
| 40 | Message.from_role_and_content(Role.ASSISTANT, "print('hello')") |
| 41 | .with_channel("commentary") |
| 42 | .with_recipient("python") |
| 43 | .with_content_type("code") |
| 44 | ) |
| 45 | |
| 46 | rendered = get_encoding().decode(render_for_completion([tool_call])) |
| 47 | |
| 48 | assert ( |
| 49 | "<|start|>assistant to=python<|channel|>commentary code<|message|>" in rendered |
| 50 | ) |
| 51 | assert "commentary to=python <|constrain|>code" not in rendered |
| 52 | |
| 53 | |
| 54 | # Use a fixed date so the system message is deterministic across both paths. |
| 55 | _DATE = "2025-01-01" |
| 56 | |
| 57 | |
| 58 | def _system(): |
| 59 | return get_system_message(start_date=_DATE) |
| 60 | |
| 61 | |
| 62 | class TestResponseInputToHarmonyRenderParity: |
| 63 | """Each test drives the same conversation through both APIs and asserts |
| 64 | identical Harmony messages and rendered token sequences.""" |
| 65 | |
| 66 | # ----------------------------------------------------------------------- |
| 67 | # Single-message cases |
| 68 | # ----------------------------------------------------------------------- |
| 69 | |
| 70 | def test_developer_message(self): |
| 71 | """Both APIs must render developer messages identically using |
| 72 | DeveloperContent (with the '# Instructions' header).""" |
| 73 | chat_msgs = parse_chat_input_to_harmony_message( |
| 74 | {"role": "developer", "content": "Be concise."} |
| 75 | ) |
| 76 | resp_msgs = [ |
| 77 | response_input_to_harmony( |
| 78 | { |
| 79 | "type": "message", |
| 80 | "role": "developer", |
| 81 | "content": "Be concise.", |
| 82 | }, |
| 83 | prev_responses=[], |
| 84 | ) |
| 85 | ] |
| 86 | |
| 87 | expected = [{"role": "developer", "instructions": "Be concise."}] |
| 88 | verify_harmony_messages(chat_msgs, expected) |
| 89 | verify_harmony_messages(resp_msgs, expected) |
| 90 | |
| 91 | assert render_for_completion([_system()] + chat_msgs) == render_for_completion( |
| 92 | [_system()] + resp_msgs |
| 93 | ) |
| 94 | |
| 95 | def test_user_message(self): |
| 96 | chat_msgs = parse_chat_input_to_harmony_message( |
| 97 | {"role": "user", "content": "What's the weather in Paris?"} |
| 98 | ) |
| 99 | resp_msgs = [ |
| 100 | response_input_to_harmony( |
| 101 | { |
| 102 | "type": "message", |
| 103 | "role": "user", |
| 104 | "content": "What's the weather in Paris?", |
| 105 | }, |
| 106 | prev_responses=[], |
| 107 | ) |
| 108 | ] |
| 109 | |
| 110 | expected = [{"role": "user", "content": "What's the weather in Paris?"}] |
| 111 | verify_harmony_messages(chat_msgs, expected) |
| 112 | verify_harmony_messages(resp_msgs, expected) |
| 113 | |
| 114 | assert render_for_completion([_system()] + chat_msgs) == render_for_completion( |
| 115 | [_system()] + resp_msgs |
| 116 | ) |
| 117 | |
| 118 | def test_assistant_final_message(self): |
| 119 | chat_msgs = parse_chat_input_to_harmony_message( |
| 120 | {"role": "assistant", "content": "It is 18°C in Paris."} |
| 121 | ) |
| 122 | resp_msgs = [ |
| 123 | response_input_to_harmony( |
| 124 | { |
| 125 | "type": "message", |
| 126 | "role": "assistant", |
| 127 | "content": "It is 18°C in Paris.", |
| 128 | }, |
| 129 | prev_responses=[], |
| 130 | ) |
| 131 | ] |
| 132 | |
| 133 | expected = [ |
| 134 | {"role": "assistant", "channel": "final", "content": "It is 18°C in Paris."} |
| 135 | ] |
| 136 | verify_harmony_messages(chat_msgs, expected) |
| 137 | verify_harmony_messages(resp_msgs, expected) |
| 138 | |
| 139 | assert render_for_completion([_system()] + chat_msgs) == render_for_completion( |
| 140 | [_system()] + resp_msgs |
| 141 | ) |
| 142 | |
| 143 | def test_reasoning_item(self): |
| 144 | # Chat path: assistant message with only a reasoning field and no content. |
| 145 | chat_msgs = parse_chat_input_to_harmony_message( |
| 146 | { |
| 147 | "role": "assistant", |
| 148 | "reasoning": "I should call get_weather.", |
| 149 | "content": "", |
| 150 | } |
| 151 | ) |
| 152 | resp_msgs = [ |
| 153 | response_input_to_harmony( |
| 154 | { |
| 155 | "type": "reasoning", |
| 156 | "content": [ |
| 157 | {"type": "reasoning_text", "text": "I should call get_weather."} |
| 158 | ], |
| 159 | }, |
| 160 | prev_responses=[], |
| 161 | ) |
| 162 | ] |
| 163 | |
| 164 | expected = [ |
| 165 | { |
| 166 | "role": "assistant", |
| 167 | "channel": "analysis", |
| 168 | "content": "I should call get_weather.", |
| 169 | } |
| 170 | ] |
| 171 | verify_harmony_messages(chat_msgs, expected) |
| 172 | verify_harmony_messages(resp_msgs, expected) |
| 173 | |
| 174 | assert render_for_completion([_system()] + chat_msgs) == render_for_completion( |
| 175 | [_system()] + resp_msgs |
| 176 | ) |
| 177 | |
| 178 | def test_function_call(self): |
| 179 | chat_msgs = parse_chat_input_to_harmony_message( |
| 180 | { |
| 181 | "role": "assistant", |
| 182 | "tool_calls": [ |
| 183 | { |
| 184 | "id": "call_1", |
| 185 | "function": { |
| 186 | "name": "get_weather", |
| 187 | "arguments": '{"location": "Paris"}', |
| 188 | }, |
| 189 | } |
| 190 | ], |
| 191 | } |
| 192 | ) |
| 193 | resp_msgs = [ |
| 194 | response_input_to_harmony( |
| 195 | { |
| 196 | "type": "function_call", |
| 197 | "name": "get_weather", |
| 198 | "arguments": '{"location": "Paris"}', |
| 199 | }, |
| 200 | prev_responses=[], |
| 201 | ) |
| 202 | ] |
| 203 | |
| 204 | expected = [ |
| 205 | { |
| 206 | "role": "assistant", |
| 207 | "channel": "commentary", |
| 208 | "recipient": "functions.get_weather", |
| 209 | "content": '{"location": "Paris"}', |
| 210 | "content_type": "json", |
| 211 | } |
| 212 | ] |
| 213 | verify_harmony_messages(chat_msgs, expected) |
| 214 | verify_harmony_messages(resp_msgs, expected) |
| 215 | |
| 216 | assert render_for_completion([_system()] + chat_msgs) == render_for_completion( |
| 217 | [_system()] + resp_msgs |
| 218 | ) |
| 219 | |
| 220 | def test_tool_output(self): |
| 221 | prev_call = ResponseFunctionToolCall( |
| 222 | id="fc_1", |
| 223 | call_id="call_1", |
| 224 | name="get_weather", |
| 225 | arguments='{"location": "Paris"}', |
| 226 | type="function_call", |
| 227 | ) |
| 228 | |
| 229 | chat_msgs = parse_chat_input_to_harmony_message( |
| 230 | {"role": "tool", "tool_call_id": "call_1", "content": "18°C, clear skies."}, |
| 231 | tool_id_names={"call_1": "get_weather"}, |
| 232 | ) |
| 233 | resp_msgs = [ |
| 234 | response_input_to_harmony( |
| 235 | { |
| 236 | "type": "function_call_output", |
| 237 | "call_id": "call_1", |
| 238 | "output": "18°C, clear skies.", |
| 239 | }, |
| 240 | prev_responses=[prev_call], |
| 241 | ) |
| 242 | ] |
| 243 | |
| 244 | expected = [ |
| 245 | { |
| 246 | "role": "tool", |
| 247 | "author_name": "functions.get_weather", |
| 248 | "channel": "commentary", |
| 249 | "recipient": "assistant", |
| 250 | "content": "18°C, clear skies.", |
| 251 | } |
| 252 | ] |
| 253 | verify_harmony_messages(chat_msgs, expected) |
| 254 | verify_harmony_messages(resp_msgs, expected) |
| 255 | |
| 256 | assert render_for_completion([_system()] + chat_msgs) == render_for_completion( |
| 257 | [_system()] + resp_msgs |
| 258 | ) |
| 259 | |
| 260 | # ----------------------------------------------------------------------- |
| 261 | # Combined and multi-turn cases |
| 262 | # ----------------------------------------------------------------------- |
| 263 | |
| 264 | def test_reasoning_combined_with_function_call(self): |
| 265 | """Chat API packs reasoning + tool_calls into one dict; responses API |
| 266 | represents them as two separate items. Both must produce the same two |
| 267 | Harmony messages in the same order: analysis then commentary.""" |
| 268 | chat_msgs = parse_chat_input_to_harmony_message( |
| 269 | { |
| 270 | "role": "assistant", |
| 271 | "reasoning": "I should get the weather for Paris.", |
| 272 | "tool_calls": [ |
| 273 | { |
| 274 | "id": "call_1", |
| 275 | "function": { |
| 276 | "name": "get_weather", |
| 277 | "arguments": '{"location": "Paris"}', |
| 278 | }, |
| 279 | } |
| 280 | ], |
| 281 | } |
| 282 | ) |
| 283 | resp_msgs = [ |
| 284 | response_input_to_harmony( |
| 285 | { |
| 286 | "type": "reasoning", |
| 287 | "content": [ |
| 288 | { |
| 289 | "type": "reasoning_text", |
| 290 | "text": "I should get the weather for Paris.", |
| 291 | } |
| 292 | ], |
| 293 | }, |
| 294 | prev_responses=[], |
| 295 | ), |
| 296 | response_input_to_harmony( |
| 297 | { |
| 298 | "type": "function_call", |
| 299 | "name": "get_weather", |
| 300 | "arguments": '{"location": "Paris"}', |
| 301 | }, |
| 302 | prev_responses=[], |
| 303 | ), |
| 304 | ] |
| 305 | |
| 306 | expected = [ |
| 307 | { |
| 308 | "role": "assistant", |
| 309 | "channel": "analysis", |
| 310 | "content": "I should get the weather for Paris.", |
| 311 | }, |
| 312 | { |
| 313 | "role": "assistant", |
| 314 | "channel": "commentary", |
| 315 | "recipient": "functions.get_weather", |
| 316 | "content": '{"location": "Paris"}', |
| 317 | "content_type": "json", |
| 318 | }, |
| 319 | ] |
| 320 | verify_harmony_messages(chat_msgs, expected) |
| 321 | verify_harmony_messages(resp_msgs, expected) |
| 322 | |
| 323 | assert render_for_completion([_system()] + chat_msgs) == render_for_completion( |
| 324 | [_system()] + resp_msgs |
| 325 | ) |
| 326 | |
| 327 | def test_full_multi_turn_tool_call_conversation(self): |
| 328 | """Full conversation: user -> reasoning + tool_call -> tool_output -> final. |
| 329 | |
| 330 | Both APIs must render the complete conversation to identical token sequences. |
| 331 | This exercises the entire input pipeline including all message types and |
| 332 | the Rust harmony encoder. |
| 333 | """ |
| 334 | prev_call = ResponseFunctionToolCall( |
| 335 | id="fc_1", |
| 336 | call_id="call_1", |
| 337 | name="get_weather", |
| 338 | arguments='{"location": "Paris"}', |
| 339 | type="function_call", |
| 340 | ) |
| 341 | |
| 342 | # --- Chat completion API path --- |
| 343 | tool_id_names = {"call_1": "get_weather"} |
| 344 | chat_msgs = [] |
| 345 | chat_msgs += parse_chat_input_to_harmony_message( |
| 346 | {"role": "user", "content": "What's the weather in Paris?"} |
| 347 | ) |
| 348 | chat_msgs += parse_chat_input_to_harmony_message( |
| 349 | { |
| 350 | "role": "assistant", |
| 351 | "reasoning": "I should call get_weather for Paris.", |
| 352 | "tool_calls": [ |
| 353 | { |
| 354 | "id": "call_1", |
| 355 | "function": { |
| 356 | "name": "get_weather", |
| 357 | "arguments": '{"location": "Paris"}', |
| 358 | }, |
| 359 | } |
| 360 | ], |
| 361 | } |
| 362 | ) |
| 363 | chat_msgs += parse_chat_input_to_harmony_message( |
| 364 | {"role": "tool", "tool_call_id": "call_1", "content": "18°C, clear skies."}, |
| 365 | tool_id_names=tool_id_names, |
| 366 | ) |
| 367 | chat_msgs += parse_chat_input_to_harmony_message( |
| 368 | { |
| 369 | "role": "assistant", |
| 370 | "content": "It is currently 18°C in Paris with clear skies.", |
| 371 | } |
| 372 | ) |
| 373 | |
| 374 | # --- Responses API path --- |
| 375 | resp_input = [ |
| 376 | { |
| 377 | "type": "message", |
| 378 | "role": "user", |
| 379 | "content": "What's the weather in Paris?", |
| 380 | }, |
| 381 | { |
| 382 | "type": "reasoning", |
| 383 | "content": [ |
| 384 | { |
| 385 | "type": "reasoning_text", |
| 386 | "text": "I should call get_weather for Paris.", |
| 387 | } |
| 388 | ], |
| 389 | }, |
| 390 | { |
| 391 | "type": "function_call", |
| 392 | "name": "get_weather", |
| 393 | "arguments": '{"location": "Paris"}', |
| 394 | }, |
| 395 | { |
| 396 | "type": "function_call_output", |
| 397 | "call_id": "call_1", |
| 398 | "output": "18°C, clear skies.", |
| 399 | }, |
| 400 | { |
| 401 | "type": "message", |
| 402 | "role": "assistant", |
| 403 | "content": "It is currently 18°C in Paris with clear skies.", |
| 404 | }, |
| 405 | ] |
| 406 | resp_msgs = [ |
| 407 | response_input_to_harmony(item, prev_responses=[prev_call]) |
| 408 | for item in resp_input |
| 409 | ] |
| 410 | |
| 411 | assert render_for_completion([_system()] + chat_msgs) == render_for_completion( |
| 412 | [_system()] + resp_msgs |
| 413 | ) |
| 414 | |
| 415 | def test_multi_turn_two_tool_calls_with_reasoning_between(self): |
| 416 | """Validates parity for a chain of two tool calls, each with its own |
| 417 | reasoning trace. Reasoning traces in between commentary-channel tool |
| 418 | calls must survive as analysis-channel messages in both paths. |
| 419 | """ |
| 420 | first_reasoning = "I need current weather first." |
| 421 | second_reasoning = "Now I need the weekly forecast." |
| 422 | |
| 423 | prev_call_1 = ResponseFunctionToolCall( |
| 424 | id="fc_1", |
| 425 | call_id="call_1", |
| 426 | name="get_weather", |
| 427 | arguments='{"location": "Paris"}', |
| 428 | type="function_call", |
| 429 | ) |
| 430 | prev_call_2 = ResponseFunctionToolCall( |
| 431 | id="fc_2", |
| 432 | call_id="call_2", |
| 433 | name="get_forecast", |
| 434 | arguments='{"location": "Paris", "days": 7}', |
| 435 | type="function_call", |
| 436 | ) |
| 437 | |
| 438 | # --- Chat completion API path --- |
| 439 | tool_id_names = {"call_1": "get_weather", "call_2": "get_forecast"} |
| 440 | chat_msgs = [] |
| 441 | chat_msgs += parse_chat_input_to_harmony_message( |
| 442 | {"role": "user", "content": "What's the weather and forecast for Paris?"} |
| 443 | ) |
| 444 | # First reasoning + tool call |
| 445 | chat_msgs += parse_chat_input_to_harmony_message( |
| 446 | { |
| 447 | "role": "assistant", |
| 448 | "reasoning": first_reasoning, |
| 449 | "tool_calls": [ |
| 450 | { |
| 451 | "id": "call_1", |
| 452 | "function": { |
| 453 | "name": "get_weather", |
| 454 | "arguments": '{"location": "Paris"}', |
| 455 | }, |
| 456 | } |
| 457 | ], |
| 458 | } |
| 459 | ) |
| 460 | chat_msgs += parse_chat_input_to_harmony_message( |
| 461 | {"role": "tool", "tool_call_id": "call_1", "content": "18°C, clear skies."}, |
| 462 | tool_id_names=tool_id_names, |
| 463 | ) |
| 464 | # Second reasoning + tool call |
| 465 | chat_msgs += parse_chat_input_to_harmony_message( |
| 466 | { |
| 467 | "role": "assistant", |
| 468 | "reasoning": second_reasoning, |
| 469 | "tool_calls": [ |
| 470 | { |
| 471 | "id": "call_2", |
| 472 | "function": { |
| 473 | "name": "get_forecast", |
| 474 | "arguments": '{"location": "Paris", "days": 7}', |
| 475 | }, |
| 476 | } |
| 477 | ], |
| 478 | } |
| 479 | ) |
| 480 | chat_msgs += parse_chat_input_to_harmony_message( |
| 481 | { |
| 482 | "role": "tool", |
| 483 | "tool_call_id": "call_2", |
| 484 | "content": "Mon 17°C, Tue 19°C, Wed 16°C", |
| 485 | }, |
| 486 | tool_id_names=tool_id_names, |
| 487 | ) |
| 488 | |
| 489 | # --- Responses API path --- |
| 490 | prev_responses = [prev_call_1, prev_call_2] |
| 491 | resp_input = [ |
| 492 | { |
| 493 | "type": "message", |
| 494 | "role": "user", |
| 495 | "content": "What's the weather and forecast for Paris?", |
| 496 | }, |
| 497 | # First reasoning + tool call |
| 498 | { |
| 499 | "type": "reasoning", |
| 500 | "content": [{"type": "reasoning_text", "text": first_reasoning}], |
| 501 | }, |
| 502 | { |
| 503 | "type": "function_call", |
| 504 | "name": "get_weather", |
| 505 | "arguments": '{"location": "Paris"}', |
| 506 | }, |
| 507 | { |
| 508 | "type": "function_call_output", |
| 509 | "call_id": "call_1", |
| 510 | "output": "18°C, clear skies.", |
| 511 | }, |
| 512 | # Second reasoning + tool call |
| 513 | { |
| 514 | "type": "reasoning", |
| 515 | "content": [ |
| 516 | { |
| 517 | "type": "reasoning_text", |
| 518 | "text": second_reasoning, |
| 519 | } |
| 520 | ], |
| 521 | }, |
| 522 | { |
| 523 | "type": "function_call", |
| 524 | "name": "get_forecast", |
| 525 | "arguments": '{"location": "Paris", "days": 7}', |
| 526 | }, |
| 527 | { |
| 528 | "type": "function_call_output", |
| 529 | "call_id": "call_2", |
| 530 | "output": "Mon 17°C, Tue 19°C, Wed 16°C", |
| 531 | }, |
| 532 | ] |
| 533 | resp_msgs = [ |
| 534 | response_input_to_harmony(item, prev_responses=prev_responses) |
| 535 | for item in resp_input |
| 536 | ] |
| 537 | |
| 538 | chat_completion_tokens = render_for_completion([_system()] + chat_msgs) |
| 539 | responses_tokens = render_for_completion([_system()] + resp_msgs) |
| 540 | |
| 541 | assert chat_completion_tokens == responses_tokens |
| 542 | |
| 543 | rendered_prompt = get_encoding().decode(chat_completion_tokens) |
| 544 | assert first_reasoning in rendered_prompt |
| 545 | assert second_reasoning in rendered_prompt |
| 546 | |
| 547 | def test_completed_turns_drop_reasoning(self): |
| 548 | """Validates that reasoning from completed turns is dropped, while |
| 549 | reasoning from the current in-progress tool-call turn is preserved |
| 550 | in both chat completions and responses previous_input_messages.""" |
| 551 | first_turn_reasoning = "FIRST_TURN_REASONING" |
| 552 | second_turn_reasoning = "SECOND_TURN_REASONING" |
| 553 | |
| 554 | chat_completion_msgs = [] |
| 555 | for chat_message in [ |
| 556 | {"role": "user", "content": "What is 2+2?"}, |
| 557 | { |
| 558 | "role": "assistant", |
| 559 | "reasoning": first_turn_reasoning, |
| 560 | "content": "The answer is 4.", |
| 561 | }, |
| 562 | {"role": "user", "content": "Now what is 3+3?"}, |
| 563 | { |
| 564 | "role": "assistant", |
| 565 | "reasoning": second_turn_reasoning, |
| 566 | "tool_calls": [ |
| 567 | { |
| 568 | "id": "call_1", |
| 569 | "function": { |
| 570 | "name": "calc", |
| 571 | "arguments": '{"a":3,"b":3}', |
| 572 | }, |
| 573 | } |
| 574 | ], |
| 575 | }, |
| 576 | ]: |
| 577 | chat_completion_msgs.extend( |
| 578 | parse_chat_input_to_harmony_message(chat_message) |
| 579 | ) |
| 580 | |
| 581 | responses_prev_input_msgs = [] |
| 582 | for responses_message in [ |
| 583 | { |
| 584 | "author": {"role": "user"}, |
| 585 | "content": [{"type": "text", "text": "What is 2+2?"}], |
| 586 | }, |
| 587 | { |
| 588 | "author": {"role": "assistant"}, |
| 589 | "channel": "analysis", |
| 590 | "content": [{"type": "text", "text": first_turn_reasoning}], |
| 591 | }, |
| 592 | { |
| 593 | "author": {"role": "assistant"}, |
| 594 | "channel": "final", |
| 595 | "content": [{"type": "text", "text": "The answer is 4."}], |
| 596 | }, |
| 597 | { |
| 598 | "author": {"role": "user"}, |
| 599 | "content": [{"type": "text", "text": "Now what is 3+3?"}], |
| 600 | }, |
| 601 | { |
| 602 | "author": {"role": "assistant"}, |
| 603 | "channel": "analysis", |
| 604 | "content": [{"type": "text", "text": second_turn_reasoning}], |
| 605 | }, |
| 606 | { |
| 607 | "author": {"role": "assistant"}, |
| 608 | "channel": "commentary", |
| 609 | "recipient": "functions.calc", |
| 610 | "content_type": "json", |
| 611 | "content": [{"type": "text", "text": '{"a":3,"b":3}'}], |
| 612 | }, |
| 613 | ]: |
| 614 | responses_prev_input_msgs.extend( |
| 615 | response_previous_input_to_harmony(responses_message) |
| 616 | ) |
| 617 | |
| 618 | chat_completion_tokens = render_for_completion( |
| 619 | [_system()] + chat_completion_msgs |
| 620 | ) |
| 621 | responses_tokens = render_for_completion( |
| 622 | [_system()] + responses_prev_input_msgs |
| 623 | ) |
| 624 | |
| 625 | assert chat_completion_tokens == responses_tokens |
| 626 | |
| 627 | rendered_prompt = get_encoding().decode(responses_tokens) |
| 628 | assert first_turn_reasoning not in rendered_prompt |
| 629 | assert second_turn_reasoning in rendered_prompt |