| 1 | import json |
| 2 | import sys |
| 3 | from pathlib import Path |
| 4 | |
| 5 | import pytest |
| 6 | from langchain_core.messages import AIMessage, HumanMessage, SystemMessage |
| 7 | |
| 8 | |
| 9 | PROJECT_ROOT = Path(__file__).resolve().parents[1] |
| 10 | if str(PROJECT_ROOT) not in sys.path: |
| 11 | sys.path.insert(0, str(PROJECT_ROOT)) |
| 12 | |
| 13 | import models |
| 14 | from helpers import extract_tools |
| 15 | from helpers import litellm_transport |
| 16 | from helpers.dirty_json import DirtyJson |
| 17 | |
| 18 | |
| 19 | @pytest.fixture(autouse=True) |
| 20 | def _clear_transport_capability_cache(): |
| 21 | litellm_transport.clear_transport_capability_cache() |
| 22 | |
| 23 | |
| 24 | def _chunk(content: str) -> dict: |
| 25 | return {"choices": [{"delta": {"content": content}, "message": {}}]} |
| 26 | |
| 27 | |
| 28 | def _response_event(delta: str) -> dict: |
| 29 | return {"type": "response.output_text.delta", "delta": delta} |
| 30 | |
| 31 | |
| 32 | class _AsyncChunkStream: |
| 33 | def __init__(self, chunks: list[dict]): |
| 34 | self._chunks = chunks |
| 35 | self.index = 0 |
| 36 | self.closed = False |
| 37 | |
| 38 | def __aiter__(self): |
| 39 | return self |
| 40 | |
| 41 | async def __anext__(self): |
| 42 | if self.index >= len(self._chunks): |
| 43 | raise StopAsyncIteration |
| 44 | chunk = self._chunks[self.index] |
| 45 | self.index += 1 |
| 46 | return chunk |
| 47 | |
| 48 | async def aclose(self): |
| 49 | self.closed = True |
| 50 | |
| 51 | |
| 52 | class _FailingAsyncChunkStream: |
| 53 | def __init__(self, exc: Exception): |
| 54 | self.exc = exc |
| 55 | self.closed = False |
| 56 | |
| 57 | def __aiter__(self): |
| 58 | return self |
| 59 | |
| 60 | async def __anext__(self): |
| 61 | raise self.exc |
| 62 | |
| 63 | async def aclose(self): |
| 64 | self.closed = True |
| 65 | |
| 66 | |
| 67 | class _DumpOnly: |
| 68 | def __init__(self, **data): |
| 69 | self._data = data |
| 70 | |
| 71 | def model_dump(self): |
| 72 | return dict(self._data) |
| 73 | |
| 74 | |
| 75 | def test_extract_json_root_string_returns_canonical_snapshot(): |
| 76 | text = ( |
| 77 | 'prefix {"tool_name":"response","tool_args":{"text":"brace } inside"}} ' |
| 78 | "trailing noise" |
| 79 | ) |
| 80 | |
| 81 | root = extract_tools.extract_json_root_string(text) |
| 82 | |
| 83 | assert root == '{"tool_name":"response","tool_args":{"text":"brace } inside"}}' |
| 84 | assert extract_tools.json_parse_dirty(root)["tool_args"]["text"] == "brace } inside" |
| 85 | assert extract_tools.extract_json_root_string( |
| 86 | '{"tool_name":"response","tool_args":{"text":"missing"' |
| 87 | ) is None |
| 88 | assert extract_tools.extract_json_root_string('[{"tool_name":"response"}]') is None |
| 89 | |
| 90 | |
| 91 | @pytest.mark.parametrize( |
| 92 | "content", |
| 93 | [ |
| 94 | '{"tool_name":"response","tool_args":{"text":"partial"', |
| 95 | 'prefix {"tool_name":"response","tool_args":{}}', |
| 96 | '[{"tool_name":"response","tool_args":{}}]', |
| 97 | '```json\n{"tool_name":"response","tool_args":{}}\n```', |
| 98 | ], |
| 99 | ) |
| 100 | def test_extract_tool_request_skips_noncanonical_boundaries(monkeypatch, content): |
| 101 | monkeypatch.setattr( |
| 102 | extract_tools, |
| 103 | "extract_json_root_string", |
| 104 | lambda _content: pytest.fail("noncanonical content reached the root scanner"), |
| 105 | ) |
| 106 | |
| 107 | assert extract_tools.extract_tool_request(content) is None |
| 108 | |
| 109 | |
| 110 | def test_json_parse_dirty_prefers_valid_tool_request_after_preamble_object(): |
| 111 | text = ( |
| 112 | 'I will call the tool after this note {"note":"not the tool"}.\n' |
| 113 | '{"tool_name":"response","tool_args":{"text":"ok"}} trailing text' |
| 114 | ) |
| 115 | |
| 116 | assert extract_tools.json_parse_dirty(text) == { |
| 117 | "tool_name": "response", |
| 118 | "tool_args": {"text": "ok"}, |
| 119 | } |
| 120 | |
| 121 | |
| 122 | def test_extract_json_root_string_prefers_valid_tool_request(): |
| 123 | text = ( |
| 124 | 'I will call the tool after this note {"note":"not the tool"}.\n' |
| 125 | '{"tool_name":"response","tool_args":{"text":"ok"}} trailing text' |
| 126 | ) |
| 127 | |
| 128 | assert extract_tools.extract_json_root_string(text) == ( |
| 129 | '{"tool_name":"response","tool_args":{"text":"ok"}}' |
| 130 | ) |
| 131 | assert extract_tools.extract_json_root_string( |
| 132 | 'Only a note {"note":"not the tool"}' |
| 133 | ) == '{"note":"not the tool"}' |
| 134 | |
| 135 | |
| 136 | def test_extract_json_root_string_waits_for_complete_parallel_parent(): |
| 137 | partial = ( |
| 138 | '{"tool_name":"parallel","tool_args":{"tool_calls":[' |
| 139 | '{"tool_name":"code_execution_tool","tool_args":{"code":"first"}}' |
| 140 | ) |
| 141 | |
| 142 | assert extract_tools.extract_json_root_string(partial) is None |
| 143 | |
| 144 | full = ( |
| 145 | partial |
| 146 | + ',{"tool_name":"code_execution_tool","tool_args":{"code":"second"}}' |
| 147 | '],"wait":true}} trailing text' |
| 148 | ) |
| 149 | |
| 150 | root = extract_tools.extract_json_root_string(full) |
| 151 | assert root == ( |
| 152 | '{"tool_name":"parallel","tool_args":{"tool_calls":[' |
| 153 | '{"tool_name":"code_execution_tool","tool_args":{"code":"first"}},' |
| 154 | '{"tool_name":"code_execution_tool","tool_args":{"code":"second"}}' |
| 155 | '],"wait":true}}' |
| 156 | ) |
| 157 | parsed = extract_tools.json_parse_dirty(root) |
| 158 | assert parsed["tool_name"] == "parallel" |
| 159 | assert len(parsed["tool_args"]["tool_calls"]) == 2 |
| 160 | |
| 161 | |
| 162 | def test_litellm_global_kwargs_merge_defaults_and_config(monkeypatch): |
| 163 | monkeypatch.setattr( |
| 164 | models.settings, |
| 165 | "get_settings", |
| 166 | lambda: {"litellm_global_kwargs": {}}, |
| 167 | ) |
| 168 | |
| 169 | assert models._merge_litellm_call_kwargs({})["drop_params"] is True |
| 170 | assert models._merge_litellm_call_kwargs({"temperature": 0}) == { |
| 171 | "drop_params": True, |
| 172 | "temperature": 0, |
| 173 | } |
| 174 | |
| 175 | monkeypatch.setattr( |
| 176 | models.settings, |
| 177 | "get_settings", |
| 178 | lambda: { |
| 179 | "litellm_global_kwargs": { |
| 180 | "drop_params": "false", |
| 181 | "timeout": "30", |
| 182 | "additional_drop_params": ["response_format"], |
| 183 | } |
| 184 | }, |
| 185 | ) |
| 186 | |
| 187 | assert models._merge_litellm_call_kwargs({}) == { |
| 188 | "drop_params": False, |
| 189 | "timeout": 30, |
| 190 | "additional_drop_params": ["response_format"], |
| 191 | } |
| 192 | |
| 193 | original_drop_params = getattr(models.litellm, "drop_params", None) |
| 194 | had_timeout = hasattr(models.litellm, "timeout") |
| 195 | original_timeout = getattr(models.litellm, "timeout", None) |
| 196 | had_additional_drop_params = hasattr(models.litellm, "additional_drop_params") |
| 197 | original_additional_drop_params = getattr( |
| 198 | models.litellm, "additional_drop_params", None |
| 199 | ) |
| 200 | try: |
| 201 | assert models.set_litellm_params() == { |
| 202 | "drop_params": False, |
| 203 | "timeout": 30, |
| 204 | "additional_drop_params": ["response_format"], |
| 205 | } |
| 206 | assert models.litellm.drop_params is False |
| 207 | if had_timeout: |
| 208 | assert models.litellm.timeout == original_timeout |
| 209 | else: |
| 210 | assert not hasattr(models.litellm, "timeout") |
| 211 | if had_additional_drop_params: |
| 212 | assert ( |
| 213 | models.litellm.additional_drop_params |
| 214 | == original_additional_drop_params |
| 215 | ) |
| 216 | else: |
| 217 | assert not hasattr(models.litellm, "additional_drop_params") |
| 218 | finally: |
| 219 | setattr(models.litellm, "drop_params", original_drop_params) |
| 220 | if had_timeout: |
| 221 | setattr(models.litellm, "timeout", original_timeout) |
| 222 | elif hasattr(models.litellm, "timeout"): |
| 223 | delattr(models.litellm, "timeout") |
| 224 | if had_additional_drop_params: |
| 225 | setattr( |
| 226 | models.litellm, |
| 227 | "additional_drop_params", |
| 228 | original_additional_drop_params, |
| 229 | ) |
| 230 | elif hasattr(models.litellm, "additional_drop_params"): |
| 231 | delattr(models.litellm, "additional_drop_params") |
| 232 | |
| 233 | |
| 234 | def test_provider_defaults_do_not_freeze_litellm_global_kwargs(monkeypatch): |
| 235 | monkeypatch.setattr(models, "get_provider_config", lambda *args, **kwargs: None) |
| 236 | monkeypatch.setattr(models, "get_api_key", lambda *_args, **_kwargs: None) |
| 237 | monkeypatch.setattr( |
| 238 | models.settings, |
| 239 | "get_settings", |
| 240 | lambda: {"litellm_global_kwargs": {"drop_params": "true"}}, |
| 241 | ) |
| 242 | |
| 243 | _, provider_kwargs = models._merge_provider_defaults("chat", "openai", {}) |
| 244 | |
| 245 | assert "drop_params" not in provider_kwargs |
| 246 | assert models._merge_litellm_call_kwargs(provider_kwargs)["drop_params"] is True |
| 247 | |
| 248 | monkeypatch.setattr( |
| 249 | models.settings, |
| 250 | "get_settings", |
| 251 | lambda: {"litellm_global_kwargs": {"drop_params": "false"}}, |
| 252 | ) |
| 253 | |
| 254 | assert models._merge_litellm_call_kwargs(provider_kwargs)["drop_params"] is False |
| 255 | |
| 256 | |
| 257 | @pytest.mark.asyncio |
| 258 | async def test_unified_call_stops_chat_after_canonical_root_snapshot(monkeypatch): |
| 259 | stream = _AsyncChunkStream( |
| 260 | [ |
| 261 | _chunk('{"tool_name":"response","tool_args":{"text":"hello"}}'), |
| 262 | _chunk(" unreachable"), |
| 263 | ] |
| 264 | ) |
| 265 | |
| 266 | async def fake_acompletion(*args, **kwargs): |
| 267 | assert kwargs["stream"] is True |
| 268 | return stream |
| 269 | |
| 270 | async def fake_rate_limiter(*args, **kwargs): |
| 271 | return None |
| 272 | |
| 273 | monkeypatch.setattr(litellm_transport, "acompletion", fake_acompletion) |
| 274 | monkeypatch.setattr(models, "apply_rate_limiter", fake_rate_limiter) |
| 275 | monkeypatch.setattr( |
| 276 | models.settings, |
| 277 | "get_settings", |
| 278 | lambda: {"litellm_global_kwargs": {}}, |
| 279 | ) |
| 280 | |
| 281 | wrapper = models.LiteLLMChatWrapper( |
| 282 | model="test-model", |
| 283 | provider="openai", |
| 284 | model_config=None, |
| 285 | a0_api_mode="chat", |
| 286 | ) |
| 287 | |
| 288 | seen: list[tuple[str, str]] = [] |
| 289 | |
| 290 | async def response_callback(chunk: str, full: str): |
| 291 | seen.append((chunk, full)) |
| 292 | return full.strip() if extract_tools.extract_tool_request(full) else None |
| 293 | |
| 294 | response, reasoning = await wrapper.unified_call( |
| 295 | messages=[], |
| 296 | response_callback=response_callback, |
| 297 | ) |
| 298 | |
| 299 | assert response == '{"tool_name":"response","tool_args":{"text":"hello"}}' |
| 300 | assert reasoning == "" |
| 301 | assert stream.index == 1 |
| 302 | assert stream.closed is True |
| 303 | assert len(seen) == 1 |
| 304 | assert seen[0][1] == '{"tool_name":"response","tool_args":{"text":"hello"}}' |
| 305 | |
| 306 | |
| 307 | @pytest.mark.asyncio |
| 308 | async def test_unified_call_does_not_stop_for_embedded_tool_json(monkeypatch): |
| 309 | stream = _AsyncChunkStream( |
| 310 | [ |
| 311 | _chunk('Preamble {"note":"not the tool"}.\n'), |
| 312 | _chunk( |
| 313 | '{"tool_name":"response","tool_args":{"text":"ok"}} trailing text' |
| 314 | ), |
| 315 | _chunk(" unreachable"), |
| 316 | ] |
| 317 | ) |
| 318 | |
| 319 | async def fake_acompletion(*args, **kwargs): |
| 320 | assert kwargs["stream"] is True |
| 321 | return stream |
| 322 | |
| 323 | async def fake_rate_limiter(*args, **kwargs): |
| 324 | return None |
| 325 | |
| 326 | monkeypatch.setattr(litellm_transport, "acompletion", fake_acompletion) |
| 327 | monkeypatch.setattr(models, "apply_rate_limiter", fake_rate_limiter) |
| 328 | monkeypatch.setattr( |
| 329 | models.settings, |
| 330 | "get_settings", |
| 331 | lambda: {"litellm_global_kwargs": {}}, |
| 332 | ) |
| 333 | |
| 334 | wrapper = models.LiteLLMChatWrapper( |
| 335 | model="test-model", |
| 336 | provider="openai", |
| 337 | model_config=None, |
| 338 | a0_api_mode="chat", |
| 339 | ) |
| 340 | |
| 341 | seen: list[tuple[str, str]] = [] |
| 342 | |
| 343 | async def response_callback(chunk: str, full: str): |
| 344 | seen.append((chunk, full)) |
| 345 | return full.strip() if extract_tools.extract_tool_request(full) else None |
| 346 | |
| 347 | response, reasoning = await wrapper.unified_call( |
| 348 | messages=[], |
| 349 | response_callback=response_callback, |
| 350 | ) |
| 351 | |
| 352 | assert response == ( |
| 353 | 'Preamble {"note":"not the tool"}.\n' |
| 354 | '{"tool_name":"response","tool_args":{"text":"ok"}} trailing text unreachable' |
| 355 | ) |
| 356 | assert reasoning == "" |
| 357 | assert stream.index == 3 |
| 358 | assert stream.closed is False |
| 359 | assert len(seen) == 3 |
| 360 | assert seen[0][1] == 'Preamble {"note":"not the tool"}.\n' |
| 361 | assert ( |
| 362 | seen[1][1] |
| 363 | == 'Preamble {"note":"not the tool"}.\n' |
| 364 | '{"tool_name":"response","tool_args":{"text":"ok"}} trailing text' |
| 365 | ) |
| 366 | |
| 367 | |
| 368 | @pytest.mark.asyncio |
| 369 | async def test_unified_call_closes_responses_stream_when_callback_raises(monkeypatch): |
| 370 | stream = _AsyncChunkStream([_response_event("interrupt me")]) |
| 371 | |
| 372 | class ExpectedIntervention(Exception): |
| 373 | pass |
| 374 | |
| 375 | async def fake_aresponses(*args, **kwargs): |
| 376 | assert kwargs["stream"] is True |
| 377 | return stream |
| 378 | |
| 379 | async def fake_rate_limiter(*args, **kwargs): |
| 380 | return None |
| 381 | |
| 382 | monkeypatch.setattr(litellm_transport, "aresponses", fake_aresponses) |
| 383 | monkeypatch.setattr(models, "apply_rate_limiter", fake_rate_limiter) |
| 384 | |
| 385 | wrapper = models.LiteLLMChatWrapper( |
| 386 | model="test-model", |
| 387 | provider="openai", |
| 388 | model_config=None, |
| 389 | a0_api_mode="responses", |
| 390 | ) |
| 391 | |
| 392 | async def response_callback(chunk: str, full: str): |
| 393 | raise ExpectedIntervention() |
| 394 | |
| 395 | with pytest.raises(ExpectedIntervention): |
| 396 | await wrapper.unified_call( |
| 397 | messages=[], |
| 398 | response_callback=response_callback, |
| 399 | ) |
| 400 | |
| 401 | assert stream.closed is True |
| 402 | |
| 403 | |
| 404 | @pytest.mark.asyncio |
| 405 | async def test_chat_completions_default_uses_acompletion(monkeypatch): |
| 406 | stream = _AsyncChunkStream([_chunk("hello")]) |
| 407 | calls: list[str] = [] |
| 408 | |
| 409 | async def fake_acompletion(*args, **kwargs): |
| 410 | calls.append("chat") |
| 411 | assert kwargs["stream"] is True |
| 412 | assert "a0_api_mode" not in kwargs |
| 413 | return stream |
| 414 | |
| 415 | async def fake_aresponses(*args, **kwargs): |
| 416 | raise AssertionError("Responses path should not be used") |
| 417 | |
| 418 | async def fake_rate_limiter(*args, **kwargs): |
| 419 | return None |
| 420 | |
| 421 | monkeypatch.setattr(litellm_transport, "acompletion", fake_acompletion) |
| 422 | monkeypatch.setattr(litellm_transport, "aresponses", fake_aresponses) |
| 423 | monkeypatch.setattr(models, "apply_rate_limiter", fake_rate_limiter) |
| 424 | |
| 425 | wrapper = models.LiteLLMChatWrapper( |
| 426 | model="test-model", |
| 427 | provider="openai", |
| 428 | model_config=None, |
| 429 | ) |
| 430 | |
| 431 | async def response_callback(chunk: str, full: str): |
| 432 | return None |
| 433 | |
| 434 | response, reasoning = await wrapper.unified_call( |
| 435 | messages=[], |
| 436 | response_callback=response_callback, |
| 437 | ) |
| 438 | |
| 439 | assert response == "hello" |
| 440 | assert reasoning == "" |
| 441 | assert calls == ["chat"] |
| 442 | |
| 443 | |
| 444 | @pytest.mark.asyncio |
| 445 | async def test_unified_turn_stops_chat_stream_after_text_tool_request(monkeypatch): |
| 446 | message = ( |
| 447 | '{"thoughts":["test"],"actions":[' |
| 448 | '{"tool_name":"response","tool_args":{"text":"ok"}}]}' |
| 449 | ) |
| 450 | stream = _AsyncChunkStream([_chunk(message), _chunk(" unreachable")]) |
| 451 | |
| 452 | async def fake_acompletion(*args, **kwargs): |
| 453 | assert kwargs["stream"] is True |
| 454 | return stream |
| 455 | |
| 456 | async def fake_rate_limiter(*args, **kwargs): |
| 457 | return None |
| 458 | |
| 459 | monkeypatch.setattr(litellm_transport, "acompletion", fake_acompletion) |
| 460 | monkeypatch.setattr(models, "apply_rate_limiter", fake_rate_limiter) |
| 461 | |
| 462 | wrapper = models.LiteLLMChatWrapper( |
| 463 | model="test-model", |
| 464 | provider="openai", |
| 465 | model_config=None, |
| 466 | a0_api_mode="chat", |
| 467 | ) |
| 468 | |
| 469 | async def response_callback(chunk: str, full: str): |
| 470 | return full if extract_tools.extract_tool_request(full) else None |
| 471 | |
| 472 | result = await wrapper.unified_turn.__wrapped__( |
| 473 | wrapper, |
| 474 | messages=[], |
| 475 | response_callback=response_callback, |
| 476 | ) |
| 477 | |
| 478 | assert result.response == message |
| 479 | assert stream.index == 1 |
| 480 | assert stream.closed is True |
| 481 | |
| 482 | |
| 483 | @pytest.mark.asyncio |
| 484 | async def test_unified_call_retries_responses_with_high_reasoning(monkeypatch): |
| 485 | validation_error = ValueError( |
| 486 | "1 validation error for ResponseCreatedEvent\n" |
| 487 | "response.reasoning.effort\n" |
| 488 | "Input should be 'minimal', 'low', 'medium' or 'high' " |
| 489 | "[type=literal_error, input_value='none', input_type=str]" |
| 490 | ) |
| 491 | failing_stream = _FailingAsyncChunkStream(validation_error) |
| 492 | working_stream = _AsyncChunkStream([_response_event("ok")]) |
| 493 | calls: list[dict] = [] |
| 494 | |
| 495 | async def fake_aresponses(*args, **kwargs): |
| 496 | calls.append(kwargs) |
| 497 | if len(calls) == 1: |
| 498 | return failing_stream |
| 499 | return working_stream |
| 500 | |
| 501 | async def fake_rate_limiter(*args, **kwargs): |
| 502 | return None |
| 503 | |
| 504 | monkeypatch.setattr(litellm_transport, "aresponses", fake_aresponses) |
| 505 | monkeypatch.setattr(models, "apply_rate_limiter", fake_rate_limiter) |
| 506 | |
| 507 | wrapper = models.LiteLLMChatWrapper( |
| 508 | model="gpt-5.4", |
| 509 | provider="openai", |
| 510 | model_config=None, |
| 511 | a0_api_mode="responses", |
| 512 | ) |
| 513 | |
| 514 | async def response_callback(chunk: str, full: str): |
| 515 | return None |
| 516 | |
| 517 | response, reasoning = await wrapper.unified_call( |
| 518 | messages=[], |
| 519 | response_callback=response_callback, |
| 520 | ) |
| 521 | |
| 522 | assert response == "ok" |
| 523 | assert reasoning == "" |
| 524 | assert failing_stream.closed is True |
| 525 | assert len(calls) == 2 |
| 526 | assert "reasoning" not in calls[0] |
| 527 | assert calls[1]["reasoning"] == {"effort": "high"} |
| 528 | |
| 529 | |
| 530 | @pytest.mark.asyncio |
| 531 | async def test_unified_call_falls_back_to_chat_when_responses_endpoint_missing( |
| 532 | monkeypatch, |
| 533 | ): |
| 534 | calls: list[str] = [] |
| 535 | |
| 536 | async def fake_aresponses(*args, **kwargs): |
| 537 | calls.append("responses") |
| 538 | raise RuntimeError( |
| 539 | "Client error '404 Not Found' for url " |
| 540 | "'https://llm.agent-zero.ai/v1/responses'" |
| 541 | ) |
| 542 | |
| 543 | async def fake_acompletion(*args, **kwargs): |
| 544 | calls.append("chat") |
| 545 | assert kwargs["stream"] is True |
| 546 | assert kwargs["drop_params"] is True |
| 547 | assert "tool_choice" not in kwargs |
| 548 | assert "parallel_tool_calls" not in kwargs |
| 549 | return _AsyncChunkStream([_chunk("fallback")]) |
| 550 | |
| 551 | async def fake_rate_limiter(*args, **kwargs): |
| 552 | return None |
| 553 | |
| 554 | monkeypatch.setattr(litellm_transport, "aresponses", fake_aresponses) |
| 555 | monkeypatch.setattr(litellm_transport, "acompletion", fake_acompletion) |
| 556 | monkeypatch.setattr(models, "apply_rate_limiter", fake_rate_limiter) |
| 557 | |
| 558 | wrapper = models.LiteLLMChatWrapper( |
| 559 | model="claude-opus-4.7", |
| 560 | provider="openai", |
| 561 | model_config=None, |
| 562 | a0_api_mode="responses", |
| 563 | tool_choice="auto", |
| 564 | parallel_tool_calls=True, |
| 565 | ) |
| 566 | |
| 567 | async def response_callback(chunk: str, full: str): |
| 568 | return None |
| 569 | |
| 570 | response, reasoning = await wrapper.unified_call( |
| 571 | messages=[], |
| 572 | response_callback=response_callback, |
| 573 | ) |
| 574 | |
| 575 | assert response == "fallback" |
| 576 | assert reasoning == "" |
| 577 | assert calls == ["responses", "chat"] |
| 578 | |
| 579 | response, reasoning = await wrapper.unified_call( |
| 580 | messages=[], |
| 581 | response_callback=response_callback, |
| 582 | ) |
| 583 | |
| 584 | assert response == "fallback" |
| 585 | assert reasoning == "" |
| 586 | assert calls == ["responses", "chat", "chat"] |
| 587 | |
| 588 | |
| 589 | @pytest.mark.asyncio |
| 590 | async def test_unified_call_falls_back_when_litellm_hides_responses_404_url( |
| 591 | monkeypatch, |
| 592 | ): |
| 593 | class NotFoundError(Exception): |
| 594 | status_code = 404 |
| 595 | |
| 596 | calls: list[str] = [] |
| 597 | |
| 598 | async def fake_aresponses(*args, **kwargs): |
| 599 | calls.append("responses") |
| 600 | raise NotFoundError( |
| 601 | 'litellm.NotFoundError: NotFoundError: OpenAIException - {"detail":"Not Found"}' |
| 602 | ) |
| 603 | |
| 604 | async def fake_acompletion(*args, **kwargs): |
| 605 | calls.append("chat") |
| 606 | assert kwargs["stream"] is True |
| 607 | assert kwargs["drop_params"] is True |
| 608 | return _AsyncChunkStream([_chunk("fallback")]) |
| 609 | |
| 610 | async def fake_rate_limiter(*args, **kwargs): |
| 611 | return None |
| 612 | |
| 613 | monkeypatch.setattr(litellm_transport, "aresponses", fake_aresponses) |
| 614 | monkeypatch.setattr(litellm_transport, "acompletion", fake_acompletion) |
| 615 | monkeypatch.setattr(models, "apply_rate_limiter", fake_rate_limiter) |
| 616 | |
| 617 | wrapper = models.LiteLLMChatWrapper( |
| 618 | model="claude-opus-4.7", |
| 619 | provider="openai", |
| 620 | model_config=None, |
| 621 | a0_api_mode="responses", |
| 622 | ) |
| 623 | |
| 624 | async def response_callback(chunk: str, full: str): |
| 625 | return None |
| 626 | |
| 627 | response, reasoning = await wrapper.unified_call( |
| 628 | messages=[], |
| 629 | response_callback=response_callback, |
| 630 | ) |
| 631 | |
| 632 | assert response == "fallback" |
| 633 | assert reasoning == "" |
| 634 | assert calls == ["responses", "chat"] |
| 635 | |
| 636 | |
| 637 | @pytest.mark.parametrize( |
| 638 | "responses_error", |
| 639 | [ |
| 640 | "litellm.exceptions.APIError: Path /api/v1/responses is not " |
| 641 | "available through this proxy.", |
| 642 | "MaskedHTTPStatusError: Server error '500 Internal Server Error' " |
| 643 | "for url 'https://api.venice.ai/api/v1/responses'", |
| 644 | "InternalServerError: OpenAIException - '<=' not supported between " |
| 645 | "instances of 'str' and 'int' for url 'http://192.168.200.52:4000/responses'", |
| 646 | "ImportError Missing dependency No module named 'fastapi'. " |
| 647 | "Run `pip install 'litellm[proxy]'`", |
| 648 | ], |
| 649 | ) |
| 650 | @pytest.mark.asyncio |
| 651 | async def test_unified_call_falls_back_for_proxy_responses_failures( |
| 652 | monkeypatch, |
| 653 | responses_error, |
| 654 | ): |
| 655 | calls: list[str] = [] |
| 656 | |
| 657 | async def fake_aresponses(*args, **kwargs): |
| 658 | calls.append("responses") |
| 659 | raise RuntimeError(responses_error) |
| 660 | |
| 661 | async def fake_acompletion(*args, **kwargs): |
| 662 | calls.append("chat") |
| 663 | assert kwargs["stream"] is True |
| 664 | assert kwargs["drop_params"] is True |
| 665 | return _AsyncChunkStream([_chunk("fallback")]) |
| 666 | |
| 667 | async def fake_rate_limiter(*args, **kwargs): |
| 668 | return None |
| 669 | |
| 670 | monkeypatch.setattr(litellm_transport, "aresponses", fake_aresponses) |
| 671 | monkeypatch.setattr(litellm_transport, "acompletion", fake_acompletion) |
| 672 | monkeypatch.setattr(models, "apply_rate_limiter", fake_rate_limiter) |
| 673 | |
| 674 | wrapper = models.LiteLLMChatWrapper( |
| 675 | model="test-model", |
| 676 | provider="openai", |
| 677 | model_config=None, |
| 678 | a0_api_mode="responses", |
| 679 | ) |
| 680 | |
| 681 | async def response_callback(chunk: str, full: str): |
| 682 | return None |
| 683 | |
| 684 | response, reasoning = await wrapper.unified_call( |
| 685 | messages=[], |
| 686 | response_callback=response_callback, |
| 687 | ) |
| 688 | |
| 689 | assert response == "fallback" |
| 690 | assert reasoning == "" |
| 691 | assert calls == ["responses", "chat"] |
| 692 | |
| 693 | |
| 694 | @pytest.mark.asyncio |
| 695 | async def test_unified_call_falls_back_when_responses_mock_reads_sse_as_json( |
| 696 | monkeypatch, |
| 697 | ): |
| 698 | calls: list[str] = [] |
| 699 | sse_error = json.JSONDecodeError( |
| 700 | "Expecting value", |
| 701 | 'event: response.output_text.delta\ndata: {"delta":"hello"}\n\n', |
| 702 | 0, |
| 703 | ) |
| 704 | failing_stream = _FailingAsyncChunkStream(sse_error) |
| 705 | |
| 706 | async def fake_aresponses(*args, **kwargs): |
| 707 | calls.append("responses") |
| 708 | return failing_stream |
| 709 | |
| 710 | async def fake_acompletion(*args, **kwargs): |
| 711 | calls.append("chat") |
| 712 | assert kwargs["stream"] is True |
| 713 | assert kwargs["drop_params"] is True |
| 714 | return _AsyncChunkStream([_chunk("fallback")]) |
| 715 | |
| 716 | async def fake_rate_limiter(*args, **kwargs): |
| 717 | return None |
| 718 | |
| 719 | monkeypatch.setattr(litellm_transport, "aresponses", fake_aresponses) |
| 720 | monkeypatch.setattr(litellm_transport, "acompletion", fake_acompletion) |
| 721 | monkeypatch.setattr(models, "apply_rate_limiter", fake_rate_limiter) |
| 722 | |
| 723 | wrapper = models.LiteLLMChatWrapper( |
| 724 | model="omniroute/test-model", |
| 725 | provider="openai", |
| 726 | model_config=None, |
| 727 | a0_api_mode="responses", |
| 728 | ) |
| 729 | |
| 730 | async def response_callback(chunk: str, full: str): |
| 731 | return None |
| 732 | |
| 733 | response, reasoning = await wrapper.unified_call( |
| 734 | messages=[], |
| 735 | response_callback=response_callback, |
| 736 | ) |
| 737 | |
| 738 | assert response == "fallback" |
| 739 | assert reasoning == "" |
| 740 | assert calls == ["responses", "chat"] |
| 741 | assert failing_stream.closed is True |
| 742 | |
| 743 | |
| 744 | @pytest.mark.asyncio |
| 745 | async def test_unified_call_falls_back_when_responses_bad_request_rejects_shape( |
| 746 | monkeypatch, |
| 747 | ): |
| 748 | class BadRequestError(Exception): |
| 749 | status_code = 400 |
| 750 | |
| 751 | calls: list[str] = [] |
| 752 | |
| 753 | async def fake_aresponses(*args, **kwargs): |
| 754 | calls.append("responses") |
| 755 | raise BadRequestError( |
| 756 | 'BadRequestError: Zod validation error: input_image Expected object, ' |
| 757 | 'received string; Expected string, received array' |
| 758 | ) |
| 759 | |
| 760 | async def fake_acompletion(*args, **kwargs): |
| 761 | calls.append("chat") |
| 762 | assert kwargs["stream"] is True |
| 763 | assert kwargs["drop_params"] is True |
| 764 | return _AsyncChunkStream([_chunk("fallback")]) |
| 765 | |
| 766 | async def fake_rate_limiter(*args, **kwargs): |
| 767 | return None |
| 768 | |
| 769 | monkeypatch.setattr(litellm_transport, "aresponses", fake_aresponses) |
| 770 | monkeypatch.setattr(litellm_transport, "acompletion", fake_acompletion) |
| 771 | monkeypatch.setattr(models, "apply_rate_limiter", fake_rate_limiter) |
| 772 | |
| 773 | wrapper = models.LiteLLMChatWrapper( |
| 774 | model="venice-model", |
| 775 | provider="openai", |
| 776 | model_config=None, |
| 777 | a0_api_mode="responses", |
| 778 | ) |
| 779 | |
| 780 | async def response_callback(chunk: str, full: str): |
| 781 | return None |
| 782 | |
| 783 | response, reasoning = await wrapper.unified_call( |
| 784 | messages=[ |
| 785 | HumanMessage( |
| 786 | content=[ |
| 787 | {"type": "text", "text": "describe it"}, |
| 788 | { |
| 789 | "type": "image_url", |
| 790 | "image_url": {"url": "https://example.test/a.png"}, |
| 791 | }, |
| 792 | ] |
| 793 | ) |
| 794 | ], |
| 795 | response_callback=response_callback, |
| 796 | ) |
| 797 | |
| 798 | assert response == "fallback" |
| 799 | assert reasoning == "" |
| 800 | assert calls == ["responses", "chat"] |
| 801 | |
| 802 | |
| 803 | @pytest.mark.asyncio |
| 804 | async def test_unified_call_raises_generic_responses_bad_request(monkeypatch): |
| 805 | class BadRequestError(Exception): |
| 806 | status_code = 400 |
| 807 | |
| 808 | calls: list[str] = [] |
| 809 | |
| 810 | async def fake_aresponses(*args, **kwargs): |
| 811 | calls.append("responses") |
| 812 | raise BadRequestError( |
| 813 | "BadRequestError: validation error: invalid request: max_tokens is too high" |
| 814 | ) |
| 815 | |
| 816 | async def fake_acompletion(*args, **kwargs): |
| 817 | calls.append("chat") |
| 818 | raise AssertionError("generic 400 should not fallback to chat") |
| 819 | |
| 820 | async def fake_rate_limiter(*args, **kwargs): |
| 821 | return None |
| 822 | |
| 823 | monkeypatch.setattr(litellm_transport, "aresponses", fake_aresponses) |
| 824 | monkeypatch.setattr(litellm_transport, "acompletion", fake_acompletion) |
| 825 | monkeypatch.setattr(models, "apply_rate_limiter", fake_rate_limiter) |
| 826 | |
| 827 | wrapper = models.LiteLLMChatWrapper( |
| 828 | model="test-model", |
| 829 | provider="openai", |
| 830 | model_config=None, |
| 831 | a0_api_mode="responses", |
| 832 | ) |
| 833 | |
| 834 | async def response_callback(chunk: str, full: str): |
| 835 | return None |
| 836 | |
| 837 | with pytest.raises(BadRequestError): |
| 838 | await wrapper.unified_call( |
| 839 | messages=[], |
| 840 | response_callback=response_callback, |
| 841 | ) |
| 842 | |
| 843 | assert calls == ["responses"] |
| 844 | |
| 845 | |
| 846 | @pytest.mark.asyncio |
| 847 | async def test_unified_call_preserves_cache_control_with_chat_for_non_native_responses( |
| 848 | monkeypatch, |
| 849 | ): |
| 850 | calls: list[str] = [] |
| 851 | |
| 852 | async def fake_aresponses(*args, **kwargs): |
| 853 | raise AssertionError("cache_control should keep Anthropic-family calls on chat") |
| 854 | |
| 855 | async def fake_acompletion(*args, **kwargs): |
| 856 | calls.append("chat") |
| 857 | assert kwargs["stream"] is True |
| 858 | messages = kwargs["messages"] |
| 859 | assert "cache_control" not in messages[0] |
| 860 | assert messages[0]["content"][-1]["cache_control"] == { |
| 861 | "type": "ephemeral" |
| 862 | } |
| 863 | assert messages[1]["content"][-1]["cache_control"] == { |
| 864 | "type": "ephemeral" |
| 865 | } |
| 866 | assert "cache_control" not in messages[2] |
| 867 | assert messages[3]["content"][-1]["cache_control"] == { |
| 868 | "type": "ephemeral" |
| 869 | } |
| 870 | return _AsyncChunkStream([_chunk("cached")]) |
| 871 | |
| 872 | async def fake_rate_limiter(*args, **kwargs): |
| 873 | return None |
| 874 | |
| 875 | monkeypatch.setattr(litellm_transport, "aresponses", fake_aresponses) |
| 876 | monkeypatch.setattr(litellm_transport, "acompletion", fake_acompletion) |
| 877 | monkeypatch.setattr(models, "apply_rate_limiter", fake_rate_limiter) |
| 878 | |
| 879 | wrapper = models.LiteLLMChatWrapper( |
| 880 | model="claude-sonnet-4-5", |
| 881 | provider="anthropic", |
| 882 | model_config=None, |
| 883 | a0_api_mode="responses", |
| 884 | ) |
| 885 | |
| 886 | async def response_callback(chunk: str, full: str): |
| 887 | return None |
| 888 | |
| 889 | response, reasoning = await wrapper.unified_call( |
| 890 | messages=[ |
| 891 | SystemMessage(content="static instructions"), |
| 892 | HumanMessage(content="question"), |
| 893 | AIMessage(content="previous answer"), |
| 894 | HumanMessage(content="follow up"), |
| 895 | ], |
| 896 | response_callback=response_callback, |
| 897 | explicit_caching=True, |
| 898 | ) |
| 899 | |
| 900 | assert response == "cached" |
| 901 | assert reasoning == "" |
| 902 | assert calls == ["chat"] |
| 903 | |
| 904 | |
| 905 | def test_responses_request_translates_messages_and_params(): |
| 906 | messages = [ |
| 907 | {"role": "system", "content": "You are precise."}, |
| 908 | { |
| 909 | "role": "user", |
| 910 | "content": [ |
| 911 | {"type": "text", "text": "Inspect this."}, |
| 912 | { |
| 913 | "type": "image_url", |
| 914 | "image_url": {"url": "https://example.test/a.png"}, |
| 915 | }, |
| 916 | ], |
| 917 | }, |
| 918 | { |
| 919 | "role": "assistant", |
| 920 | "content": "empty", |
| 921 | "tool_calls": [ |
| 922 | { |
| 923 | "id": "call_1", |
| 924 | "type": "function", |
| 925 | "function": {"name": "lookup", "arguments": '{"q":"a0"}'}, |
| 926 | } |
| 927 | ], |
| 928 | }, |
| 929 | {"role": "tool", "tool_call_id": "call_1", "content": "done"}, |
| 930 | ] |
| 931 | kwargs = { |
| 932 | "max_tokens": 42, |
| 933 | "reasoning_effort": "high", |
| 934 | "response_format": { |
| 935 | "type": "json_schema", |
| 936 | "json_schema": { |
| 937 | "name": "answer", |
| 938 | "schema": {"type": "object"}, |
| 939 | "strict": True, |
| 940 | }, |
| 941 | }, |
| 942 | "tools": [ |
| 943 | { |
| 944 | "type": "function", |
| 945 | "function": { |
| 946 | "name": "lookup", |
| 947 | "description": "Search", |
| 948 | "parameters": {"type": "object"}, |
| 949 | "strict": True, |
| 950 | }, |
| 951 | } |
| 952 | ], |
| 953 | } |
| 954 | |
| 955 | request = litellm_transport.ResponsesTransport.from_chat(messages, kwargs) |
| 956 | |
| 957 | assert "instructions" not in request |
| 958 | assert request["store"] is True |
| 959 | assert request["max_output_tokens"] == 42 |
| 960 | assert request["reasoning"] == {"effort": "high"} |
| 961 | assert request["text"] == { |
| 962 | "format": { |
| 963 | "type": "json_schema", |
| 964 | "name": "answer", |
| 965 | "schema": {"type": "object"}, |
| 966 | "strict": True, |
| 967 | } |
| 968 | } |
| 969 | assert request["tools"] == [ |
| 970 | { |
| 971 | "type": "function", |
| 972 | "name": "lookup", |
| 973 | "description": "Search", |
| 974 | "parameters": {"type": "object", "properties": {}}, |
| 975 | "strict": True, |
| 976 | } |
| 977 | ] |
| 978 | assert request["input"] == [ |
| 979 | {"role": "system", "content": "You are precise."}, |
| 980 | { |
| 981 | "role": "user", |
| 982 | "content": [ |
| 983 | {"type": "input_text", "text": "Inspect this."}, |
| 984 | { |
| 985 | "type": "input_image", |
| 986 | "image_url": "https://example.test/a.png", |
| 987 | }, |
| 988 | ], |
| 989 | }, |
| 990 | { |
| 991 | "type": "function_call", |
| 992 | "call_id": "call_1", |
| 993 | "id": "call_1", |
| 994 | "name": "lookup", |
| 995 | "arguments": '{"q":"a0"}', |
| 996 | "status": "completed", |
| 997 | }, |
| 998 | {"type": "function_call_output", "call_id": "call_1", "output": "done"}, |
| 999 | ] |
| 1000 | |
| 1001 | |
| 1002 | def test_responses_request_normalizes_reasoning_and_orphan_tool_choice(): |
| 1003 | request = litellm_transport.ResponsesTransport.from_chat( |
| 1004 | [], |
| 1005 | { |
| 1006 | "reasoning_effort": "none", |
| 1007 | "tools": [], |
| 1008 | "tool_choice": "auto", |
| 1009 | "parallel_tool_calls": True, |
| 1010 | }, |
| 1011 | ) |
| 1012 | |
| 1013 | assert "reasoning" not in request |
| 1014 | assert "tools" not in request |
| 1015 | assert "tool_choice" not in request |
| 1016 | assert "parallel_tool_calls" not in request |
| 1017 | |
| 1018 | request = litellm_transport.ResponsesTransport.from_chat( |
| 1019 | [], |
| 1020 | {"reasoning": {"effort": "xhigh"}}, |
| 1021 | ) |
| 1022 | |
| 1023 | assert request["reasoning"] == {"effort": "high"} |
| 1024 | |
| 1025 | request = litellm_transport.ResponsesTransport.from_chat( |
| 1026 | [], |
| 1027 | {"reasoning_effort": "off"}, |
| 1028 | ) |
| 1029 | |
| 1030 | assert "reasoning" not in request |
| 1031 | |
| 1032 | |
| 1033 | def test_responses_request_normalizes_function_tool_parameter_shapes(): |
| 1034 | request = litellm_transport.ResponsesTransport.from_chat( |
| 1035 | [], |
| 1036 | { |
| 1037 | "functions": [ |
| 1038 | { |
| 1039 | "name": "legacy_noop", |
| 1040 | "description": "Legacy function", |
| 1041 | "parameters": {}, |
| 1042 | } |
| 1043 | ], |
| 1044 | }, |
| 1045 | ) |
| 1046 | |
| 1047 | assert request["tools"] == [ |
| 1048 | { |
| 1049 | "type": "function", |
| 1050 | "name": "legacy_noop", |
| 1051 | "description": "Legacy function", |
| 1052 | "parameters": { |
| 1053 | "type": "object", |
| 1054 | "properties": {}, |
| 1055 | }, |
| 1056 | } |
| 1057 | ] |
| 1058 | |
| 1059 | request = litellm_transport.ResponsesTransport.from_chat( |
| 1060 | [], |
| 1061 | { |
| 1062 | "a0_responses_function_tools": [ |
| 1063 | { |
| 1064 | "type": "function", |
| 1065 | "name": "native_noop", |
| 1066 | "description": "Native function", |
| 1067 | "parameters": {"type": "object"}, |
| 1068 | } |
| 1069 | ], |
| 1070 | "responses_builtin_tools": [{"type": "web_search"}], |
| 1071 | }, |
| 1072 | ) |
| 1073 | |
| 1074 | assert request["tools"] == [ |
| 1075 | { |
| 1076 | "type": "function", |
| 1077 | "name": "native_noop", |
| 1078 | "description": "Native function", |
| 1079 | "parameters": { |
| 1080 | "type": "object", |
| 1081 | "properties": {}, |
| 1082 | }, |
| 1083 | }, |
| 1084 | {"type": "web_search"}, |
| 1085 | ] |
| 1086 | assert request["tool_choice"] == "required" |
| 1087 | assert request["parallel_tool_calls"] is False |
| 1088 | |
| 1089 | |
| 1090 | def test_responses_request_preserves_explicit_a0_tool_controls(): |
| 1091 | request = litellm_transport.ResponsesTransport.from_chat( |
| 1092 | [], |
| 1093 | { |
| 1094 | "a0_responses_function_tools": [ |
| 1095 | { |
| 1096 | "type": "function", |
| 1097 | "name": "native_noop", |
| 1098 | "parameters": {"type": "object"}, |
| 1099 | } |
| 1100 | ], |
| 1101 | "tool_choice": "auto", |
| 1102 | "parallel_tool_calls": True, |
| 1103 | }, |
| 1104 | ) |
| 1105 | |
| 1106 | assert request["tool_choice"] == "auto" |
| 1107 | assert request["parallel_tool_calls"] is True |
| 1108 | |
| 1109 | |
| 1110 | def test_chat_completions_kwargs_omit_empty_tools(): |
| 1111 | kwargs = litellm_transport.ChatCompletionsTransport.prepare_kwargs( |
| 1112 | { |
| 1113 | "tools": [], |
| 1114 | "tool_choice": "auto", |
| 1115 | "parallel_tool_calls": True, |
| 1116 | "max_tokens": 8, |
| 1117 | } |
| 1118 | ) |
| 1119 | |
| 1120 | assert kwargs == {"max_tokens": 8} |
| 1121 | |
| 1122 | kwargs = litellm_transport.ChatCompletionsTransport.prepare_kwargs( |
| 1123 | { |
| 1124 | "tools": [ |
| 1125 | { |
| 1126 | "type": "function", |
| 1127 | "function": { |
| 1128 | "name": "lookup", |
| 1129 | "parameters": {"type": "object"}, |
| 1130 | }, |
| 1131 | } |
| 1132 | ], |
| 1133 | "tool_choice": "auto", |
| 1134 | } |
| 1135 | ) |
| 1136 | |
| 1137 | assert kwargs["tools"][0]["function"]["name"] == "lookup" |
| 1138 | assert kwargs["tool_choice"] == "auto" |
| 1139 | |
| 1140 | |
| 1141 | def test_complete_falls_back_to_chat_when_responses_shim_sends_empty_tools( |
| 1142 | monkeypatch, |
| 1143 | ): |
| 1144 | calls: list[str] = [] |
| 1145 | |
| 1146 | def fake_responses(*args, **kwargs): |
| 1147 | calls.append("responses") |
| 1148 | raise RuntimeError( |
| 1149 | "Value error, `tools` must not be an empty array. " |
| 1150 | "Either provide at least one tool or omit the field entirely." |
| 1151 | ) |
| 1152 | |
| 1153 | def fake_completion(*args, **kwargs): |
| 1154 | calls.append("chat") |
| 1155 | assert kwargs["drop_params"] is True |
| 1156 | assert "tools" not in kwargs |
| 1157 | assert "tool_choice" not in kwargs |
| 1158 | assert "parallel_tool_calls" not in kwargs |
| 1159 | return {"choices": [{"message": {"content": "ok"}}]} |
| 1160 | |
| 1161 | monkeypatch.setattr(litellm_transport, "responses", fake_responses) |
| 1162 | monkeypatch.setattr(litellm_transport, "completion", fake_completion) |
| 1163 | |
| 1164 | transport = litellm_transport.LiteLLMTransport( |
| 1165 | model="hosted_vllm/qwen", |
| 1166 | messages=[{"role": "user", "content": "hi"}], |
| 1167 | kwargs={ |
| 1168 | "a0_api_mode": "responses", |
| 1169 | "tools": [], |
| 1170 | "tool_choice": "auto", |
| 1171 | "parallel_tool_calls": True, |
| 1172 | "max_tokens": 8, |
| 1173 | }, |
| 1174 | ) |
| 1175 | |
| 1176 | parsed = transport.complete() |
| 1177 | |
| 1178 | assert parsed["response_delta"] == "ok" |
| 1179 | assert calls == ["responses", "chat"] |
| 1180 | |
| 1181 | |
| 1182 | def test_responses_request_adds_openai_prompt_cache_key_for_static_prefix(): |
| 1183 | request = litellm_transport.ResponsesTransport.from_chat( |
| 1184 | [ |
| 1185 | {"role": "system", "content": "stable system prompt"}, |
| 1186 | {"role": "user", "content": "dynamic question"}, |
| 1187 | ], |
| 1188 | { |
| 1189 | "tools": [ |
| 1190 | { |
| 1191 | "type": "function", |
| 1192 | "function": { |
| 1193 | "name": "lookup", |
| 1194 | "description": "Search", |
| 1195 | "parameters": {"type": "object"}, |
| 1196 | }, |
| 1197 | } |
| 1198 | ], |
| 1199 | }, |
| 1200 | model="openai/gpt-5.4", |
| 1201 | ) |
| 1202 | |
| 1203 | assert request["prompt_cache_key"].startswith("a0-") |
| 1204 | assert len(request["prompt_cache_key"]) == 35 |
| 1205 | assert "stable system prompt" not in request["prompt_cache_key"] |
| 1206 | |
| 1207 | request_again = litellm_transport.ResponsesTransport.from_chat( |
| 1208 | [ |
| 1209 | {"role": "system", "content": "stable system prompt"}, |
| 1210 | {"role": "user", "content": "different dynamic question"}, |
| 1211 | ], |
| 1212 | { |
| 1213 | "tools": [ |
| 1214 | { |
| 1215 | "type": "function", |
| 1216 | "function": { |
| 1217 | "name": "lookup", |
| 1218 | "description": "Search", |
| 1219 | "parameters": {"type": "object"}, |
| 1220 | }, |
| 1221 | } |
| 1222 | ], |
| 1223 | }, |
| 1224 | model="openai/gpt-5.4", |
| 1225 | ) |
| 1226 | |
| 1227 | assert request_again["prompt_cache_key"] == request["prompt_cache_key"] |
| 1228 | |
| 1229 | |
| 1230 | def test_responses_request_respects_explicit_prompt_cache_and_retention(): |
| 1231 | request = litellm_transport.ResponsesTransport.from_chat( |
| 1232 | [{"role": "system", "content": "stable system prompt"}], |
| 1233 | { |
| 1234 | "prompt_cache_key": "user-provided-key", |
| 1235 | "prompt_cache_retention": "24h", |
| 1236 | "extra_body": {"prompt_cache_retention": "in_memory"}, |
| 1237 | }, |
| 1238 | model="openai/gpt-5.4", |
| 1239 | ) |
| 1240 | |
| 1241 | assert request["prompt_cache_key"] == "user-provided-key" |
| 1242 | assert "prompt_cache_retention" not in request |
| 1243 | assert request["extra_body"]["prompt_cache_retention"] == "in_memory" |
| 1244 | |
| 1245 | |
| 1246 | def test_responses_request_adds_azure_prompt_cache_params(): |
| 1247 | request = litellm_transport.ResponsesTransport.from_chat( |
| 1248 | [{"role": "system", "content": "stable system prompt"}], |
| 1249 | {"prompt_cache_retention": "24h"}, |
| 1250 | model="azure/gpt-4.1", |
| 1251 | ) |
| 1252 | |
| 1253 | assert request["prompt_cache_key"].startswith("a0-") |
| 1254 | assert "prompt_cache_retention" not in request |
| 1255 | assert request["extra_body"]["prompt_cache_retention"] == "24h" |
| 1256 | |
| 1257 | |
| 1258 | def test_responses_request_does_not_add_openai_cache_key_to_custom_api_base(): |
| 1259 | request = litellm_transport.ResponsesTransport.from_chat( |
| 1260 | [{"role": "system", "content": "stable system prompt"}], |
| 1261 | {"api_base": "https://llm.agent-zero.ai/v1"}, |
| 1262 | model="openai/gpt-5.4", |
| 1263 | ) |
| 1264 | |
| 1265 | assert "prompt_cache_key" not in request |
| 1266 | |
| 1267 | |
| 1268 | def test_chat_kwargs_add_openai_prompt_cache_key_for_chat_completions(): |
| 1269 | kwargs = litellm_transport.ChatCompletionsTransport.prepare_kwargs( |
| 1270 | {"max_tokens": 10}, |
| 1271 | model="openai/gpt-5.4", |
| 1272 | messages=[ |
| 1273 | {"role": "system", "content": "stable system prompt"}, |
| 1274 | {"role": "user", "content": "dynamic question"}, |
| 1275 | ], |
| 1276 | ) |
| 1277 | |
| 1278 | assert kwargs["prompt_cache_key"].startswith("a0-") |
| 1279 | assert kwargs["max_tokens"] == 10 |
| 1280 | |
| 1281 | |
| 1282 | def test_chat_messages_strip_cache_control_for_openai_prompt_cache(): |
| 1283 | messages = [ |
| 1284 | { |
| 1285 | "role": "system", |
| 1286 | "cache_control": {"type": "ephemeral"}, |
| 1287 | "content": [ |
| 1288 | { |
| 1289 | "type": "text", |
| 1290 | "text": "stable system prompt", |
| 1291 | "cache_control": {"type": "ephemeral"}, |
| 1292 | } |
| 1293 | ], |
| 1294 | } |
| 1295 | ] |
| 1296 | |
| 1297 | prepared = litellm_transport.ChatCompletionsTransport.prepare_messages( |
| 1298 | messages, |
| 1299 | model="openai/gpt-5.4", |
| 1300 | kwargs={}, |
| 1301 | ) |
| 1302 | |
| 1303 | assert "cache_control" not in prepared[0] |
| 1304 | assert "cache_control" not in prepared[0]["content"][0] |
| 1305 | assert messages[0]["content"][0]["cache_control"] == {"type": "ephemeral"} |
| 1306 | |
| 1307 | |
| 1308 | def test_chat_kwargs_mark_cached_tools_for_cache_control_providers(): |
| 1309 | kwargs = litellm_transport.ChatCompletionsTransport.prepare_kwargs( |
| 1310 | { |
| 1311 | "tools": [ |
| 1312 | { |
| 1313 | "type": "function", |
| 1314 | "function": { |
| 1315 | "name": "lookup", |
| 1316 | "description": "Search", |
| 1317 | "parameters": {"type": "object"}, |
| 1318 | }, |
| 1319 | } |
| 1320 | ], |
| 1321 | }, |
| 1322 | model="anthropic/claude-sonnet-4-5", |
| 1323 | messages=[ |
| 1324 | { |
| 1325 | "role": "system", |
| 1326 | "content": [ |
| 1327 | { |
| 1328 | "type": "text", |
| 1329 | "text": "static instructions", |
| 1330 | "cache_control": {"type": "ephemeral"}, |
| 1331 | } |
| 1332 | ], |
| 1333 | } |
| 1334 | ], |
| 1335 | explicit_prompt_caching=True, |
| 1336 | ) |
| 1337 | |
| 1338 | assert kwargs["tools"][0]["function"]["cache_control"] == { |
| 1339 | "type": "ephemeral" |
| 1340 | } |
| 1341 | |
| 1342 | |
| 1343 | def test_chat_kwargs_strip_orphan_tool_choice_and_enable_fallback_drop_params(): |
| 1344 | kwargs = litellm_transport.ChatCompletionsTransport.prepare_kwargs( |
| 1345 | { |
| 1346 | "tool_choice": "auto", |
| 1347 | "parallel_tool_calls": True, |
| 1348 | "max_tokens": 10, |
| 1349 | }, |
| 1350 | fallback_error=RuntimeError("This model does not support Responses API"), |
| 1351 | ) |
| 1352 | |
| 1353 | assert kwargs["max_tokens"] == 10 |
| 1354 | assert kwargs["drop_params"] is True |
| 1355 | assert "tool_choice" not in kwargs |
| 1356 | assert "parallel_tool_calls" not in kwargs |
| 1357 | |
| 1358 | |
| 1359 | def test_cache_control_policy_keeps_native_responses_first(): |
| 1360 | messages = [ |
| 1361 | { |
| 1362 | "role": "system", |
| 1363 | "content": "static instructions", |
| 1364 | "cache_control": {"type": "ephemeral"}, |
| 1365 | } |
| 1366 | ] |
| 1367 | |
| 1368 | openai_policy = litellm_transport.TransportPolicy.from_request( |
| 1369 | "openai/gpt-5.4", |
| 1370 | {"a0_api_mode": "responses"}, |
| 1371 | messages=messages, |
| 1372 | ) |
| 1373 | anthropic_policy = litellm_transport.TransportPolicy.from_request( |
| 1374 | "anthropic/claude-sonnet-4-5", |
| 1375 | {"a0_api_mode": "responses"}, |
| 1376 | messages=messages, |
| 1377 | ) |
| 1378 | |
| 1379 | assert openai_policy.mode is litellm_transport.TransportMode.RESPONSES |
| 1380 | assert anthropic_policy.mode is litellm_transport.TransportMode.CHAT_COMPLETIONS |
| 1381 | |
| 1382 | |
| 1383 | def test_responses_fallback_does_not_mask_rate_limits(): |
| 1384 | exc = RuntimeError( |
| 1385 | "RateLimitError: 429 Too Many Requests for url " |
| 1386 | "https://provider.example/v1/responses" |
| 1387 | ) |
| 1388 | |
| 1389 | policy = litellm_transport.TransportPolicy( |
| 1390 | mode=litellm_transport.TransportMode.RESPONSES |
| 1391 | ) |
| 1392 | |
| 1393 | assert ( |
| 1394 | policy.recover(exc, got_any_chunk=False) |
| 1395 | is litellm_transport.TransportRecovery.RAISE |
| 1396 | ) |
| 1397 | |
| 1398 | |
| 1399 | def test_responses_fallback_on_untyped_input_item_rejection(): |
| 1400 | class BadRequestError(RuntimeError): |
| 1401 | status_code = 400 |
| 1402 | |
| 1403 | policy = litellm_transport.TransportPolicy( |
| 1404 | mode=litellm_transport.TransportMode.RESPONSES |
| 1405 | ) |
| 1406 | |
| 1407 | assert policy.recover( |
| 1408 | BadRequestError("Cannot determine type of item"), got_any_chunk=False |
| 1409 | ) is litellm_transport.TransportRecovery.FALLBACK_TO_CHAT |
| 1410 | assert policy.mode is litellm_transport.TransportMode.CHAT_COMPLETIONS |
| 1411 | |
| 1412 | |
| 1413 | def test_responses_response_parser_extracts_text_reasoning_and_function_calls(): |
| 1414 | text_response = { |
| 1415 | "output": [ |
| 1416 | {"type": "reasoning", "summary": [{"text": "because"}]}, |
| 1417 | { |
| 1418 | "type": "message", |
| 1419 | "content": [{"type": "output_text", "text": "answer"}], |
| 1420 | }, |
| 1421 | ] |
| 1422 | } |
| 1423 | |
| 1424 | parsed = litellm_transport.ResponsesTransport.parse_response(text_response) |
| 1425 | |
| 1426 | assert parsed == {"response_delta": "answer", "reasoning_delta": "because"} |
| 1427 | |
| 1428 | tool_response = { |
| 1429 | "output": [ |
| 1430 | { |
| 1431 | "type": "function_call", |
| 1432 | "name": "lookup", |
| 1433 | "arguments": '{"q":"a0"}', |
| 1434 | } |
| 1435 | ] |
| 1436 | } |
| 1437 | |
| 1438 | parsed_tool = litellm_transport.ResponsesTransport.parse_response(tool_response) |
| 1439 | |
| 1440 | assert extract_tools.json_parse_dirty(parsed_tool["response_delta"]) == { |
| 1441 | "tool_name": "lookup", |
| 1442 | "tool_args": {"q": "a0"}, |
| 1443 | } |
| 1444 | |
| 1445 | |
| 1446 | def test_chat_completions_response_parser_extracts_tool_calls(): |
| 1447 | parsed = litellm_transport.ChatCompletionsTransport.parse( |
| 1448 | { |
| 1449 | "choices": [ |
| 1450 | { |
| 1451 | "message": { |
| 1452 | "tool_calls": [ |
| 1453 | { |
| 1454 | "id": "call_1", |
| 1455 | "type": "function", |
| 1456 | "function": { |
| 1457 | "name": "lookup", |
| 1458 | "arguments": '{"q":"a0"}', |
| 1459 | }, |
| 1460 | } |
| 1461 | ] |
| 1462 | } |
| 1463 | } |
| 1464 | ] |
| 1465 | } |
| 1466 | ) |
| 1467 | |
| 1468 | assert extract_tools.json_parse_dirty(parsed["response_delta"]) == { |
| 1469 | "tool_name": "lookup", |
| 1470 | "tool_args": {"q": "a0"}, |
| 1471 | } |
| 1472 | assert parsed["_output_items"][0]["name"] == "lookup" |
| 1473 | |
| 1474 | |
| 1475 | def test_chat_completions_stream_parser_accumulates_tool_call_arguments(): |
| 1476 | parser = litellm_transport.ChatCompletionsStreamParser() |
| 1477 | |
| 1478 | assert parser.parse( |
| 1479 | { |
| 1480 | "choices": [ |
| 1481 | { |
| 1482 | "delta": { |
| 1483 | "tool_calls": [ |
| 1484 | { |
| 1485 | "index": 0, |
| 1486 | "id": "call_1", |
| 1487 | "type": "function", |
| 1488 | "function": { |
| 1489 | "name": "lookup", |
| 1490 | "arguments": '{"q":', |
| 1491 | }, |
| 1492 | } |
| 1493 | ] |
| 1494 | } |
| 1495 | } |
| 1496 | ] |
| 1497 | } |
| 1498 | ) == {"reasoning_delta": "", "response_delta": ""} |
| 1499 | parsed = parser.parse( |
| 1500 | { |
| 1501 | "choices": [ |
| 1502 | { |
| 1503 | "delta": { |
| 1504 | "tool_calls": [ |
| 1505 | { |
| 1506 | "index": 0, |
| 1507 | "function": {"arguments": '"a0"}'}, |
| 1508 | } |
| 1509 | ] |
| 1510 | }, |
| 1511 | "finish_reason": "tool_calls", |
| 1512 | } |
| 1513 | ] |
| 1514 | } |
| 1515 | ) |
| 1516 | |
| 1517 | assert extract_tools.json_parse_dirty(parsed["response_delta"]) == { |
| 1518 | "tool_name": "lookup", |
| 1519 | "tool_args": {"q": "a0"}, |
| 1520 | } |
| 1521 | assert parser.output_items()[0]["name"] == "lookup" |
| 1522 | assert parser.flush() == {"reasoning_delta": "", "response_delta": ""} |
| 1523 | |
| 1524 | |
| 1525 | def test_chat_completions_stream_parser_reads_dumped_tool_calls(): |
| 1526 | parser = litellm_transport.ChatCompletionsStreamParser() |
| 1527 | |
| 1528 | assert parser.parse( |
| 1529 | _DumpOnly( |
| 1530 | choices=[ |
| 1531 | _DumpOnly( |
| 1532 | delta=_DumpOnly( |
| 1533 | tool_calls=[ |
| 1534 | { |
| 1535 | "index": 0, |
| 1536 | "id": "call_1", |
| 1537 | "type": "function", |
| 1538 | "function": _DumpOnly( |
| 1539 | name="lookup", |
| 1540 | arguments='{"q":"a0"}', |
| 1541 | ), |
| 1542 | } |
| 1543 | ] |
| 1544 | ) |
| 1545 | ) |
| 1546 | ] |
| 1547 | ) |
| 1548 | ) == {"reasoning_delta": "", "response_delta": ""} |
| 1549 | |
| 1550 | parsed = parser.parse( |
| 1551 | _DumpOnly(choices=[_DumpOnly(delta=_DumpOnly(), finish_reason="tool_calls")]) |
| 1552 | ) |
| 1553 | |
| 1554 | assert extract_tools.json_parse_dirty(parsed["response_delta"]) == { |
| 1555 | "tool_name": "lookup", |
| 1556 | "tool_args": {"q": "a0"}, |
| 1557 | } |
| 1558 | |
| 1559 | |
| 1560 | def test_chat_completions_stream_parser_preserves_optional_usage(): |
| 1561 | parser = litellm_transport.ChatCompletionsStreamParser() |
| 1562 | parser.parse( |
| 1563 | { |
| 1564 | "choices": [], |
| 1565 | "usage": {"prompt_tokens": 240}, |
| 1566 | "_hidden_params": {"response_cost": 0.0084}, |
| 1567 | } |
| 1568 | ) |
| 1569 | parser.parse( |
| 1570 | { |
| 1571 | "choices": [], |
| 1572 | "usage": {"completion_tokens": 16, "total_tokens": 256}, |
| 1573 | } |
| 1574 | ) |
| 1575 | transport = litellm_transport.LiteLLMTransport( |
| 1576 | model="custom/model", |
| 1577 | messages=[{"role": "user", "content": "question"}], |
| 1578 | kwargs={"a0_api_mode": "chat_completions"}, |
| 1579 | ) |
| 1580 | |
| 1581 | result = transport._stream_result_from_chat_parser(parser) |
| 1582 | |
| 1583 | assert result is not None |
| 1584 | assert result.usage == { |
| 1585 | "prompt_tokens": 240, |
| 1586 | "completion_tokens": 16, |
| 1587 | "total_tokens": 256, |
| 1588 | "cost": 0.0084, |
| 1589 | } |
| 1590 | |
| 1591 | |
| 1592 | @pytest.mark.asyncio |
| 1593 | async def test_unified_turn_preserves_chat_streaming_tool_calls(monkeypatch): |
| 1594 | async def fake_acompletion(*args, **kwargs): |
| 1595 | return _AsyncChunkStream( |
| 1596 | [ |
| 1597 | { |
| 1598 | "choices": [ |
| 1599 | { |
| 1600 | "delta": { |
| 1601 | "tool_calls": [ |
| 1602 | { |
| 1603 | "index": 0, |
| 1604 | "id": "call_1", |
| 1605 | "type": "function", |
| 1606 | "function": { |
| 1607 | "name": "lookup", |
| 1608 | "arguments": '{"q":', |
| 1609 | }, |
| 1610 | } |
| 1611 | ] |
| 1612 | } |
| 1613 | } |
| 1614 | ] |
| 1615 | }, |
| 1616 | { |
| 1617 | "choices": [ |
| 1618 | { |
| 1619 | "delta": { |
| 1620 | "tool_calls": [ |
| 1621 | { |
| 1622 | "index": 0, |
| 1623 | "function": {"arguments": '"a0"}'}, |
| 1624 | } |
| 1625 | ] |
| 1626 | }, |
| 1627 | "finish_reason": "tool_calls", |
| 1628 | } |
| 1629 | ] |
| 1630 | }, |
| 1631 | ] |
| 1632 | ) |
| 1633 | |
| 1634 | async def fake_rate_limiter(*args, **kwargs): |
| 1635 | return None |
| 1636 | |
| 1637 | monkeypatch.setattr(litellm_transport, "acompletion", fake_acompletion) |
| 1638 | monkeypatch.setattr(models, "apply_rate_limiter", fake_rate_limiter) |
| 1639 | |
| 1640 | wrapper = models.LiteLLMChatWrapper( |
| 1641 | model="test-model", |
| 1642 | provider="openai", |
| 1643 | model_config=None, |
| 1644 | ) |
| 1645 | |
| 1646 | async def response_callback(chunk: str, full: str): |
| 1647 | return None |
| 1648 | |
| 1649 | result = await wrapper.unified_turn( |
| 1650 | messages=[], |
| 1651 | response_callback=response_callback, |
| 1652 | a0_api_mode="chat", |
| 1653 | ) |
| 1654 | |
| 1655 | assert extract_tools.json_parse_dirty(result.response) == { |
| 1656 | "tool_name": "lookup", |
| 1657 | "tool_args": {"q": "a0"}, |
| 1658 | } |
| 1659 | assert result.function_calls[0].name == "lookup" |
| 1660 | assert result.function_calls[0].arguments == {"q": "a0"} |
| 1661 | |
| 1662 | |
| 1663 | def test_responses_stream_parser_accumulates_function_call_arguments(): |
| 1664 | parser = litellm_transport.ResponsesEventParser() |
| 1665 | |
| 1666 | assert parser.parse( |
| 1667 | { |
| 1668 | "type": "response.output_item.added", |
| 1669 | "output_index": 0, |
| 1670 | "item": { |
| 1671 | "type": "function_call", |
| 1672 | "id": "fc_1", |
| 1673 | "call_id": "call_1", |
| 1674 | "name": "lookup", |
| 1675 | "arguments": "", |
| 1676 | }, |
| 1677 | } |
| 1678 | ) == {"reasoning_delta": "", "response_delta": ""} |
| 1679 | assert parser.parse( |
| 1680 | { |
| 1681 | "type": "response.function_call_arguments.delta", |
| 1682 | "item_id": "fc_1", |
| 1683 | "output_index": 0, |
| 1684 | "delta": '{"q":', |
| 1685 | } |
| 1686 | ) == {"reasoning_delta": "", "response_delta": ""} |
| 1687 | |
| 1688 | parsed = parser.parse( |
| 1689 | { |
| 1690 | "type": "response.function_call_arguments.done", |
| 1691 | "item_id": "fc_1", |
| 1692 | "output_index": 0, |
| 1693 | "name": "lookup", |
| 1694 | "arguments": '{"q":"a0"}', |
| 1695 | } |
| 1696 | ) |
| 1697 | |
| 1698 | assert extract_tools.json_parse_dirty(parsed["response_delta"]) == { |
| 1699 | "tool_name": "lookup", |
| 1700 | "tool_args": {"q": "a0"}, |
| 1701 | } |
| 1702 | assert parser.parse( |
| 1703 | { |
| 1704 | "type": "response.output_item.done", |
| 1705 | "output_index": 0, |
| 1706 | "item": { |
| 1707 | "type": "function_call", |
| 1708 | "id": "fc_1", |
| 1709 | "call_id": "call_1", |
| 1710 | "name": "lookup", |
| 1711 | "arguments": '{"q":"a0"}', |
| 1712 | }, |
| 1713 | } |
| 1714 | ) == {"reasoning_delta": "", "response_delta": ""} |
| 1715 | |
| 1716 | |
| 1717 | def test_responses_stream_parser_streams_response_function_arguments(): |
| 1718 | parser = litellm_transport.ResponsesEventParser() |
| 1719 | |
| 1720 | parser.parse( |
| 1721 | { |
| 1722 | "type": "response.output_item.added", |
| 1723 | "output_index": 0, |
| 1724 | "item": { |
| 1725 | "type": "function_call", |
| 1726 | "id": "fc_1", |
| 1727 | "name": "response", |
| 1728 | "arguments": "", |
| 1729 | }, |
| 1730 | } |
| 1731 | ) |
| 1732 | chunks = [ |
| 1733 | parser.parse( |
| 1734 | { |
| 1735 | "type": "response.function_call_arguments.delta", |
| 1736 | "item_id": "fc_1", |
| 1737 | "delta": '{"text":"Hello', |
| 1738 | } |
| 1739 | )["response_delta"], |
| 1740 | parser.parse( |
| 1741 | { |
| 1742 | "type": "response.function_call_arguments.delta", |
| 1743 | "item_id": "fc_1", |
| 1744 | "delta": ' world"}', |
| 1745 | } |
| 1746 | )["response_delta"], |
| 1747 | parser.parse( |
| 1748 | { |
| 1749 | "type": "response.function_call_arguments.done", |
| 1750 | "item_id": "fc_1", |
| 1751 | "name": "response", |
| 1752 | "arguments": '{"text":"Hello world"}', |
| 1753 | } |
| 1754 | )["response_delta"], |
| 1755 | ] |
| 1756 | |
| 1757 | assert chunks[0] == '{"tool_name":"response","tool_args":{"text":"Hello' |
| 1758 | assert DirtyJson.parse_string(chunks[0]) == { |
| 1759 | "tool_name": "response", |
| 1760 | "tool_args": {"text": "Hello"}, |
| 1761 | } |
| 1762 | assert extract_tools.json_parse_dirty("".join(chunks)) == { |
| 1763 | "tool_name": "response", |
| 1764 | "tool_args": {"text": "Hello world"}, |
| 1765 | } |
| 1766 | assert chunks[-1] == "}" |
| 1767 | assert parser.parse( |
| 1768 | { |
| 1769 | "type": "response.output_item.done", |
| 1770 | "item": { |
| 1771 | "type": "function_call", |
| 1772 | "id": "fc_1", |
| 1773 | "name": "response", |
| 1774 | "arguments": '{"text":"Hello world"}', |
| 1775 | }, |
| 1776 | } |
| 1777 | ) == {"reasoning_delta": "", "response_delta": ""} |
| 1778 | |
| 1779 | |
| 1780 | def test_responses_stream_parser_uses_completed_response_when_no_deltas_arrive(): |
| 1781 | parser = litellm_transport.ResponsesEventParser() |
| 1782 | |
| 1783 | parsed = parser.parse( |
| 1784 | { |
| 1785 | "type": "response.completed", |
| 1786 | "response": { |
| 1787 | "output": [ |
| 1788 | { |
| 1789 | "type": "message", |
| 1790 | "content": [{"type": "output_text", "text": "done"}], |
| 1791 | } |
| 1792 | ] |
| 1793 | }, |
| 1794 | } |
| 1795 | ) |
| 1796 | |
| 1797 | assert parsed == {"reasoning_delta": "", "response_delta": "done"} |
| 1798 | |
| 1799 | |
| 1800 | def test_responses_stream_parser_handles_refusal_and_failed_events(): |
| 1801 | parser = litellm_transport.ResponsesEventParser() |
| 1802 | |
| 1803 | assert parser.parse( |
| 1804 | {"type": "response.refusal.delta", "delta": "no"} |
| 1805 | ) == {"reasoning_delta": "", "response_delta": "no"} |
| 1806 | |
| 1807 | with pytest.raises(RuntimeError, match="policy"): |
| 1808 | parser.parse( |
| 1809 | { |
| 1810 | "type": "response.failed", |
| 1811 | "response": {"error": {"message": "policy"}}, |
| 1812 | } |
| 1813 | ) |
| 1814 | |
| 1815 | |
| 1816 | def test_responses_response_parser_groups_parallel_function_calls(): |
| 1817 | response = { |
| 1818 | "output": [ |
| 1819 | { |
| 1820 | "type": "function_call", |
| 1821 | "name": "lookup", |
| 1822 | "arguments": '{"q":"a0"}', |
| 1823 | }, |
| 1824 | { |
| 1825 | "type": "function_call", |
| 1826 | "name": "rank", |
| 1827 | "arguments": '{"limit":2}', |
| 1828 | }, |
| 1829 | ] |
| 1830 | } |
| 1831 | |
| 1832 | parsed = litellm_transport.ResponsesTransport.parse_response(response) |
| 1833 | |
| 1834 | assert extract_tools.json_parse_dirty(parsed["response_delta"]) == { |
| 1835 | "tool_name": "parallel_tool_calls", |
| 1836 | "tool_args": { |
| 1837 | "calls": [ |
| 1838 | {"tool_name": "lookup", "tool_args": {"q": "a0"}}, |
| 1839 | {"tool_name": "rank", "tool_args": {"limit": 2}}, |
| 1840 | ] |
| 1841 | }, |
| 1842 | } |
| 1843 | |
| 1844 | |
| 1845 | def test_responses_stream_parser_preserves_non_ascii_function_call_arguments(): |
| 1846 | parser = litellm_transport.ResponsesEventParser() |
| 1847 | |
| 1848 | parser.parse( |
| 1849 | { |
| 1850 | "type": "response.output_item.added", |
| 1851 | "output_index": 0, |
| 1852 | "item": { |
| 1853 | "type": "function_call", |
| 1854 | "id": "fc_1", |
| 1855 | "name": "response", |
| 1856 | "arguments": "", |
| 1857 | }, |
| 1858 | } |
| 1859 | ) |
| 1860 | parsed = parser.parse( |
| 1861 | { |
| 1862 | "type": "response.function_call_arguments.done", |
| 1863 | "item_id": "fc_1", |
| 1864 | "name": "response", |
| 1865 | "arguments": '{"text":"привет"}', |
| 1866 | } |
| 1867 | ) |
| 1868 | |
| 1869 | assert parsed["response_delta"] == '{"tool_name": "response", "tool_args": {"text": "привет"}}' |