Skip to content

pydantic_ai.models.openai

Setup

For details on how to set up authentication with this model, see model configuration for OpenAI.

DEPRECATED_OPENAI_MODELS module-attribute 

DEPRECATED_OPENAI_MODELS: frozenset[str] = frozenset(
    {
        "chatgpt-4o-latest",
        "codex-mini-latest",
        "gpt-4-0125-preview",
        "gpt-4-1106-preview",
        "gpt-4-turbo-preview",
        "gpt-4-32k",
        "gpt-4-32k-0314",
        "gpt-4-32k-0613",
        "gpt-4-vision-preview",
        "gpt-4o-audio-preview-2024-10-01",
        "gpt-5.1-mini",
        "o1-mini",
        "o1-mini-2024-09-12",
        "o1-preview",
        "o1-preview-2024-09-12",
    }
)

Models that are deprecated or don't exist but are still present in the OpenAI SDK's type definitions.

OpenAIModelName module-attribute 

OpenAIModelName = str | AllModels

Possible OpenAI model names.

Since OpenAI supports a variety of date-stamped models, we explicitly list the latest models but allow any name in the type hints. See the OpenAI docs for a full list.

Using this more broad type for the model name instead of the ChatModel definition allows this model to be used more easily with other model types (ie, Ollama, Deepseek).

MCP_SERVER_TOOL_CONNECTOR_URI_SCHEME module-attribute 

MCP_SERVER_TOOL_CONNECTOR_URI_SCHEME: Literal[
    "x-openai-connector"
] = "x-openai-connector"

Prefix for OpenAI connector IDs. OpenAI supports either a URL or a connector ID when passing MCP configuration to a model, by using that prefix like x-openai-connector:<connector-id> in a URL, you can pass a connector ID to a model.

OpenAIChatModelSettings

Bases: ModelSettings

Settings used for an OpenAI model request.

Source code in pydantic_ai_slim/pydantic_ai/models/openai.py 
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
class OpenAIChatModelSettings(ModelSettings, total=False):
    """Settings used for an OpenAI model request."""

    # ALL FIELDS MUST BE `openai_` PREFIXED SO YOU CAN MERGE THEM WITH OTHER MODELS.

    openai_reasoning_effort: ReasoningEffort
    """Constrains effort on reasoning for [reasoning models](https://platform.openai.com/docs/guides/reasoning).

    Currently supported values are `low`, `medium`, and `high`. Reducing reasoning effort can
    result in faster responses and fewer tokens used on reasoning in a response.
    """

    openai_logprobs: bool
    """Include log probabilities in the response.

    For Chat models, these will be included in `ModelResponse.provider_details['logprobs']`.
    For Responses models, these will be included in the response output parts `TextPart.provider_details['logprobs']`.
    """

    openai_top_logprobs: int
    """Include log probabilities of the top n tokens in the response."""

    openai_store: bool | None
    """Whether or not to store the output of this request in OpenAI's systems.

    If `False`, OpenAI will not store the request for its own internal review or training.
    See [OpenAI API reference](https://platform.openai.com/docs/api-reference/chat/create#chat-create-store).

    When used with `OpenAIResponsesModel`, stored responses appear in OpenAI's dashboard and
    can be referenced via [`openai_previous_response_id`][pydantic_ai.models.openai.OpenAIResponsesModelSettings.openai_previous_response_id].
    Pair this with `openai_previous_response_id='auto'` to avoid storing duplicate copies of
    the conversation history across retries and subsequent requests within the same run.
    """

    openai_user: str
    """A unique identifier representing the end-user, which can help OpenAI monitor and detect abuse.

    See [OpenAI's safety best practices](https://platform.openai.com/docs/guides/safety-best-practices#end-user-ids) for more details.
    """

    openai_service_tier: Literal['auto', 'default', 'flex', 'priority']
    """The service tier to use for the model request.

    Currently supported values are `auto`, `default`, `flex`, and `priority`.
    For more information, see [OpenAI's service tiers documentation](https://platform.openai.com/docs/api-reference/chat/object#chat/object-service_tier).
    """

    openai_prediction: ChatCompletionPredictionContentParam
    """Enables [predictive outputs](https://platform.openai.com/docs/guides/predicted-outputs).

    This feature is currently only supported for some OpenAI models.
    """

    openai_prompt_cache_key: str
    """Used by OpenAI to cache responses for similar requests to optimize your cache hit rates.

    See the [OpenAI Prompt Caching documentation](https://platform.openai.com/docs/guides/prompt-caching#how-it-works) for more information.
    """

    openai_prompt_cache_retention: Literal['in_memory', '24h']
    """The retention policy for the prompt cache. Set to 24h to enable extended prompt caching, which keeps cached prefixes active for longer, up to a maximum of 24 hours.

    See the [OpenAI Prompt Caching documentation](https://platform.openai.com/docs/guides/prompt-caching#how-it-works) for more information.
    """

    openai_continuous_usage_stats: bool
    """When True, enables continuous usage statistics in streaming responses.

    When enabled, the API returns cumulative usage data with each chunk rather than only at the end.
    This setting correctly handles the cumulative nature of these stats by using only the final
    usage values rather than summing all intermediate values.

    See [OpenAI's streaming documentation](https://platform.openai.com/docs/api-reference/chat/create#stream_options) for more information.
    """

openai_reasoning_effort instance-attribute 

openai_reasoning_effort: ReasoningEffort

Constrains effort on reasoning for reasoning models.

Currently supported values are low, medium, and high. Reducing reasoning effort can result in faster responses and fewer tokens used on reasoning in a response.

openai_logprobs instance-attribute 

openai_logprobs: bool

Include log probabilities in the response.

For Chat models, these will be included in ModelResponse.provider_details['logprobs']. For Responses models, these will be included in the response output parts TextPart.provider_details['logprobs'].

openai_top_logprobs instance-attribute 

openai_top_logprobs: int

Include log probabilities of the top n tokens in the response.

openai_store instance-attribute 

openai_store: bool | None

Whether or not to store the output of this request in OpenAI's systems.

If False, OpenAI will not store the request for its own internal review or training. See OpenAI API reference.

When used with OpenAIResponsesModel, stored responses appear in OpenAI's dashboard and can be referenced via openai_previous_response_id. Pair this with openai_previous_response_id='auto' to avoid storing duplicate copies of the conversation history across retries and subsequent requests within the same run.

openai_user instance-attribute 

openai_user: str

A unique identifier representing the end-user, which can help OpenAI monitor and detect abuse.

See OpenAI's safety best practices for more details.

openai_service_tier instance-attribute 

openai_service_tier: Literal[
    "auto", "default", "flex", "priority"
]

The service tier to use for the model request.

Currently supported values are auto, default, flex, and priority. For more information, see OpenAI's service tiers documentation.

openai_prediction instance-attribute 

openai_prediction: ChatCompletionPredictionContentParam

Enables predictive outputs.

This feature is currently only supported for some OpenAI models.

openai_prompt_cache_key instance-attribute 

openai_prompt_cache_key: str

Used by OpenAI to cache responses for similar requests to optimize your cache hit rates.

See the OpenAI Prompt Caching documentation for more information.

openai_prompt_cache_retention instance-attribute 

openai_prompt_cache_retention: Literal['in_memory', '24h']

The retention policy for the prompt cache. Set to 24h to enable extended prompt caching, which keeps cached prefixes active for longer, up to a maximum of 24 hours.

See the OpenAI Prompt Caching documentation for more information.

openai_continuous_usage_stats instance-attribute 

openai_continuous_usage_stats: bool

When True, enables continuous usage statistics in streaming responses.

When enabled, the API returns cumulative usage data with each chunk rather than only at the end. This setting correctly handles the cumulative nature of these stats by using only the final usage values rather than summing all intermediate values.

See OpenAI's streaming documentation for more information.

OpenAIResponsesModelSettings

Bases: OpenAIChatModelSettings

Settings used for an OpenAI Responses model request.

ALL FIELDS MUST BE openai_ PREFIXED SO YOU CAN MERGE THEM WITH OTHER MODELS.

Source code in pydantic_ai_slim/pydantic_ai/models/openai.py 
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
class OpenAIResponsesModelSettings(OpenAIChatModelSettings, total=False):
    """Settings used for an OpenAI Responses model request.

    ALL FIELDS MUST BE `openai_` PREFIXED SO YOU CAN MERGE THEM WITH OTHER MODELS.
    """

    openai_native_tools: Sequence[FileSearchToolParam | WebSearchToolParam | ComputerToolParam]
    """The provided OpenAI built-in tools to use.

    See [OpenAI's built-in tools](https://platform.openai.com/docs/guides/tools?api-mode=responses) for more details.
    """

    openai_reasoning_summary: Literal['detailed', 'concise', 'auto']
    """A summary of the reasoning performed by the model.

    This can be useful for debugging and understanding the model's reasoning process.
    One of `concise`, `detailed`, or `auto`.

    Check the [OpenAI Reasoning documentation](https://platform.openai.com/docs/guides/reasoning?api-mode=responses#reasoning-summaries)
    for more details.
    """

    openai_send_reasoning_ids: bool
    """Whether to send the unique IDs of reasoning, text, and function call parts from the message history to the model. Enabled by default for reasoning models.

    This can result in errors like `"Item 'rs_123' of type 'reasoning' was provided without its required following item."`
    if the message history you're sending does not match exactly what was received from the Responses API in a previous response,
    for example if you're using a [history processor](../../message-history.md#processing-message-history).
    In that case, you'll want to disable this.
    """

    openai_truncation: Literal['disabled', 'auto']
    """The truncation strategy to use for the model response.

    It can be either:
    - `disabled` (default): If a model response will exceed the context window size for a model, the
        request will fail with a 400 error.
    - `auto`: If the context of this response and previous ones exceeds the model's context window size,
        the model will truncate the response to fit the context window by dropping input items in the
        middle of the conversation.
    """

    openai_text_verbosity: Literal['low', 'medium', 'high']
    """Constrains the verbosity of the model's text response.

    Lower values will result in more concise responses, while higher values will
    result in more verbose responses. Currently supported values are `low`,
    `medium`, and `high`.
    """

    openai_previous_response_id: Literal['auto'] | str
    """Reference a prior OpenAI response to continue a conversation server-side, omitting already-stored messages from the input.

    - `'auto'`: chain to the most recent `provider_response_id` in the message history.
      If the history contains no such response, no `previous_response_id` is sent.
    - A concrete response ID string: use it as the seed for the first request in the run
      (e.g. to continue from a prior turn). On subsequent in-run requests (retries,
      tool-call continuations), the most recent `provider_response_id` from the message
      history takes precedence so the chain extends correctly without re-sending messages
      that are already server-side.

    In both cases, messages that precede the chosen response in the history are omitted
    from the input, since OpenAI reconstructs them from server-side state.

    Requires the referenced response to have been stored (see
    [`openai_store`][pydantic_ai.models.openai.OpenAIResponsesModelSettings.openai_store],
    which defaults to `True` on OpenAI's side). Not compatible with Zero Data Retention.

    See the [OpenAI Responses API documentation](https://platform.openai.com/docs/guides/reasoning#keeping-reasoning-items-in-context)
    for more information.
    """

    openai_conversation_id: Literal['auto'] | str
    """Reference an OpenAI conversation to continue durable conversation state server-side.

    - `'auto'`: use the most recent OpenAI conversation ID from `ModelResponse.provider_details['conversation_id']`
      in the message history with the same Pydantic AI `conversation_id`, when available. If the history
      contains no such response, no `conversation` is sent.
    - A concrete conversation ID string: use it as the OpenAI Responses API `conversation` parameter.

    When a matching conversation ID is found in message history, messages that precede that response
    are omitted from the input, since OpenAI reconstructs them from the server-side conversation.

    Not compatible with
    [`openai_previous_response_id`][pydantic_ai.models.openai.OpenAIResponsesModelSettings.openai_previous_response_id].

    See the [OpenAI conversation state documentation](https://platform.openai.com/docs/guides/conversation-state)
    for more information.
    """

    openai_include_code_execution_outputs: bool
    """Whether to include the code execution results in the response.

    Corresponds to the `code_interpreter_call.outputs` value of the `include` parameter in the Responses API.
    """

    openai_include_web_search_sources: bool
    """Whether to include the web search results in the response.

    Corresponds to the `web_search_call.action.sources` value of the `include` parameter in the Responses API.
    """

    openai_include_file_search_results: bool
    """Whether to include the file search results in the response.

    Corresponds to the `file_search_call.results` value of the `include` parameter in the Responses API.
    """

    openai_include_raw_annotations: bool
    """Whether to include the raw annotations in `TextPart.provider_details`.

    When enabled, any annotations (e.g., citations from web search) will be available
    in the `provider_details['annotations']` field of text parts.
    This is opt-in since there may be overlap with native annotation support once
    added via https://github.com/pydantic/pydantic-ai/issues/3126.
    """

    openai_context_management: list[ContextManagement]
    """Context management configuration for the request.

    This enables OpenAI's server-side automatic compaction inside the regular
    `/responses` call, as opposed to the standalone `/responses/compact` endpoint.
    See [OpenAI's compaction guide](https://developers.openai.com/api/docs/guides/compaction)
    for details.

    The [`OpenAICompaction`][pydantic_ai.models.openai.OpenAICompaction] capability
    sets this automatically in its default (stateful) mode.
    """

openai_native_tools instance-attribute 

openai_native_tools: Sequence[
    FileSearchToolParam
    | WebSearchToolParam
    | ComputerToolParam
]

The provided OpenAI built-in tools to use.

See OpenAI's built-in tools for more details.

openai_reasoning_summary instance-attribute 

openai_reasoning_summary: Literal[
    "detailed", "concise", "auto"
]

A summary of the reasoning performed by the model.

This can be useful for debugging and understanding the model's reasoning process. One of concise, detailed, or auto.

Check the OpenAI Reasoning documentation for more details.

openai_send_reasoning_ids instance-attribute 

openai_send_reasoning_ids: bool

Whether to send the unique IDs of reasoning, text, and function call parts from the message history to the model. Enabled by default for reasoning models.

This can result in errors like "Item 'rs_123' of type 'reasoning' was provided without its required following item." if the message history you're sending does not match exactly what was received from the Responses API in a previous response, for example if you're using a history processor. In that case, you'll want to disable this.

openai_truncation instance-attribute 

openai_truncation: Literal['disabled', 'auto']

The truncation strategy to use for the model response.

It can be either: - disabled (default): If a model response will exceed the context window size for a model, the request will fail with a 400 error. - auto: If the context of this response and previous ones exceeds the model's context window size, the model will truncate the response to fit the context window by dropping input items in the middle of the conversation.

openai_text_verbosity instance-attribute 

openai_text_verbosity: Literal['low', 'medium', 'high']

Constrains the verbosity of the model's text response.

Lower values will result in more concise responses, while higher values will result in more verbose responses. Currently supported values are low, medium, and high.

openai_previous_response_id instance-attribute 

openai_previous_response_id: Literal['auto'] | str

Reference a prior OpenAI response to continue a conversation server-side, omitting already-stored messages from the input.

  • 'auto': chain to the most recent provider_response_id in the message history. If the history contains no such response, no previous_response_id is sent.
  • A concrete response ID string: use it as the seed for the first request in the run (e.g. to continue from a prior turn). On subsequent in-run requests (retries, tool-call continuations), the most recent provider_response_id from the message history takes precedence so the chain extends correctly without re-sending messages that are already server-side.

In both cases, messages that precede the chosen response in the history are omitted from the input, since OpenAI reconstructs them from server-side state.

Requires the referenced response to have been stored (see openai_store, which defaults to True on OpenAI's side). Not compatible with Zero Data Retention.

See the OpenAI Responses API documentation for more information.

openai_conversation_id instance-attribute 

openai_conversation_id: Literal['auto'] | str

Reference an OpenAI conversation to continue durable conversation state server-side.

  • 'auto': use the most recent OpenAI conversation ID from ModelResponse.provider_details['conversation_id'] in the message history with the same Pydantic AI conversation_id, when available. If the history contains no such response, no conversation is sent.
  • A concrete conversation ID string: use it as the OpenAI Responses API conversation parameter.

When a matching conversation ID is found in message history, messages that precede that response are omitted from the input, since OpenAI reconstructs them from the server-side conversation.

Not compatible with openai_previous_response_id.

See the OpenAI conversation state documentation for more information.

openai_include_code_execution_outputs instance-attribute 

openai_include_code_execution_outputs: bool

Whether to include the code execution results in the response.

Corresponds to the code_interpreter_call.outputs value of the include parameter in the Responses API.

openai_include_web_search_sources instance-attribute 

openai_include_web_search_sources: bool

Whether to include the web search results in the response.

Corresponds to the web_search_call.action.sources value of the include parameter in the Responses API.

openai_include_file_search_results instance-attribute 

openai_include_file_search_results: bool

Whether to include the file search results in the response.

Corresponds to the file_search_call.results value of the include parameter in the Responses API.

openai_include_raw_annotations instance-attribute 

openai_include_raw_annotations: bool

Whether to include the raw annotations in TextPart.provider_details.

When enabled, any annotations (e.g., citations from web search) will be available in the provider_details['annotations'] field of text parts. This is opt-in since there may be overlap with native annotation support once added via https://github.com/pydantic/pydantic-ai/issues/3126.

openai_context_management instance-attribute 

openai_context_management: list[ContextManagement]

Context management configuration for the request.

This enables OpenAI's server-side automatic compaction inside the regular /responses call, as opposed to the standalone /responses/compact endpoint. See OpenAI's compaction guide for details.

The OpenAICompaction capability sets this automatically in its default (stateful) mode.

OpenAIChatModel dataclass

Bases: Model[AsyncOpenAI]

A model that uses the OpenAI API.

Internally, this uses the OpenAI Python client to interact with the API.

Apart from __init__, all methods are private or match those of the base class.

Source code in pydantic_ai_slim/pydantic_ai/models/openai.py 
 716
 717
 718
 719
 720
 721
 722
 723
 724
 725
 726
 727
 728
 729
 730
 731
 732
 733
 734
 735
 736
 737
 738
 739
 740
 741
 742
 743
 744
 745
 746
 747
 748
 749
 750
 751
 752
 753
 754
 755
 756
 757
 758
 759
 760
 761
 762
 763
 764
 765
 766
 767
 768
 769
 770
 771
 772
 773
 774
 775
 776
 777
 778
 779
 780
 781
 782
 783
 784
 785
 786
 787
 788
 789
 790
 791
 792
 793
 794
 795
 796
 797
 798
 799
 800
 801
 802
 803
 804
 805
 806
 807
 808
 809
 810
 811
 812
 813
 814
 815
 816
 817
 818
 819
 820
 821
 822
 823
 824
 825
 826
 827
 828
 829
 830
 831
 832
 833
 834
 835
 836
 837
 838
 839
 840
 841
 842
 843
 844
 845
 846
 847
 848
 849
 850
 851
 852
 853
 854
 855
 856
 857
 858
 859
 860
 861
 862
 863
 864
 865
 866
 867
 868
 869
 870
 871
 872
 873
 874
 875
 876
 877
 878
 879
 880
 881
 882
 883
 884
 885
 886
 887
 888
 889
 890
 891
 892
 893
 894
 895
 896
 897
 898
 899
 900
 901
 902
 903
 904
 905
 906
 907
 908
 909
 910
 911
 912
 913
 914
 915
 916
 917
 918
 919
 920
 921
 922
 923
 924
 925
 926
 927
 928
 929
 930
 931
 932
 933
 934
 935
 936
 937
 938
 939
 940
 941
 942
 943
 944
 945
 946
 947
 948
 949
 950
 951
 952
 953
 954
 955
 956
 957
 958
 959
 960
 961
 962
 963
 964
 965
 966
 967
 968
 969
 970
 971
 972
 973
 974
 975
 976
 977
 978
 979
 980
 981
 982
 983
 984
 985
 986
 987
 988
 989
 990
 991
 992
 993
 994
 995
 996
 997
 998
 999
1000
1001
1002
1003
1004
1005
1006
1007
1008
1009
1010
1011
1012
1013
1014
1015
1016
1017
1018
1019
1020
1021
1022
1023
1024
1025
1026
1027
1028
1029
1030
1031
1032
1033
1034
1035
1036
1037
1038
1039
1040
1041
1042
1043
1044
1045
1046
1047
1048
1049
1050
1051
1052
1053
1054
1055
1056
1057
1058
1059
1060
1061
1062
1063
1064
1065
1066
1067
1068
1069
1070
1071
1072
1073
1074
1075
1076
1077
1078
1079
1080
1081
1082
1083
1084
1085
1086
1087
1088
1089
1090
1091
1092
1093
1094
1095
1096
1097
1098
1099
1100
1101
1102
1103
1104
1105
1106
1107
1108
1109
1110
1111
1112
1113
1114
1115
1116
1117
1118
1119
1120
1121
1122
1123
1124
1125
1126
1127
1128
1129
1130
1131
1132
1133
1134
1135
1136
1137
1138
1139
1140
1141
1142
1143
1144
1145
1146
1147
1148
1149
1150
1151
1152
1153
1154
1155
1156
1157
1158
1159
1160
1161
1162
1163
1164
1165
1166
1167
1168
1169
1170
1171
1172
1173
1174
1175
1176
1177
1178
1179
1180
1181
1182
1183
1184
1185
1186
1187
1188
1189
1190
1191
1192
1193
1194
1195
1196
1197
1198
1199
1200
1201
1202
1203
1204
1205
1206
1207
1208
1209
1210
1211
1212
1213
1214
1215
1216
1217
1218
1219
1220
1221
1222
1223
1224
1225
1226
1227
1228
1229
1230
1231
1232
1233
1234
1235
1236
1237
1238
1239
1240
1241
1242
1243
1244
1245
1246
1247
1248
1249
1250
1251
1252
1253
1254
1255
1256
1257
1258
1259
1260
1261
1262
1263
1264
1265
1266
1267
1268
1269
1270
1271
1272
1273
1274
1275
1276
1277
1278
1279
1280
1281
1282
1283
1284
1285
1286
1287
1288
1289
1290
1291
1292
1293
1294
1295
1296
1297
1298
1299
1300
1301
1302
1303
1304
1305
1306
1307
1308
1309
1310
1311
1312
1313
1314
1315
1316
1317
1318
1319
1320
1321
1322
1323
1324
1325
1326
1327
1328
1329
1330
1331
1332
1333
1334
1335
1336
1337
1338
1339
1340
1341
1342
1343
1344
1345
1346
1347
1348
1349
1350
1351
1352
1353
1354
1355
1356
1357
1358
1359
1360
1361
1362
1363
1364
1365
1366
1367
1368
1369
1370
1371
1372
1373
1374
1375
1376
1377
1378
1379
1380
1381
1382
1383
1384
1385
1386
1387
1388
1389
1390
1391
1392
1393
1394
1395
1396
1397
1398
1399
1400
1401
1402
1403
1404
1405
1406
1407
1408
1409
1410
1411
1412
1413
1414
1415
1416
1417
1418
1419
1420
1421
1422
1423
1424
1425
1426
1427
1428
1429
1430
1431
1432
1433
1434
1435
1436
1437
1438
1439
1440
1441
1442
1443
1444
1445
1446
1447
1448
1449
1450
1451
1452
1453
1454
1455
1456
1457
1458
1459
1460
1461
1462
1463
1464
1465
1466
1467
1468
1469
1470
1471
1472
1473
1474
1475
1476
1477
1478
1479
1480
1481
1482
1483
1484
1485
1486
1487
1488
1489
1490
1491
1492
1493
1494
1495
1496
1497
1498
1499
1500
1501
1502
1503
1504
1505
1506
1507
1508
1509
1510
1511
1512
1513
1514
1515
1516
1517
1518
1519
1520
1521
1522
1523
1524
1525
1526
1527
1528
1529
1530
1531
1532
1533
1534
1535
1536
1537
1538
1539
1540
1541
1542
1543
1544
1545
1546
1547
1548
1549
1550
1551
1552
1553
1554
1555
1556
1557
1558
1559
1560
1561
1562
1563
1564
1565
1566
1567
1568
1569
1570
1571
1572
1573
1574
1575
1576
1577
1578
1579
1580
1581
1582
1583
1584
1585
1586
1587
1588
1589
1590
1591
1592
1593
1594
1595
1596
1597
1598
1599
1600
1601
1602
1603
1604
1605
1606
1607
1608
1609
1610
1611
1612
1613
1614
1615
1616
1617
1618
1619
1620
1621
1622
1623
1624
1625
1626
1627
1628
1629
1630
1631
1632
1633
1634
@dataclass(init=False)
class OpenAIChatModel(Model[AsyncOpenAI]):
    """A model that uses the OpenAI API.

    Internally, this uses the [OpenAI Python client](https://github.com/openai/openai-python) to interact with the API.

    Apart from `__init__`, all methods are private or match those of the base class.
    """

    _model_name: OpenAIModelName = field(repr=False)
    _provider: Provider[AsyncOpenAI] = field(repr=False)

    def __init__(
        self,
        model_name: OpenAIModelName,
        *,
        provider: OpenAIChatCompatibleProvider
        | Literal[
            'openai',
            'openai-chat',
            'gateway',
        ]
        | Provider[AsyncOpenAI] = 'openai',
        profile: ModelProfileSpec | None = None,
        settings: ModelSettings | None = None,
    ):
        """Initialize an OpenAI model.

        Args:
            model_name: The name of the OpenAI model to use. List of model names available
                [here](https://github.com/openai/openai-python/blob/v1.54.3/src/openai/types/chat_model.py#L7)
                (Unfortunately, despite being ask to do so, OpenAI do not provide `.inv` files for their API).
            provider: The provider to use. Defaults to `'openai'`.
            profile: The model profile to use. Defaults to a profile picked by the provider based on the model name.
            settings: Default model settings for this model instance.
        """
        self._model_name = model_name

        if isinstance(provider, str):
            provider = infer_provider('gateway/openai' if provider == 'gateway' else provider)
        self._provider = provider

        super().__init__(settings=settings, profile=profile)

        validate_openai_profile(self.profile)

    @property
    def client(self) -> AsyncOpenAI:
        return self._provider.client

    @property
    def base_url(self) -> str:
        return str(self.client.base_url)

    @property
    def model_name(self) -> OpenAIModelName:
        """The model name."""
        return self._model_name

    @property
    def system(self) -> str:
        """The model provider."""
        return self._provider.name

    @classmethod
    def supported_native_tools(cls) -> frozenset[type[AbstractNativeTool]]:
        """Return the set of builtin tool types this model can handle."""
        return frozenset({WebSearchTool})

    @cached_property
    def profile(self) -> OpenAIModelProfile:
        """The model profile.

        WebSearchTool is only supported if openai_chat_supports_web_search is True.
        """
        _profile = super().profile
        if not _profile.get('openai_chat_supports_web_search', False):
            new_tools = _profile.get('supported_native_tools', SUPPORTED_NATIVE_TOOLS) - {WebSearchTool}
            _profile = merge_profile(_profile, ModelProfile(supported_native_tools=new_tools))
        return cast(OpenAIModelProfile, _profile)

    def prepare_request(
        self,
        model_settings: ModelSettings | None,
        model_request_parameters: ModelRequestParameters,
    ) -> tuple[ModelSettings | None, ModelRequestParameters]:
        # Check for WebSearchTool before base validation to provide a helpful error message
        if (
            any(isinstance(tool, WebSearchTool) for tool in model_request_parameters.native_tools)
            and not self.profile.get('openai_chat_supports_web_search', False)
            and not any(t.unless_native == 'web_search' for t in model_request_parameters.function_tools)
        ):
            raise UserError(
                f'WebSearchTool is not supported with `OpenAIChatModel` and model {self.model_name!r}. '
                f'Please use `OpenAIResponsesModel` instead.'
            )
        return super().prepare_request(model_settings, model_request_parameters)

    async def request(
        self,
        messages: list[ModelMessage],
        model_settings: ModelSettings | None,
        model_request_parameters: ModelRequestParameters,
    ) -> ModelResponse:
        check_allow_model_requests()
        model_settings, model_request_parameters = self.prepare_request(
            model_settings,
            model_request_parameters,
        )
        response = await self._completions_create(
            messages, False, cast(OpenAIChatModelSettings, model_settings or {}), model_request_parameters
        )

        # Handle ModelResponse returned directly (for content filters)
        if isinstance(response, ModelResponse):
            return response

        model_response = self._process_response(response)
        return model_response

    def _translate_thinking(
        self,
        model_settings: OpenAIChatModelSettings,
        model_request_parameters: ModelRequestParameters,
    ) -> ReasoningEffort | Omit:
        """Get reasoning effort, falling back to unified thinking when provider-specific setting is not set."""
        if effort := model_settings.get('openai_reasoning_effort'):
            return effort
        thinking = model_request_parameters.thinking
        if thinking is None:
            return OMIT
        return OPENAI_REASONING_EFFORT_MAP[thinking]  # type: ignore[return-value]

    @asynccontextmanager
    async def request_stream(
        self,
        messages: list[ModelMessage],
        model_settings: ModelSettings | None,
        model_request_parameters: ModelRequestParameters,
        run_context: RunContext[Any] | None = None,
    ) -> AsyncIterator[StreamedResponse]:
        check_allow_model_requests()
        model_settings, model_request_parameters = self.prepare_request(
            model_settings,
            model_request_parameters,
        )
        model_settings_cast = cast(OpenAIChatModelSettings, model_settings or {})
        response = await self._completions_create(messages, True, model_settings_cast, model_request_parameters)
        async with response:
            yield await self._process_streamed_response(response, model_request_parameters, model_settings_cast)

    @overload
    async def _completions_create(
        self,
        messages: list[ModelMessage],
        stream: Literal[True],
        model_settings: OpenAIChatModelSettings,
        model_request_parameters: ModelRequestParameters,
    ) -> AsyncStream[ChatCompletionChunk]: ...

    @overload
    async def _completions_create(
        self,
        messages: list[ModelMessage],
        stream: Literal[False],
        model_settings: OpenAIChatModelSettings,
        model_request_parameters: ModelRequestParameters,
    ) -> chat.ChatCompletion | ModelResponse: ...

    async def _completions_create(
        self,
        messages: list[ModelMessage],
        stream: bool,
        model_settings: OpenAIChatModelSettings,
        model_request_parameters: ModelRequestParameters,
    ) -> chat.ChatCompletion | AsyncStream[ChatCompletionChunk] | ModelResponse:
        tool_choice: ChatCompletionToolChoiceOptionParam | None

        tools, tool_choice = self._get_tool_choice(model_settings, model_request_parameters)
        web_search_options = self._get_web_search_options(model_request_parameters)
        profile = self.profile

        openai_messages = await self._map_messages(messages, model_request_parameters)

        response_format: chat.completion_create_params.ResponseFormat | None = None
        if model_request_parameters.output_mode == 'native':
            output_object = model_request_parameters.output_object
            assert output_object is not None
            response_format = self._map_json_schema(output_object)
        elif model_request_parameters.output_mode == 'prompted' and self.profile.get(
            'supports_json_object_output', False
        ):  # pragma: no branch
            response_format = {'type': 'json_object'}

        _drop_sampling_params_for_reasoning(profile, model_settings, model_request_parameters)

        _drop_unsupported_params(profile, model_settings)

        with _map_api_errors(self.model_name):
            try:
                extra_headers = model_settings.get('extra_headers', {})
                extra_headers.setdefault('User-Agent', get_user_agent())

                # OpenAI SDK type stubs incorrectly use 'in-memory' but API requires 'in_memory', so we have to use `Any` to not hit type errors
                prompt_cache_retention: Any = model_settings.get('openai_prompt_cache_retention', OMIT)
                return await self.client.chat.completions.create(
                    model=self.model_name,
                    messages=openai_messages,
                    parallel_tool_calls=model_settings.get('parallel_tool_calls', OMIT) if tools else OMIT,
                    tools=tools or OMIT,
                    tool_choice=tool_choice or OMIT,
                    stream=stream,
                    stream_options=self._get_stream_options(model_settings) if stream else OMIT,
                    stop=model_settings.get('stop_sequences', OMIT),
                    max_completion_tokens=model_settings.get('max_tokens', OMIT),
                    timeout=model_settings.get('timeout', NOT_GIVEN),
                    response_format=response_format or OMIT,
                    seed=model_settings.get('seed', OMIT),
                    reasoning_effort=self._translate_thinking(model_settings, model_request_parameters),
                    user=model_settings.get('openai_user', OMIT),
                    web_search_options=web_search_options or OMIT,
                    service_tier=_resolve_openai_service_tier(model_settings),
                    prediction=model_settings.get('openai_prediction', OMIT),
                    temperature=model_settings.get('temperature', OMIT),
                    top_p=model_settings.get('top_p', OMIT),
                    presence_penalty=model_settings.get('presence_penalty', OMIT),
                    frequency_penalty=model_settings.get('frequency_penalty', OMIT),
                    logit_bias=model_settings.get('logit_bias', OMIT),
                    logprobs=model_settings.get('openai_logprobs', OMIT),
                    top_logprobs=model_settings.get('openai_top_logprobs', OMIT),
                    store=model_settings.get('openai_store', OMIT),
                    prompt_cache_key=model_settings.get('openai_prompt_cache_key', OMIT),
                    prompt_cache_retention=prompt_cache_retention,
                    extra_headers=extra_headers,
                    extra_body=model_settings.get('extra_body'),
                )
            except APIStatusError as e:
                if model_response := _check_azure_content_filter(e, self.system, self.model_name):
                    return model_response
                raise

    def _validate_completion(self, response: chat.ChatCompletion) -> _ChatCompletion:
        """Hook that validates chat completions before processing.

        This method may be overridden by subclasses of `OpenAIChatModel` to apply custom completion validations.
        """
        return _ChatCompletion.model_validate(response.model_dump())

    def _process_provider_details(self, response: chat.ChatCompletion) -> dict[str, Any] | None:
        """Hook that response content to provider details.

        This method may be overridden by subclasses of `OpenAIChatModel` to apply custom mappings.
        """
        return _map_provider_details(response.choices[0])

    def _process_response(self, response: chat.ChatCompletion | str) -> ModelResponse:
        """Process a non-streamed response, and prepare a message to return."""
        # Although the OpenAI SDK claims to return a Pydantic model (`ChatCompletion`) from the chat completions function:
        # * it hasn't actually performed validation (presumably they're creating the model with `model_construct` or something?!)
        # * if the endpoint returns plain text, the return type is a string
        # Thus we validate it fully here.
        if not isinstance(response, chat.ChatCompletion):
            raise UnexpectedModelBehavior(
                f'Invalid response from {self.system} chat completions endpoint, expected JSON data'
            )

        timestamp = _now_utc()
        if not response.created:
            response.created = int(timestamp.timestamp())

        # Workaround for local Ollama which sometimes returns a `None` finish reason.
        if response.choices and (choice := response.choices[0]) and choice.finish_reason is None:  # pyright: ignore[reportUnnecessaryComparison]
            choice.finish_reason = 'stop'

        try:
            response = self._validate_completion(response)
        except ValidationError as e:
            raise UnexpectedModelBehavior(f'Invalid response from {self.system} chat completions endpoint: {e}') from e

        choice = response.choices[0]

        # Handle refusal responses (structured output safety filter)
        if choice.message.refusal:
            provider_details = self._process_provider_details(response) or {}
            provider_details.pop('finish_reason', None)
            provider_details['refusal'] = choice.message.refusal
            if response.created:  # pragma: no branch
                provider_details['timestamp'] = number_to_datetime(response.created)
            return ModelResponse(
                parts=[],
                usage=self._map_usage(response),
                model_name=response.model,
                timestamp=_now_utc(),
                provider_details=provider_details or None,
                provider_response_id=response.id,
                provider_name=self._provider.name,
                provider_url=self._provider.base_url,
                finish_reason='content_filter',
            )

        items: list[ModelResponsePart] = []

        if thinking_parts := self._process_thinking(choice.message):
            items.extend(thinking_parts)

        if choice.message.content:
            items.extend(
                (replace(part, id='content', provider_name=self.system) if isinstance(part, ThinkingPart) else part)
                for part in split_content_into_text_and_thinking(
                    choice.message.content, self.profile.get('thinking_tags', DEFAULT_THINKING_TAGS)
                )
            )
        if choice.message.tool_calls is not None:
            for c in choice.message.tool_calls:
                if isinstance(c, ChatCompletionMessageFunctionToolCall):
                    part = ToolCallPart(c.function.name, c.function.arguments, tool_call_id=c.id)
                elif isinstance(c, ChatCompletionMessageCustomToolCall):  # pragma: no cover
                    # NOTE: Custom tool calls are not supported.
                    # See <https://github.com/pydantic/pydantic-ai/issues/2513> for more details.
                    raise RuntimeError('Custom tool calls are not supported')
                else:
                    assert_never(c)
                part.tool_call_id = _guard_tool_call_id(part)
                items.append(part)

        provider_details = self._process_provider_details(response)
        if response.created:  # pragma: no branch
            if provider_details is None:
                provider_details = {}
            provider_details['timestamp'] = number_to_datetime(response.created)

        return ModelResponse(
            parts=items,
            usage=self._map_usage(response),
            model_name=response.model,
            timestamp=timestamp,
            provider_details=provider_details or None,
            provider_response_id=response.id,
            provider_name=self._provider.name,
            provider_url=self._provider.base_url,
            finish_reason=self._map_finish_reason(choice.finish_reason),
        )

    def _process_thinking(self, message: chat.ChatCompletionMessage) -> list[ThinkingPart] | None:
        """Hook that maps reasoning tokens to thinking parts.

        This method may be overridden by subclasses of `OpenAIChatModel` to apply custom mappings.
        """
        profile = self.profile
        custom_field = profile.get('openai_chat_thinking_field', None)
        items: list[ThinkingPart] = []

        # Prefer the configured custom reasoning field, if present in profile.
        # Fall back to built-in fields if no custom field result was found.

        # The `reasoning_content` field is typically present in DeepSeek and Moonshot models.
        # https://api-docs.deepseek.com/guides/reasoning_model

        # The `reasoning` field is typically present in gpt-oss via Ollama and OpenRouter.
        # - https://cookbook.openai.com/articles/gpt-oss/handle-raw-cot#chat-completions-api
        # - https://openrouter.ai/docs/use-cases/reasoning-tokens#basic-usage-with-reasoning-tokens
        for field_name in (custom_field, 'reasoning', 'reasoning_content'):
            if not field_name:
                continue
            reasoning: object = getattr(message, field_name, None)
            if not reasoning:
                continue
            if not isinstance(reasoning, str):
                warnings.warn(
                    f'Unexpected non-string value for {field_name!r}: {type(reasoning).__name__}. '
                    'Please open an issue at https://github.com/pydantic/pydantic-ai/issues.',
                    UserWarning,
                )
                continue
            items.append(ThinkingPart(id=field_name, content=reasoning, provider_name=self.system))
            return items

        return items or None

    async def _process_streamed_response(
        self,
        response: AsyncStream[ChatCompletionChunk],
        model_request_parameters: ModelRequestParameters,
        model_settings: OpenAIChatModelSettings | None = None,
    ) -> OpenAIStreamedResponse:
        """Process a streamed response, and prepare a streaming response to return."""
        peekable_response: _utils.PeekableAsyncStream[ChatCompletionChunk, AsyncStream[ChatCompletionChunk]] = (
            _utils.PeekableAsyncStream(response)
        )
        with _map_api_errors(self.model_name):
            first_chunk = await peekable_response.peek()
        if isinstance(first_chunk, _utils.Unset):
            raise UnexpectedModelBehavior(  # pragma: no cover
                'Streamed response ended without content or tool calls'
            )

        # When using Azure OpenAI and a content filter is enabled, the first chunk will contain a `''` model name,
        # so we set it from a later chunk in `OpenAIChatStreamedResponse`.
        model_name = first_chunk.model or self.model_name

        return self._streamed_response_cls(
            model_request_parameters=model_request_parameters,
            _model_name=model_name,
            _model_profile=self.profile,
            _response=peekable_response,
            _provider_name=self._provider.name,
            _provider_url=self._provider.base_url,
            _provider_timestamp=number_to_datetime(first_chunk.created) if first_chunk.created else None,
            _model_settings=model_settings,
        )

    @property
    def _streamed_response_cls(self) -> type[OpenAIStreamedResponse]:
        """Returns the `StreamedResponse` type that will be used for streamed responses.

        This method may be overridden by subclasses of `OpenAIChatModel` to provide their own `StreamedResponse` type.
        """
        return OpenAIStreamedResponse

    def _map_usage(self, response: chat.ChatCompletion) -> usage.RequestUsage:
        return _map_usage(response, self._provider.name, self._provider.base_url, self.model_name)

    def _get_tool_choice(
        self,
        model_settings: OpenAIChatModelSettings,
        model_request_parameters: ModelRequestParameters,
    ) -> tuple[list[chat.ChatCompletionToolParam], ChatCompletionToolChoiceOptionParam | None]:
        """Determine which tools to send and the API tool_choice value.

        Returns:
            A tuple of (filtered_tools, tool_choice).
        """
        openai_profile = self.profile

        resolved_tool_choice = resolve_tool_choice(model_settings, model_request_parameters)
        tool_defs = model_request_parameters.tool_defs

        tool_choice: ChatCompletionToolChoiceOptionParam
        if resolved_tool_choice in ('auto', 'none'):
            tool_choice = resolved_tool_choice
        elif resolved_tool_choice == 'required':
            supports = _support_tool_forcing(self.model_name, openai_profile, model_settings)
            tool_choice = 'required' if supports else 'auto'
        elif isinstance(resolved_tool_choice, tuple):
            tool_choice_mode, tool_names = resolved_tool_choice
            supports = _support_tool_forcing(self.model_name, openai_profile, model_settings)
            if tool_choice_mode == 'required' and len(tool_names) == 1:
                if supports:
                    tool_choice = {'type': 'function', 'function': {'name': next(iter(tool_names))}}
                else:
                    # Forcing not supported: filter so the model can only see the requested tool.
                    # Breaks caching, but OpenAI Chat doesn't support limiting tools via API arg.
                    tool_defs = {k: v for k, v in tool_defs.items() if k in tool_names}
                    tool_choice = 'auto'
            else:
                # Breaks caching, but OpenAI Chat doesn't support limiting tools via API arg
                tool_defs = {k: v for k, v in tool_defs.items() if k in tool_names}
                tool_choice = 'auto' if tool_choice_mode == 'auto' or not supports else tool_choice_mode
        else:
            assert_never(resolved_tool_choice)

        tools: list[chat.ChatCompletionToolParam] = [self._map_tool_definition(t) for t in tool_defs.values()]
        if not tools:
            return tools, None

        return tools, tool_choice

    def _get_stream_options(self, model_settings: OpenAIChatModelSettings) -> chat.ChatCompletionStreamOptionsParam:
        """Build stream_options for the API request.

        Returns a dict with include_usage=True and optionally continuous_usage_stats if configured.
        """
        options: dict[str, bool] = {'include_usage': True}
        if model_settings.get('openai_continuous_usage_stats'):
            options['continuous_usage_stats'] = True
        return cast(chat.ChatCompletionStreamOptionsParam, options)

    def _get_web_search_options(self, model_request_parameters: ModelRequestParameters) -> WebSearchOptions | None:
        for tool in model_request_parameters.native_tools:
            if isinstance(tool, WebSearchTool):  # pragma: no branch
                if tool.user_location:
                    return WebSearchOptions(
                        search_context_size=tool.search_context_size,
                        user_location=WebSearchOptionsUserLocation(
                            type='approximate',
                            approximate=WebSearchOptionsUserLocationApproximate(**tool.user_location),
                        ),
                    )
                return WebSearchOptions(search_context_size=tool.search_context_size)
        return None

    @dataclass
    class _MapModelResponseContext:
        """Context object for mapping a `ModelResponse` to OpenAI chat completion parameters.

        This class is designed to be subclassed to add new fields for custom logic,
        collecting various parts of the model response (like text and tool calls)
        to form a single assistant message.
        """

        _model: OpenAIChatModel

        texts: list[str] = field(default_factory=list[str])
        thinkings: dict[str, list[str]] = field(default_factory=dict[str, list[str]])
        tool_calls: list[ChatCompletionMessageFunctionToolCallParam] = field(
            default_factory=list[ChatCompletionMessageFunctionToolCallParam]
        )

        def map_assistant_message(self, message: ModelResponse) -> chat.ChatCompletionAssistantMessageParam | None:
            for item in message.parts:
                if isinstance(item, TextPart):
                    self._map_response_text_part(item)
                elif isinstance(item, ThinkingPart):
                    self._map_response_thinking_part(item)
                elif isinstance(item, ToolCallPart):
                    self._map_response_tool_call_part(item)
                elif isinstance(item, NativeToolCallPart | NativeToolReturnPart):  # pragma: no cover
                    self._map_response_builtin_part(item)
                elif isinstance(item, FilePart):  # pragma: no cover
                    self._map_response_file_part(item)
                elif isinstance(item, CompactionPart):  # pragma: no cover
                    # Compaction parts are not sent back to the Chat Completions API.
                    pass
                else:
                    assert_never(item)
            return self._into_message_param()

        def _into_message_param(self) -> chat.ChatCompletionAssistantMessageParam | None:
            """Converts the collected texts and tool calls into a single OpenAI `ChatCompletionAssistantMessageParam`.

            This method serves as a hook that can be overridden by subclasses
            to implement custom logic for how collected parts are transformed into the final message parameter.

            Returns:
                An OpenAI `ChatCompletionAssistantMessageParam` representing the assistant's response,
                or `None` if there is nothing to send (e.g. the `ModelResponse` had no parts because
                the model returned an empty response). Returning `None` ensures we don't emit an
                assistant message with `content=None` and no `tool_calls`, which the Chat Completions
                API rejects with a 400 error.
            """
            if not self.texts and not self.thinkings and not self.tool_calls:
                return None
            message_param = chat.ChatCompletionAssistantMessageParam(role='assistant')
            # Note: model responses from this model should only have one text item, so the following
            # shouldn't merge multiple texts into one unless you switch models between runs:
            if self.thinkings:
                for field_name, contents in self.thinkings.items():
                    message_param[field_name] = '\n\n'.join(contents)
            if self.texts:
                message_param['content'] = '\n\n'.join(self.texts)
            else:
                message_param['content'] = None
            if self.tool_calls:
                message_param['tool_calls'] = self.tool_calls
            return message_param

        def _map_response_text_part(self, item: TextPart) -> None:
            """Maps a `TextPart` to the response context.

            This method serves as a hook that can be overridden by subclasses
            to implement custom logic for handling text parts.
            """
            self.texts.append(item.content)

        def _map_response_thinking_part(self, item: ThinkingPart) -> None:
            """Maps a `ThinkingPart` to the response context.

            This method serves as a hook that can be overridden by subclasses
            to implement custom logic for handling thinking parts.
            """
            profile = self._model.profile
            include_method = profile.get('openai_chat_send_back_thinking_parts', 'auto')

            # Auto-detect: if thinking came from a custom field and from the same provider, use field mode
            # id='content' means it came from tags in content, not a custom field
            if include_method == 'auto':
                # Check if thinking came from a custom field from the same provider
                custom_field = profile.get('openai_chat_thinking_field', None)
                matches_custom_field = (not custom_field) or (item.id == custom_field)

                if (
                    item.id
                    and item.id != 'content'
                    and item.provider_name == self._model.system
                    and matches_custom_field
                ):
                    # Store both content and field name for later use in _into_message_param
                    self.thinkings.setdefault(item.id, []).append(item.content)
                else:
                    # Fall back to tags mode
                    start_tag, end_tag = self._model.profile.get('thinking_tags', DEFAULT_THINKING_TAGS)
                    self.texts.append('\n'.join([start_tag, item.content, end_tag]))
            elif include_method == 'tags':
                start_tag, end_tag = self._model.profile.get('thinking_tags', DEFAULT_THINKING_TAGS)
                self.texts.append('\n'.join([start_tag, item.content, end_tag]))
            elif include_method == 'field':
                field = profile.get('openai_chat_thinking_field', None)
                if field:  # pragma: no branch
                    self.thinkings.setdefault(field, []).append(item.content)

        def _map_response_tool_call_part(self, item: ToolCallPart) -> None:
            """Maps a `ToolCallPart` to the response context.

            This method serves as a hook that can be overridden by subclasses
            to implement custom logic for handling tool call parts.
            """
            self.tool_calls.append(self._model._map_tool_call(item))

        def _map_response_builtin_part(self, item: NativeToolCallPart | NativeToolReturnPart) -> None:
            """Maps a built-in tool call or return part to the response context.

            This method serves as a hook that can be overridden by subclasses
            to implement custom logic for handling built-in tool parts.
            """
            # OpenAI doesn't return built-in tool calls
            pass

        def _map_response_file_part(self, item: FilePart) -> None:
            """Maps a `FilePart` to the response context.

            This method serves as a hook that can be overridden by subclasses
            to implement custom logic for handling file parts.
            """
            # Files generated by models are not sent back to models that don't themselves generate files.
            pass

    def _map_model_response(self, message: ModelResponse) -> chat.ChatCompletionMessageParam | None:
        """Hook that determines how `ModelResponse` is mapped into `ChatCompletionMessageParam` objects before sending.

        Returns `None` to skip emitting any message for this `ModelResponse` (e.g. when it has no parts).

        Subclasses of `OpenAIChatModel` may override this method to provide their own mapping logic.
        """
        return self._MapModelResponseContext(self).map_assistant_message(message)

    def _map_finish_reason(
        self, key: Literal['stop', 'length', 'tool_calls', 'content_filter', 'function_call']
    ) -> FinishReason | None:
        """Hooks that maps a finish reason key to a [FinishReason][pydantic_ai.messages.FinishReason].

        This method may be overridden by subclasses of `OpenAIChatModel` to accommodate custom keys.
        """
        return _CHAT_FINISH_REASON_MAP.get(key)

    async def _map_messages(
        self, messages: Sequence[ModelMessage], model_request_parameters: ModelRequestParameters
    ) -> list[chat.ChatCompletionMessageParam]:
        """Just maps a `pydantic_ai.Message` to a `openai.types.ChatCompletionMessageParam`."""
        openai_messages: list[chat.ChatCompletionMessageParam] = []
        for message in messages:
            if isinstance(message, ModelRequest):
                async for item in self._map_user_message(message):
                    openai_messages.append(item)
            elif isinstance(message, ModelResponse):
                if (mapped := self._map_model_response(message)) is not None:
                    openai_messages.append(mapped)
            else:
                assert_never(message)
        system_prompt_role = self.profile.get('openai_system_prompt_role', None) or 'system'
        if instruction_parts := self._get_instruction_parts(messages, model_request_parameters):
            system_prompt_count = next(
                (i for i, m in enumerate(openai_messages) if m.get('role') != system_prompt_role), len(openai_messages)
            )
            if system_prompt_role == 'developer':
                instruction_messages: list[chat.ChatCompletionMessageParam] = [
                    chat.ChatCompletionDeveloperMessageParam(role='developer', content=part.content)
                    for part in instruction_parts
                ]
            elif system_prompt_role == 'user':
                instruction_messages = [
                    chat.ChatCompletionUserMessageParam(role='user', content=part.content) for part in instruction_parts
                ]
            else:
                instruction_messages = [
                    chat.ChatCompletionSystemMessageParam(role='system', content=part.content)
                    for part in instruction_parts
                ]
            openai_messages[system_prompt_count:system_prompt_count] = instruction_messages
        if not self.profile.get('openai_chat_supports_multiple_system_messages', True):
            openai_messages = _merge_leading_system_messages(openai_messages, system_prompt_role)
        return openai_messages

    @staticmethod
    def _map_tool_call(t: ToolCallPart) -> ChatCompletionMessageFunctionToolCallParam:
        return ChatCompletionMessageFunctionToolCallParam(
            id=_guard_tool_call_id(t=t),
            type='function',
            function={'name': t.tool_name, 'arguments': t.args_as_json_str()},
        )

    def _map_json_schema(self, o: OutputObjectDefinition) -> chat.completion_create_params.ResponseFormat:
        response_format_param: chat.completion_create_params.ResponseFormatJSONSchema = {  # pyright: ignore[reportPrivateImportUsage]
            'type': 'json_schema',
            'json_schema': {'name': o.name or DEFAULT_OUTPUT_TOOL_NAME, 'schema': o.json_schema},
        }
        if o.description:
            response_format_param['json_schema']['description'] = o.description
        if self.profile.get('openai_supports_strict_tool_definition', True):  # pragma: no branch
            response_format_param['json_schema']['strict'] = o.strict
        return response_format_param

    def _map_tool_definition(self, f: ToolDefinition) -> chat.ChatCompletionToolParam:
        tool_param: chat.ChatCompletionToolParam = {
            'type': 'function',
            'function': {
                'name': f.name,
                'description': f.description or '',
                'parameters': f.parameters_json_schema,
            },
        }
        if f.strict and self.profile.get('openai_supports_strict_tool_definition', True):
            tool_param['function']['strict'] = f.strict
        return tool_param

    async def _map_user_message(self, message: ModelRequest) -> AsyncIterable[chat.ChatCompletionMessageParam]:
        file_content: list[UserContent] = []
        for part in message.parts:
            if isinstance(part, SystemPromptPart):
                system_prompt_role = self.profile.get('openai_system_prompt_role', None)
                if system_prompt_role == 'developer':
                    yield chat.ChatCompletionDeveloperMessageParam(role='developer', content=part.content)
                elif system_prompt_role == 'user':
                    yield chat.ChatCompletionUserMessageParam(role='user', content=part.content)
                else:
                    yield chat.ChatCompletionSystemMessageParam(role='system', content=part.content)
            elif isinstance(part, UserPromptPart):
                yield await self._map_user_prompt(part)
            elif isinstance(part, ToolReturnPart):
                tool_text, tool_file_content = part.model_response_str_and_user_content()
                file_content.extend(tool_file_content)
                yield chat.ChatCompletionToolMessageParam(
                    role='tool',
                    tool_call_id=_guard_tool_call_id(t=part),
                    content=tool_text,
                )
            elif isinstance(part, RetryPromptPart):
                if part.tool_name is None:
                    yield chat.ChatCompletionUserMessageParam(role='user', content=part.model_response())
                else:
                    yield chat.ChatCompletionToolMessageParam(
                        role='tool',
                        tool_call_id=_guard_tool_call_id(t=part),
                        content=part.model_response(),
                    )
            else:
                assert_never(part)
        if file_content:
            yield await self._map_user_prompt(UserPromptPart(content=file_content))

    async def _map_image_url_item(self, item: ImageUrl) -> ChatCompletionContentPartImageParam:
        """Map an ImageUrl to a chat completion image content part."""
        image_url: ImageURL = {'url': item.url}
        if metadata := item.vendor_metadata:
            image_url['detail'] = metadata.get('detail', 'auto')
        if item.force_download:
            image_content = await download_item(item, data_format='base64_uri', type_format='extension')
            image_url['url'] = image_content['data']
        return ChatCompletionContentPartImageParam(image_url=image_url, type='image_url')

    async def _map_binary_content_item(self, item: BinaryContent) -> ChatCompletionContentPartParam:
        """Map a BinaryContent item to a chat completion content part."""
        profile = self.profile
        if _is_text_like_media_type(item.media_type):
            # Inline text-like binary content as a text block
            return self._inline_text_file_part(
                item.data.decode('utf-8'),
                media_type=item.media_type,
                identifier=item.identifier,
            )
        elif item.is_image:
            image_url = ImageURL(url=item.data_uri)
            if metadata := item.vendor_metadata:
                image_url['detail'] = metadata.get('detail', 'auto')
            return ChatCompletionContentPartImageParam(image_url=image_url, type='image_url')
        elif item.is_audio:
            assert item.format in ('wav', 'mp3')
            if profile.get('openai_chat_audio_input_encoding', 'base64') == 'uri':
                audio = InputAudio(data=item.data_uri, format=item.format)
            else:
                audio = InputAudio(data=item.base64, format=item.format)
            return ChatCompletionContentPartInputAudioParam(input_audio=audio, type='input_audio')
        elif item.is_document:
            if not profile.get('openai_chat_supports_document_input', True):
                self._raise_document_input_not_supported_error()
            return File(
                file=FileFile(
                    file_data=item.data_uri,
                    filename=f'filename.{item.format}',
                ),
                type='file',
            )
        elif item.is_video:
            raise NotImplementedError('VideoUrl is not supported in OpenAI Chat Completions user prompts')
        else:  # pragma: no cover
            raise RuntimeError(f'Unsupported binary content type: {item.media_type}')

    async def _map_audio_url_item(self, item: AudioUrl) -> ChatCompletionContentPartInputAudioParam:
        """Map an AudioUrl to a chat completion audio content part."""
        profile = self.profile
        data_format = 'base64_uri' if profile.get('openai_chat_audio_input_encoding', 'base64') == 'uri' else 'base64'
        downloaded_item = await download_item(item, data_format=data_format, type_format='extension')
        assert downloaded_item['data_type'] in (
            'wav',
            'mp3',
        ), f'Unsupported audio format: {downloaded_item["data_type"]}'
        audio = InputAudio(data=downloaded_item['data'], format=downloaded_item['data_type'])
        return ChatCompletionContentPartInputAudioParam(input_audio=audio, type='input_audio')

    async def _map_document_url_item(self, item: DocumentUrl) -> ChatCompletionContentPartParam:
        """Map a DocumentUrl to a chat completion content part."""
        profile = self.profile
        # OpenAI Chat API's FileFile only supports base64-encoded data, not URLs.
        # Some providers (e.g., OpenRouter) support URLs via the profile flag.
        if not item.force_download and profile.get('openai_chat_supports_file_urls', False):
            return File(
                file=FileFile(
                    file_data=item.url,
                    filename=f'filename.{item.format}',
                ),
                type='file',
            )
        if _is_text_like_media_type(item.media_type):
            downloaded_text = await download_item(item, data_format='text')
            return self._inline_text_file_part(
                downloaded_text['data'],
                media_type=item.media_type,
                identifier=item.identifier,
            )
        else:
            if not profile.get('openai_chat_supports_document_input', True):
                self._raise_document_input_not_supported_error()
            downloaded_item = await download_item(item, data_format='base64_uri', type_format='extension')
            return File(
                file=FileFile(
                    file_data=downloaded_item['data'],
                    filename=f'filename.{downloaded_item["data_type"]}',
                ),
                type='file',
            )

    async def _map_video_url_item(self, item: VideoUrl) -> ChatCompletionContentPartParam:
        """Map a VideoUrl to a chat completion content part."""
        raise NotImplementedError('VideoUrl is not supported in OpenAI Chat Completions user prompts')

    async def _map_content_item(
        self,
        item: str
        | TextContent
        | ImageUrl
        | BinaryContent
        | AudioUrl
        | DocumentUrl
        | VideoUrl
        | UploadedFile
        | CachePoint,
    ) -> ChatCompletionContentPartParam | None:
        """Map a single content item to a chat completion content part, or None to filter it out."""
        if isinstance(item, str | TextContent):
            text = item if isinstance(item, str) else item.content
            return ChatCompletionContentPartTextParam(text=text, type='text')
        elif isinstance(item, ImageUrl):
            return await self._map_image_url_item(item)
        elif isinstance(item, BinaryContent):
            return await self._map_binary_content_item(item)
        elif isinstance(item, AudioUrl):
            return await self._map_audio_url_item(item)
        elif isinstance(item, DocumentUrl):
            return await self._map_document_url_item(item)
        elif isinstance(item, VideoUrl):
            return await self._map_video_url_item(item)
        elif isinstance(item, UploadedFile):
            # Verify provider matches
            if item.provider_name != self.system:
                raise UserError(
                    f'UploadedFile with `provider_name={item.provider_name!r}` cannot be used with OpenAIChatModel. '
                    f'Expected `provider_name` to be `{self.system!r}`.'
                )
            return File(
                file=FileFile(file_id=item.file_id),
                type='file',
            )
        elif isinstance(item, CachePoint):
            # OpenAI doesn't support prompt caching via CachePoint, so we filter it out
            return None
        else:
            assert_never(item)

    async def _map_user_prompt(self, part: UserPromptPart) -> chat.ChatCompletionUserMessageParam:
        content: str | list[ChatCompletionContentPartParam]
        if isinstance(part.content, str):
            content = part.content
        else:
            content = []
            for item in part.content:
                mapped_item = await self._map_content_item(item)
                if mapped_item is not None:
                    content.append(mapped_item)
        return chat.ChatCompletionUserMessageParam(role='user', content=content)

    def _raise_document_input_not_supported_error(self) -> Never:
        if self._provider.name == 'azure':
            raise UserError(
                "Azure's Chat Completions API does not support document input. "
                'Use `OpenAIResponsesModel` with `AzureProvider` instead.'
            )
        raise UserError(
            f'The {self._provider.name!r} provider does not support document input via the Chat Completions API.'
        )

    @staticmethod
    def _inline_text_file_part(text: str, *, media_type: str, identifier: str) -> ChatCompletionContentPartTextParam:
        text = '\n'.join(
            [
                f'-----BEGIN FILE id="{identifier}" type="{media_type}"-----',
                text,
                f'-----END FILE id="{identifier}"-----',
            ]
        )
        return ChatCompletionContentPartTextParam(text=text, type='text')

__init__ 

__init__(
    model_name: OpenAIModelName,
    *,
    provider: (
        OpenAIChatCompatibleProvider
        | Literal["openai", "openai-chat", "gateway"]
        | Provider[AsyncOpenAI]
    ) = "openai",
    profile: ModelProfileSpec | None = None,
    settings: ModelSettings | None = None
)

Initialize an OpenAI model.

Parameters:

Name Type Description Default
model_name OpenAIModelName

The name of the OpenAI model to use. List of model names available here (Unfortunately, despite being ask to do so, OpenAI do not provide .inv files for their API).

 required
provider OpenAIChatCompatibleProvider | Literal['openai', 'openai-chat', 'gateway'] | Provider[AsyncOpenAI]

The provider to use. Defaults to 'openai'.

 'openai'
profile ModelProfileSpec | None

The model profile to use. Defaults to a profile picked by the provider based on the model name.

 None
settings ModelSettings | None

Default model settings for this model instance.

 None
Source code in pydantic_ai_slim/pydantic_ai/models/openai.py 
728
729
730
731
732
733
734
735
736
737
738
739
740
741
742
743
744
745
746
747
748
749
750
751
752
753
754
755
756
757
758
759
760
def __init__(
    self,
    model_name: OpenAIModelName,
    *,
    provider: OpenAIChatCompatibleProvider
    | Literal[
        'openai',
        'openai-chat',
        'gateway',
    ]
    | Provider[AsyncOpenAI] = 'openai',
    profile: ModelProfileSpec | None = None,
    settings: ModelSettings | None = None,
):
    """Initialize an OpenAI model.

    Args:
        model_name: The name of the OpenAI model to use. List of model names available
            [here](https://github.com/openai/openai-python/blob/v1.54.3/src/openai/types/chat_model.py#L7)
            (Unfortunately, despite being ask to do so, OpenAI do not provide `.inv` files for their API).
        provider: The provider to use. Defaults to `'openai'`.
        profile: The model profile to use. Defaults to a profile picked by the provider based on the model name.
        settings: Default model settings for this model instance.
    """
    self._model_name = model_name

    if isinstance(provider, str):
        provider = infer_provider('gateway/openai' if provider == 'gateway' else provider)
    self._provider = provider

    super().__init__(settings=settings, profile=profile)

    validate_openai_profile(self.profile)

model_name property 

model_name: OpenAIModelName

The model name.

system property 

system: str

The model provider.

supported_native_tools classmethod 

supported_native_tools() -> (
    frozenset[type[AbstractNativeTool]]
)

Return the set of builtin tool types this model can handle.

Source code in pydantic_ai_slim/pydantic_ai/models/openai.py 
780
781
782
783
@classmethod
def supported_native_tools(cls) -> frozenset[type[AbstractNativeTool]]:
    """Return the set of builtin tool types this model can handle."""
    return frozenset({WebSearchTool})

profile cached property 

profile: OpenAIModelProfile

The model profile.

WebSearchTool is only supported if openai_chat_supports_web_search is True.

OpenAIResponsesModel dataclass

Bases: Model[AsyncOpenAI]

A model that uses the OpenAI Responses API.

The OpenAI Responses API is the new API for OpenAI models.

If you are interested in the differences between the Responses API and the Chat Completions API, see the OpenAI API docs.

Source code in pydantic_ai_slim/pydantic_ai/models/openai.py 
1640
1641
1642
1643
1644
1645
1646
1647
1648
1649
1650
1651
1652
1653
1654
1655
1656
1657
1658
1659
1660
1661
1662
1663
1664
1665
1666
1667
1668
1669
1670
1671
1672
1673
1674
1675
1676
1677
1678
1679
1680
1681
1682
1683
1684
1685
1686
1687
1688
1689
1690
1691
1692
1693
1694
1695
1696
1697
1698
1699
1700
1701
1702
1703
1704
1705
1706
1707
1708
1709
1710
1711
1712
1713
1714
1715
1716
1717
1718
1719
1720
1721
1722
1723
1724
1725
1726
1727
1728
1729
1730
1731
1732
1733
1734
1735
1736
1737
1738
1739
1740
1741
1742
1743
1744
1745
1746
1747
1748
1749
1750
1751
1752
1753
1754
1755
1756
1757
1758
1759
1760
1761
1762
1763
1764
1765
1766
1767
1768
1769
1770
1771
1772
1773
1774
1775
1776
1777
1778
1779
1780
1781
1782
1783
1784
1785
1786
1787
1788
1789
1790
1791
1792
1793
1794
1795
1796
1797
1798
1799
1800
1801
1802
1803
1804
1805
1806
1807
1808
1809
1810
1811
1812
1813
1814
1815
1816
1817
1818
1819
1820
1821
1822
1823
1824
1825
1826
1827
1828
1829
1830
1831
1832
1833
1834
1835
1836
1837
1838
1839
1840
1841
1842
1843
1844
1845
1846
1847
1848
1849
1850
1851
1852
1853
1854
1855
1856
1857
1858
1859
1860
1861
1862
1863
1864
1865
1866
1867
1868
1869
1870
1871
1872
1873
1874
1875
1876
1877
1878
1879
1880
1881
1882
1883
1884
1885
1886
1887
1888
1889
1890
1891
1892
1893
1894
1895
1896
1897
1898
1899
1900
1901
1902
1903
1904
1905
1906
1907
1908
1909
1910
1911
1912
1913
1914
1915
1916
1917
1918
1919
1920
1921
1922
1923
1924
1925
1926
1927
1928
1929
1930
1931
1932
1933
1934
1935
1936
1937
1938
1939
1940
1941
1942
1943
1944
1945
1946
1947
1948
1949
1950
1951
1952
1953
1954
1955
1956
1957
1958
1959
1960
1961
1962
1963
1964
1965
1966
1967
1968
1969
1970
1971
1972
1973
1974
1975
1976
1977
1978
1979
1980
1981
1982
1983
1984
1985
1986
1987
1988
1989
1990
1991
1992
1993
1994
1995
1996
1997
1998
1999
2000
2001
2002
2003
2004
2005
2006
2007
2008
2009
2010
2011
2012
2013
2014
2015
2016
2017
2018
2019
2020
2021
2022
2023
2024
2025
2026
2027
2028
2029
2030
2031
2032
2033
2034
2035
2036
2037
2038
2039
2040
2041
2042
2043
2044
2045
2046
2047
2048
2049
2050
2051
2052
2053
2054
2055
2056
2057
2058
2059
2060
2061
2062
2063
2064
2065
2066
2067
2068
2069
2070
2071
2072
2073
2074
2075
2076
2077
2078
2079
2080
2081
2082
2083
2084
2085
2086
2087
2088
2089
2090
2091
2092
2093
2094
2095
2096
2097
2098
2099
2100
2101
2102
2103
2104
2105
2106
2107
2108
2109
2110
2111
2112
2113
2114
2115
2116
2117
2118
2119
2120
2121
2122
2123
2124
2125
2126
2127
2128
2129
2130
2131
2132
2133
2134
2135
2136
2137
2138
2139
2140
2141
2142
2143
2144
2145
2146
2147
2148
2149
2150
2151
2152
2153
2154
2155
2156
2157
2158
2159
2160
2161
2162
2163
2164
2165
2166
2167
2168
2169
2170
2171
2172
2173
2174
2175
2176
2177
2178
2179
2180
2181
2182
2183
2184
2185
2186
2187
2188
2189
2190
2191
2192
2193
2194
2195
2196
2197
2198
2199
2200
2201
2202
2203
2204
2205
2206
2207
2208
2209
2210
2211
2212
2213
2214
2215
2216
2217
2218
2219
2220
2221
2222
2223
2224
2225
2226
2227
2228
2229
2230
2231
2232
2233
2234
2235
2236
2237
2238
2239
2240
2241
2242
2243
2244
2245
2246
2247
2248
2249
2250
2251
2252
2253
2254
2255
2256
2257
2258
2259
2260
2261
2262
2263
2264
2265
2266
2267
2268
2269
2270
2271
2272
2273
2274
2275
2276
2277
2278
2279
2280
2281
2282
2283
2284
2285
2286
2287
2288
2289
2290
2291
2292
2293
2294
2295
2296
2297
2298
2299
2300
2301
2302
2303
2304
2305
2306
2307
2308
2309
2310
2311
2312
2313
2314
2315
2316
2317
2318
2319
2320
2321
2322
2323
2324
2325
2326
2327
2328
2329
2330
2331
2332
2333
2334
2335
2336
2337
2338
2339
2340
2341
2342
2343
2344
2345
2346
2347
2348
2349
2350
2351
2352
2353
2354
2355
2356
2357
2358
2359
2360
2361
2362
2363
2364
2365
2366
2367
2368
2369
2370
2371
2372
2373
2374
2375
2376
2377
2378
2379
2380
2381
2382
2383
2384
2385
2386
2387
2388
2389
2390
2391
2392
2393
2394
2395
2396
2397
2398
2399
2400
2401
2402
2403
2404
2405
2406
2407
2408
2409
2410
2411
2412
2413
2414
2415
2416
2417
2418
2419
2420
2421
2422
2423
2424
2425
2426
2427
2428
2429
2430
2431
2432
2433
2434
2435
2436
2437
2438
2439
2440
2441
2442
2443
2444
2445
2446
2447
2448
2449
2450
2451
2452
2453
2454
2455
2456
2457
2458
2459
2460
2461
2462
2463
2464
2465
2466
2467
2468
2469
2470
2471
2472
2473
2474
2475
2476
2477
2478
2479
2480
2481
2482
2483
2484
2485
2486
2487
2488
2489
2490
2491
2492
2493
2494
2495
2496
2497
2498
2499
2500
2501
2502
2503
2504
2505
2506
2507
2508
2509
2510
2511
2512
2513
2514
2515
2516
2517
2518
2519
2520
2521
2522
2523
2524
2525
2526
2527
2528
2529
2530
2531
2532
2533
2534
2535
2536
2537
2538
2539
2540
2541
2542
2543
2544
2545
2546
2547
2548
2549
2550
2551
2552
2553
2554
2555
2556
2557
2558
2559
2560
2561
2562
2563
2564
2565
2566
2567
2568
2569
2570
2571
2572
2573
2574
2575
2576
2577
2578
2579
2580
2581
2582
2583
2584
2585
2586
2587
2588
2589
2590
2591
2592
2593
2594
2595
2596
2597
2598
2599
2600
2601
2602
2603
2604
2605
2606
2607
2608
2609
2610
2611
2612
2613
2614
2615
2616
2617
2618
2619
2620
2621
2622
2623
2624
2625
2626
2627
2628
2629
2630
2631
2632
2633
2634
2635
2636
2637
2638
2639
2640
2641
2642
2643
2644
2645
2646
2647
2648
2649
2650
2651
2652
2653
2654
2655
2656
2657
2658
2659
2660
2661
2662
2663
2664
2665
2666
2667
2668
2669
2670
2671
2672
2673
2674
2675
2676
2677
2678
2679
2680
2681
2682
2683
2684
2685
2686
2687
2688
2689
2690
2691
2692
2693
2694
2695
2696
2697
2698
2699
2700
2701
2702
2703
2704
2705
2706
2707
2708
2709
2710
2711
2712
2713
2714
2715
2716
2717
2718
2719
2720
2721
2722
2723
2724
2725
2726
2727
2728
2729
2730
2731
2732
2733
2734
2735
2736
2737
2738
2739
2740
2741
2742
2743
2744
2745
2746
2747
2748
2749
2750
2751
2752
2753
2754
2755
2756
2757
2758
2759
2760
2761
2762
2763
2764
2765
2766
2767
2768
2769
2770
2771
2772
2773
2774
2775
2776
2777
2778
2779
2780
2781
2782
2783
2784
2785
2786
2787
2788
2789
2790
2791
2792
2793
2794
2795
2796
2797
2798
2799
2800
2801
2802
2803
2804
2805
2806
2807
2808
2809
2810
2811
2812
2813
2814
2815
2816
2817
2818
2819
2820
2821
2822
2823
2824
2825
2826
2827
2828
2829
2830
2831
2832
2833
2834
2835
2836
2837
2838
2839
2840
2841
2842
2843
2844
2845
2846
2847
2848
2849
2850
2851
2852
2853
2854
2855
2856
2857
2858
2859
2860
2861
2862
2863
2864
2865
2866
2867
2868
2869
2870
2871
2872
2873
2874
2875
2876
2877
2878
2879
2880
2881
2882
2883
2884
2885
2886
2887
2888
2889
2890
2891
2892
2893
2894
2895
2896
2897
2898
2899
2900
2901
2902
2903
2904
2905
2906
2907
2908
2909
2910
2911
2912
2913
2914
2915
2916
2917
2918
2919
2920
2921
2922
2923
2924
2925
2926
2927
2928
2929
2930
2931
2932
2933
2934
2935
2936
2937
2938
2939
2940
2941
2942
2943
2944
2945
2946
2947
2948
2949
2950
2951
2952
2953
2954
2955
2956
2957
2958
2959
2960
2961
2962
2963
2964
2965
2966
2967
2968
2969
2970
2971
2972
2973
2974
2975
2976
2977
2978
2979
2980
2981
2982
2983
2984
2985
2986
2987
2988
2989
2990
2991
2992
2993
2994
2995
2996
2997
2998
2999
3000
3001
3002
3003
3004
3005
3006
3007
3008
3009
3010
3011
3012
3013
3014
3015
3016
3017
3018
3019
3020
3021
3022
3023
3024
3025
3026
3027
3028
3029
3030
3031
3032
3033
3034
3035
3036
3037
3038
3039
3040
3041
3042
3043
3044
3045
3046
3047
3048
3049
3050
3051
3052
3053
3054
3055
3056
3057
3058
3059
3060
3061
3062
3063
3064
3065
3066
3067
3068
3069
3070
@dataclass(init=False)
class OpenAIResponsesModel(Model[AsyncOpenAI]):
    """A model that uses the OpenAI Responses API.

    The [OpenAI Responses API](https://platform.openai.com/docs/api-reference/responses) is the
    new API for OpenAI models.

    If you are interested in the differences between the Responses API and the Chat Completions API,
    see the [OpenAI API docs](https://platform.openai.com/docs/guides/responses-vs-chat-completions).
    """

    _model_name: OpenAIModelName = field(repr=False)
    _provider: Provider[AsyncOpenAI] = field(repr=False)

    def __init__(
        self,
        model_name: OpenAIModelName,
        *,
        provider: OpenAIResponsesCompatibleProvider
        | Literal[
            'openai',
            'gateway',
        ]
        | Provider[AsyncOpenAI] = 'openai',
        profile: ModelProfileSpec | None = None,
        settings: ModelSettings | None = None,
    ):
        """Initialize an OpenAI Responses model.

        Args:
            model_name: The name of the OpenAI model to use.
            provider: The provider to use. Defaults to `'openai'`.
            profile: The model profile to use. Defaults to a profile picked by the provider based on the model name.
            settings: Default model settings for this model instance.
        """
        self._model_name = model_name

        if isinstance(provider, str):
            provider = infer_provider('gateway/openai' if provider == 'gateway' else provider)
        self._provider = provider

        super().__init__(settings=settings, profile=profile)

    @property
    def client(self) -> AsyncOpenAI:
        return self._provider.client

    @property
    def base_url(self) -> str:
        return str(self.client.base_url)

    @property
    def model_name(self) -> OpenAIModelName:
        """The model name."""
        return self._model_name

    @property
    def system(self) -> str:
        """The model provider."""
        return self._provider.name

    @cached_property
    def profile(self) -> OpenAIModelProfile:
        return cast(OpenAIModelProfile, super().profile)

    @classmethod
    def supported_native_tools(cls) -> frozenset[type[AbstractNativeTool]]:
        """Return the set of builtin tool types this model can handle."""
        return frozenset(
            {WebSearchTool, CodeExecutionTool, FileSearchTool, MCPServerTool, ImageGenerationTool, ToolSearchTool}
        )

    async def compact_messages(
        self,
        request_context: ModelRequestContext,
        *,
        instructions: str | None = None,
    ) -> ModelResponse:
        """Compact messages using the OpenAI Responses compaction endpoint.

        This calls OpenAI's `responses.compact` API to produce an encrypted compaction
        that summarizes the conversation history. The returned `ModelResponse` contains
        a single `CompactionPart` that must be round-tripped in subsequent requests.

        Args:
            request_context: The model request context containing messages, settings, and parameters.
            instructions: Optional custom instructions for the compaction summarization.
                If provided, these override the agent-level instructions.

        Returns:
            A `ModelResponse` with a single `CompactionPart` containing the encrypted compaction data.
        """
        check_allow_model_requests()
        model_settings, model_request_parameters = self.prepare_request(
            request_context.model_settings,
            request_context.model_request_parameters,
        )
        response = await self._responses_compact(
            request_context.messages,
            cast(OpenAIResponsesModelSettings, model_settings or {}),
            model_request_parameters,
            instructions_override=instructions,
        )

        # Handle ModelResponse (e.g. from content filter)
        if isinstance(response, ModelResponse):  # pragma: no cover
            return response

        if not response.output:  # pragma: no cover
            raise UnexpectedModelBehavior('CompactedResponse returned with no output items')

        compaction = response.output[-1]
        if not isinstance(compaction, ResponseCompactionItem):  # pragma: no cover
            raise UnexpectedModelBehavior(f'Last item in response is not a compaction, got: {compaction.type}')

        part = _map_compaction_item(compaction, self.system)
        return ModelResponse(
            parts=[part],
            usage=_map_usage(response, self._provider.name, self._provider.base_url, self.model_name),
            model_name=self._model_name,
            provider_response_id=response.id,
            # Marks this ModelResponse as coming from the stateless `/compact` endpoint.
            # `_get_previous_response_id_and_new_messages` uses this to break the auto-chain,
            # since compaction response ids cannot be used as `previous_response_id`.
            provider_details={'compaction': True},
            timestamp=_now_utc(),
            provider_name=self._provider.name,
            provider_url=self._provider.base_url,
        )

    async def _responses_compact(
        self,
        messages: list[ModelMessage],
        model_settings: OpenAIResponsesModelSettings,
        model_request_parameters: ModelRequestParameters,
        *,
        instructions_override: str | None = None,
    ) -> responses.CompactedResponse | ModelResponse:
        """Call the OpenAI Responses compaction endpoint."""
        previous_response_id, messages = self._resolve_previous_response_id(
            model_settings.get('openai_previous_response_id'), messages, allow_no_new_messages=True
        )

        instructions, openai_messages = await self._map_messages(messages, model_settings, model_request_parameters)
        if instructions_override is not None:
            instructions = instructions_override

        try:
            return await self.client.responses.compact(
                input=openai_messages,
                model=self.model_name,
                instructions=instructions,
                previous_response_id=previous_response_id or OMIT,
            )
        except APIStatusError as e:  # pragma: no cover
            if model_response := _check_azure_content_filter(e, self.system, self.model_name):
                return model_response
            if (status_code := e.status_code) >= 400:
                raise ModelHTTPError(status_code=status_code, model_name=self.model_name, body=e.body) from e
            raise
        except APIConnectionError as e:  # pragma: no cover
            raise ModelAPIError(model_name=self.model_name, message=e.message) from e

    async def request(
        self,
        messages: list[ModelRequest | ModelResponse],
        model_settings: ModelSettings | None,
        model_request_parameters: ModelRequestParameters,
    ) -> ModelResponse:
        check_allow_model_requests()
        model_settings, model_request_parameters = self.prepare_request(
            model_settings,
            model_request_parameters,
        )
        response = await self._responses_create(
            messages, False, cast(OpenAIResponsesModelSettings, model_settings or {}), model_request_parameters
        )

        # Handle ModelResponse
        if isinstance(response, ModelResponse):
            return response

        return self._process_response(
            response, cast(OpenAIResponsesModelSettings, model_settings or {}), model_request_parameters
        )

    async def count_tokens(
        self,
        messages: list[ModelRequest | ModelResponse],
        model_settings: ModelSettings | None,
        model_request_parameters: ModelRequestParameters,
    ) -> usage.RequestUsage:
        check_allow_model_requests()
        model_settings, model_request_parameters = self.prepare_request(
            model_settings,
            model_request_parameters,
        )
        settings = cast(OpenAIResponsesModelSettings, model_settings or {})

        # Validate that we have something meaningful to count tokens for.
        if (
            not messages
            and not settings.get('openai_previous_response_id')
            and not settings.get('openai_conversation_id')
        ):
            raise UserError('Cannot count tokens without any messages or a previous response ID.')

        profile = self.profile
        request_params = await self._build_responses_request_params(
            messages,
            settings,
            model_request_parameters,
            profile,
        )

        extra_headers, timeout = self._build_request_options(settings)
        with _map_api_errors(self.model_name):
            response = await self.client.responses.input_tokens.count(
                model=request_params.model,
                input=request_params.input,
                instructions=request_params.instructions,
                parallel_tool_calls=request_params.parallel_tool_calls,
                tools=request_params.tools,
                tool_choice=request_params.tool_choice,
                previous_response_id=request_params.previous_response_id,
                conversation=request_params.conversation,
                reasoning=request_params.reasoning,
                text=request_params.text,
                truncation=request_params.truncation,
                extra_headers=extra_headers,
                extra_body=settings.get('extra_body'),
                timeout=timeout,
            )

        return usage.RequestUsage(
            input_tokens=response.input_tokens,
        )

    @asynccontextmanager
    async def request_stream(
        self,
        messages: list[ModelMessage],
        model_settings: ModelSettings | None,
        model_request_parameters: ModelRequestParameters,
        run_context: RunContext[Any] | None = None,
    ) -> AsyncIterator[StreamedResponse]:
        check_allow_model_requests()
        model_settings, model_request_parameters = self.prepare_request(
            model_settings,
            model_request_parameters,
        )
        response = await self._responses_create(
            messages, True, cast(OpenAIResponsesModelSettings, model_settings or {}), model_request_parameters
        )
        async with response:
            yield await self._process_streamed_response(
                response, cast(OpenAIResponsesModelSettings, model_settings or {}), model_request_parameters
            )

    def _process_response(  # noqa: C901
        self,
        response: responses.Response,
        model_settings: OpenAIResponsesModelSettings,
        model_request_parameters: ModelRequestParameters,
    ) -> ModelResponse:
        """Process a non-streamed response, and prepare a message to return."""
        items: list[ModelResponsePart] = []
        refusal_text: str | None = None
        # Index tool_search_output items by call_id so we can pair them with the
        # corresponding tool_search_call when we encounter it (they may come in either
        # order).
        tool_search_outputs: dict[str, responses.ResponseToolSearchOutputItem] = {
            output_item.call_id: output_item
            for output_item in response.output
            if isinstance(output_item, responses.ResponseToolSearchOutputItem) and output_item.call_id is not None
        }
        for item in response.output:
            if isinstance(item, responses.ResponseReasoningItem):
                signature = item.encrypted_content
                # Handle raw CoT content from gpt-oss models
                provider_details: dict[str, Any] = {}
                raw_content: list[str] | None = [c.text for c in item.content] if item.content else None
                if raw_content:
                    provider_details['raw_content'] = raw_content

                if item.summary:
                    for summary in item.summary:
                        # We use the same id for all summaries so that we can merge them on the round trip.
                        items.append(
                            ThinkingPart(
                                content=summary.text,
                                id=item.id,
                                signature=signature,
                                provider_name=self.system,
                                provider_details=provider_details or None,
                            )
                        )
                        # We only need to store the signature and raw_content once.
                        signature = None
                        provider_details = {}
                elif signature or provider_details:
                    items.append(
                        ThinkingPart(
                            content='',
                            id=item.id,
                            signature=signature,
                            provider_name=self.system,
                            provider_details=provider_details or None,
                        )
                    )
            elif isinstance(item, responses.ResponseOutputMessage):
                for content in item.content:
                    if isinstance(content, responses.ResponseOutputRefusal):
                        refusal_text = content.refusal
                    elif isinstance(content, responses.ResponseOutputText):  # pragma: no branch
                        part_provider_details: dict[str, Any] | None = None
                        if content.logprobs:
                            part_provider_details = {'logprobs': _map_logprobs(content.logprobs)}
                        if model_settings.get('openai_include_raw_annotations') and content.annotations:
                            part_provider_details = part_provider_details or {}
                            part_provider_details['annotations'] = responses_output_text_annotations_ta.dump_python(
                                list(content.annotations), warnings=False
                            )
                        if item.phase is not None:
                            part_provider_details = part_provider_details or {}
                            part_provider_details['phase'] = item.phase
                        # Some OpenAI-compatible gateways (e.g. Bifrost) return text=null;
                        # coalesce to '' so the part (and its ID) is preserved for round-tripping.
                        items.append(
                            TextPart(
                                content.text or '',
                                id=item.id,
                                provider_name=self.system,
                                provider_details=part_provider_details,
                            )
                        )
            elif isinstance(item, responses.ResponseFunctionToolCall):
                # When the Responses API wraps a discovered deferred tool in a namespace
                # (tool-search flow), it requires the namespace to be round-tripped on
                # subsequent turns. Preserve it via `provider_details`.
                fn_provider_details: dict[str, Any] | None = None
                if item.namespace:
                    fn_provider_details = {'namespace': item.namespace}
                items.append(
                    ToolCallPart(
                        item.name,
                        item.arguments,
                        tool_call_id=item.call_id,
                        id=item.id,
                        provider_name=self.system,
                        provider_details=fn_provider_details,
                    )
                )
            elif isinstance(item, responses.ResponseCodeInterpreterToolCall):
                call_part, return_part, file_parts = _map_code_interpreter_tool_call(item, self.system)
                items.append(call_part)
                if file_parts:
                    items.extend(file_parts)
                items.append(return_part)
            elif isinstance(item, responses.ResponseFunctionWebSearch):
                call_part, return_part = _map_web_search_tool_call(item, self.system)
                items.append(call_part)
                items.append(return_part)
            elif isinstance(item, responses.ResponseToolSearchCall):
                if item.execution == 'client':
                    # Client-executed: emit a regular ToolCallPart so the standard agent-graph
                    # tool-execution path runs the local `search_tools` function. The matching
                    # `ToolReturnPart` is produced by the tool runner, not here.
                    items.append(_map_client_tool_search_call(item, self.system))
                else:
                    output_item = tool_search_outputs.get(item.call_id) if item.call_id else None
                    call_part, return_part = _map_tool_search_call(item, output_item, self.system)
                    items.append(call_part)
                    items.append(return_part)
            elif isinstance(item, responses.ResponseToolSearchOutputItem):
                # The output item is paired with its call via `tool_search_outputs`; skip here.
                pass
            elif isinstance(item, responses.response_output_item.ImageGenerationCall):
                call_part, return_part, file_part = _map_image_generation_tool_call(item, self.system)
                items.append(call_part)
                if file_part:  # pragma: no branch
                    items.append(file_part)
                items.append(return_part)
            elif isinstance(item, ResponseCompactionItem):
                items.append(_map_compaction_item(item, self.system))
            elif isinstance(item, responses.ResponseComputerToolCall):  # pragma: no cover
                # Pydantic AI doesn't yet support the ComputerUse built-in tool
                pass
            elif isinstance(item, responses.ResponseCustomToolCall):  # pragma: no cover
                # Support is being implemented in https://github.com/pydantic/pydantic-ai/pull/2572
                pass
            elif isinstance(item, responses.response_output_item.LocalShellCall):  # pragma: no cover
                # Pydantic AI doesn't yet support the `codex-mini-latest` LocalShell built-in tool
                pass
            elif isinstance(item, responses.ResponseFileSearchToolCall):
                call_part, return_part = _map_file_search_tool_call(item, self.system)
                items.append(call_part)
                items.append(return_part)
            elif isinstance(item, responses.response_output_item.McpCall):
                call_part, return_part = _map_mcp_call(item, self.system)
                items.append(call_part)
                items.append(return_part)
            elif isinstance(item, responses.response_output_item.McpListTools):
                call_part, return_part = _map_mcp_list_tools(item, self.system)
                items.append(call_part)
                items.append(return_part)
            elif isinstance(item, responses.response_output_item.McpApprovalRequest):  # pragma: no cover
                # Pydantic AI doesn't yet support McpApprovalRequest (explicit tool usage approval)
                pass

        finish_reason: FinishReason | None = None
        provider_details: dict[str, Any] = {}
        raw_finish_reason = details.reason if (details := response.incomplete_details) else response.status
        if raw_finish_reason:
            provider_details['finish_reason'] = raw_finish_reason
            finish_reason = _RESPONSES_FINISH_REASON_MAP.get(raw_finish_reason)
        if response.created_at:  # pragma: no branch
            provider_details['timestamp'] = number_to_datetime(response.created_at)
        if response.conversation:
            provider_details['conversation_id'] = response.conversation.id

        if refusal_text is not None:
            items = []
            finish_reason = 'content_filter'
            provider_details.pop('finish_reason', None)
            provider_details['refusal'] = refusal_text

        return ModelResponse(
            parts=items,
            usage=_map_usage(response, self._provider.name, self._provider.base_url, self.model_name),
            model_name=response.model,
            provider_response_id=response.id,
            timestamp=_now_utc(),
            provider_name=self._provider.name,
            provider_url=self._provider.base_url,
            finish_reason=finish_reason,
            provider_details=provider_details or None,
        )

    async def _process_streamed_response(
        self,
        response: AsyncStream[responses.ResponseStreamEvent],
        model_settings: OpenAIResponsesModelSettings,
        model_request_parameters: ModelRequestParameters,
    ) -> OpenAIResponsesStreamedResponse:
        """Process a streamed response, and prepare a streaming response to return."""
        peekable_response: _utils.PeekableAsyncStream[
            responses.ResponseStreamEvent, AsyncStream[responses.ResponseStreamEvent]
        ] = _utils.PeekableAsyncStream(response)
        with _map_api_errors(self.model_name):
            first_chunk = await peekable_response.peek()
        if isinstance(first_chunk, _utils.Unset):  # pragma: no cover
            raise UnexpectedModelBehavior('Streamed response ended without content or tool calls')

        assert isinstance(first_chunk, responses.ResponseCreatedEvent)
        return OpenAIResponsesStreamedResponse(
            model_request_parameters=model_request_parameters,
            _model_name=first_chunk.response.model,
            _model_settings=model_settings,
            _response=peekable_response,
            _provider_name=self._provider.name,
            _provider_url=self._provider.base_url,
            _provider_timestamp=number_to_datetime(first_chunk.response.created_at)
            if first_chunk.response.created_at
            else None,
        )

    async def _build_responses_request_params(
        self,
        messages: list[ModelRequest | ModelResponse],
        model_settings: OpenAIResponsesModelSettings,
        model_request_parameters: ModelRequestParameters,
        profile: OpenAIModelProfile,
    ) -> _ResponsesRequestParams:
        """Build typed request parameters shared by Responses API calls."""
        function_tools, tool_choice = self._get_responses_tool_choice(model_settings, model_request_parameters)
        extra_native_tools = model_settings.get('openai_native_tools', ())
        tools: list[responses.ToolParam] = (
            self._get_native_tools(model_request_parameters) + list(extra_native_tools) + function_tools
        )
        if not tools:
            tool_choice = None

        previous_response_id, conversation_id, messages = self._resolve_server_side_state(model_settings, messages)

        instructions, openai_messages = await self._map_messages(messages, model_settings, model_request_parameters)
        reasoning = self._translate_thinking(model_settings, model_request_parameters)

        text: responses.ResponseTextConfigParam | Omit = OMIT
        if model_request_parameters.output_mode == 'native':
            output_object = model_request_parameters.output_object
            assert output_object is not None
            text = {'format': self._map_json_schema(output_object)}
        elif model_request_parameters.output_mode == 'prompted' and profile.get(
            'supports_json_object_output', False
        ):  # pragma: no branch
            text = {'format': {'type': 'json_object'}}

            # Without this trick, we'd hit this error:
            # > Response input messages must contain the word 'json' in some form to use 'text.format' of type 'json_object'.
            # Apparently they're only checking input messages for "JSON", not instructions.
            assert isinstance(instructions, str)
            system_prompt_role = profile.get('openai_system_prompt_role', None) or 'system'
            system_prompt_count = next(
                (i for i, m in enumerate(openai_messages) if m.get('role') != system_prompt_role), len(openai_messages)
            )
            openai_messages.insert(
                system_prompt_count, responses.EasyInputMessageParam(role=system_prompt_role, content=instructions)
            )
            instructions = OMIT

        if verbosity := model_settings.get('openai_text_verbosity'):
            text_with_verbosity: responses.ResponseTextConfigParam = text if isinstance(text, dict) else {}
            text_with_verbosity['verbosity'] = verbosity
            text = text_with_verbosity

        # When there are no input messages and we're not reusing server-side state,
        # the OpenAI API will reject a request without any input,
        # even if there are instructions.
        # To avoid this provide an explicit empty user message.
        if not openai_messages and not previous_response_id and not conversation_id:
            openai_messages.append(
                responses.EasyInputMessageParam(
                    role='user',
                    content='',
                )
            )

        return _ResponsesRequestParams(
            model=self.model_name,
            input=openai_messages,
            instructions=instructions,
            parallel_tool_calls=model_settings.get('parallel_tool_calls', OMIT) if tools else OMIT,
            tools=tools or OMIT,
            tool_choice=tool_choice or OMIT,
            previous_response_id=previous_response_id or OMIT,
            conversation=conversation_id or OMIT,
            reasoning=reasoning,
            text=text,
            truncation=model_settings.get('openai_truncation', OMIT),
            context_management=model_settings.get('openai_context_management', OMIT),
        )

    @staticmethod
    def _build_request_options(
        model_settings: OpenAIResponsesModelSettings,
    ) -> tuple[dict[str, str], float | Timeout | NotGiven]:
        extra_headers = dict(model_settings.get('extra_headers', {}))
        extra_headers.setdefault('User-Agent', get_user_agent())
        timeout = model_settings.get('timeout', NOT_GIVEN)
        return extra_headers, timeout

    @overload
    async def _responses_create(
        self,
        messages: list[ModelRequest | ModelResponse],
        stream: Literal[False],
        model_settings: OpenAIResponsesModelSettings,
        model_request_parameters: ModelRequestParameters,
    ) -> responses.Response: ...

    @overload
    async def _responses_create(
        self,
        messages: list[ModelRequest | ModelResponse],
        stream: Literal[True],
        model_settings: OpenAIResponsesModelSettings,
        model_request_parameters: ModelRequestParameters,
    ) -> AsyncStream[responses.ResponseStreamEvent]: ...

    async def _responses_create(
        self,
        messages: list[ModelRequest | ModelResponse],
        stream: bool,
        model_settings: OpenAIResponsesModelSettings,
        model_request_parameters: ModelRequestParameters,
    ) -> responses.Response | AsyncStream[responses.ResponseStreamEvent] | ModelResponse:
        profile = self.profile

        include: list[responses.ResponseIncludable] = []
        if profile.get('openai_supports_encrypted_reasoning_content', False):
            include.append('reasoning.encrypted_content')
        if model_settings.get('openai_include_code_execution_outputs'):
            include.append('code_interpreter_call.outputs')
        if model_settings.get('openai_include_web_search_sources'):
            include.append('web_search_call.action.sources')
        if model_settings.get('openai_include_file_search_results'):
            include.append('file_search_call.results')
        if model_settings.get('openai_logprobs'):
            include.append('message.output_text.logprobs')

        request_params = await self._build_responses_request_params(
            messages,
            model_settings,
            model_request_parameters,
            profile,
        )
        _drop_sampling_params_for_reasoning(profile, model_settings, model_request_parameters)
        _drop_unsupported_params(profile, model_settings)
        extra_headers, timeout = self._build_request_options(model_settings)

        # OpenAI SDK type stubs incorrectly use 'in-memory' but API requires 'in_memory', so we have to use `Any` to not hit type errors
        prompt_cache_retention: Any = model_settings.get('openai_prompt_cache_retention', OMIT)

        with _map_api_errors(self.model_name):
            try:
                return await self.client.responses.create(
                    model=request_params.model,
                    input=request_params.input,
                    instructions=request_params.instructions,
                    parallel_tool_calls=request_params.parallel_tool_calls,
                    tools=request_params.tools,
                    tool_choice=request_params.tool_choice,
                    previous_response_id=request_params.previous_response_id,
                    reasoning=request_params.reasoning,
                    text=request_params.text,
                    truncation=request_params.truncation,
                    context_management=request_params.context_management,
                    max_output_tokens=model_settings.get('max_tokens', OMIT),
                    stream=stream,
                    temperature=model_settings.get('temperature', OMIT),
                    top_p=model_settings.get('top_p', OMIT),
                    service_tier=_resolve_openai_service_tier(model_settings),
                    conversation=request_params.conversation,
                    top_logprobs=model_settings.get('openai_top_logprobs', OMIT),
                    store=model_settings.get('openai_store', OMIT),
                    user=model_settings.get('openai_user', OMIT),
                    include=include or OMIT,
                    prompt_cache_key=model_settings.get('openai_prompt_cache_key', OMIT),
                    prompt_cache_retention=prompt_cache_retention,
                    timeout=timeout,
                    extra_headers=extra_headers,
                    extra_body=model_settings.get('extra_body'),
                )
            except APIStatusError as e:
                if model_response := _check_azure_content_filter(e, self.system, self.model_name):
                    return model_response
                raise

    def _translate_thinking(
        self,
        model_settings: OpenAIResponsesModelSettings,
        model_request_parameters: ModelRequestParameters,
    ) -> Reasoning | Omit:
        reasoning_effort = model_settings.get('openai_reasoning_effort', None)
        reasoning_summary = model_settings.get('openai_reasoning_summary', None)

        # Fall back to unified thinking when openai_reasoning_effort is not set
        if reasoning_effort is None and (thinking := model_request_parameters.thinking) is not None:
            reasoning_effort = OPENAI_REASONING_EFFORT_MAP[thinking]

        reasoning: Reasoning = {}
        if reasoning_effort:
            reasoning['effort'] = reasoning_effort  # type: ignore[typeddict-item]
        if reasoning_summary:
            reasoning['summary'] = reasoning_summary
        return reasoning or OMIT

    def _get_responses_tool_choice(
        self,
        model_settings: OpenAIResponsesModelSettings,
        model_request_parameters: ModelRequestParameters,
    ) -> tuple[list[responses.FunctionToolParam], ResponsesToolChoice | None]:
        """Determine which tools to send and the API tool_choice value.

        Returns:
            A tuple of (filtered_function_tools, tool_choice).
            Note: builtin tools are handled separately and should be added to this list.
        """
        openai_profile = self.profile

        resolved_tool_choice = resolve_tool_choice(model_settings, model_request_parameters)

        tool_choice: ResponsesToolChoice | None
        if resolved_tool_choice in ('auto', 'none'):
            tool_choice = resolved_tool_choice
        elif resolved_tool_choice == 'required':
            supports = _support_tool_forcing(self.model_name, openai_profile, model_settings)
            tool_choice = 'required' if supports else 'auto'
        elif isinstance(resolved_tool_choice, tuple):
            tool_choice_mode, tool_names = resolved_tool_choice
            supports = _support_tool_forcing(self.model_name, openai_profile, model_settings)
            if tool_choice_mode == 'required' and len(tool_names) == 1 and supports:
                tool_choice = ToolChoiceFunctionParam(type='function', name=next(iter(tool_names)))
            else:
                # `allowed_tools` filters via the API rather than the tools list,
                # so caching is preserved. Used for the multi-tool case and as the
                # cache-preserving fallback when forcing isn't supported.
                effective_mode = 'auto' if tool_choice_mode == 'auto' or not supports else tool_choice_mode
                tool_choice = ToolChoiceAllowedParam(
                    type='allowed_tools',
                    mode=effective_mode,
                    tools=[{'type': 'function', 'name': n} for n in tool_names],
                )
        else:
            assert_never(resolved_tool_choice)

        # In client-executed tool search mode the `search_tools` function tool is
        # represented on the wire by `ToolSearchToolParam(execution='client')` — sending
        # it as a regular function tool too makes OpenAI emit `function_call` items
        # instead of `tool_search_call` items. The local dispatch path still sees the
        # function tool via `ModelRequestParameters.function_tools` (unaffected by this
        # filtering), so `_search_tools` continues to run normally.
        client_tool_search = _has_tool_search(model_request_parameters)
        tools: list[responses.FunctionToolParam] = [
            self._map_tool_definition(t)
            for t in model_request_parameters.tool_defs.values()
            if not (client_tool_search and t.name == TOOL_SEARCH_FUNCTION_TOOL_NAME)
        ]
        return tools, tool_choice

    def _get_native_tools(  # noqa: C901
        self, model_request_parameters: ModelRequestParameters
    ) -> list[responses.ToolParam]:
        tools: list[responses.ToolParam] = []
        has_image_generating_tool = False
        for tool in model_request_parameters.native_tools:
            if isinstance(tool, WebSearchTool):
                web_search_tool = responses.WebSearchToolParam(
                    type='web_search', search_context_size=tool.search_context_size
                )
                if tool.user_location:
                    web_search_tool['user_location'] = responses.web_search_tool_param.UserLocation(
                        type='approximate', **tool.user_location
                    )
                if tool.allowed_domains:
                    web_search_tool['filters'] = responses.web_search_tool_param.Filters(
                        allowed_domains=tool.allowed_domains
                    )
                tools.append(web_search_tool)
            elif isinstance(tool, FileSearchTool):
                file_search_tool = cast(
                    responses.FileSearchToolParam,
                    {'type': 'file_search', 'vector_store_ids': list(tool.file_store_ids)},
                )
                tools.append(file_search_tool)
            elif isinstance(tool, CodeExecutionTool):
                has_image_generating_tool = True
                tools.append({'type': 'code_interpreter', 'container': {'type': 'auto'}})
            elif isinstance(tool, MCPServerTool):
                mcp_tool = responses.tool_param.Mcp(
                    type='mcp',
                    server_label=tool.id,
                    require_approval='never',
                )

                if tool.authorization_token:  # pragma: no branch
                    mcp_tool['authorization'] = tool.authorization_token

                if tool.allowed_tools is not None:  # pragma: no branch
                    mcp_tool['allowed_tools'] = tool.allowed_tools

                if tool.description:  # pragma: no branch
                    mcp_tool['server_description'] = tool.description

                if tool.headers:  # pragma: no branch
                    mcp_tool['headers'] = tool.headers

                if tool.url.startswith(MCP_SERVER_TOOL_CONNECTOR_URI_SCHEME + ':'):
                    _, connector_id = tool.url.split(':', maxsplit=1)
                    mcp_tool['connector_id'] = connector_id  # pyright: ignore[reportGeneralTypeIssues]
                else:
                    mcp_tool['server_url'] = tool.url

                tools.append(mcp_tool)
            elif isinstance(tool, ImageGenerationTool):
                has_image_generating_tool = True
                tools.append(_map_openai_image_generation_tool(tool))
            elif isinstance(tool, ToolSearchTool):  # pragma: no branch
                # OpenAI's Responses API has no concept of named native strategies — the
                # `tool_search` builtin is one knob, decided server-side. Honor the user's
                # explicit choice or fail loudly; do not silently run a different algorithm.
                # `'custom'` is the internal marker for callable strategies and is handled
                # below — only named native strategies (`'bm25'`/`'regex'`) raise here.
                if tool.strategy is None:
                    tools.append(ToolSearchToolParam(type='tool_search'))
                elif tool.strategy == 'custom':
                    # With `strategy='custom'`, the search runs client-side via our local
                    # `search_tools` function tool: the provider surfaces the search as a
                    # `tool_search_call` with `execution='client'`, we dispatch it through the
                    # normal function-call path, and reply with a `tool_search_output` that
                    # carries the discovered tool definitions.
                    search_tool_def = _find_search_tool_definition(model_request_parameters)
                    parameters = dict(search_tool_def.parameters_json_schema) if search_tool_def else {}
                    # OpenAI's strict JSON schema mode is opt-in for client tool search — the
                    # `additionalProperties: False` is what it prefers for closed-world schemas.
                    parameters.setdefault('additionalProperties', False)
                    tool_search_param: ToolSearchToolParam = {
                        'type': 'tool_search',
                        'execution': 'client',
                        'description': (search_tool_def.description if search_tool_def else None),
                        'parameters': parameters,
                    }
                    tools.append(tool_search_param)
                else:
                    raise UserError(
                        f'`ToolSearch(strategy={tool.strategy!r})` is an Anthropic-native strategy '
                        'and is not supported by OpenAI Responses. Use `strategy=None` (default, '
                        "provider-managed), `strategy='keywords'` (local keyword matching), or "
                        'a callable strategy.'
                    )
            else:
                raise UserError(  # pragma: no cover
                    f'`{tool.__class__.__name__}` is not supported by `OpenAIResponsesModel`. If it should be, please file an issue.'
                )

        if model_request_parameters.allow_image_output and not has_image_generating_tool:
            tools.append({'type': 'image_generation'})
        return tools

    def _map_tool_definition(self, f: ToolDefinition) -> responses.FunctionToolParam:
        tool_param: responses.FunctionToolParam = {
            'name': f.name,
            'parameters': f.parameters_json_schema,
            'type': 'function',
            'description': f.description,
            'strict': bool(f.strict and self.profile.get('openai_supports_strict_tool_definition', True)),
        }
        if f.with_native == ToolSearchTool.kind:
            # `defer_loading` on the wire controls OpenAI's native tool search caching.
            # `ToolDefinition.defer_loading` is the local discovery flag — separate
            # concern — so we gate on `with_native` to know we're on the native
            # path.
            tool_param['defer_loading'] = True
        return tool_param

    def _resolve_previous_response_id(
        self,
        setting: str | None,
        messages: list[ModelMessage],
        *,
        allow_no_new_messages: bool = False,
    ) -> tuple[str | None, list[ModelMessage]]:
        # Resolve the effective `previous_response_id` and trim already-stored messages.
        #
        # A concrete ID in `setting` acts as a seed for the first request in a run (to
        # continue from a prior turn's stored response). On subsequent in-run requests
        # (retries, tool-call continuations), the most recent `provider_response_id`
        # from the message history takes precedence, so we chain to the latest stored
        # response instead of re-sending messages that are already server-side.
        #
        # A compaction response in the tail is a hard chain boundary even with a
        # concrete seed: crossing it would re-inject the context that compaction was
        # meant to replace.
        if setting is None:
            return None, messages
        auto_id, trimmed = self._get_previous_response_id_and_new_messages(
            messages, allow_no_new_messages=allow_no_new_messages
        )
        if auto_id is not None:
            return auto_id, trimmed
        if setting == 'auto' or self._is_at_compaction_boundary(messages):
            return None, messages
        return setting, messages

    def _is_at_compaction_boundary(self, messages: list[ModelMessage]) -> bool:
        for m in reversed(messages):
            if isinstance(m, ModelResponse) and m.provider_name == self.system:
                return bool(m.provider_details and m.provider_details.get('compaction'))
        return False

    def _get_previous_response_id_and_new_messages(
        self, messages: list[ModelMessage], *, allow_no_new_messages: bool = False
    ) -> tuple[str | None, list[ModelMessage]]:
        # Find the most recent `provider_response_id` from messages produced by this
        # provider, and return it along with the messages that came after it (which
        # still need to be sent as new input). When nothing suitable is found, returns
        # `(None, messages)` with the full list unchanged.
        previous_response_id = None
        trimmed_messages: list[ModelMessage] = []
        for m in reversed(messages):
            if isinstance(m, ModelResponse) and m.provider_name == self.system:
                # Responses from the stateless `/compact` endpoint can't be used as
                # `previous_response_id`, so the compaction acts as a hard chain boundary:
                # the next request must pass the `CompactionPart` via the `input` array
                # (handled by `_map_messages`) without a `previous_response_id`.
                if m.provider_details and m.provider_details.get('compaction'):
                    return None, messages
                previous_response_id = m.provider_response_id
                break
            else:
                trimmed_messages.append(m)

        if previous_response_id and (allow_no_new_messages or trimmed_messages):
            return previous_response_id, list(reversed(trimmed_messages))
        else:
            return None, messages

    def _resolve_server_side_state(
        self, model_settings: OpenAIResponsesModelSettings, messages: list[ModelMessage]
    ) -> tuple[str | None, str | None, list[ModelMessage]]:
        previous_response_id_setting = model_settings.get('openai_previous_response_id')
        conversation_id_setting = model_settings.get('openai_conversation_id')
        if previous_response_id_setting is not None and conversation_id_setting is not None:
            raise UserError(
                '`openai_previous_response_id` and `openai_conversation_id` cannot both be set because '
                'the OpenAI Responses API does not support `previous_response_id` with `conversation`.'
            )

        if conversation_id_setting is not None:
            conversation_id, messages = self._resolve_conversation_id(conversation_id_setting, messages)
            return None, conversation_id, messages

        previous_response_id, messages = self._resolve_previous_response_id(previous_response_id_setting, messages)
        return previous_response_id, None, messages

    def _resolve_conversation_id(
        self, setting: Literal['auto'] | str, messages: list[ModelMessage]
    ) -> tuple[str | None, list[ModelMessage]]:
        if setting == 'auto':
            # Agent runs stamp the active Pydantic AI conversation ID on the final request.
            # Direct model calls may still pass an empty message list.
            pydantic_ai_conversation_id = next((m.conversation_id for m in messages[-1:]), None)
            return self._get_conversation_id_and_new_messages(
                messages, pydantic_ai_conversation_id=pydantic_ai_conversation_id
            )

        conversation_id, trimmed = self._get_conversation_id_and_new_messages(messages, openai_conversation_id=setting)
        if conversation_id is not None:
            return conversation_id, trimmed
        return setting, messages

    def _get_conversation_id_and_new_messages(
        self,
        messages: list[ModelMessage],
        *,
        openai_conversation_id: str | None = None,
        pydantic_ai_conversation_id: str | None = None,
    ) -> tuple[str | None, list[ModelMessage]]:
        trimmed_messages: list[ModelMessage] = []
        for m in reversed(messages):
            if isinstance(m, ModelResponse) and m.provider_name == self.system:
                candidate = m.provider_details and m.provider_details.get('conversation_id')
                if (
                    pydantic_ai_conversation_id is not None
                    and m.conversation_id is not None
                    and m.conversation_id != pydantic_ai_conversation_id
                ):
                    trimmed_messages.append(m)
                    continue
                if isinstance(candidate, str) and (
                    openai_conversation_id is None or candidate == openai_conversation_id
                ):
                    return candidate, list(reversed(trimmed_messages))
            trimmed_messages.append(m)

        return None, messages

    async def _map_messages(  # noqa: C901
        self,
        messages: list[ModelMessage],
        model_settings: OpenAIResponsesModelSettings,
        model_request_parameters: ModelRequestParameters,
    ) -> tuple[str | Omit, list[responses.ResponseInputItemParam]]:
        """Maps a `pydantic_ai.Message` to a `openai.types.responses.ResponseInputParam` i.e. the OpenAI Responses API input format.

        For `ThinkingParts`, this method:
        - Sends `signature` back as `encrypted_content` (for official OpenAI reasoning)
        - Sends `content` back as `summary` text
        - Sends `provider_details['raw_content']` back as `content` items (for gpt-oss raw CoT)

        Raw CoT is sent back to improve model performance in multi-turn conversations.
        """
        profile = self.profile
        send_item_ids = model_settings.get(
            'openai_send_reasoning_ids', profile.get('openai_supports_encrypted_reasoning_content', False)
        )
        # With `execution='client'` tool search, `search_tools` calls/returns need to be
        # replayed as `tool_search_call` / `tool_search_output` items so the provider can
        # pair them with the builtin and unlock the discovered tools. The current
        # request's builtin configuration is the source of truth: a `ToolCallPart` for
        # `search_tools` originating from this provider (`provider_name == self.system`)
        # belongs in the client-execution flow whenever a custom-callable tool search is
        # active.
        client_tool_search_active = _has_tool_search(model_request_parameters)
        client_replay_call_ids: set[str] = set()

        openai_messages: list[responses.ResponseInputItemParam] = []
        for message in messages:
            if isinstance(message, ModelRequest):
                for part in message.parts:
                    if isinstance(part, SystemPromptPart):
                        openai_messages.append(
                            responses.EasyInputMessageParam(
                                role=profile.get('openai_system_prompt_role', None) or 'system', content=part.content
                            )
                        )
                    elif isinstance(part, UserPromptPart):
                        openai_messages.append(await self._map_user_prompt(part))
                    elif isinstance(part, ToolReturnPart):
                        call_id = _guard_tool_call_id(t=part)
                        call_id, _ = _split_combined_tool_call_id(call_id)
                        if call_id in client_replay_call_ids and isinstance(part, ToolSearchReturnPart):
                            # Replay the local `search_tools` result as a
                            # `tool_search_output` so the provider rehydrates the
                            # discovered-tool state tied to the `tool_search_call`.
                            openai_messages.append(
                                _build_client_tool_search_output_param(
                                    part, call_id, model_request_parameters, self._map_tool_definition
                                )
                            )
                        else:
                            output = await self._map_tool_return_output(part)
                            item = FunctionCallOutput(
                                type='function_call_output',
                                call_id=call_id,
                                output=output,
                            )
                            openai_messages.append(item)
                    elif isinstance(part, RetryPromptPart):
                        if part.tool_name is None:
                            openai_messages.append(
                                Message(role='user', content=[{'type': 'input_text', 'text': part.model_response()}])
                            )
                        else:
                            call_id = _guard_tool_call_id(t=part)
                            call_id, _ = _split_combined_tool_call_id(call_id)
                            item = FunctionCallOutput(
                                type='function_call_output',
                                call_id=call_id,
                                output=part.model_response(),
                            )
                            openai_messages.append(item)
                    else:
                        assert_never(part)
            elif isinstance(message, ModelResponse):
                message_item: responses.ResponseOutputMessageParam | None = None
                reasoning_item: responses.ResponseReasoningItemParam | None = None
                web_search_item: responses.ResponseFunctionWebSearchParam | None = None
                file_search_item: responses.ResponseFileSearchToolCallParam | None = None
                code_interpreter_item: responses.ResponseCodeInterpreterToolCallParam | None = None
                for item in message.parts:
                    should_send_item_id = send_item_ids and (
                        item.provider_name == self.system
                        or (item.provider_name is None and message.provider_name == self.system)
                    )

                    if isinstance(item, TextPart):
                        phase = (item.provider_details or {}).get('phase')
                        send_phase = (
                            profile.get('openai_supports_phase', False)
                            and item.provider_name == self.system
                            and phase in ('commentary', 'final_answer')
                        )
                        if item.id and should_send_item_id:
                            if message_item is None or message_item['id'] != item.id:  # pragma: no branch
                                message_item = responses.ResponseOutputMessageParam(
                                    role='assistant',
                                    id=item.id,
                                    content=[],
                                    type='message',
                                    status='completed',
                                )
                                openai_messages.append(message_item)

                            message_item['content'] = [
                                *message_item['content'],
                                responses.ResponseOutputTextParam(
                                    text=item.content, type='output_text', annotations=[]
                                ),
                            ]
                            if send_phase:
                                message_item['phase'] = phase
                        else:
                            easy_message_item = responses.EasyInputMessageParam(role='assistant', content=item.content)
                            if send_phase:
                                easy_message_item['phase'] = phase
                            openai_messages.append(easy_message_item)
                    elif isinstance(item, ToolCallPart):
                        call_id = _guard_tool_call_id(t=item)
                        call_id, id = _split_combined_tool_call_id(call_id)
                        id = id or item.id

                        if client_tool_search_active and item.tool_name == TOOL_SEARCH_FUNCTION_TOOL_NAME:
                            # Replay the local `search_tools` call as a `tool_search_call`
                            # with `execution='client'` so OpenAI re-attaches it to the
                            # builtin and unlocks the discovered tools' schemas. Fires for
                            # any `provider_name`, including `None` (cross-provider history
                            # from a prior local-fallback turn) and other-provider names
                            # (Google, etc.) — `execution='client'` is OpenAI's accepted
                            # historical-replay shape regardless of the call's origin.
                            client_replay_call_ids.add(call_id)
                            tool_search_call: ToolSearchCallParam = {
                                'type': 'tool_search_call',
                                'execution': 'client',
                                'arguments': item.args_as_dict() or {},
                                'call_id': call_id,
                                'status': 'completed',
                            }
                            if id and should_send_item_id:  # pragma: no branch
                                tool_search_call['id'] = id
                            openai_messages.append(tool_search_call)
                        else:
                            param = responses.ResponseFunctionToolCallParam(
                                name=item.tool_name,
                                arguments=item.args_as_json_str(),
                                call_id=call_id,
                                type='function_call',
                            )
                            if profile.get('openai_responses_requires_function_call_status_none', False):
                                param['status'] = None  # type: ignore[reportGeneralTypeIssues]
                            if id and should_send_item_id:  # pragma: no branch
                                param['id'] = id
                            # The Responses API attaches a `namespace` to function calls
                            # for discovered deferred tools (tool-search flow). Round-trip
                            # it on replay or the API rejects the call as missing namespace.
                            if (
                                item.provider_name == self.system
                                and item.provider_details
                                and (ns := item.provider_details.get('namespace'))
                            ):
                                param['namespace'] = ns
                            openai_messages.append(param)
                    elif isinstance(item, NativeToolCallPart):
                        if should_send_item_id:  # pragma: no branch
                            if (
                                item.tool_name == CodeExecutionTool.kind
                                and item.tool_call_id
                                and (args := item.args_as_dict())
                                and (container_id := args.get('container_id'))
                            ):
                                code_interpreter_item = responses.ResponseCodeInterpreterToolCallParam(
                                    id=item.tool_call_id,
                                    code=args.get('code'),
                                    container_id=container_id,
                                    outputs=None,  # These can be read server-side
                                    status='completed',
                                    type='code_interpreter_call',
                                )
                                openai_messages.append(code_interpreter_item)
                            elif (
                                item.tool_name == WebSearchTool.kind
                                and item.tool_call_id
                                and (args := item.args_as_dict())
                            ):
                                # We need to exclude None values because of https://github.com/pydantic/pydantic-ai/issues/3653
                                args = {k: v for k, v in args.items() if v is not None}
                                web_search_item = responses.ResponseFunctionWebSearchParam(
                                    id=item.id or item.tool_call_id,
                                    action=cast(responses.response_function_web_search_param.Action, args),
                                    status='completed',
                                    type='web_search_call',
                                )
                                openai_messages.append(web_search_item)
                            elif (  # pragma: no cover
                                item.tool_name == FileSearchTool.kind
                                and item.tool_call_id
                                and (args := item.args_as_dict())
                            ):
                                file_search_item = cast(
                                    responses.ResponseFileSearchToolCallParam,
                                    {
                                        'id': item.id or item.tool_call_id,
                                        'queries': args.get('queries', []),
                                        'status': 'completed',
                                        'type': 'file_search_call',
                                    },
                                )
                                openai_messages.append(file_search_item)
                            elif item.tool_name == ImageGenerationTool.kind and item.tool_call_id:
                                # The cast is necessary because of https://github.com/openai/openai-python/issues/2648
                                image_generation_item = cast(
                                    responses.response_input_item_param.ImageGenerationCall,
                                    {
                                        'id': item.tool_call_id,
                                        'type': 'image_generation_call',
                                    },
                                )
                                openai_messages.append(image_generation_item)
                            elif item.tool_name == ToolSearchTool.kind and item.tool_call_id:
                                openai_messages.append(
                                    responses.response_input_item_param.ToolSearchCall(
                                        id=item.id or item.tool_call_id,
                                        call_id=item.tool_call_id,
                                        arguments=item.args_as_dict() or {},
                                        type='tool_search_call',
                                        execution='server',
                                        status='completed',
                                    )
                                )
                            elif (  # pragma: no branch
                                item.tool_name.startswith(MCPServerTool.kind)
                                and item.tool_call_id
                                and (server_id := item.tool_name.split(':', 1)[1])
                                and (args := item.args_as_dict())
                                and (action := args.get('action'))
                            ):
                                if action == 'list_tools':
                                    mcp_list_tools_item = responses.response_input_item_param.McpListTools(
                                        id=item.tool_call_id,
                                        type='mcp_list_tools',
                                        server_label=server_id,
                                        tools=[],  # These can be read server-side
                                    )
                                    openai_messages.append(mcp_list_tools_item)
                                elif (  # pragma: no branch
                                    action == 'call_tool'
                                    and (tool_name := args.get('tool_name'))
                                    and (tool_args := args.get('tool_args')) is not None
                                ):
                                    mcp_call_item = responses.response_input_item_param.McpCall(
                                        id=item.tool_call_id,
                                        server_label=server_id,
                                        name=tool_name,
                                        arguments=to_json(tool_args).decode(),
                                        error=None,  # These can be read server-side
                                        output=None,  # These can be read server-side
                                        type='mcp_call',
                                    )
                                    openai_messages.append(mcp_call_item)

                    elif isinstance(item, NativeToolReturnPart):
                        if should_send_item_id:  # pragma: no branch
                            status = item.content.get('status') if _is_str_dict(item.content) else None
                            kind_to_item = {
                                CodeExecutionTool.kind: code_interpreter_item,
                                WebSearchTool.kind: web_search_item,
                                FileSearchTool.kind: file_search_item,
                            }
                            if status and (builtin_item := kind_to_item.get(item.tool_name)) is not None:
                                builtin_item['status'] = status
                            elif item.tool_name == ImageGenerationTool.kind:
                                # Image generation result does not need to be sent back, just the `id` off of `NativeToolCallPart`.
                                pass
                            elif item.tool_name.startswith(MCPServerTool.kind):
                                # MCP call result does not need to be sent back, just the fields off of `NativeToolCallPart`.
                                pass
                            elif item.tool_name == ToolSearchTool.kind:  # pragma: no branch
                                # Tool search output (discovered tool definitions) is
                                # reconstructed server-side from the `tool_search_call` —
                                # same pattern as web search — so nothing to send back.
                                pass
                    elif isinstance(item, FilePart):
                        # This was generated by the `ImageGenerationTool` or `CodeExecutionTool`,
                        # and does not need to be sent back separately from the corresponding `NativeToolReturnPart`.
                        # If `send_item_ids` is false, we won't send the `NativeToolReturnPart`, but OpenAI does not have a type for files from the assistant.
                        pass
                    elif isinstance(item, ThinkingPart):
                        # Get raw CoT content from provider_details if present and from this provider
                        raw_content: list[str] | None = None
                        if item.provider_name == self.system:
                            raw_content = (item.provider_details or {}).get('raw_content')

                        if item.id and (should_send_item_id or raw_content):
                            signature: str | None = None
                            if (
                                item.signature
                                and item.provider_name == self.system
                                and profile.get('openai_supports_encrypted_reasoning_content', False)
                            ):
                                signature = item.signature

                            if (reasoning_item is None or reasoning_item['id'] != item.id) and (
                                signature or item.content or raw_content
                            ):  # pragma: no branch
                                reasoning_item = responses.ResponseReasoningItemParam(
                                    id=item.id,
                                    summary=[],
                                    encrypted_content=signature,
                                    type='reasoning',
                                )
                                openai_messages.append(reasoning_item)

                            if item.content:
                                # The check above guarantees that `reasoning_item` is not None
                                assert reasoning_item is not None
                                reasoning_item['summary'] = [
                                    *reasoning_item['summary'],
                                    ReasoningSummary(text=item.content, type='summary_text'),
                                ]

                            if raw_content:
                                # Send raw CoT back
                                assert reasoning_item is not None
                                reasoning_item['content'] = [
                                    ReasoningContent(text=text, type='reasoning_text') for text in raw_content
                                ]
                        else:
                            start_tag, end_tag = profile.get('thinking_tags', DEFAULT_THINKING_TAGS)
                            openai_messages.append(
                                responses.EasyInputMessageParam(
                                    role='assistant', content='\n'.join([start_tag, item.content, end_tag])
                                )
                            )
                    elif isinstance(item, CompactionPart):
                        if (
                            item.provider_name == self.system
                            and item.provider_details
                            and 'encrypted_content' in item.provider_details
                        ):  # pragma: no branch
                            openai_messages.append(
                                ResponseCompactionItemParamParam(
                                    id=item.id,
                                    encrypted_content=item.provider_details['encrypted_content'],
                                    type='compaction',
                                )
                            )
                    else:
                        assert_never(item)
            else:
                assert_never(message)
        instructions = get_instructions(messages, model_request_parameters) or OMIT
        return instructions, openai_messages

    def _map_json_schema(self, o: OutputObjectDefinition) -> responses.ResponseFormatTextJSONSchemaConfigParam:
        response_format_param: responses.ResponseFormatTextJSONSchemaConfigParam = {
            'type': 'json_schema',
            'name': o.name or DEFAULT_OUTPUT_TOOL_NAME,
            'schema': o.json_schema,
        }
        if o.description:
            response_format_param['description'] = o.description
        if self.profile.get('openai_supports_strict_tool_definition', True):  # pragma: no branch
            response_format_param['strict'] = o.strict
        return response_format_param

    async def _map_user_prompt(self, part: UserPromptPart) -> responses.EasyInputMessageParam:
        content: str | list[responses.ResponseInputContentParam]
        if isinstance(part.content, str):
            content = part.content
        else:
            content = []
            for item in part.content:
                if isinstance(item, str | TextContent):
                    text = item if isinstance(item, str) else item.content
                    content.append(responses.ResponseInputTextParam(text=text, type='input_text'))
                elif isinstance(item, UploadedFile):
                    if item.provider_name != self.system:
                        raise UserError(
                            f'UploadedFile with `provider_name={item.provider_name!r}` cannot be used with OpenAIResponsesModel. '
                            f'Expected `provider_name` to be `{self.system!r}`.'
                        )
                    content.append(
                        responses.ResponseInputFileParam(
                            type='input_file',
                            file_id=item.file_id,
                        )
                    )
                elif isinstance(item, CachePoint):
                    pass
                elif is_multi_modal_content(item):
                    content.append(await OpenAIResponsesModel._map_file_to_response_content(item, 'user prompts'))  # pyright: ignore[reportArgumentType]
                else:
                    raise RuntimeError(f'Unsupported content type: {type(item)}')  # pragma: no cover
        return responses.EasyInputMessageParam(role='user', content=content)

    @staticmethod
    async def _map_file_to_response_content(
        item: BinaryContent | ImageUrl | DocumentUrl | AudioUrl | VideoUrl,
        context: str,
    ) -> ResponseInputImageContentParam | ResponseInputFileContentParam:
        """Map a multimodal file item to its OpenAI Responses API content param."""
        if isinstance(item, BinaryContent):
            if item.is_image:
                detail: Literal['auto', 'low', 'high'] = 'auto'
                if metadata := item.vendor_metadata:
                    detail = metadata.get('detail', 'auto')
                return ResponseInputImageContentParam(
                    image_url=item.data_uri,
                    type='input_image',
                    detail=detail,
                )
            elif item.is_document:
                return ResponseInputFileContentParam(
                    type='input_file',
                    file_data=item.data_uri,
                    filename=f'filename.{item.format}',
                )
            elif item.is_audio:
                raise NotImplementedError(f'BinaryContent with audio is not supported in OpenAI Responses {context}')
            elif item.is_video:
                raise NotImplementedError(f'BinaryContent with video is not supported in OpenAI Responses {context}')
            else:  # pragma: no cover
                raise RuntimeError(f'Unsupported binary content type: {item.media_type}')
        elif isinstance(item, ImageUrl):
            detail = 'auto'
            image_url = item.url
            if metadata := item.vendor_metadata:
                detail = metadata.get('detail', 'auto')
            if item.force_download:
                downloaded = await download_item(item, data_format='base64_uri', type_format='extension')
                image_url = downloaded['data']
            return ResponseInputImageContentParam(
                image_url=image_url,
                type='input_image',
                detail=detail,
            )
        elif isinstance(item, (AudioUrl, DocumentUrl)):
            if item.force_download:
                downloaded = await download_item(item, data_format='base64_uri', type_format='extension')
                return ResponseInputFileContentParam(
                    type='input_file',
                    file_data=downloaded['data'],
                    filename=f'filename.{downloaded["data_type"]}',
                )
            return ResponseInputFileContentParam(
                type='input_file',
                file_url=item.url,
            )
        else:
            raise NotImplementedError(f'VideoUrl is not supported in OpenAI Responses {context}')

    @staticmethod
    async def _map_tool_return_output(
        part: ToolReturnPart,
    ) -> str | list[ResponseInputTextContentParam | ResponseInputImageContentParam | ResponseInputFileContentParam]:
        """Map a `ToolReturnPart` to OpenAI Responses API output format, supporting multimodal content.

        Iterates content directly to preserve order of mixed file/data content.
        """
        if not part.files:
            return part.model_response_str()

        output: list[
            ResponseInputTextContentParam | ResponseInputImageContentParam | ResponseInputFileContentParam
        ] = []

        for item in part.content_items(mode='str'):
            if isinstance(item, UploadedFile):
                output.append(
                    ResponseInputFileContentParam(
                        type='input_file',
                        file_id=item.file_id,
                    )
                )
            elif is_multi_modal_content(item):
                output.append(await OpenAIResponsesModel._map_file_to_response_content(item, 'tool returns'))  # pyright: ignore[reportArgumentType]
            elif isinstance(item, str):  # pragma: no branch
                output.append(ResponseInputTextContentParam(type='input_text', text=item))

        return output

__init__ 

__init__(
    model_name: OpenAIModelName,
    *,
    provider: (
        OpenAIResponsesCompatibleProvider
        | Literal["openai", "gateway"]
        | Provider[AsyncOpenAI]
    ) = "openai",
    profile: ModelProfileSpec | None = None,
    settings: ModelSettings | None = None
)

Initialize an OpenAI Responses model.

Parameters:

Name Type Description Default
model_name OpenAIModelName

The name of the OpenAI model to use.

 required
provider OpenAIResponsesCompatibleProvider | Literal['openai', 'gateway'] | Provider[AsyncOpenAI]

The provider to use. Defaults to 'openai'.

 'openai'
profile ModelProfileSpec | None

The model profile to use. Defaults to a profile picked by the provider based on the model name.

 None
settings ModelSettings | None

Default model settings for this model instance.

 None
Source code in pydantic_ai_slim/pydantic_ai/models/openai.py 
1654
1655
1656
1657
1658
1659
1660
1661
1662
1663
1664
1665
1666
1667
1668
1669
1670
1671
1672
1673
1674
1675
1676
1677
1678
1679
1680
1681
def __init__(
    self,
    model_name: OpenAIModelName,
    *,
    provider: OpenAIResponsesCompatibleProvider
    | Literal[
        'openai',
        'gateway',
    ]
    | Provider[AsyncOpenAI] = 'openai',
    profile: ModelProfileSpec | None = None,
    settings: ModelSettings | None = None,
):
    """Initialize an OpenAI Responses model.

    Args:
        model_name: The name of the OpenAI model to use.
        provider: The provider to use. Defaults to `'openai'`.
        profile: The model profile to use. Defaults to a profile picked by the provider based on the model name.
        settings: Default model settings for this model instance.
    """
    self._model_name = model_name

    if isinstance(provider, str):
        provider = infer_provider('gateway/openai' if provider == 'gateway' else provider)
    self._provider = provider

    super().__init__(settings=settings, profile=profile)

model_name property 

model_name: OpenAIModelName

The model name.

system property 

system: str

The model provider.

supported_native_tools classmethod 

supported_native_tools() -> (
    frozenset[type[AbstractNativeTool]]
)

Return the set of builtin tool types this model can handle.

Source code in pydantic_ai_slim/pydantic_ai/models/openai.py 
1705
1706
1707
1708
1709
1710
@classmethod
def supported_native_tools(cls) -> frozenset[type[AbstractNativeTool]]:
    """Return the set of builtin tool types this model can handle."""
    return frozenset(
        {WebSearchTool, CodeExecutionTool, FileSearchTool, MCPServerTool, ImageGenerationTool, ToolSearchTool}
    )

compact_messages async 

compact_messages(
    request_context: ModelRequestContext,
    *,
    instructions: str | None = None
) -> ModelResponse

Compact messages using the OpenAI Responses compaction endpoint.

This calls OpenAI's responses.compact API to produce an encrypted compaction that summarizes the conversation history. The returned ModelResponse contains a single CompactionPart that must be round-tripped in subsequent requests.

Parameters:

Name Type Description Default
request_context ModelRequestContext

The model request context containing messages, settings, and parameters.

 required
instructions str | None

Optional custom instructions for the compaction summarization. If provided, these override the agent-level instructions.

 None

Returns:

Type Description
ModelResponse

A ModelResponse with a single CompactionPart containing the encrypted compaction data.
Source code in pydantic_ai_slim/pydantic_ai/models/openai.py 
1712
1713
1714
1715
1716
1717
1718
1719
1720
1721
1722
1723
1724
1725
1726
1727
1728
1729
1730
1731
1732
1733
1734
1735
1736
1737
1738
1739
1740
1741
1742
1743
1744
1745
1746
1747
1748
1749
1750
1751
1752
1753
1754
1755
1756
1757
1758
1759
1760
1761
1762
1763
1764
1765
1766
1767
1768
async def compact_messages(
    self,
    request_context: ModelRequestContext,
    *,
    instructions: str | None = None,
) -> ModelResponse:
    """Compact messages using the OpenAI Responses compaction endpoint.

    This calls OpenAI's `responses.compact` API to produce an encrypted compaction
    that summarizes the conversation history. The returned `ModelResponse` contains
    a single `CompactionPart` that must be round-tripped in subsequent requests.

    Args:
        request_context: The model request context containing messages, settings, and parameters.
        instructions: Optional custom instructions for the compaction summarization.
            If provided, these override the agent-level instructions.

    Returns:
        A `ModelResponse` with a single `CompactionPart` containing the encrypted compaction data.
    """
    check_allow_model_requests()
    model_settings, model_request_parameters = self.prepare_request(
        request_context.model_settings,
        request_context.model_request_parameters,
    )
    response = await self._responses_compact(
        request_context.messages,
        cast(OpenAIResponsesModelSettings, model_settings or {}),
        model_request_parameters,
        instructions_override=instructions,
    )

    # Handle ModelResponse (e.g. from content filter)
    if isinstance(response, ModelResponse):  # pragma: no cover
        return response

    if not response.output:  # pragma: no cover
        raise UnexpectedModelBehavior('CompactedResponse returned with no output items')

    compaction = response.output[-1]
    if not isinstance(compaction, ResponseCompactionItem):  # pragma: no cover
        raise UnexpectedModelBehavior(f'Last item in response is not a compaction, got: {compaction.type}')

    part = _map_compaction_item(compaction, self.system)
    return ModelResponse(
        parts=[part],
        usage=_map_usage(response, self._provider.name, self._provider.base_url, self.model_name),
        model_name=self._model_name,
        provider_response_id=response.id,
        # Marks this ModelResponse as coming from the stateless `/compact` endpoint.
        # `_get_previous_response_id_and_new_messages` uses this to break the auto-chain,
        # since compaction response ids cannot be used as `previous_response_id`.
        provider_details={'compaction': True},
        timestamp=_now_utc(),
        provider_name=self._provider.name,
        provider_url=self._provider.base_url,
    )

OpenAIStreamedResponse dataclass

Bases: StreamedResponse

Implementation of StreamedResponse for OpenAI models.

Source code in pydantic_ai_slim/pydantic_ai/models/openai.py 
3073
3074
3075
3076
3077
3078
3079
3080
3081
3082
3083
3084
3085
3086
3087
3088
3089
3090
3091
3092
3093
3094
3095
3096
3097
3098
3099
3100
3101
3102
3103
3104
3105
3106
3107
3108
3109
3110
3111
3112
3113
3114
3115
3116
3117
3118
3119
3120
3121
3122
3123
3124
3125
3126
3127
3128
3129
3130
3131
3132
3133
3134
3135
3136
3137
3138
3139
3140
3141
3142
3143
3144
3145
3146
3147
3148
3149
3150
3151
3152
3153
3154
3155
3156
3157
3158
3159
3160
3161
3162
3163
3164
3165
3166
3167
3168
3169
3170
3171
3172
3173
3174
3175
3176
3177
3178
3179
3180
3181
3182
3183
3184
3185
3186
3187
3188
3189
3190
3191
3192
3193
3194
3195
3196
3197
3198
3199
3200
3201
3202
3203
3204
3205
3206
3207
3208
3209
3210
3211
3212
3213
3214
3215
3216
3217
3218
3219
3220
3221
3222
3223
3224
3225
3226
3227
3228
3229
3230
3231
3232
3233
3234
3235
3236
3237
3238
3239
3240
3241
3242
3243
3244
3245
3246
3247
3248
3249
3250
3251
3252
3253
3254
3255
3256
3257
3258
3259
3260
3261
3262
3263
3264
3265
3266
3267
3268
3269
3270
3271
@dataclass
class OpenAIStreamedResponse(StreamedResponse):
    """Implementation of `StreamedResponse` for OpenAI models."""

    _model_name: OpenAIModelName
    _model_profile: OpenAIModelProfile
    _response: _utils.PeekableAsyncStream[ChatCompletionChunk, AsyncStream[ChatCompletionChunk]]
    _provider_name: str
    _provider_url: str
    _provider_timestamp: datetime | None = None
    _timestamp: datetime = field(default_factory=_now_utc)
    _model_settings: OpenAIChatModelSettings | None = None
    _has_refusal: bool = field(default=False, init=False)
    _refusal_text: str = field(default='', init=False)

    async def close_stream(self) -> None:
        await self._response.source.close()

    async def _get_event_iterator(self) -> AsyncIterator[ModelResponseStreamEvent]:
        with _map_api_errors(self._model_name):
            if self._provider_timestamp is not None:  # pragma: no branch
                self.provider_details = {'timestamp': self._provider_timestamp}
            async for chunk in self._validate_response():
                chunk_usage = self._map_usage(chunk)
                if self._model_settings and self._model_settings.get('openai_continuous_usage_stats'):
                    # When continuous_usage_stats is enabled, each chunk contains cumulative usage,
                    # so we replace rather than increment to avoid double-counting.
                    self._usage = chunk_usage
                else:
                    self._usage += chunk_usage

                if chunk.id:  # pragma: no branch
                    self.provider_response_id = chunk.id

                if chunk.model:
                    self._model_name = chunk.model

                # Empty on the final usage-only chunk; `None` from OpenAI-compatible providers emitting
                # malformed chunks that the openai SDK's loose constructor lets through (#5165).
                if not chunk.choices:
                    continue
                choice = chunk.choices[0]

                # When using Azure OpenAI and an async content filter is enabled, the openai SDK can return None deltas.
                if choice.delta is None:  # pyright: ignore[reportUnnecessaryComparison]
                    continue

                # Handle refusal responses (structured output safety filter).
                # Note: OpenAI sends refusal instead of content (not alongside it), so in practice
                # text parts won't have been yielded before _has_refusal is set.
                if choice.delta.refusal:
                    self._has_refusal = True
                    self.finish_reason = 'content_filter'
                    self._refusal_text += choice.delta.refusal
                    continue

                if raw_finish_reason := choice.finish_reason:
                    if not self._has_refusal:
                        self.finish_reason = self._map_finish_reason(raw_finish_reason)

                if provider_details := self._map_provider_details(chunk):  # pragma: no branch
                    if self._has_refusal:
                        provider_details.pop('finish_reason', None)
                    self.provider_details = {**(self.provider_details or {}), **provider_details}

                for event in self._map_part_delta(choice):
                    yield event

            if self._refusal_text:
                self.provider_details = {**(self.provider_details or {}), 'refusal': self._refusal_text}

    def _validate_response(self) -> AsyncIterable[ChatCompletionChunk]:
        """Hook that validates incoming chunks.

        This method may be overridden by subclasses of `OpenAIStreamedResponse` to apply custom chunk validations.

        By default, this is a no-op since `ChatCompletionChunk` is already validated.
        """
        return self._response

    def _map_part_delta(self, choice: chat_completion_chunk.Choice) -> Iterable[ModelResponseStreamEvent]:
        """Hook that determines the sequence of mappings that will be called to produce events.

        This method may be overridden by subclasses of `OpenAIStreamResponse` to customize the mapping.
        """
        return itertools.chain(
            self._map_thinking_delta(choice), self._map_text_delta(choice), self._map_tool_call_delta(choice)
        )

    def _map_thinking_delta(self, choice: chat_completion_chunk.Choice) -> Iterable[ModelResponseStreamEvent]:
        """Hook that maps thinking delta content to events.

        This method may be overridden by subclasses of `OpenAIStreamResponse` to customize the mapping.
        """
        profile = self._model_profile
        custom_field = profile.get('openai_chat_thinking_field', None)

        # Prefer the configured custom reasoning field, if present in profile.
        # Fall back to built-in fields if no custom field result was found.

        # The `reasoning_content` field is typically present in DeepSeek and Moonshot models.
        # https://api-docs.deepseek.com/guides/reasoning_model

        # The `reasoning` field is typically present in gpt-oss via Ollama and OpenRouter.
        # - https://cookbook.openai.com/articles/gpt-oss/handle-raw-cot#chat-completions-api
        # - https://openrouter.ai/docs/use-cases/reasoning-tokens#basic-usage-with-reasoning-tokens
        for field_name in (custom_field, 'reasoning', 'reasoning_content'):
            if not field_name:
                continue
            reasoning: object = getattr(choice.delta, field_name, None)
            if not reasoning:
                continue
            if not isinstance(reasoning, str):
                warnings.warn(
                    f'Unexpected non-string value for {field_name!r}: {type(reasoning).__name__}. '
                    'Please open an issue at https://github.com/pydantic/pydantic-ai/issues.',
                    UserWarning,
                )
                continue
            yield from self._parts_manager.handle_thinking_delta(
                vendor_part_id=field_name,
                id=field_name,
                content=reasoning,
                provider_name=self.provider_name,
            )
            break

    def _map_text_delta(self, choice: chat_completion_chunk.Choice) -> Iterable[ModelResponseStreamEvent]:
        """Hook that maps text delta content to events.

        This method may be overridden by subclasses of `OpenAIStreamResponse` to customize the mapping.
        """
        # Handle the text part of the response
        content = choice.delta.content
        if content:
            for event in self._parts_manager.handle_text_delta(
                vendor_part_id='content',
                content=content,
                thinking_tags=self._model_profile.get('thinking_tags', DEFAULT_THINKING_TAGS),
                ignore_leading_whitespace=self._model_profile.get('ignore_streamed_leading_whitespace', False),
            ):
                if isinstance(event, PartStartEvent) and isinstance(event.part, ThinkingPart):
                    event.part.id = 'content'
                    event.part.provider_name = self.provider_name
                yield event

    def _map_tool_call_delta(self, choice: chat_completion_chunk.Choice) -> Iterable[ModelResponseStreamEvent]:
        """Hook that maps tool call delta content to events.

        This method may be overridden by subclasses of `OpenAIStreamResponse` to customize the mapping.
        """
        for dtc in choice.delta.tool_calls or []:
            maybe_event = self._parts_manager.handle_tool_call_delta(
                vendor_part_id=dtc.index,
                tool_name=dtc.function and dtc.function.name,
                args=dtc.function and dtc.function.arguments,
                tool_call_id=dtc.id,
            )
            if maybe_event is not None:
                yield maybe_event

    def _map_provider_details(self, chunk: ChatCompletionChunk) -> dict[str, Any] | None:
        """Hook that generates the provider details from chunk content.

        This method may be overridden by subclasses of `OpenAIStreamResponse` to customize the provider details.
        """
        return _map_provider_details(chunk.choices[0])

    def _map_usage(self, response: ChatCompletionChunk) -> usage.RequestUsage:
        return _map_usage(response, self._provider_name, self._provider_url, self.model_name)

    def _map_finish_reason(
        self, key: Literal['stop', 'length', 'tool_calls', 'content_filter', 'function_call']
    ) -> FinishReason | None:
        """Hooks that maps a finish reason key to a [FinishReason](pydantic_ai.messages.FinishReason).

        This method may be overridden by subclasses of `OpenAIChatModel` to accommodate custom keys.
        """
        return _CHAT_FINISH_REASON_MAP.get(key)

    @property
    def model_name(self) -> OpenAIModelName:
        """Get the model name of the response."""
        return self._model_name

    @property
    def provider_name(self) -> str:
        """Get the provider name."""
        return self._provider_name

    @property
    def provider_url(self) -> str:
        """Get the provider base URL."""
        return self._provider_url

    @property
    def timestamp(self) -> datetime:
        """Get the timestamp of the response."""
        return self._timestamp

model_name property 

model_name: OpenAIModelName

Get the model name of the response.

provider_name property 

provider_name: str

Get the provider name.

provider_url property 

provider_url: str

Get the provider base URL.

timestamp property 

timestamp: datetime

Get the timestamp of the response.

OpenAIResponsesStreamedResponse dataclass

Bases: StreamedResponse

Implementation of StreamedResponse for OpenAI Responses API.

Source code in pydantic_ai_slim/pydantic_ai/models/openai.py 
3274
3275
3276
3277
3278
3279
3280
3281
3282
3283
3284
3285
3286
3287
3288
3289
3290
3291
3292
3293
3294
3295
3296
3297
3298
3299
3300
3301
3302
3303
3304
3305
3306
3307
3308
3309
3310
3311
3312
3313
3314
3315
3316
3317
3318
3319
3320
3321
3322
3323
3324
3325
3326
3327
3328
3329
3330
3331
3332
3333
3334
3335
3336
3337
3338
3339
3340
3341
3342
3343
3344
3345
3346
3347
3348
3349
3350
3351
3352
3353
3354
3355
3356
3357
3358
3359
3360
3361
3362
3363
3364
3365
3366
3367
3368
3369
3370
3371
3372
3373
3374
3375
3376
3377
3378
3379
3380
3381
3382
3383
3384
3385
3386
3387
3388
3389
3390
3391
3392
3393
3394
3395
3396
3397
3398
3399
3400
3401
3402
3403
3404
3405
3406
3407
3408
3409
3410
3411
3412
3413
3414
3415
3416
3417
3418
3419
3420
3421
3422
3423
3424
3425
3426
3427
3428
3429
3430
3431
3432
3433
3434
3435
3436
3437
3438
3439
3440
3441
3442
3443
3444
3445
3446
3447
3448
3449
3450
3451
3452
3453
3454
3455
3456
3457
3458
3459
3460
3461
3462
3463
3464
3465
3466
3467
3468
3469
3470
3471
3472
3473
3474
3475
3476
3477
3478
3479
3480
3481
3482
3483
3484
3485
3486
3487
3488
3489
3490
3491
3492
3493
3494
3495
3496
3497
3498
3499
3500
3501
3502
3503
3504
3505
3506
3507
3508
3509
3510
3511
3512
3513
3514
3515
3516
3517
3518
3519
3520
3521
3522
3523
3524
3525
3526
3527
3528
3529
3530
3531
3532
3533
3534
3535
3536
3537
3538
3539
3540
3541
3542
3543
3544
3545
3546
3547
3548
3549
3550
3551
3552
3553
3554
3555
3556
3557
3558
3559
3560
3561
3562
3563
3564
3565
3566
3567
3568
3569
3570
3571
3572
3573
3574
3575
3576
3577
3578
3579
3580
3581
3582
3583
3584
3585
3586
3587
3588
3589
3590
3591
3592
3593
3594
3595
3596
3597
3598
3599
3600
3601
3602
3603
3604
3605
3606
3607
3608
3609
3610
3611
3612
3613
3614
3615
3616
3617
3618
3619
3620
3621
3622
3623
3624
3625
3626
3627
3628
3629
3630
3631
3632
3633
3634
3635
3636
3637
3638
3639
3640
3641
3642
3643
3644
3645
3646
3647
3648
3649
3650
3651
3652
3653
3654
3655
3656
3657
3658
3659
3660
3661
3662
3663
3664
3665
3666
3667
3668
3669
3670
3671
3672
3673
3674
3675
3676
3677
3678
3679
3680
3681
3682
3683
3684
3685
3686
3687
3688
3689
3690
3691
3692
3693
3694
3695
3696
3697
3698
3699
3700
3701
3702
3703
3704
3705
3706
3707
3708
3709
3710
3711
3712
3713
3714
3715
3716
3717
3718
3719
3720
3721
3722
3723
3724
3725
3726
3727
3728
3729
3730
3731
3732
3733
3734
3735
3736
3737
3738
3739
3740
3741
3742
3743
3744
3745
3746
3747
3748
3749
3750
3751
3752
3753
3754
3755
3756
3757
3758
3759
3760
3761
3762
3763
3764
3765
3766
3767
3768
3769
3770
3771
3772
3773
3774
3775
3776
3777
3778
3779
3780
3781
3782
3783
3784
3785
3786
3787
3788
3789
3790
3791
3792
3793
3794
3795
3796
3797
3798
3799
3800
3801
3802
3803
3804
3805
@dataclass
class OpenAIResponsesStreamedResponse(StreamedResponse):
    """Implementation of `StreamedResponse` for OpenAI Responses API."""

    _model_name: OpenAIModelName
    _model_settings: OpenAIResponsesModelSettings
    _response: _utils.PeekableAsyncStream[responses.ResponseStreamEvent, AsyncStream[responses.ResponseStreamEvent]]
    _provider_name: str
    _provider_url: str
    _provider_timestamp: datetime | None = None
    _timestamp: datetime = field(default_factory=_now_utc)
    _has_refusal: bool = field(default=False, init=False)
    _refusal_text: str = field(default='', init=False)

    async def close_stream(self) -> None:
        await self._response.source.close()

    async def _get_event_iterator(self) -> AsyncIterator[ModelResponseStreamEvent]:  # noqa: C901
        with _map_api_errors(self._model_name):
            # Track annotations by item_id and content_index
            _annotations_by_item: dict[str, list[Any]] = {}
            # Track `phase` (commentary | final_answer) on assistant message items, captured
            # from the `output_item.added` event and merged into the corresponding
            # `TextPart.provider_details` on `output_text.done`.
            _phase_by_item: dict[str, Literal['commentary', 'final_answer']] = {}

            if self._provider_timestamp is not None:  # pragma: no branch
                self.provider_details = {'timestamp': self._provider_timestamp}

            async for chunk in self._response:
                # NOTE: You can inspect the builtin tools used checking the `ResponseCompletedEvent`.
                if isinstance(chunk, responses.ResponseCompletedEvent):
                    self._usage += self._map_usage(chunk.response)
                    self._store_conversation_id(chunk.response)

                    raw_finish_reason = (
                        details.reason if (details := chunk.response.incomplete_details) else chunk.response.status
                    )

                    if raw_finish_reason:  # pragma: no branch
                        if not self._has_refusal:
                            self.provider_details = {
                                **(self.provider_details or {}),
                                'finish_reason': raw_finish_reason,
                            }
                            self.finish_reason = _RESPONSES_FINISH_REASON_MAP.get(raw_finish_reason)

                elif isinstance(chunk, responses.ResponseContentPartAddedEvent):
                    pass  # there's nothing we need to do here

                elif isinstance(chunk, responses.ResponseContentPartDoneEvent):
                    pass  # there's nothing we need to do here

                elif isinstance(chunk, responses.ResponseCreatedEvent):
                    if chunk.response.id:  # pragma: no branch
                        self.provider_response_id = chunk.response.id
                    self._store_conversation_id(chunk.response)

                elif isinstance(chunk, responses.ResponseFailedEvent):  # pragma: no cover
                    self._usage += self._map_usage(chunk.response)

                elif isinstance(chunk, responses.ResponseFunctionCallArgumentsDeltaEvent):
                    maybe_event = self._parts_manager.handle_tool_call_delta(
                        vendor_part_id=chunk.item_id,
                        args=chunk.delta,
                    )
                    if maybe_event is not None:  # pragma: no branch
                        yield maybe_event

                elif isinstance(chunk, responses.ResponseFunctionCallArgumentsDoneEvent):
                    pass  # there's nothing we need to do here

                elif isinstance(chunk, responses.ResponseIncompleteEvent):  # pragma: no cover
                    self._usage += self._map_usage(chunk.response)

                elif isinstance(chunk, responses.ResponseInProgressEvent):
                    self._usage += self._map_usage(chunk.response)

                elif isinstance(chunk, responses.ResponseOutputItemAddedEvent):
                    if isinstance(chunk.item, responses.ResponseFunctionToolCall):
                        # Preserve any `namespace` the Responses API attaches to a
                        # discovered deferred tool so it can be round-tripped on replay.
                        fn_provider_details: dict[str, Any] | None = None
                        if chunk.item.namespace:
                            fn_provider_details = {'namespace': chunk.item.namespace}
                        yield self._parts_manager.handle_tool_call_part(
                            vendor_part_id=chunk.item.id,
                            tool_name=chunk.item.name,
                            args=chunk.item.arguments,
                            tool_call_id=chunk.item.call_id,
                            id=chunk.item.id,
                            provider_name=self.provider_name,
                            provider_details=fn_provider_details,
                        )
                    elif isinstance(chunk.item, responses.ResponseReasoningItem):
                        pass
                    elif isinstance(chunk.item, responses.ResponseOutputMessage):
                        if chunk.item.phase is not None:
                            _phase_by_item[chunk.item.id] = chunk.item.phase
                    elif isinstance(chunk.item, responses.ResponseFunctionWebSearch):
                        call_part, _ = _map_web_search_tool_call(chunk.item, self.provider_name)
                        yield self._parts_manager.handle_part(
                            vendor_part_id=f'{chunk.item.id}-call', part=replace(call_part, args=None)
                        )
                    elif isinstance(chunk.item, responses.ResponseToolSearchCall):
                        if chunk.item.execution == 'client':
                            # Emit a regular function-tool call so the standard
                            # tool-execution path runs our local `search_tools`.
                            client_call_part = _map_client_tool_search_call(chunk.item, self.provider_name)
                            yield self._parts_manager.handle_tool_call_part(
                                vendor_part_id=chunk.item.id,
                                tool_name=client_call_part.tool_name,
                                args=None,
                                tool_call_id=client_call_part.tool_call_id,
                                id=client_call_part.id,
                                provider_name=self.provider_name,
                            )
                        else:
                            call_part, _ = _map_tool_search_call(chunk.item, None, self.provider_name)
                            yield self._parts_manager.handle_part(
                                vendor_part_id=f'{chunk.item.id}-call', part=replace(call_part, args=None)
                            )
                    elif isinstance(chunk.item, responses.ResponseFileSearchToolCall):
                        call_part, _ = _map_file_search_tool_call(chunk.item, self.provider_name)
                        yield self._parts_manager.handle_part(
                            vendor_part_id=f'{chunk.item.id}-call', part=replace(call_part, args=None)
                        )
                    elif isinstance(chunk.item, responses.ResponseToolSearchOutputItem):
                        # Added carries the full discovered-tools payload; emit the return
                        # part here. Done repeats the same item — handled as a no-op
                        # below to avoid double-emission.
                        call_id = chunk.item.call_id or chunk.item.id
                        yield self._parts_manager.handle_part(
                            vendor_part_id=f'{chunk.item.id}-return',
                            part=_build_tool_search_return_part(
                                call_id, chunk.item.status or 'completed', chunk.item, self.provider_name
                            ),
                        )
                    elif isinstance(chunk.item, responses.ResponseCodeInterpreterToolCall):
                        call_part, _, _ = _map_code_interpreter_tool_call(chunk.item, self.provider_name)

                        args_json = call_part.args_as_json_str()
                        # Drop the final `"}` so that we can add code deltas
                        args_json_delta = args_json[:-2]
                        assert args_json_delta.endswith('"code":"'), (
                            f'Expected {args_json_delta!r} to end in `"code":"`'
                        )

                        yield self._parts_manager.handle_part(
                            vendor_part_id=f'{chunk.item.id}-call', part=replace(call_part, args=None)
                        )
                        maybe_event = self._parts_manager.handle_tool_call_delta(
                            vendor_part_id=f'{chunk.item.id}-call',
                            args=args_json_delta,
                        )
                        if maybe_event is not None:  # pragma: no branch
                            yield maybe_event
                    elif isinstance(chunk.item, responses.response_output_item.ImageGenerationCall):
                        call_part, _, _ = _map_image_generation_tool_call(chunk.item, self.provider_name)
                        yield self._parts_manager.handle_part(vendor_part_id=f'{chunk.item.id}-call', part=call_part)
                    elif isinstance(chunk.item, responses.response_output_item.McpCall):
                        call_part, _ = _map_mcp_call(chunk.item, self.provider_name)

                        args_json = call_part.args_as_json_str()
                        # Drop the final `{}}` so that we can add tool args deltas
                        args_json_delta = args_json[:-3]
                        assert args_json_delta.endswith('"tool_args":'), (
                            f'Expected {args_json_delta!r} to end in `"tool_args":"`'
                        )

                        yield self._parts_manager.handle_part(
                            vendor_part_id=f'{chunk.item.id}-call', part=replace(call_part, args=None)
                        )
                        maybe_event = self._parts_manager.handle_tool_call_delta(
                            vendor_part_id=f'{chunk.item.id}-call',
                            args=args_json_delta,
                        )
                        if maybe_event is not None:  # pragma: no branch
                            yield maybe_event
                    elif isinstance(chunk.item, responses.response_output_item.McpListTools):
                        call_part, _ = _map_mcp_list_tools(chunk.item, self.provider_name)
                        yield self._parts_manager.handle_part(vendor_part_id=f'{chunk.item.id}-call', part=call_part)
                    elif isinstance(chunk.item, ResponseCompactionItem):
                        # Emit a PartStartEvent so UIs can render compaction in progress.
                        # The "done" event replaces this with the final encrypted_content.
                        yield self._parts_manager.handle_part(
                            vendor_part_id=chunk.item.id,
                            part=_map_compaction_item(chunk.item, self.provider_name),
                        )
                    else:
                        warnings.warn(  # pragma: no cover
                            f'Handling of this item type is not yet implemented. Please report on our GitHub: {chunk}',
                            UserWarning,
                        )

                elif isinstance(chunk, responses.ResponseOutputItemDoneEvent):
                    if isinstance(chunk.item, responses.ResponseReasoningItem):
                        if signature := chunk.item.encrypted_content:  # pragma: no branch
                            # Add the signature to the part corresponding to the first summary/raw CoT
                            for event in self._parts_manager.handle_thinking_delta(
                                vendor_part_id=chunk.item.id,
                                id=chunk.item.id,
                                signature=signature,
                                provider_name=self.provider_name,
                            ):
                                yield event
                    elif isinstance(chunk.item, responses.ResponseCodeInterpreterToolCall):
                        _, return_part, file_parts = _map_code_interpreter_tool_call(chunk.item, self.provider_name)
                        for i, file_part in enumerate(file_parts):
                            yield self._parts_manager.handle_part(
                                vendor_part_id=f'{chunk.item.id}-file-{i}', part=file_part
                            )
                        yield self._parts_manager.handle_part(
                            vendor_part_id=f'{chunk.item.id}-return', part=return_part
                        )
                    elif isinstance(chunk.item, responses.ResponseFunctionWebSearch):
                        call_part, return_part = _map_web_search_tool_call(chunk.item, self.provider_name)

                        maybe_event = self._parts_manager.handle_tool_call_delta(
                            vendor_part_id=f'{chunk.item.id}-call',
                            args=call_part.args,
                        )
                        if maybe_event is not None:  # pragma: no branch
                            yield maybe_event

                        yield self._parts_manager.handle_part(
                            vendor_part_id=f'{chunk.item.id}-return', part=return_part
                        )
                    elif isinstance(chunk.item, responses.ResponseToolSearchCall):
                        if chunk.item.execution == 'client':
                            # Feed the final args through the tool-call delta path so the
                            # part manager resolves its pending args; there's no paired
                            # `tool_search_output` — the return part is produced by the
                            # local tool runner after the stream completes.
                            #
                            # OpenAI Responses streaming sometimes assigns a different
                            # `call_id` between the `output_item.added` and `output_item.done`
                            # frames for client-executed `tool_search_call`s; the server's
                            # log of record is the latter. Pass it through here so the
                            # `tool_search_output` we replay on the next turn references the
                            # call_id the API actually expects.
                            client_call_part = _map_client_tool_search_call(chunk.item, self.provider_name)
                            client_args_json = client_call_part.args_as_json_str()
                            maybe_event = self._parts_manager.handle_tool_call_delta(
                                vendor_part_id=chunk.item.id,
                                args=client_args_json,
                                tool_call_id=client_call_part.tool_call_id,
                            )
                            if maybe_event is not None:  # pragma: no branch
                                yield maybe_event
                        else:
                            call_part, _ = _map_tool_search_call(chunk.item, None, self.provider_name)

                            maybe_event = self._parts_manager.handle_tool_call_delta(
                                vendor_part_id=f'{chunk.item.id}-call',
                                args=cast('str | dict[str, Any] | None', call_part.args),
                            )
                            if maybe_event is not None:  # pragma: no branch
                                yield maybe_event
                    elif isinstance(chunk.item, responses.ResponseToolSearchOutputItem):
                        # Already emitted on the matching Added event — the Done payload
                        # is a duplicate.
                        pass
                    elif isinstance(chunk.item, responses.ResponseFileSearchToolCall):
                        call_part, return_part = _map_file_search_tool_call(chunk.item, self.provider_name)

                        maybe_event = self._parts_manager.handle_tool_call_delta(
                            vendor_part_id=f'{chunk.item.id}-call',
                            args=call_part.args,
                        )
                        if maybe_event is not None:  # pragma: no branch
                            yield maybe_event

                        yield self._parts_manager.handle_part(
                            vendor_part_id=f'{chunk.item.id}-return', part=return_part
                        )
                    elif isinstance(chunk.item, responses.response_output_item.ImageGenerationCall):
                        _, return_part, file_part = _map_image_generation_tool_call(chunk.item, self.provider_name)
                        if file_part:  # pragma: no branch
                            yield self._parts_manager.handle_part(
                                vendor_part_id=f'{chunk.item.id}-file', part=file_part
                            )
                        yield self._parts_manager.handle_part(
                            vendor_part_id=f'{chunk.item.id}-return', part=return_part
                        )

                    elif isinstance(chunk.item, responses.response_output_item.McpCall):
                        _, return_part = _map_mcp_call(chunk.item, self.provider_name)
                        yield self._parts_manager.handle_part(
                            vendor_part_id=f'{chunk.item.id}-return', part=return_part
                        )
                    elif isinstance(chunk.item, responses.response_output_item.McpListTools):
                        _, return_part = _map_mcp_list_tools(chunk.item, self.provider_name)
                        yield self._parts_manager.handle_part(
                            vendor_part_id=f'{chunk.item.id}-return', part=return_part
                        )
                    elif isinstance(chunk.item, ResponseCompactionItem):
                        # Replace the preliminary part from the "added" event with the
                        # final encrypted_content for round-tripping.
                        yield self._parts_manager.handle_part(
                            vendor_part_id=chunk.item.id,
                            part=_map_compaction_item(chunk.item, self.provider_name),
                        )

                elif isinstance(chunk, responses.ResponseReasoningSummaryPartAddedEvent):
                    # Use same vendor_part_id as raw CoT for first summary (index 0) so they merge into one ThinkingPart
                    vendor_id = chunk.item_id if chunk.summary_index == 0 else f'{chunk.item_id}-{chunk.summary_index}'
                    for event in self._parts_manager.handle_thinking_delta(
                        vendor_part_id=vendor_id,
                        content=chunk.part.text,
                        id=chunk.item_id,
                        provider_name=self.provider_name,
                    ):
                        yield event

                elif isinstance(chunk, responses.ResponseReasoningSummaryPartDoneEvent):
                    pass  # there's nothing we need to do here

                elif isinstance(chunk, responses.ResponseReasoningSummaryTextDoneEvent):
                    pass  # there's nothing we need to do here

                elif isinstance(chunk, responses.ResponseReasoningSummaryTextDeltaEvent):
                    # Use same vendor_part_id as raw CoT for first summary (index 0) so they merge into one ThinkingPart
                    vendor_id = chunk.item_id if chunk.summary_index == 0 else f'{chunk.item_id}-{chunk.summary_index}'
                    for event in self._parts_manager.handle_thinking_delta(
                        vendor_part_id=vendor_id,
                        content=chunk.delta,
                        id=chunk.item_id,
                        provider_name=self.provider_name,
                    ):
                        yield event

                elif isinstance(chunk, responses.ResponseReasoningTextDeltaEvent):
                    # Handle raw CoT from gpt-oss models using callback pattern
                    for event in self._parts_manager.handle_thinking_delta(
                        vendor_part_id=chunk.item_id,
                        id=chunk.item_id,
                        provider_name=self.provider_name,
                        provider_details=_make_raw_content_updater(chunk.delta, chunk.content_index),
                    ):
                        yield event

                elif isinstance(chunk, responses.ResponseReasoningTextDoneEvent):
                    pass  # content already accumulated via delta events

                elif isinstance(chunk, responses.ResponseOutputTextAnnotationAddedEvent):
                    # Collect annotations if the setting is enabled
                    if self._model_settings.get('openai_include_raw_annotations'):
                        _annotations_by_item.setdefault(chunk.item_id, []).append(chunk.annotation)

                elif isinstance(chunk, responses.ResponseTextDeltaEvent):
                    # Guard against delta=null from OpenAI-compatible gateways (e.g. Bifrost).
                    if chunk.delta is not None:  # pyright: ignore[reportUnnecessaryComparison]
                        for event in self._parts_manager.handle_text_delta(
                            vendor_part_id=chunk.item_id,
                            content=chunk.delta,
                            id=chunk.item_id,
                            provider_name=self.provider_name,
                        ):
                            yield event

                elif isinstance(chunk, responses.ResponseTextDoneEvent):
                    # Add annotations to provider_details if available
                    provider_details: dict[str, Any] = {}
                    annotations = _annotations_by_item.get(chunk.item_id)
                    if annotations:
                        provider_details['annotations'] = responses_output_text_annotations_ta.dump_python(
                            list(annotations), warnings=False
                        )
                    if chunk.logprobs:
                        provider_details['logprobs'] = _map_logprobs(chunk.logprobs)
                    if (phase := _phase_by_item.get(chunk.item_id)) is not None:
                        provider_details['phase'] = phase
                    if provider_details:
                        for event in self._parts_manager.handle_text_delta(
                            vendor_part_id=chunk.item_id,
                            content='',
                            provider_name=self.provider_name,
                            provider_details=provider_details,
                        ):
                            yield event

                elif isinstance(chunk, responses.ResponseRefusalDeltaEvent):
                    # Accumulate refusal text from deltas as a fallback in case the done event is missing.
                    self._has_refusal = True
                    self.finish_reason = 'content_filter'
                    self._refusal_text += chunk.delta

                elif isinstance(chunk, responses.ResponseRefusalDoneEvent):
                    # The done event contains the full refusal text, replacing any accumulated deltas.
                    self._has_refusal = True
                    self.finish_reason = 'content_filter'
                    self._refusal_text = chunk.refusal

                elif isinstance(chunk, responses.ResponseWebSearchCallInProgressEvent):
                    pass  # there's nothing we need to do here

                elif isinstance(chunk, responses.ResponseWebSearchCallSearchingEvent):
                    pass  # there's nothing we need to do here

                elif isinstance(chunk, responses.ResponseWebSearchCallCompletedEvent):
                    pass  # there's nothing we need to do here

                elif isinstance(chunk, responses.ResponseAudioDeltaEvent):  # pragma: lax no cover
                    pass  # there's nothing we need to do here

                elif isinstance(chunk, responses.ResponseCodeInterpreterCallCodeDeltaEvent):
                    json_args_delta = to_json(chunk.delta).decode()[1:-1]  # Drop the surrounding `"`
                    maybe_event = self._parts_manager.handle_tool_call_delta(
                        vendor_part_id=f'{chunk.item_id}-call',
                        args=json_args_delta,
                    )
                    if maybe_event is not None:  # pragma: no branch
                        yield maybe_event

                elif isinstance(chunk, responses.ResponseCodeInterpreterCallCodeDoneEvent):
                    maybe_event = self._parts_manager.handle_tool_call_delta(
                        vendor_part_id=f'{chunk.item_id}-call',
                        args='"}',
                    )
                    if maybe_event is not None:  # pragma: no branch
                        yield maybe_event

                elif isinstance(chunk, responses.ResponseCodeInterpreterCallCompletedEvent):
                    pass  # there's nothing we need to do here

                elif isinstance(chunk, responses.ResponseCodeInterpreterCallInProgressEvent):
                    pass  # there's nothing we need to do here

                elif isinstance(chunk, responses.ResponseCodeInterpreterCallInterpretingEvent):
                    pass  # there's nothing we need to do here

                elif isinstance(chunk, responses.ResponseImageGenCallCompletedEvent):  # pragma: no cover
                    pass  # there's nothing we need to do here

                elif isinstance(chunk, responses.ResponseImageGenCallGeneratingEvent):
                    pass  # there's nothing we need to do here

                elif isinstance(chunk, responses.ResponseImageGenCallInProgressEvent):
                    pass  # there's nothing we need to do here

                elif isinstance(chunk, responses.ResponseImageGenCallPartialImageEvent):
                    # Not present on the type, but present on the actual object.
                    # See https://github.com/openai/openai-python/issues/2649
                    output_format = getattr(chunk, 'output_format', 'png')
                    file_part = FilePart(
                        content=BinaryImage(
                            data=base64.b64decode(chunk.partial_image_b64),
                            media_type=f'image/{output_format}',
                        ),
                        id=chunk.item_id,
                    )
                    yield self._parts_manager.handle_part(vendor_part_id=f'{chunk.item_id}-file', part=file_part)

                elif isinstance(chunk, responses.ResponseMcpCallArgumentsDoneEvent):
                    maybe_event = self._parts_manager.handle_tool_call_delta(
                        vendor_part_id=f'{chunk.item_id}-call',
                        args='}',
                    )
                    if maybe_event is not None:  # pragma: no branch
                        yield maybe_event

                elif isinstance(chunk, responses.ResponseMcpCallArgumentsDeltaEvent):
                    maybe_event = self._parts_manager.handle_tool_call_delta(
                        vendor_part_id=f'{chunk.item_id}-call',
                        args=chunk.delta,
                    )
                    if maybe_event is not None:  # pragma: no branch
                        yield maybe_event

                elif isinstance(chunk, responses.ResponseMcpListToolsInProgressEvent):
                    pass  # there's nothing we need to do here

                elif isinstance(chunk, responses.ResponseMcpListToolsCompletedEvent):
                    pass  # there's nothing we need to do here

                elif isinstance(chunk, responses.ResponseMcpListToolsFailedEvent):  # pragma: no cover
                    pass  # there's nothing we need to do here

                elif isinstance(chunk, responses.ResponseMcpCallInProgressEvent):
                    pass  # there's nothing we need to do here

                elif isinstance(chunk, responses.ResponseMcpCallFailedEvent):  # pragma: no cover
                    pass  # there's nothing we need to do here

                elif isinstance(chunk, responses.ResponseMcpCallCompletedEvent):
                    pass  # there's nothing we need to do here

                elif isinstance(chunk, responses.ResponseFileSearchCallCompletedEvent):
                    pass  # there's nothing we need to do here

                elif isinstance(chunk, responses.ResponseFileSearchCallSearchingEvent):
                    pass  # there's nothing we need to do here

                elif isinstance(chunk, responses.ResponseFileSearchCallInProgressEvent):
                    pass  # there's nothing we need to do here

                else:  # pragma: no cover
                    warnings.warn(
                        f'Handling of this event type is not yet implemented. Please report on our GitHub: {chunk}',
                        UserWarning,
                    )

            if self._refusal_text:
                self.provider_details = {**(self.provider_details or {}), 'refusal': self._refusal_text}

    def _store_conversation_id(self, response: responses.Response) -> None:
        if response.conversation:
            self.provider_details = {**(self.provider_details or {}), 'conversation_id': response.conversation.id}

    def _map_usage(self, response: responses.Response) -> usage.RequestUsage:
        return _map_usage(response, self._provider_name, self._provider_url, self.model_name)

    @property
    def model_name(self) -> OpenAIModelName:
        """Get the model name of the response."""
        return self._model_name

    @property
    def provider_name(self) -> str:
        """Get the provider name."""
        return self._provider_name

    @property
    def provider_url(self) -> str:
        """Get the provider base URL."""
        return self._provider_url

    @property
    def timestamp(self) -> datetime:
        """Get the timestamp of the response."""
        return self._timestamp

model_name property 

model_name: OpenAIModelName

Get the model name of the response.

provider_name property 

provider_name: str

Get the provider name.

provider_url property 

provider_url: str

Get the provider base URL.

timestamp property 

timestamp: datetime

Get the timestamp of the response.

OpenAICompaction

Bases: AbstractCapability[AgentDepsT]

Compaction capability for OpenAI Responses API.

Automatically compacts conversation history to keep long-running agent runs within manageable context limits. Two modes are supported, selected by the stateless flag:

  • Stateful mode (default, stateless=False): configures OpenAI's server-side auto-compaction via the context_management field on the regular /responses request. The server triggers compaction when input tokens cross a threshold, and the compacted item is returned alongside the normal response. Compatible with openai_previous_response_id='auto' and server-side conversation state.

Configurable with token_threshold (compact_threshold on the API). If omitted, OpenAI picks a server-side default.

  • Stateless mode (stateless=True): calls the stateless /responses/compact endpoint from a before_model_request hook when your trigger condition is met. Use this in ZDR environments where OpenAI must not retain conversation data, when you set openai_store=False, or when you need explicit out-of-band control over when compaction runs.

Requires either message_count_threshold or a custom trigger callable.

If stateless is not set, it is inferred from which parameters you provide: passing any stateless-only parameter (message_count_threshold or trigger) implies stateless=True; otherwise stateful mode is used.

Example usage:

from pydantic_ai import Agent
from pydantic_ai.models.openai import OpenAICompaction

# Stateful mode with OpenAI's server-side default threshold:
agent = Agent(
    'openai-responses:gpt-5.2',
    capabilities=[OpenAICompaction()],
)

# Stateful mode with a custom token threshold:
agent = Agent(
    'openai-responses:gpt-5.2',
    capabilities=[OpenAICompaction(token_threshold=100_000)],
)

# Stateless mode for ZDR environments or explicit control:
agent = Agent(
    'openai-responses:gpt-5.2',
    capabilities=[OpenAICompaction(message_count_threshold=20)],
)
Source code in pydantic_ai_slim/pydantic_ai/models/openai.py 
3808
3809
3810
3811
3812
3813
3814
3815
3816
3817
3818
3819
3820
3821
3822
3823
3824
3825
3826
3827
3828
3829
3830
3831
3832
3833
3834
3835
3836
3837
3838
3839
3840
3841
3842
3843
3844
3845
3846
3847
3848
3849
3850
3851
3852
3853
3854
3855
3856
3857
3858
3859
3860
3861
3862
3863
3864
3865
3866
3867
3868
3869
3870
3871
3872
3873
3874
3875
3876
3877
3878
3879
3880
3881
3882
3883
3884
3885
3886
3887
3888
3889
3890
3891
3892
3893
3894
3895
3896
3897
3898
3899
3900
3901
3902
3903
3904
3905
3906
3907
3908
3909
3910
3911
3912
3913
3914
3915
3916
3917
3918
3919
3920
3921
3922
3923
3924
3925
3926
3927
3928
3929
3930
3931
3932
3933
3934
3935
3936
3937
3938
3939
3940
3941
3942
3943
3944
3945
3946
3947
3948
3949
3950
3951
3952
3953
3954
3955
3956
3957
3958
3959
3960
3961
3962
3963
3964
3965
3966
3967
3968
3969
3970
3971
3972
3973
3974
3975
3976
3977
3978
3979
3980
3981
3982
3983
3984
3985
3986
3987
3988
class OpenAICompaction(AbstractCapability[AgentDepsT]):
    """Compaction capability for OpenAI Responses API.

    Automatically compacts conversation history to keep long-running agent
    runs within manageable context limits. Two modes are supported, selected
    by the `stateless` flag:

    - **Stateful mode** (default, `stateless=False`): configures
      [OpenAI's server-side auto-compaction](https://developers.openai.com/api/docs/guides/compaction)
      via the `context_management` field on the regular `/responses` request.
      The server triggers compaction when input tokens cross a threshold,
      and the compacted item is returned alongside the normal response.
      Compatible with [`openai_previous_response_id='auto'`][pydantic_ai.models.openai.OpenAIResponsesModelSettings.openai_previous_response_id]
      and server-side conversation state.

      Configurable with `token_threshold` (`compact_threshold` on the API).
      If omitted, OpenAI picks a server-side default.

    - **Stateless mode** (`stateless=True`): calls the stateless
      `/responses/compact` endpoint from a `before_model_request` hook when
      your trigger condition is met. Use this in
      [ZDR](https://openai.com/enterprise-privacy/) environments where
      OpenAI must not retain conversation data, when you set
      [`openai_store=False`][pydantic_ai.models.openai.OpenAIResponsesModelSettings.openai_store],
      or when you need explicit out-of-band control over when compaction runs.

      Requires either `message_count_threshold` or a custom `trigger` callable.

    If `stateless` is not set, it is inferred from which parameters you
    provide: passing any stateless-only parameter (`message_count_threshold`
    or `trigger`) implies `stateless=True`; otherwise stateful mode is used.

    Example usage:

    ```python {test="skip"}
    from pydantic_ai import Agent
    from pydantic_ai.models.openai import OpenAICompaction

    # Stateful mode with OpenAI's server-side default threshold:
    agent = Agent(
        'openai-responses:gpt-5.2',
        capabilities=[OpenAICompaction()],
    )

    # Stateful mode with a custom token threshold:
    agent = Agent(
        'openai-responses:gpt-5.2',
        capabilities=[OpenAICompaction(token_threshold=100_000)],
    )

    # Stateless mode for ZDR environments or explicit control:
    agent = Agent(
        'openai-responses:gpt-5.2',
        capabilities=[OpenAICompaction(message_count_threshold=20)],
    )
    ```
    """

    def __init__(
        self,
        *,
        stateless: bool | None = None,
        token_threshold: int | None = None,
        message_count_threshold: int | None = None,
        trigger: Callable[[list[ModelMessage]], bool] | None = None,
    ) -> None:
        """Initialize the OpenAI compaction capability.

        Args:
            stateless: Select the compaction mode explicitly. If `None` (the
                default), the mode is inferred from the other parameters:
                passing any stateless-only parameter (`message_count_threshold`
                or `trigger`) implies `stateless=True`; otherwise stateful
                mode is used.
            token_threshold: Stateful-mode only. Input token threshold at which
                OpenAI's server-side compaction is triggered. Corresponds to
                `compact_threshold` in the `context_management` API field. If
                `None`, OpenAI picks a server-side default.
            message_count_threshold: Stateless-mode only. Compact when the
                message count exceeds this threshold.
            trigger: Stateless-mode only. Custom callable that decides whether
                to compact based on the current messages. Takes precedence
                over `message_count_threshold`.
        """
        has_stateless_only = message_count_threshold is not None or trigger is not None
        has_stateful_only = token_threshold is not None

        if stateless is None:
            stateless = has_stateless_only

        if stateless:
            if has_stateful_only:
                raise UserError(
                    '`token_threshold` is only valid for stateful compaction (`stateless=False`). '
                    'For stateless `/compact` endpoint compaction, use `message_count_threshold` or `trigger`.'
                )
            if not has_stateless_only:
                raise UserError(
                    '`stateless=True` requires `message_count_threshold` or `trigger` '
                    'to determine when to invoke the `/compact` endpoint.'
                )
        else:
            if has_stateless_only:
                raise UserError(
                    '`message_count_threshold` and `trigger` are only valid for stateless compaction '
                    '(`stateless=True`). For stateful server-side compaction, use `token_threshold` '
                    '(or omit it to use the OpenAI-managed default).'
                )

        self.stateless = stateless
        self.token_threshold = token_threshold
        self.message_count_threshold = message_count_threshold
        self.trigger = trigger

    def get_model_settings(self) -> Callable[[RunContext[AgentDepsT]], ModelSettings] | None:
        if self.stateless:
            return None
        edit: ContextManagement = {'type': 'compaction'}
        if self.token_threshold is not None:
            edit['compact_threshold'] = self.token_threshold

        def resolve(ctx: RunContext[AgentDepsT]) -> ModelSettings:
            # If the user already set `openai_context_management` on their model settings,
            # defer to it entirely — we don't want to end up with two conflicting `compaction`
            # entries, since OpenAI's context_management list only meaningfully supports one.
            if ctx.model_settings:
                existing = cast(dict[str, Any], ctx.model_settings).get('openai_context_management')
                if existing:
                    return cast(ModelSettings, {})
            return cast(ModelSettings, {'openai_context_management': [edit]})

        return resolve

    def _should_compact(self, messages: list[ModelMessage]) -> bool:
        if not self.stateless:
            return False
        if self.trigger is not None:
            return self.trigger(messages)
        if self.message_count_threshold is not None:
            return len(messages) > self.message_count_threshold
        return False  # pragma: no cover

    async def before_model_request(
        self,
        ctx: RunContext[AgentDepsT],
        request_context: ModelRequestContext,
    ) -> ModelRequestContext:
        if not self._should_compact(request_context.messages):
            return request_context

        from .wrapper import WrapperModel

        model = request_context.model
        while isinstance(model, WrapperModel):
            model = model.wrapped
        if not isinstance(model, OpenAIResponsesModel):
            raise UserError(
                f'OpenAICompaction requires OpenAIResponsesModel, got {type(model).__name__}. '
                f'Use the provider-specific compaction capability for your model.'
            )

        # Need at least 2 messages (history + current request) to compact
        if len(request_context.messages) < 2:  # pragma: no cover
            return request_context

        # Compact all messages except the last (current) request
        compact_ctx = ModelRequestContext(
            model=request_context.model,
            messages=request_context.messages[:-1],
            model_settings=request_context.model_settings,
            model_request_parameters=request_context.model_request_parameters,
        )
        compacted_response = await request_context.model.compact_messages(compact_ctx)

        # Replace message history with compaction + last request
        request_context.messages = [compacted_response, request_context.messages[-1]]
        return request_context

    @classmethod
    def get_serialization_name(cls) -> str | None:
        return 'OpenAICompaction'

__init__ 

__init__(
    *,
    stateless: bool | None = None,
    token_threshold: int | None = None,
    message_count_threshold: int | None = None,
    trigger: (
        Callable[[list[ModelMessage]], bool] | None
    ) = None
) -> None

Initialize the OpenAI compaction capability.

Parameters:

Name Type Description Default
stateless bool | None

Select the compaction mode explicitly. If None (the default), the mode is inferred from the other parameters: passing any stateless-only parameter (message_count_threshold or trigger) implies stateless=True; otherwise stateful mode is used.

 None
token_threshold int | None

Stateful-mode only. Input token threshold at which OpenAI's server-side compaction is triggered. Corresponds to compact_threshold in the context_management API field. If None, OpenAI picks a server-side default.

 None
message_count_threshold int | None

Stateless-mode only. Compact when the message count exceeds this threshold.

 None
trigger Callable[[list[ModelMessage]], bool] | None

Stateless-mode only. Custom callable that decides whether to compact based on the current messages. Takes precedence over message_count_threshold.

 None
Source code in pydantic_ai_slim/pydantic_ai/models/openai.py 
3866
3867
3868
3869
3870
3871
3872
3873
3874
3875
3876
3877
3878
3879
3880
3881
3882
3883
3884
3885
3886
3887
3888
3889
3890
3891
3892
3893
3894
3895
3896
3897
3898
3899
3900
3901
3902
3903
3904
3905
3906
3907
3908
3909
3910
3911
3912
3913
3914
3915
3916
3917
3918
3919
3920
def __init__(
    self,
    *,
    stateless: bool | None = None,
    token_threshold: int | None = None,
    message_count_threshold: int | None = None,
    trigger: Callable[[list[ModelMessage]], bool] | None = None,
) -> None:
    """Initialize the OpenAI compaction capability.

    Args:
        stateless: Select the compaction mode explicitly. If `None` (the
            default), the mode is inferred from the other parameters:
            passing any stateless-only parameter (`message_count_threshold`
            or `trigger`) implies `stateless=True`; otherwise stateful
            mode is used.
        token_threshold: Stateful-mode only. Input token threshold at which
            OpenAI's server-side compaction is triggered. Corresponds to
            `compact_threshold` in the `context_management` API field. If
            `None`, OpenAI picks a server-side default.
        message_count_threshold: Stateless-mode only. Compact when the
            message count exceeds this threshold.
        trigger: Stateless-mode only. Custom callable that decides whether
            to compact based on the current messages. Takes precedence
            over `message_count_threshold`.
    """
    has_stateless_only = message_count_threshold is not None or trigger is not None
    has_stateful_only = token_threshold is not None

    if stateless is None:
        stateless = has_stateless_only

    if stateless:
        if has_stateful_only:
            raise UserError(
                '`token_threshold` is only valid for stateful compaction (`stateless=False`). '
                'For stateless `/compact` endpoint compaction, use `message_count_threshold` or `trigger`.'
            )
        if not has_stateless_only:
            raise UserError(
                '`stateless=True` requires `message_count_threshold` or `trigger` '
                'to determine when to invoke the `/compact` endpoint.'
            )
    else:
        if has_stateless_only:
            raise UserError(
                '`message_count_threshold` and `trigger` are only valid for stateless compaction '
                '(`stateless=True`). For stateful server-side compaction, use `token_threshold` '
                '(or omit it to use the OpenAI-managed default).'
            )

    self.stateless = stateless
    self.token_threshold = token_threshold
    self.message_count_threshold = message_count_threshold
    self.trigger = trigger