Merge 7b52423380 into 635085ac14

litellm/anthropic_beta_headers_config.json

@@ -201,6 +201,7 @@
     "mcp-client-2025-11-20": null,
     "mcp-client-2025-04-04": null,
     "mcp-servers-2025-12-04": null,
+    "mid-conversation-tool-changes-2026-07-01": "mid-conversation-tool-changes-2026-07-01",
     "output-128k-2025-02-19": null,
     "structured-output-2024-03-01": null,
     "per-turn-control-2026-07-01": "per-turn-control-2026-07-01",

litellm/llms/anthropic/common_utils.py

@@ -46,10 +46,10 @@ from litellm.types.llms.anthropic import (
     ANTHROPIC_OAUTH_BETA_HEADER,
     ANTHROPIC_OAUTH_TOKEN_PREFIX,
     ANTHROPIC_THINKING_DISPLAY_UPDATES_BETA_HEADER,
+    LEGACY_THINKING_DISPLAY_UPDATES_PROVIDERS,
     AllAnthropicToolsValues,
     AnthropicMcpServerTool,
     AnthropicMessagesToolChoice,
-    AnthropicThinkingParam,
 )
 from litellm.types.llms.openai import AllMessageValues
 from litellm.types.proxy.model_listing import ModelInfoResponse
@@ -459,7 +459,7 @@ class AnthropicModelInfo(BaseLLMModelInfo):
         file_ids: Final = get_file_ids_from_messages(messages)
         return len(file_ids) > 0
 
-    def is_thinking_display_updates_used(self, thinking: AnthropicThinkingParam | None) -> bool:
+    def is_thinking_display_updates_used(self, thinking: object) -> bool:
         if not isinstance(thinking, dict):
             return False
         return thinking.get("type") in ("adaptive", "enabled") and thinking.get("display") == "updates"
@@ -889,7 +889,11 @@ class AnthropicModelInfo(BaseLLMModelInfo):
         )
         existing_output_config: Final = optional_params.get("output_config")
         display: Final = thinking.get("display")
-        if display in ("summarized", "omitted"):
+        preserve_display: Final = display in ("summarized", "omitted")
+        preserve_updates: Final = (
+            display == "updates" and custom_llm_provider in LEGACY_THINKING_DISPLAY_UPDATES_PROVIDERS
+        )
+        if preserve_display or preserve_updates:
             optional_params["thinking"] = {"type": "adaptive", "display": display}
         else:
             optional_params["thinking"] = {"type": "adaptive"}

litellm/llms/azure_ai/anthropic/transformation.py

@@ -96,9 +96,14 @@ class AzureAnthropicConfig(AnthropicConfig):
             is_vertex_request=optional_params.get("is_vertex_request", False),
             user_anthropic_beta_headers=user_anthropic_beta_headers,
             mcp_server_used=mcp_server_used,
+            is_thinking_display_updates_used=self.is_thinking_display_updates_used(
+                optional_params.get("thinking"),  # pyright: ignore[reportUnknownArgumentType]  # detector accepts unvalidated input
+            ),
         )
         # Merge headers - Azure auth (api-key or Authorization) takes precedence
         headers = {**anthropic_headers, **headers}
+        if "anthropic-beta" in anthropic_headers:
+            headers["anthropic-beta"] = anthropic_headers["anthropic-beta"]
 
         # Ensure anthropic-version header is set
         if "anthropic-version" not in headers:

litellm/llms/vertex_ai/vertex_ai_partner_models/anthropic/experimental_pass_through/transformation.py

@@ -8,6 +8,8 @@ from litellm.llms.anthropic.pass_through.messages.transformation import (
 from litellm.types.llms.anthropic import (
     ANTHROPIC_BETA_HEADER_VALUES,
     ANTHROPIC_HOSTED_TOOLS,
+    ANTHROPIC_MID_CONVERSATION_TOOL_CHANGES_BETA_HEADER,
+    ANTHROPIC_THINKING_DISPLAY_UPDATES_BETA_HEADER,
 )
 from litellm.types.llms.anthropic_tool_search import get_tool_search_beta_header
 from litellm.types.llms.vertex_ai import VertexPartnerProvider
@@ -115,8 +117,21 @@ class VertexAIPartnerModelsAnthropicMessagesConfig(AnthropicMessagesConfig, Vert
         if _messages_carry_output_config(messages):
             beta_values.add(ANTHROPIC_BETA_HEADER_VALUES.PER_TURN_CONTROL_2026_07_01.value)
 
-        if beta_values:
-            headers["anthropic-beta"] = ",".join(beta_values)
+        thinking_display_betas: Final = (
+            (ANTHROPIC_THINKING_DISPLAY_UPDATES_BETA_HEADER,)
+            if anthropic_model_info.is_thinking_display_updates_used(
+                optional_params.get("thinking"),  # pyright: ignore[reportUnknownArgumentType]  # detector accepts unvalidated input
+            )
+            else ()
+        )
+        tool_change_betas: Final = (
+            (ANTHROPIC_MID_CONVERSATION_TOOL_CHANGES_BETA_HEADER,)
+            if anthropic_model_info.is_mid_conversation_tool_change_used(messages)
+            else ()
+        )
+        all_beta_values: Final = beta_values.union(thinking_display_betas, tool_change_betas)
+        if all_beta_values:
+            headers["anthropic-beta"] = ",".join(sorted(all_beta_values))
 
         return headers, api_base
 

litellm/llms/vertex_ai/vertex_ai_partner_models/anthropic/transformation.py

@@ -120,6 +120,9 @@ class VertexAIAnthropicConfig(AnthropicConfig):
             file_id_used=self.is_file_id_used(messages),
             mcp_server_used=self.is_mcp_server_used(optional_params.get("mcp_servers")),
             custom_llm_provider="vertex_ai",
+            is_thinking_display_updates_used=self.is_thinking_display_updates_used(
+                data.get("thinking"),  # pyright: ignore[reportUnknownArgumentType]  # detector accepts unvalidated input
+            ),
         )
 
         beta_set: Final = set(auto_betas)

litellm/types/llms/anthropic.py

@@ -736,6 +736,9 @@ class AnthropicThinkingParam(TypedDict, total=False):
     display: ReadOnly[Literal["summarized", "omitted", "updates"]]
 
 
+LEGACY_THINKING_DISPLAY_UPDATES_PROVIDERS: Final = frozenset(("azure_ai", "vertex_ai"))
+
+
 class ANTHROPIC_HOSTED_TOOLS(str, Enum):
     WEB_SEARCH = "web_search"
     BASH = "bash"

tests/unit/llms/anthropic/test_anthropic_common_utils.py

@@ -4048,7 +4048,7 @@ def test_thinking_display_beta_requires_active_thinking(thinking: object, expect
     (
         ("summarized", {"type": "adaptive", "display": "summarized"}),
         ("omitted", {"type": "adaptive", "display": "omitted"}),
-        ("updates", {"type": "adaptive"}),
+        ("updates", {"type": "adaptive", "display": "updates"}),
     ),
 )
 def test_shared_legacy_thinking_translation_preserves_supported_display(

tests/unit/llms/azure_ai/claude/test_azure_anthropic_messages_transformation.py

@@ -470,3 +470,45 @@ def test_azure_claude_4_8_plus_cost_map_entries_carry_mid_conversation_system_fl
         and info.get("supports_mid_conversation_system") is not True
     ]
     assert missing == []
+
+
+@pytest.mark.usefixtures("local_model_cost_map")
+@pytest.mark.parametrize("feature", ("thinking", "output_config", "tool_addition", "tool_removal", "none"))
+@pytest.mark.parametrize("explicit_beta", (False, True))
+def test_messages_feature_beta_headers(feature: str, explicit_beta: bool) -> None:
+    from typing import Final
+
+    from litellm.llms.azure_ai.anthropic.messages_transformation import AzureAnthropicMessagesConfig
+    from litellm.types.llms.anthropic import (
+        ANTHROPIC_BETA_HEADER_VALUES,
+        ANTHROPIC_MID_CONVERSATION_TOOL_CHANGES_BETA_HEADER,
+        ANTHROPIC_THINKING_DISPLAY_UPDATES_BETA_HEADER,
+    )
+
+    beta: Final = (
+        ANTHROPIC_THINKING_DISPLAY_UPDATES_BETA_HEADER
+        if feature == "thinking"
+        else ANTHROPIC_BETA_HEADER_VALUES.PER_TURN_CONTROL_2026_07_01.value
+        if feature == "output_config"
+        else ANTHROPIC_MID_CONVERSATION_TOOL_CHANGES_BETA_HEADER
+    )
+    message: Final = (
+        {"role": "system", "content": [], "output_config": {"effort": "high"}}
+        if feature == "output_config"
+        else {"role": "system", "content": [{"type": feature, "tool": {"type": "tool_reference", "name": "ping"}}]}
+        if feature in ("tool_addition", "tool_removal")
+        else {"role": "user", "content": "Reply with OK"}
+    )
+    headers, _ = AzureAnthropicMessagesConfig().validate_anthropic_messages_environment(
+        headers={"anthropic-beta": f"existing-beta,{beta}" if explicit_beta else "existing-beta"},
+        model="claude-fable-5-1",
+        messages=[{"role": "user", "content": "Hello"}, message],
+        optional_params={"thinking": {"type": "adaptive", "display": "updates"}} if feature == "thinking" else {},
+        litellm_params={"vertex_ai_project": "test-project", "vertex_ai_location": "us-east5"},
+        api_key="test-key",
+        api_base="https://example.com/anthropic",
+    )
+
+    betas: Final = headers["anthropic-beta"].split(",")
+    assert betas.count(beta) == int(feature != "none" or explicit_beta)
+    assert betas.count("existing-beta") == 1

tests/unit/llms/azure_ai/claude/test_azure_anthropic_transformation.py

@@ -485,3 +485,57 @@ def test_chat_flagged_model_keeps_mid_conversation_system_role_in_place(local_mo
         "content": [{"type": "text", "text": "<system-reminder>Answer with exactly one word.</system-reminder>"}],
     }
 
+
+
+@pytest.mark.usefixtures("local_model_cost_map")
+@pytest.mark.parametrize("thinking_type", ("adaptive", "enabled"))
+@pytest.mark.parametrize("display", ("summarized", "omitted", "updates"))
+def test_chat_preserves_thinking_display_and_beta(thinking_type: str, display: str) -> None:
+    from typing import Final
+
+    from litellm.llms.azure_ai.anthropic.transformation import AzureAnthropicConfig
+    from litellm.types.llms.anthropic import ANTHROPIC_THINKING_DISPLAY_UPDATES_BETA_HEADER
+
+    config: Final = AzureAnthropicConfig()
+    optional_params: Final = config.map_openai_params(
+        non_default_params={"thinking": {"type": thinking_type, "display": display, "budget_tokens": 2048}},
+        optional_params={},
+        model="claude-fable-5-1",
+        drop_params=False,
+    )
+    headers: Final = config.validate_environment(
+        headers={"anthropic-beta": "existing-beta"},
+        model="claude-fable-5-1",
+        messages=[{"role": "user", "content": "Reply with OK"}],
+        optional_params=optional_params,
+        litellm_params={},
+        api_key="test-key",
+    )
+    result: Final = config.transform_request(
+        model="claude-fable-5-1",
+        messages=[{"role": "user", "content": "Reply with OK"}],
+        optional_params=optional_params,
+        litellm_params={},
+        headers=headers,
+    )
+
+    assert result["thinking"]["type"] == "adaptive"
+    assert result["thinking"]["display"] == display
+    assert (ANTHROPIC_THINKING_DISPLAY_UPDATES_BETA_HEADER in headers["anthropic-beta"].split(",")) == (display == "updates")
+    assert "existing-beta" in headers["anthropic-beta"].split(",")
+
+
+@pytest.mark.parametrize("thinking", ("enabled", {"type": "bogus"}, {"type": "adaptive", "display": "future-mode"}))
+def test_beta_detection_preserves_unrecognized_thinking(thinking: object) -> None:
+    from typing import Final
+
+    from litellm.types.llms.anthropic import ANTHROPIC_THINKING_DISPLAY_UPDATES_BETA_HEADER
+
+    params: Final = {"thinking": thinking}
+    headers: Final = AzureAnthropicConfig().validate_environment(
+        headers={}, model="claude-fable-5-1", messages=[], optional_params=params,
+        litellm_params={}, api_key="test-key",
+    )
+
+    assert params["thinking"] == thinking
+    assert ANTHROPIC_THINKING_DISPLAY_UPDATES_BETA_HEADER not in headers.get("anthropic-beta", "").split(",")

tests/unit/llms/vertex_ai/vertex_ai_partner_models/anthropic/test_vertex_ai_partner_models_anthropic_messages_config.py

@@ -726,3 +726,79 @@ def test_vertex_claude_4_8_plus_cost_map_entries_carry_mid_conversation_system_f
         and info.get("supports_mid_conversation_system") is not True
     ]
     assert missing == []
+
+
+@pytest.mark.usefixtures("local_model_cost_map", "local_beta_headers_config")
+@pytest.mark.parametrize("feature", ("thinking", "output_config", "tool_addition", "tool_removal", "none"))
+@pytest.mark.parametrize("explicit_beta", (False, True))
+def test_messages_feature_beta_headers(feature: str, explicit_beta: bool) -> None:
+    from typing import Final
+    from unittest.mock import patch
+
+    from google.oauth2.credentials import Credentials
+    from litellm.anthropic_beta_headers_manager import update_headers_with_filtered_beta
+    from litellm.llms.vertex_ai.vertex_ai_partner_models.anthropic.experimental_pass_through.transformation import VertexAIPartnerModelsAnthropicMessagesConfig
+    from litellm.types.llms.anthropic import (
+        ANTHROPIC_BETA_HEADER_VALUES,
+        ANTHROPIC_MID_CONVERSATION_TOOL_CHANGES_BETA_HEADER,
+        ANTHROPIC_THINKING_DISPLAY_UPDATES_BETA_HEADER,
+    )
+
+    beta: Final = (
+        ANTHROPIC_THINKING_DISPLAY_UPDATES_BETA_HEADER
+        if feature == "thinking"
+        else ANTHROPIC_BETA_HEADER_VALUES.PER_TURN_CONTROL_2026_07_01.value
+        if feature == "output_config"
+        else ANTHROPIC_MID_CONVERSATION_TOOL_CHANGES_BETA_HEADER
+    )
+    message: Final = (
+        {"role": "system", "content": [], "output_config": {"effort": "high"}}
+        if feature == "output_config"
+        else {"role": "system", "content": [{"type": feature, "tool": {"type": "tool_reference", "name": "ping"}}]}
+        if feature in ("tool_addition", "tool_removal")
+        else {"role": "user", "content": "Reply with OK"}
+    )
+    with patch(
+        "google.auth.default",
+        return_value=(MagicMock(spec=Credentials, token="test-token", expired=False), "test-project"),
+    ):
+        headers, _ = VertexAIPartnerModelsAnthropicMessagesConfig().validate_anthropic_messages_environment(
+            headers={"anthropic-beta": f"existing-beta,{beta}" if explicit_beta else "existing-beta"},
+            model="claude-fable-5-1",
+            messages=[{"role": "user", "content": "Hello"}, message],
+            optional_params={"thinking": {"type": "adaptive", "display": "updates"}} if feature == "thinking" else {},
+            litellm_params={"vertex_ai_project": "test-project", "vertex_ai_location": "us-east5"},
+            api_key="test-key",
+            api_base="https://example.com/anthropic",
+        )
+
+    betas: Final = headers["anthropic-beta"].split(",")
+    assert betas.count(beta) == int(feature != "none" or explicit_beta)
+    assert betas.count("existing-beta") == 1
+
+    if feature in ("tool_addition", "tool_removal"):
+        filtered_headers: Final = update_headers_with_filtered_beta(headers=headers, provider="vertex_ai")
+        assert filtered_headers.get("anthropic-beta", "").split(",").count(beta) == 1
+
+
+@pytest.mark.parametrize("thinking", ("enabled", {"type": "bogus"}, {"type": "adaptive", "display": "future-mode"}))
+def test_beta_detection_preserves_unrecognized_thinking(thinking: object) -> None:
+    from typing import Final
+    from unittest.mock import patch
+
+    from google.oauth2.credentials import Credentials
+    from litellm.types.llms.anthropic import ANTHROPIC_THINKING_DISPLAY_UPDATES_BETA_HEADER
+
+    params: Final = {"thinking": thinking}
+    with patch(
+        "google.auth.default",
+        return_value=(MagicMock(spec=Credentials, token="test-token", expired=False), "test-project"),
+    ):
+        headers, _ = VertexAIPartnerModelsAnthropicMessagesConfig().validate_anthropic_messages_environment(
+            headers={}, model="claude-fable-5-1", messages=[], optional_params=params,
+            litellm_params={"vertex_ai_project": "test-project", "vertex_ai_location": "us-east5"},
+            api_key="test-key", api_base="https://example.com/anthropic",
+        )
+
+    assert params["thinking"] == thinking
+    assert ANTHROPIC_THINKING_DISPLAY_UPDATES_BETA_HEADER not in headers.get("anthropic-beta", "").split(",")

tests/unit/llms/vertex_ai/vertex_ai_partner_models/anthropic/test_vertex_ai_partner_models_anthropic_transformation.py

@@ -875,3 +875,51 @@ def test_chat_flagged_model_replays_a_byte_identical_prefix_around_a_mid_convers
     _assert_prefix_stable(requests)
     assert [m["role"] for m in requests[1]["messages"]] == ["user", "assistant", "user", "system"]
     assert [m["role"] for m in requests[2]["messages"]] == ["user", "assistant", "user", "system", "assistant", "user"]
+
+
+@pytest.mark.usefixtures("local_model_cost_map")
+@pytest.mark.parametrize("thinking_type", ("adaptive", "enabled"))
+@pytest.mark.parametrize("display", ("summarized", "omitted", "updates"))
+def test_chat_preserves_thinking_display_and_beta(thinking_type: str, display: str) -> None:
+    from typing import Final
+
+    from litellm.llms.vertex_ai.vertex_ai_partner_models.anthropic.transformation import VertexAIAnthropicConfig
+    from litellm.types.llms.anthropic import ANTHROPIC_THINKING_DISPLAY_UPDATES_BETA_HEADER
+
+    config: Final = VertexAIAnthropicConfig()
+    optional_params: Final = config.map_openai_params(
+        non_default_params={"thinking": {"type": thinking_type, "display": display, "budget_tokens": 2048}},
+        optional_params={},
+        model="claude-fable-5-1",
+        drop_params=False,
+    )
+    optional_params["extra_headers"] = {"anthropic-beta": "existing-beta"}
+    headers: Final = {}
+    result: Final = config.transform_request(
+        model="claude-fable-5-1",
+        messages=[{"role": "user", "content": "Reply with OK"}],
+        optional_params=optional_params,
+        litellm_params={},
+        headers=headers,
+    )
+
+    assert result["thinking"]["type"] == "adaptive"
+    assert result["thinking"]["display"] == display
+    assert (ANTHROPIC_THINKING_DISPLAY_UPDATES_BETA_HEADER in headers["anthropic-beta"].split(",")) == (display == "updates")
+    assert "existing-beta" in headers["anthropic-beta"].split(",")
+
+
+@pytest.mark.parametrize("thinking", ("enabled", {"type": "bogus"}, {"type": "adaptive", "display": "future-mode"}))
+def test_beta_detection_preserves_unrecognized_thinking(thinking: object) -> None:
+    from typing import Final
+
+    from litellm.types.llms.anthropic import ANTHROPIC_THINKING_DISPLAY_UPDATES_BETA_HEADER
+
+    headers: Final = {}
+    result: Final = VertexAIAnthropicConfig().transform_request(
+        model="claude-fable-5-1", messages=[{"role": "user", "content": "Hello"}],
+        optional_params={"thinking": thinking}, litellm_params={}, headers=headers,
+    )
+
+    assert result["thinking"] == thinking
+    assert ANTHROPIC_THINKING_DISPLAY_UPDATES_BETA_HEADER not in headers.get("anthropic-beta", "").split(",")