From fd4b15fae6d8592aacd1ee2a60cdb9738f5dfc4e Mon Sep 17 00:00:00 2001 From: mateo-berri <277851410+mateo-berri@users.noreply.github.com> Date: Wed, 2 Sep 2026 13:37:14 -0700 Subject: [PATCH] fix(anthropic): upgrade legacy thinking after the Bedrock Invoke and Vertex structured-output stub swap --- .../anthropic_claude3_transformation.py | 4 ++++ .../anthropic/transformation.py | 4 ++++ ...ations_anthropic_claude3_transformation.py | 24 +++++++++++++++++++ ...partner_models_anthropic_transformation.py | 23 ++++++++++++++++++ 4 files changed, 55 insertions(+) diff --git a/litellm/llms/bedrock/chat/invoke_transformations/anthropic_claude3_transformation.py b/litellm/llms/bedrock/chat/invoke_transformations/anthropic_claude3_transformation.py index 2a4c38e71ea..67720451c00 100644 --- a/litellm/llms/bedrock/chat/invoke_transformations/anthropic_claude3_transformation.py +++ b/litellm/llms/bedrock/chat/invoke_transformations/anthropic_claude3_transformation.py @@ -107,6 +107,10 @@ class AmazonAnthropicClaudeConfig(AmazonInvokeConfig, AnthropicConfig): # Restore original model name model = original_model + AnthropicModelInfo.translate_legacy_thinking_for_adaptive_model( + model=original_model, optional_params=optional_params, custom_llm_provider="bedrock" + ) + # The stub model hides the original model from the parent's forced-tool-use backstop response_format_tool_choice: Final = optional_params.get("tool_choice") if ( diff --git a/litellm/llms/vertex_ai/vertex_ai_partner_models/anthropic/transformation.py b/litellm/llms/vertex_ai/vertex_ai_partner_models/anthropic/transformation.py index ef03e61a858..7579bc8c02e 100644 --- a/litellm/llms/vertex_ai/vertex_ai_partner_models/anthropic/transformation.py +++ b/litellm/llms/vertex_ai/vertex_ai_partner_models/anthropic/transformation.py @@ -177,6 +177,10 @@ class VertexAIAnthropicConfig(AnthropicConfig): # Restore original model name for any other processing model = original_model + AnthropicModelInfo.translate_legacy_thinking_for_adaptive_model( + model=original_model, optional_params=optional_params, custom_llm_provider="vertex_ai" + ) + return optional_params def transform_response( diff --git a/tests/test_litellm/llms/bedrock/chat/invoke_transformations/test_bedrock_chat_invoke_transformations_anthropic_claude3_transformation.py b/tests/test_litellm/llms/bedrock/chat/invoke_transformations/test_bedrock_chat_invoke_transformations_anthropic_claude3_transformation.py index 41d82e4f960..4a1e97ca170 100644 --- a/tests/test_litellm/llms/bedrock/chat/invoke_transformations/test_bedrock_chat_invoke_transformations_anthropic_claude3_transformation.py +++ b/tests/test_litellm/llms/bedrock/chat/invoke_transformations/test_bedrock_chat_invoke_transformations_anthropic_claude3_transformation.py @@ -671,3 +671,27 @@ def test_bedrock_chat_invoke_fable_5_1_response_format_avoids_forced_tool_choice assert "output_format" not in result assert "tools" in result assert "tool_choice" not in result + + +@pytest.mark.parametrize("model", ["us.anthropic.claude-sonnet-5", "us.anthropic.claude-fable-5-1"]) +def test_bedrock_chat_invoke_tool_based_response_format_still_upgrades_legacy_thinking(local_model_cost_map, model): + result = AmazonAnthropicClaudeConfig().map_openai_params( + non_default_params={ + "response_format": { + "type": "json_schema", + "json_schema": { + "name": "test_schema", + "schema": {"type": "object", "properties": {"result": {"type": "string"}}}, + }, + }, + "thinking": {"type": "enabled", "budget_tokens": 4096}, + "max_tokens": 8192, + }, + optional_params={}, + model=model, + drop_params=False, + ) + + assert "tools" in result + assert result["thinking"] == {"type": "adaptive"} + assert result["output_config"] == {"effort": "high"} diff --git a/tests/test_litellm/llms/vertex_ai/vertex_ai_partner_models/anthropic/test_vertex_ai_partner_models_anthropic_transformation.py b/tests/test_litellm/llms/vertex_ai/vertex_ai_partner_models/anthropic/test_vertex_ai_partner_models_anthropic_transformation.py index 9419f88a981..a57672cfbfb 100644 --- a/tests/test_litellm/llms/vertex_ai/vertex_ai_partner_models/anthropic/test_vertex_ai_partner_models_anthropic_transformation.py +++ b/tests/test_litellm/llms/vertex_ai/vertex_ai_partner_models/anthropic/test_vertex_ai_partner_models_anthropic_transformation.py @@ -752,3 +752,26 @@ def test_vertex_ai_fable_5_1_response_format_uses_native_output_format(local_mod assert "output_format" in result_params assert "tool_choice" not in result_params assert "tools" not in result_params + + +def test_vertex_ai_anthropic_tool_based_response_format_still_upgrades_legacy_thinking(local_model_cost_map): + result_params = VertexAIAnthropicConfig().map_openai_params( + non_default_params={ + "response_format": { + "type": "json_schema", + "json_schema": { + "name": "test_schema", + "schema": {"type": "object", "properties": {"result": {"type": "string"}}}, + }, + }, + "thinking": {"type": "enabled", "budget_tokens": 4096}, + "max_tokens": 8192, + }, + optional_params={}, + model="claude-opus-4-8", + drop_params=False, + ) + + assert "tools" in result_params + assert result_params["thinking"] == {"type": "adaptive"} + assert result_params["output_config"] == {"effort": "high"}