fix(anthropic): keep adaptive thinking when modify_params drops missing thinking blocks (#43531)

The modify_params guard (#14194 / #18926) drops the thinking param whenever
the last assistant message with tool_calls has no thinking_blocks. For
adaptive-thinking models (Claude 4.6+ / Opus 5.x, thinking={"type": "adaptive"})
that guard fires on ordinary traffic and Anthropic does not need it: adaptive
thinking may legitimately decide not to think on a trivial first tool call.
Dropping it made effort-configured models stream no reasoning (no
reasoning_content deltas, empty signed thinking block), and at high effort the
silence could exceed intermediary timeouts.

Skip the drop for models where AnthropicConfig._is_adaptive_thinking_model()
is true; legacy budget-based thinking keeps the existing behavior.

Tests: adaptive model keeps thinking/output_config; legacy model still drops.

litellm/llms/anthropic/chat/transformation.py

@@ -1920,9 +1920,18 @@ class AnthropicConfig(AnthropicModelInfo, BaseConfig):
         # If any message has thinking_blocks, we must keep thinking enabled, otherwise
         # Anthropic errors with: "When thinking is disabled, an assistant message cannot contain thinking"
         # Related issue: https://github.com/BerriAI/litellm/issues/18926
+        #
+        # Adaptive-thinking models (Claude 4.6+ / Opus 5.x) are exempt: with
+        # thinking={"type": "adaptive"} Anthropic accepts a tool_use-only prior
+        # turn (adaptive thinking may legitimately decide not to think on a
+        # trivial tool call), so the guard applies only to legacy budget-based
+        # thinking. Dropping it there breaks streaming for effort-configured
+        # models (no reasoning_content deltas until first text/tool token).
+        # Related issue: https://github.com/BerriAI/litellm/issues/43531
         if (
             optional_params.get("thinking") is not None
             and messages is not None
+            and not AnthropicConfig._is_adaptive_thinking_model(model, self._resolved_provider)
             and last_assistant_with_tool_calls_has_no_thinking_blocks(messages)
             and not any_assistant_message_has_thinking_blocks(messages)
         ):

tests/unit/llms/anthropic/chat/test_anthropic_chat_transformation.py

@@ -6798,3 +6798,91 @@ def test_chat_dummy_tool_result_for_an_orphaned_tool_call_replays_a_byte_identic
     _assert_prefix_stable(requests)
     assert [m["role"] for m in requests[0]["messages"]] == ["user", "assistant", "user"]
     assert requests[0]["messages"][2]["content"][0]["type"] == "tool_result"
+
+
+def test_adaptive_thinking_kept_for_adaptive_model_when_tool_calls_missing_thinking_blocks():
+    """
+    modify_params must NOT drop thinking for adaptive-thinking models
+    (Claude 4.6+ / Opus 5.x): Anthropic accepts a tool_use-only prior turn
+    with thinking={"type": "adaptive"}, and dropping it breaks reasoning
+    streaming (no reasoning_content deltas until first text/tool token).
+
+    Related issue: https://github.com/BerriAI/litellm/issues/43531
+    """
+    import litellm
+
+    original_modify_params = litellm.modify_params
+    litellm.modify_params = True
+    try:
+        messages = [
+            {"role": "user", "content": "Run `echo seed` with bash, then reason."},
+            {
+                "role": "assistant",
+                "content": None,
+                "tool_calls": [
+                    {
+                        "id": "toolu_1",
+                        "type": "function",
+                        "function": {"name": "bash", "arguments": '{"command":"echo seed"}'},
+                    }
+                ],
+            },
+            {"role": "tool", "tool_call_id": "toolu_1", "content": "seed"},
+        ]
+        optional_params = {
+            "thinking": {"type": "adaptive", "display": "summarized"},
+            "output_config": {"effort": "xhigh"},
+        }
+        transformed = AnthropicConfig().transform_request(
+            model="claude-opus-5-5",
+            messages=messages,
+            optional_params=optional_params,
+            litellm_params={},
+            headers={},
+        )
+        assert "thinking" in transformed
+        assert transformed["thinking"] == {"type": "adaptive", "display": "summarized"}
+        assert transformed["output_config"] == {"effort": "xhigh"}
+    finally:
+        litellm.modify_params = original_modify_params
+
+
+def test_legacy_budget_thinking_still_dropped_when_tool_calls_missing_thinking_blocks():
+    """
+    The original guard (#14194 / #18926) must keep working for legacy
+    budget-based thinking: a tool_use-only prior turn is rejected by
+    Anthropic unless the request pairs it with thinking blocks.
+
+    Related issue: https://github.com/BerriAI/litellm/issues/43531
+    """
+    import litellm
+
+    original_modify_params = litellm.modify_params
+    litellm.modify_params = True
+    try:
+        messages = [
+            {"role": "user", "content": "Run `echo seed` with bash, then reason."},
+            {
+                "role": "assistant",
+                "content": "",
+                "tool_calls": [
+                    {
+                        "id": "toolu_1",
+                        "type": "function",
+                        "function": {"name": "bash", "arguments": '{"command":"echo seed"}'},
+                    }
+                ],
+            },
+            {"role": "tool", "tool_call_id": "toolu_1", "content": "seed"},
+        ]
+        optional_params = {"thinking": {"type": "enabled", "budget_tokens": 1024}}
+        transformed = AnthropicConfig().transform_request(
+            model="claude-3-5-sonnet-20241022",
+            messages=messages,
+            optional_params=optional_params,
+            litellm_params={},
+            headers={},
+        )
+        assert "thinking" not in transformed
+    finally:
+        litellm.modify_params = original_modify_params