Merge 26604120bc into b781d157d7

litellm/llms/anthropic/chat/transformation.py

@@ -1920,9 +1920,24 @@ 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 (thinking={"type": "adaptive"}, Claude 4.6+ / Opus 5.x)
+        # is exempt: Anthropic accepts a tool_use-only prior turn, because adaptive
+        # thinking may legitimately decide not to think on a trivial tool call.
+        # Dropping it breaks streaming for effort-configured models (no
+        # reasoning_content deltas until first text/tool token).
+        #
+        # The exemption is keyed off the *parameter shape*, not the model: a
+        # Claude 4.6+ model explicitly configured with legacy budget-based
+        # thinking ({"type": "enabled", "budget_tokens": N}) still needs the
+        # guard, exactly like older models.
+        # Related issue: https://github.com/BerriAI/litellm/issues/43531
+        _thinking_param = optional_params.get("thinking")
+        _is_adaptive_thinking = isinstance(_thinking_param, dict) and _thinking_param.get("type") == "adaptive"
         if (
-            optional_params.get("thinking") is not None
+            _thinking_param is not None
             and messages is not None
+            and not _is_adaptive_thinking
             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 +6798,95 @@ 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(monkeypatch):
+    """
+    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
+
+    monkeypatch.setattr(litellm, "modify_params", True)
+    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"}
+
+
+def test_legacy_budget_thinking_still_dropped_when_tool_calls_missing_thinking_blocks(monkeypatch):
+    """
+    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
+
+    monkeypatch.setattr(litellm, "modify_params", True)
+    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"},
+    ]
+    # Legacy budget-based thinking on an older model: guard still drops.
+    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
+    # Legacy budget-based thinking on an adaptive-capable model (Claude 4.6+):
+    # the exemption is keyed off the parameter shape, so the guard must still drop.
+    optional_params2 = {"thinking": {"type": "enabled", "budget_tokens": 2048}}
+    transformed2 = AnthropicConfig().transform_request(
+        model="claude-opus-5-5",
+        messages=messages,
+        optional_params=optional_params2,
+        litellm_params={},
+        headers={},
+    )
+    assert "thinking" not in transformed2