fix(sdk/llm): emit output_config at top level for adaptive thinking

Author: piyushchandra357

openhands-sdk/openhands/sdk/llm/options/chat_options.py

@@ -95,22 +95,18 @@ def select_chat_options(
         out["extra_body"] = llm.litellm_extra_body
 
     # Opt-in adaptive thinking for models that only accept thinking.type="adaptive"
-    # (Opus 4.7/4.8/Fable 5). Without display="summarized" the model's default is
-    # "omitted" and thinking blocks come back empty, so display is required here.
-    # output_config goes through extra_body so it survives older litellm releases
-    # that don't recognize it as a top-level kwarg. Placed after the user-provided
-    # extra_body pass-through so output_config merges into (not replaces) it.
+    # (Opus 4.7/4.8/Fable 5). display="summarized" is REQUIRED for these models to
+    # return populated thinking content (their default is "omitted"). output_config
+    # is emitted at the top level so litellm transformations preserve it; an
+    # earlier draft wrapped it in extra_body, but older litellm versions did not
+    # unpack extra_body for Vertex Claude and the API rejected the request.
     if (
         llm.force_adaptive_thinking
         and get_features(llm.model).supports_adaptive_thinking
     ):
         out["thinking"] = {"type": "adaptive", "display": "summarized"}
         if llm.reasoning_effort is not None and llm.reasoning_effort != "none":
-            existing_eb = out.get("extra_body") or {}
-            out["extra_body"] = {
-                **existing_eb,
-                "output_config": {"effort": llm.reasoning_effort},
-            }
+            out["output_config"] = {"effort": llm.reasoning_effort}
         # Strip reasoning_effort so litellm does not re-derive and overwrite
         # the thinking block via its reasoning_effort -> thinking mapping.
         out.pop("reasoning_effort", None)

tests/sdk/llm/test_chat_options.py

@@ -177,15 +177,15 @@ def test_force_adaptive_thinking_emits_adaptive_schema_for_opus_4_8():
     out = select_chat_options(llm, user_kwargs={}, has_tools=True)
 
     assert out.get("thinking") == {"type": "adaptive", "display": "summarized"}
-    assert out.get("extra_body") == {"output_config": {"effort": "high"}}
+    assert out.get("output_config") == {"effort": "high"}
     assert out.get("max_tokens") == 4096
     assert "reasoning_effort" not in out
     assert "temperature" not in out
     assert "top_p" not in out
     assert "top_k" not in out
 
 
-def test_force_adaptive_thinking_merges_into_existing_extra_body():
+def test_force_adaptive_thinking_preserves_existing_extra_body():
     llm = DummyLLM(
         model="claude-opus-4-7",
         reasoning_effort="medium",
@@ -194,10 +194,8 @@ def test_force_adaptive_thinking_merges_into_existing_extra_body():
     )
     out = select_chat_options(llm, user_kwargs={}, has_tools=True)
 
-    assert out.get("extra_body") == {
-        "return_token_ids": True,
-        "output_config": {"effort": "medium"},
-    }
+    assert out.get("extra_body") == {"return_token_ids": True}
+    assert out.get("output_config") == {"effort": "medium"}
 
 
 def test_force_adaptive_thinking_noop_when_model_not_supported():
@@ -209,7 +207,7 @@ def test_force_adaptive_thinking_noop_when_model_not_supported():
     out = select_chat_options(llm, user_kwargs={}, has_tools=True)
 
     assert "thinking" not in out
-    assert "output_config" not in (out.get("extra_body") or {})
+    assert "output_config" not in out
     assert out.get("reasoning_effort") == "high"
 
 
@@ -222,4 +220,4 @@ def test_force_adaptive_thinking_disabled_by_default_for_opus_4_8():
 
     assert "thinking" not in out
     assert out.get("reasoning_effort") == "high"
-    assert "output_config" not in (out.get("extra_body") or {})
+    assert "output_config" not in out