remove auto disable default thinking

docs/evaluating.md

@@ -273,9 +273,8 @@ See [README.md](../README.md#reasoning--extended-thinking) for the user-facing `
 
 | Quirk | Applies to | Effect |
 |-------|------------|--------|
-| `adaptive_thinking` | `opus-4-7`, `opus-4-8`, `fable-5`, `sonnet-5` | Uses `thinking={"type": "adaptive"}` + `effort=<level>` instead of manual `budget_tokens` (which 400s on these models); rejects `temperature`/`top_p`/`top_k` at any non-default value. |
-| `defaults_thinking_on` | `sonnet-5` | Runs adaptive thinking by default when `thinking` is omitted, so `thinking_effort` must be explicitly unset to disable it. Not needed for `fable-5`, where adaptive thinking is always on and can't be disabled at all. |
-| `sparse_max_tokens_profile` | `sonnet-5` | The installed `langchain-anthropic` has no model-profile entry for this model, so it silently falls back to `max_tokens=4096` instead of the usual 64k–128k auto-set — too tight for structured output. The client sets an explicit default instead. |
+| `adaptive_thinking` | `opus-4-7`, `opus-4-8`, `opus-5`, `fable-5`, `sonnet-5` | Uses `thinking={"type": "adaptive"}` + `effort=<level>` instead of manual `budget_tokens` (which 400s on these models); rejects `temperature`/`top_p`/`top_k` at any non-default value. |
+| `sparse_max_tokens_profile` | `opus-5`, `sonnet-5` | The installed `langchain-anthropic` has no model-profile entry for these models, so it silently falls back to `max_tokens=4096` instead of the usual 64k–128k auto-set — too tight for structured output. The client sets an explicit default instead. |
 
 To onboard a new Claude model that needs similar special-casing, add its marker to `_MODEL_QUIRKS` with the applicable quirks rather than adding another `is_x` check.
 

llm_clients/claude_llm.py

@@ -35,12 +35,6 @@
 #   `_unsupported_model_params`); Anthropic's guidance is to omit them and
 #   steer behavior via the system prompt instead.
 #   https://platform.claude.com/docs/en/about-claude/models/whats-new-sonnet-5
-# defaults_thinking_on: runs adaptive thinking by default when `thinking` is
-#   omitted (unlike other adaptive models, which default to no thinking); must
-#   be explicitly disabled so no thinking_effort means no thinking. Not used
-#   for fable-5: adaptive thinking is *always* on there and
-#   `thinking={"type": "disabled"}` is rejected outright, so there is nothing
-#   to force - omitting `thinking` already does the right thing.
 # sparse_max_tokens_profile: the installed langchain-anthropic has no
 #   model-profile entry for this model, so it silently falls back to
 #   max_tokens=4096 (vs 64k-128k auto-set for profiled models) - too tight for
@@ -49,10 +43,9 @@
 _MODEL_QUIRKS: Dict[str, frozenset[str]] = {
     "opus-4-7": frozenset({"adaptive_thinking"}),
     "opus-4-8": frozenset({"adaptive_thinking"}),
+    "opus-5": frozenset({"adaptive_thinking", "sparse_max_tokens_profile"}),
     "fable-5": frozenset({"adaptive_thinking"}),
-    "sonnet-5": frozenset(
-        {"adaptive_thinking", "defaults_thinking_on", "sparse_max_tokens_profile"}
-    ),
+    "sonnet-5": frozenset({"adaptive_thinking", "sparse_max_tokens_profile"}),
 }
 
 # thinking_effort labels -> Anthropic's raw `budget_tokens`, for models that
@@ -125,8 +118,6 @@ def _apply_thinking_kwargs(
                 budget = _EFFORT_BUDGETS.get(effort_str, _EFFORT_BUDGETS["medium"])
                 kwargs["thinking"] = {"type": "enabled", "budget_tokens": budget}
                 kwargs.setdefault("max_tokens", budget + 1024)  # must exceed budget
-        elif "defaults_thinking_on" in quirks:
-            kwargs.setdefault("thinking", {"type": "disabled"})
 
         if "sparse_max_tokens_profile" in quirks:
             kwargs.setdefault("max_tokens", 8192)
@@ -205,6 +196,7 @@ def __init__(
         print(f"  Model: {llm_params['model']}")
         print(f"  Temperature: {llm_params.get('temperature', 'default')}")
         print(f"  Max tokens: {llm_params.get('max_tokens', 'default')}")
+        print(f"  Thinking: {llm_params.get('thinking', 'omitted (model default)')}")
         if dropped_params:
             print(
                 f"  Dropped (unsupported for {self.model_name}): "
@@ -213,7 +205,7 @@ def __init__(
         extra_params = {
             k: v
             for k, v in llm_params.items()
-            if k not in ["model", "anthropic_api_key"]
+            if k not in ["model", "anthropic_api_key", "thinking"]
         }
         if extra_params:
             print(f"  Extra parameters: {extra_params}")

tests/unit/llm_clients/test_claude_llm.py

@@ -218,24 +218,30 @@ def test_filter_supported_params_strips_sampling_params_for_opus_4_8(
     def test_is_adaptive_thinking_model(self, model_name, expected):
         assert ClaudeLLM._is_adaptive_thinking_model(model_name) is expected
 
-    def test_apply_thinking_kwargs_no_effort_sonnet_5_disables_thinking(self):
-        """Sonnet 5 runs adaptive thinking by default; no effort means disabled."""
+    def test_apply_thinking_kwargs_no_effort_sonnet_5_omits_thinking(self):
+        """No effort means `thinking` is omitted; the model's own default applies."""
         kwargs: dict = {}
         ClaudeLLM._apply_thinking_kwargs(kwargs, "claude-sonnet-5", None)
-        assert kwargs == {"thinking": {"type": "disabled"}, "max_tokens": 8192}
+        assert kwargs == {"max_tokens": 8192}
 
     def test_apply_thinking_kwargs_no_effort_sonnet_5_keeps_explicit_thinking(self):
-        """A caller-supplied `thinking` kwarg must survive the disabled default."""
+        """A caller-supplied `thinking` kwarg must survive untouched."""
         kwargs: dict = {"thinking": {"type": "enabled", "budget_tokens": 5000}}
         ClaudeLLM._apply_thinking_kwargs(kwargs, "claude-sonnet-5", None)
         assert kwargs["thinking"] == {"type": "enabled", "budget_tokens": 5000}
 
     def test_apply_thinking_kwargs_no_effort_opus_4_8_is_untouched(self):
-        """Unlike sonnet-5, opus-4-8 defaults to no thinking already."""
         kwargs: dict = {}
         ClaudeLLM._apply_thinking_kwargs(kwargs, "claude-opus-4-8", None)
         assert kwargs == {}
 
+    def test_apply_thinking_kwargs_no_effort_opus_5_sets_max_tokens(self):
+        """opus-5 also lacks a langchain-anthropic model profile, like sonnet-5."""
+        assert ClaudeLLM._is_adaptive_thinking_model("claude-opus-5")
+        kwargs: dict = {}
+        ClaudeLLM._apply_thinking_kwargs(kwargs, "claude-opus-5", None)
+        assert kwargs == {"max_tokens": 8192}
+
     def test_apply_thinking_kwargs_no_effort_fable_5_is_untouched(self):
         """fable-5 can't disable thinking at all; omitting `thinking` already
         gives the (only) adaptive-always-on behavior, so nothing is forced."""
@@ -338,6 +344,7 @@ def test_unsupported_model_params_covers_all_adaptive_markers(self):
         assert unsupported == {
             "opus-4-7": frozenset({"temperature", "top_p", "top_k"}),
             "opus-4-8": frozenset({"temperature", "top_p", "top_k"}),
+            "opus-5": frozenset({"temperature", "top_p", "top_k"}),
             "fable-5": frozenset({"temperature", "top_p", "top_k"}),
             "sonnet-5": frozenset({"temperature", "top_p", "top_k"}),
         }