feat(openai_api_compatible): Add Strict compatibility mode option (langgenius#2416)

- Add `Strict compatibility mode` option to the LLM node.
- When set to True, do not add non-standard request extensions
  such as `thinking` and `chat_template_kwargs`.
- The default value is False.

Signed-off-by: Yuichiro Utsumi <utsumi.yuichiro@fujitsu.com>

Author: utsumi-fj

models/openai_api_compatible/manifest.yaml

@@ -1,4 +1,4 @@
-version: 0.0.30
+version: 0.0.31
 type: plugin
 author: "langgenius"
 name: "openai_api_compatible"

models/openai_api_compatible/models/llm/llm.py

@@ -82,6 +82,28 @@ def get_customizable_model_schema(
                 en_US=credentials["display_name"], zh_Hans=credentials["display_name"]
             )
 
+        entity.parameter_rules.append(
+            ParameterRule(
+                name="strict_compatibility",
+                label=I18nObject(en_US="Strict compatibility mode", zh_Hans="严格兼容模式"),
+                help=I18nObject(
+                    en_US=(
+                        "Whether to prioritize strict OpenAI compatibility. "
+                        "When True, OpenAI compatibility is prioritized and extended parameters "
+                        "(e.g., thinking, chat_template_kwargs) are not added. "
+                        "Set to False to enable these extensions."
+                    ),
+                    zh_Hans=(
+                        "是否优先严格的 OpenAI 兼容性。"
+                        "为 True 时，将优先 OpenAI 兼容性，并且不会添加扩展参数（例如 thinking、chat_template_kwargs）。"
+                        "设为 False 以启用这些扩展。"
+                    )
+                ),
+                type=ParameterType.BOOLEAN,
+                required=False,
+            )
+        )
+
         # Configure thinking mode parameter based on model support
         agent_thought_support = credentials.get("agent_thought_support", "not_supported")
 
@@ -200,9 +222,18 @@ def _invoke(
             user_enable_thinking = model_parameters.pop("enable_thinking", None)
             if user_enable_thinking is not None:
                 enable_thinking_value = bool(user_enable_thinking)
-                
-        chat_template_kwargs = model_parameters.setdefault("chat_template_kwargs", {})
-        if enable_thinking_value is not None:
+
+        user_strict_compatibility = model_parameters.pop("strict_compatibility", None)
+        # Default `strict_compatibility_value` is False.
+        strict_compatibility_value = False
+        if user_strict_compatibility is not None:
+            strict_compatibility_value = bool(user_strict_compatibility)
+
+        if enable_thinking_value is not None and strict_compatibility_value is False:
+            # Only apply when `strict_compatibility_value` is False since
+            # `chat_template_kwargs` and `thinking` are non-standard parameters.
+
+            chat_template_kwargs = model_parameters.setdefault("chat_template_kwargs", {})
             # Support vLLM/SGLang format (chat_template_kwargs)
             chat_template_kwargs["enable_thinking"] = enable_thinking_value
             chat_template_kwargs["thinking"] = enable_thinking_value
@@ -220,7 +251,11 @@ def _invoke(
             # - chat_template_kwargs for runtimes that read template kwargs (e.g., llama.cpp).
             # Only apply when thinking mode is explicitly enabled.
             model_parameters["reasoning_effort"] = reasoning_effort_value
-            chat_template_kwargs["reasoning_effort"] = reasoning_effort_value
+            if strict_compatibility_value is False:
+                # Only apply when `strict_compatibility_value` is False since
+                # `chat_template_kwargs` is a non-standard parameter.
+                chat_template_kwargs = model_parameters.setdefault("chat_template_kwargs", {})
+                chat_template_kwargs["reasoning_effort"] = reasoning_effort_value
 
         # Remove thinking content from assistant messages for better performance.
         with suppress(Exception):