fix(providers): map native reasoning efforts

docs/providers/groq.md

@@ -71,6 +71,14 @@ Use `openclaw models list --provider groq` for the most up-to-date list of
 models available on your account.
 </Tip>
 
+## Reasoning models
+
+OpenClaw maps its shared `/think` levels to Groq's model-specific
+`reasoning_effort` values. For `qwen/qwen3-32b`, disabled thinking sends
+`none` and enabled thinking sends `default`. For Groq GPT-OSS reasoning models,
+OpenClaw sends `low`, `medium`, or `high`; disabled thinking omits
+`reasoning_effort` because those models do not support a disabled value.
+
 ## Audio transcription
 
 Groq also provides fast Whisper-based audio transcription. When configured as a

docs/providers/lmstudio.md

@@ -104,7 +104,7 @@ LM Studio is streaming-usage compatible. When it does not emit an OpenAI-shaped
 `usage` object, OpenClaw recovers token counts from llama.cpp-style
 `timings.prompt_n` / `timings.predicted_n` metadata instead.
 
-Same behavior applies to these OpenAI-compatible local backends:
+Same streaming usage behavior applies to these OpenAI-compatible local backends:
 
 - vLLM
 - SGLang
@@ -114,6 +114,14 @@ Same behavior applies to these OpenAI-compatible local backends:
 - TabbyAPI
 - text-generation-webui
 
+### Thinking compatibility
+
+When LM Studio's `/api/v1/models` discovery reports model-specific reasoning
+options, OpenClaw preserves those native values in model compat metadata. For
+binary thinking models that advertise `allowed_options: ["off", "on"]`,
+OpenClaw maps disabled thinking to `off` and enabled `/think` levels to `on`
+instead of sending OpenAI-only values such as `low` or `medium`.
+
 ### Explicit configuration
 
 ```json5