feat(llm): expose temperature, top_p, and max_tokens settings

Add optional generation parameters so users can steer model behavior —
particularly useful for local / OpenAI-compatible models that need a
lower temperature for steadier tool calling (see #514).

- STRIX_LLM_TEMPERATURE, STRIX_LLM_TOP_P, STRIX_LLM_MAX_TOKENS
  (all unset by default -> provider defaults, so no behavior change).
- Threaded through make_model_settings into the SDK ModelSettings used
  for the scan. Params a given model rejects are dropped automatically
  (litellm.drop_params is already enabled).
- Documented in docs/advanced/configuration.mdx.

Closes #514

docs/advanced/configuration.mdx

@@ -31,6 +31,18 @@ Configure Strix using environment variables or a config file.
   Control thinking effort for reasoning models. Valid values: `none`, `minimal`, `low`, `medium`, `high`, `xhigh`. Defaults to `medium` for quick scan mode.
 </ParamField>
 
+<ParamField path="STRIX_LLM_TEMPERATURE" type="number">
+  Sampling temperature passed to the model. Unset by default (uses the provider default). Useful for steadier tool use on local/OpenAI-compatible models. Parameters unsupported by a given model are dropped automatically.
+</ParamField>
+
+<ParamField path="STRIX_LLM_TOP_P" type="number">
+  Nucleus sampling `top_p` passed to the model. Unset by default.
+</ParamField>
+
+<ParamField path="STRIX_LLM_MAX_TOKENS" type="integer">
+  Maximum number of tokens to generate per LLM response. Unset by default (uses the provider default).
+</ParamField>
+
 <ParamField path="STRIX_MEMORY_COMPRESSOR_TIMEOUT" default="30" type="integer">
   Timeout in seconds for memory compression operations (context summarization).
 </ParamField>

strix/config/settings.py

@@ -37,6 +37,9 @@ class LlmSettings(BaseSettings):
     )
     reasoning_effort: ReasoningEffort = Field(default="high", alias="STRIX_REASONING_EFFORT")
     timeout: int = Field(default=300, alias="LLM_TIMEOUT")
+    temperature: float | None = Field(default=None, alias="STRIX_LLM_TEMPERATURE")
+    top_p: float | None = Field(default=None, alias="STRIX_LLM_TOP_P")
+    max_tokens: int | None = Field(default=None, alias="STRIX_LLM_MAX_TOKENS")
 
 
 class RuntimeSettings(BaseSettings):

strix/core/inputs.py

@@ -110,11 +110,17 @@ def make_model_settings(
     reasoning_effort: ReasoningEffort | None,
     *,
     model_name: str,
+    temperature: float | None = None,
+    top_p: float | None = None,
+    max_tokens: int | None = None,
 ) -> ModelSettings:
     model_settings = ModelSettings(
         parallel_tool_calls=False,
         retry=DEFAULT_MODEL_RETRY,
         include_usage=True,
+        temperature=temperature,
+        top_p=top_p,
+        max_tokens=max_tokens,
     )
     if (
         reasoning_effort is not None

strix/core/runner.py

@@ -156,6 +156,9 @@ async def run_strix_scan(
         model_settings = make_model_settings(
             settings.llm.reasoning_effort,
             model_name=resolved_model,
+            temperature=settings.llm.temperature,
+            top_p=settings.llm.top_p,
+            max_tokens=settings.llm.max_tokens,
         )
         run_config = RunConfig(
             model=resolved_model,