Skip to content

armgit / strix

Public Imported from GitHub
0

Open-source AI penetration testing tool to find and fix your app’s vulnerabilities.

Drop tool_choice for registry-unknown reasoning-effort runs

0xallam712c64f6300d
diff --git a/strix/config/models.py b/strix/config/models.py
index 59a52a48..83f8e914 100644
--- a/strix/config/models.py
+++ b/strix/config/models.py
@@ -121,7 +121,7 @@ def uses_chat_completions_tool_schema(model_name: str, settings: Settings) -> bo
return bool(settings.llm.api_base)
-def model_supports_reasoning(model_name: str) -> bool:
+def _model_cost_entry(model_name: str) -> dict[str, object] | None:
import litellm
name = model_name.strip().lower()
@@ -132,4 +132,13 @@ def model_supports_reasoning(model_name: str) -> bool:
entry = litellm.model_cost.get(name)
if entry is None and "/" in name:
entry = litellm.model_cost.get(name.rsplit("/", 1)[1])
+ return entry
+
+
+def model_supports_reasoning(model_name: str) -> bool:
+ entry = _model_cost_entry(model_name)
return bool(entry and entry.get("supports_reasoning"))
+
+
+def model_known_to_registry(model_name: str) -> bool:
+ return _model_cost_entry(model_name) is not None
diff --git a/strix/core/inputs.py b/strix/core/inputs.py
index 148bf09b..58da3e60 100644
--- a/strix/core/inputs.py
+++ b/strix/core/inputs.py
@@ -8,7 +8,11 @@ from typing import TYPE_CHECKING, Any
from agents.model_settings import ModelSettings
from openai.types.shared import Reasoning
-from strix.config.models import DEFAULT_MODEL_RETRY, model_supports_reasoning
+from strix.config.models import (
+ DEFAULT_MODEL_RETRY,
+ model_known_to_registry,
+ model_supports_reasoning,
+)
if TYPE_CHECKING:
@@ -111,23 +115,25 @@ def make_model_settings(
*,
model_name: str,
) -> ModelSettings:
- # Anthropic + DeepSeek thinking reject ``tool_choice="required"`` outright
- # when reasoning is enabled; OpenAI o-series accepts both but doesn't need
- # the safety net. When reasoning is on we let the model self-select tools
- # and rely on the system prompt + the ``_finish_tool_use_behavior`` callback
- # to keep the loop converging on a lifecycle tool.
- use_reasoning = (
- reasoning_effort is not None
- and reasoning_effort != "none"
- and model_supports_reasoning(model_name)
+ # Anthropic + DeepSeek thinking reject ``tool_choice="required"`` outright;
+ # when reasoning is enabled we let the model self-select tools and rely on
+ # the system prompt + the ``_finish_tool_use_behavior`` callback to keep
+ # the loop converging. When the user opted into reasoning but the model
+ # is unknown to LiteLLM's registry (e.g. a private DeepSeek SKU, a fresh
+ # release the registry hasn't picked up), drop ``tool_choice`` too —
+ # server-side thinking-mode endpoints reject it and we can't confirm.
+ user_wants_reasoning = reasoning_effort is not None and reasoning_effort != "none"
+ confirmed_reasoning = model_supports_reasoning(model_name)
+ drop_tool_choice = user_wants_reasoning and (
+ confirmed_reasoning or not model_known_to_registry(model_name)
)
model_settings = ModelSettings(
parallel_tool_calls=False,
- tool_choice=None if use_reasoning else "required",
+ tool_choice=None if drop_tool_choice else "required",
retry=DEFAULT_MODEL_RETRY,
include_usage=True,
)
- if use_reasoning:
+ if user_wants_reasoning and confirmed_reasoning:
model_settings = model_settings.resolve(
ModelSettings(reasoning=Reasoning(effort=reasoning_effort)),
)