Skip to content

Your lightweight open-source personal agent.

fix(providers): retain GPT-6 temperature restrictions

chengyongru93df72307067
diff --git a/nanobot/providers/openai_compat_provider.py b/nanobot/providers/openai_compat_provider.py
index 6087305f3..de2e4b676 100644
--- a/nanobot/providers/openai_compat_provider.py
+++ b/nanobot/providers/openai_compat_provider.py
@@ -895,15 +895,21 @@ class OpenAICompatProvider(LLMProvider):
return model_name
@staticmethod
- def _supports_temperature(model_name: str) -> bool:
+ def _supports_temperature(
+ model_name: str,
+ reasoning_effort: str | None = None,
+ ) -> bool:
"""Return True when the model accepts a temperature parameter.
- Kimi K3 uses a fixed temperature that should be omitted. GPT-5 family
- and reasoning models (o1/o3/o4) reject temperature.
+ Temperature is omitted for fixed-temperature Kimi K3, GPT-5, and
+ o-series models. GPT-6 requires explicit ``"none"`` effort; its
+ default enables reasoning.
"""
if _model_slug(model_name) == _KIMI_K3_MODEL:
return False
name = model_name.lower()
+ if "gpt-6" in name:
+ return bool(reasoning_effort and reasoning_effort.lower() == "none")
return not any(token in name for token in ("gpt-5", "o1", "o3", "o4"))
def _opencode_affinity_headers(
@@ -949,8 +955,7 @@ class OpenAICompatProvider(LLMProvider):
),
}
- # GPT-5 and reasoning models (o1/o3/o4) reject temperature.
- if self._supports_temperature(model_name):
+ if self._supports_temperature(model_name, reasoning_effort):
kwargs["temperature"] = temperature
if (
@@ -1313,7 +1318,7 @@ class OpenAICompatProvider(LLMProvider):
"compact_threshold": compact_threshold,
}]
- supports_temperature = self._supports_temperature(model_name)
+ supports_temperature = self._supports_temperature(model_name, reasoning_effort)
if supports_temperature:
body["temperature"] = temperature
diff --git a/tests/providers/test_litellm_kwargs.py b/tests/providers/test_litellm_kwargs.py
index 05c2e88f8..53a397e10 100644
--- a/tests/providers/test_litellm_kwargs.py
+++ b/tests/providers/test_litellm_kwargs.py
@@ -1014,13 +1014,22 @@ def test_openai_compat_supports_temperature_matches_reasoning_model_rules() -> N
assert OpenAICompatProvider._supports_temperature("o3-mini") is False
[email protected]("responses", [False, True])
-def test_deepseek_keeps_temperature_when_reasoning_effort_is_set(responses: bool) -> None:
- spec = find_by_name("deepseek")
+ ("provider_name", "model", "responses"),
+ [
+ ("deepseek", "deepseek-v4-flash", False),
+ ("deepseek", "deepseek-v4-flash", True),
+ ("mistral", "mistral-medium-3-5", False),
+ ],
+)
+def test_compatible_provider_keeps_temperature_when_reasoning_effort_is_set(
+ provider_name: str, model: str, responses: bool,
+) -> None:
+ spec = find_by_name(provider_name)
with patch("nanobot.providers.openai_compat_provider.AsyncOpenAI"):
provider = OpenAICompatProvider(
api_key="sk-test-key",
- default_model="deepseek-v4-flash",
+ default_model=model,
spec=spec,
)
@@ -1028,7 +1037,7 @@ def test_deepseek_keeps_temperature_when_reasoning_effort_is_set(responses: bool
request = build_request(
messages=[{"role": "user", "content": "hello"}],
tools=None,
- model="deepseek-v4-flash",
+ model=model,
max_tokens=4096,
temperature=0.2,
reasoning_effort="high",
@@ -1042,6 +1051,51 @@ def test_deepseek_keeps_temperature_when_reasoning_effort_is_set(responses: bool
assert request["reasoning_effort"] == "high"
[email protected]("api_type", ["chat_completions", "responses"])
+ ("model", "effort", "supports_temperature"),
+ [
+ ("gpt-6.1-sol", "high", False),
+ ("gpt-6-sol", "high", False),
+ ("gpt-6-sol", None, False),
+ ("gpt-6-sol", "none", True),
+ ],
+)
+async def test_gpt6_temperature_requires_explicit_none_effort(
+ api_type: str, model: str, effort: str | None, supports_temperature: bool,
+) -> None:
+ mock_chat = AsyncMock(return_value=_fake_chat_response())
+ mock_responses = AsyncMock(return_value=_fake_responses_response())
+ with patch("nanobot.providers.openai_compat_provider.AsyncOpenAI") as mock_client:
+ mock_client.return_value.chat.completions.create = mock_chat
+ mock_client.return_value.responses.create = mock_responses
+ provider = OpenAICompatProvider(
+ api_key="sk-test-key",
+ default_model=model,
+ spec=find_by_name("openai"),
+ api_type=api_type,
+ )
+ result = await provider.chat(
+ messages=[{"role": "user", "content": "hello"}],
+ temperature=0.2,
+ reasoning_effort=effort,
+ )
+
+ assert result.content == "ok"
+ request_mock = mock_responses if api_type == "responses" else mock_chat
+ other_mock = mock_chat if api_type == "responses" else mock_responses
+ request_mock.assert_awaited_once()
+ other_mock.assert_not_awaited()
+ request = request_mock.call_args.kwargs
+ if supports_temperature:
+ assert request["temperature"] == 0.2
+ else:
+ assert "temperature" not in request
+ if api_type == "responses":
+ assert request["include"] == ["reasoning.encrypted_content"]
+
+
def test_openai_compat_build_kwargs_uses_gpt5_safe_parameters() -> None:
spec = find_by_name("openai")
with patch("nanobot.providers.openai_compat_provider.AsyncOpenAI"):