Skip to content

DeepTutor: Lifelong Personalized Tutoring. https://deeptutor.info/.

fix(llm): share Claude temperature rules across provider routes

Bingxi Zhao (Frank)880080d94f01
diff --git a/deeptutor/services/llm/provider_core/anthropic_provider.py b/deeptutor/services/llm/provider_core/anthropic_provider.py
index 712e423db..47bd029e5 100644
--- a/deeptutor/services/llm/provider_core/anthropic_provider.py
+++ b/deeptutor/services/llm/provider_core/anthropic_provider.py
@@ -16,27 +16,13 @@ from typing import Any
import json_repair
from deeptutor.services.llm.provider_core.base import LLMProvider, LLMResponse, ToolCallRequest
+from deeptutor.services.provider_registry import ANTHROPIC_EFFORT_BASED_FAMILIES
from deeptutor.services.session.provider_response_state import (
normalize_provider_response_state,
)
_ALNUM = string.ascii_letters + string.digits
-# Effort-based model families from Opus 4.7 onward. These REJECT both the
-# `temperature` parameter and `thinking: {type: "enabled", budget_tokens: N}`
-# with a 400 — adaptive is their only thinking on-mode. Opus 4.6 and
-# Sonnet 4.6 still ACCEPT both older forms; adding them here would silently
-# drop the user's settings. Extend as new families ship (a capability lookup
-# is the longer-term fix).
-_EFFORT_BASED_FAMILIES: tuple[str, ...] = (
- "opus-4-7",
- "opus-4-8",
- "opus-5",
- "sonnet-5",
- "fable-5",
- "mythos-5",
-)
-
# reasoning_effort values that mean "thinking off" (see services/config
# reasoning_params). On effort-based families the correct off/default
# expression is omitting the `thinking` param entirely — explicit
@@ -413,7 +399,9 @@ class AnthropicProvider(LLMProvider):
# value as the default budget, so a plain `bool(reasoning_effort)`
# turned `none` into thinking ON with 4096 tokens.
thinking_enabled = bool(effort) and effort not in _THINKING_OFF_EFFORTS
- effort_based = any(family in model_name for family in _EFFORT_BASED_FAMILIES)
+ effort_based = any(
+ family in model_name.lower() for family in ANTHROPIC_EFFORT_BASED_FAMILIES
+ )
kwargs: dict[str, Any] = {
"model": model_name,
diff --git a/deeptutor/services/llm/provider_core/openai_compat_provider.py b/deeptutor/services/llm/provider_core/openai_compat_provider.py
index fdadcd2d3..5d6e88533 100644
--- a/deeptutor/services/llm/provider_core/openai_compat_provider.py
+++ b/deeptutor/services/llm/provider_core/openai_compat_provider.py
@@ -742,6 +742,11 @@ class OpenAICompatProvider(LLMProvider):
if self._supports_temperature(model_name, reasoning_effort):
body["temperature"] = temperature
+ for key, value in model_overrides_for(model_name, self._spec).items():
+ if value is None:
+ body.pop(key, None)
+ else:
+ body[key] = value
if reasoning_effort and reasoning_effort.lower() != "none":
body["reasoning"] = {"effort": reasoning_effort}
body["include"] = ["reasoning.encrypted_content"]
diff --git a/deeptutor/services/provider_registry.py b/deeptutor/services/provider_registry.py
index 575a0791a..537b452ff 100644
--- a/deeptutor/services/provider_registry.py
+++ b/deeptutor/services/provider_registry.py
@@ -169,6 +169,19 @@ def canonical_provider_name(name: str | None) -> str | None:
return PROVIDER_ALIASES.get(key, key)
+# These Claude families reject explicit temperature and use adaptive thinking.
+# Keep one family list for direct Anthropic calls and OpenAI-compatible gateways.
+# Older Opus/Sonnet families still accept temperature and budget-token thinking.
+ANTHROPIC_EFFORT_BASED_FAMILIES: tuple[str, ...] = (
+ "claude-opus-4-7",
+ "claude-opus-4-8",
+ "claude-opus-5",
+ "claude-sonnet-5",
+ "claude-fable-5",
+ "claude-mythos-5",
+)
+
+
# ---------------------------------------------------------------------------
# PROVIDERS — the registry. Order = priority.
# ---------------------------------------------------------------------------
@@ -339,15 +352,10 @@ PROVIDERS: tuple[ProviderSpec, ...] = (
backend="anthropic",
default_api_base="https://api.anthropic.com/v1",
supports_prompt_caching=True,
- # Claude Sonnet 5 and Opus 4.8 reject an explicit `temperature`
- # outright ("`temperature` is deprecated for this model."), where
- # earlier Claude models still accept it. Dropping the parameter
- # (value None) lets the API apply its own default. Like the Kimi
- # case below, the limit is the vendor API's, not the route's, so it
- # must live here to survive binding="openai" through a gateway.
- model_overrides=(
- ("claude-sonnet-5", {"temperature": None}),
- ("claude-opus-4-8", {"temperature": None}),
+ # This model rule also applies when Claude is reached through a
+ # generic OpenAI-compatible binding or gateway.
+ model_overrides=tuple(
+ (family, {"temperature": None}) for family in ANTHROPIC_EFFORT_BASED_FAMILIES
),
),
ProviderSpec(
diff --git a/tests/services/llm/test_anthropic_provider_fixes.py b/tests/services/llm/test_anthropic_provider_fixes.py
index 0821abe79..132728a02 100644
--- a/tests/services/llm/test_anthropic_provider_fixes.py
+++ b/tests/services/llm/test_anthropic_provider_fixes.py
@@ -54,6 +54,7 @@ def test_temperature_omitted_for_effort_based_models() -> None:
"claude-sonnet-5",
"claude-opus-4-7",
"claude-fable-5",
+ "claude-mythos-5",
):
assert "temperature" not in _kwargs(provider, model), model
@@ -154,6 +155,7 @@ def test_effort_based_families_map_real_effort_to_adaptive_thinking() -> None:
"claude-sonnet-5",
"claude-opus-4-7",
"claude-fable-5",
+ "claude-mythos-5",
):
kwargs = _kwargs_with_effort(provider, model, "high")
assert kwargs["thinking"] == {"type": "adaptive"}, model
@@ -161,11 +163,24 @@ def test_effort_based_families_map_real_effort_to_adaptive_thinking() -> None:
assert kwargs["max_tokens"] == 1024, model # no budget headroom inflation
-def test_effort_based_families_omit_thinking_for_off_sentinels() -> None:
+ "model",
+ [
+ "claude-opus-4-7",
+ "claude-opus-4-8",
+ "claude-opus-5",
+ "claude-sonnet-5",
+ "claude-fable-5",
+ "claude-mythos-5",
+ ],
+)
[email protected]("effort", ["none", "minimal", "minimum"])
+def test_effort_based_families_omit_thinking_for_off_sentinels(model: str, effort: str) -> None:
provider = _provider()
- kwargs = _kwargs_with_effort(provider, "claude-opus-4-8", "minimal")
+ kwargs = _kwargs_with_effort(provider, model, effort)
assert "thinking" not in kwargs
assert "temperature" not in kwargs
+ assert "temperature" not in kwargs.get("extra_body", {})
def test_older_models_keep_budget_tokens_thinking() -> None:
diff --git a/tests/services/llm/test_model_overrides.py b/tests/services/llm/test_model_overrides.py
index 8f4d5abc4..06444fedb 100644
--- a/tests/services/llm/test_model_overrides.py
+++ b/tests/services/llm/test_model_overrides.py
@@ -64,24 +64,48 @@ def test_unrelated_models_are_untouched(model: str) -> None:
assert _payload("openai", model)["temperature"] == pytest.approx(0.7)
[email protected]("model", ["claude-sonnet-5", "claude-opus-4-8"])
[email protected]("binding", ["openai", "anthropic", "no-such-provider"])
-def test_claude_models_that_deprecated_temperature_never_send_it(
- binding: str, model: str
-) -> None:
- """Anthropic answers HTTP 400 ``\`temperature\` is deprecated for this
- model.`` for Sonnet 5 and Opus 4.8. Haiku 4.5 still accepts it, so the
- rule names models, not the vendor prefix.
-
- Verified against the live API on 2026-09-09. Like the Kimi case, the limit
- belongs to Anthropic's API and not to the route, so it must also fire for
- ``binding="openai"`` — which is how every OpenAI-compatible gateway in
- front of Claude reaches these models.
- """
+ "model",
+ [
+ "claude-opus-4-7",
+ "claude-opus-4-8",
+ "claude-opus-5",
+ "claude-sonnet-5",
+ "claude-fable-5",
+ "claude-mythos-5",
+ ],
+)
[email protected]("binding", ["openai", "openrouter", "anthropic", "no-such-provider"])
+def test_claude_effort_families_never_send_temperature(binding: str, model: str) -> None:
+ """The direct and gateway routes share the vendor's model restriction."""
assert "temperature" not in _payload(binding, model)
assert model_overrides_for(model, find_by_name(binding)) == {"temperature": None}
+def test_gateway_responses_body_applies_intrinsic_model_overrides() -> None:
+ provider = OpenAICompatProvider(
+ api_key="test-key",
+ api_base="https://openrouter.ai/api/v1",
+ default_model="anthropic/claude-opus-4-7",
+ spec=find_by_name("openrouter"),
+ provider_name="openrouter",
+ )
+
+ def body(model: str) -> dict:
+ return provider._build_responses_body(
+ messages=[{"role": "user", "content": "hi"}],
+ tools=None,
+ model=model,
+ max_tokens=256,
+ temperature=0.7,
+ reasoning_effort=None,
+ tool_choice=None,
+ )
+
+ assert "temperature" not in body("anthropic/claude-opus-4-7")
+ assert body("anthropic/claude-opus-4-6")["temperature"] == pytest.approx(0.7)
+
+
def test_configured_binding_wins_over_the_vendor_fallback() -> None:
"""An explicit binding is the user's statement about the route; the
vendor lookup is only consulted when it says nothing about this model."""