ai-workspace-services/litellm
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

refactor(anthropic): extract _validate_effort_for_model to prevent drift

Browse Source 
The chat completion path (`_apply_output_config`) and the /v1/messages
pass-through (`AnthropicMessagesConfig._translate_reasoning_effort_to_anthropic`)
both gate `max` / `xhigh` per model. The two sites had diverged from
near-identical copies into separately maintained blocks, creating a real
drift risk when a new model tier (e.g. Claude 4.8) lands -- a contributor
could update one site and miss the other.

Centralise the gating in `AnthropicConfig._validate_effort_for_model`,
which returns an error message string or `None`. Each call site keeps
its own provider-appropriate exception type (`BadRequestError` for the
chat path, `AnthropicError` for the /v1/messages pass-through) but the
gating decision now comes from one place. Net -11 LOC.

Adds a parametrised unit test exercising the helper directly across
4.5 / 4.6 / 4.7 model families and `max` / `xhigh` / lower-effort
inputs. Existing tests at both call sites continue to pass unchanged.

Addresses Greptile finding on PR #27074.
This commit is contained in:
mateo-berri 2026-05-04 00:47:55 -07:00
parent 4f9a3a5c9f
commit f8f07c5cb7
3 changed files with 89 additions and 62 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

71
litellm/llms/anthropic/chat/transformation.py
View File

@ -275,6 +275,41 @@ class AnthropicConfig(AnthropicModelInfo, BaseConfig):
pass
return False
@staticmethod
def _validate_effort_for_model(model: str, effort: Optional[str]) -> Optional[str]:
"""Return ``None`` if ``effort`` is allowed on ``model``, else an error message.
Centralises per-model gating for ``max`` and ``xhigh`` so the chat
completion path (``_apply_output_config``) and the /v1/messages
pass-through (``AnthropicMessagesConfig._translate_reasoning_effort_to_anthropic``)
can't drift when a new model tier is added. Caller raises the
provider-appropriate exception type using the returned message.
``max`` is supported on Claude 4.6 (Opus + Sonnet) and Claude 4.7
adaptive-thinking models per
https://platform.claude.com/docs/en/build-with-claude/effort. The
data-driven ``supports_max_reasoning_effort`` flag in
``model_prices_and_context_window.json`` is the source of truth;
family-level ``_is_claude_4_6_model`` / ``_is_claude_4_7_model``
checks remain as a fallback for OpenRouter / GitHub Copilot /
Vercel / Bedrock variants whose model-map entries don't yet carry
the flag.
``xhigh`` is purely data-driven via ``supports_xhigh_reasoning_effort``
so enabling it for a new model is a model-map-only change.
"""
if effort == "max" and not (
AnthropicConfig._is_claude_4_6_model(model)
or AnthropicConfig._is_claude_4_7_model(model)
or AnthropicConfig._supports_effort_level(model, "max")
):
return f"effort='max' is not supported by this model. Got model: {model}"
if effort == "xhigh" and not AnthropicConfig._supports_effort_level(
model, "xhigh"
):
return f"effort='xhigh' is not supported by this model. Got model: {model}"
return None
@staticmethod
def _model_supports_effort_param(model: str) -> bool:
"""Whether the model accepts ``output_config.effort`` at all.
@ -1710,36 +1745,14 @@ class AnthropicConfig(AnthropicModelInfo, BaseConfig):
model=model,
llm_provider=self.custom_llm_provider or "anthropic",
)
# ``max`` is supported on Claude 4.6 (Opus + Sonnet) and Claude 4.7
# adaptive-thinking models (per
# https://platform.claude.com/docs/en/build-with-claude/effort).
# Prefer the data-driven ``supports_max_reasoning_effort`` flag in
# ``model_prices_and_context_window.json`` so new variants only
# require a model-map update. Family-level ``_is_claude_4_6_model``
# / ``_is_claude_4_7_model`` checks remain as a fallback for
# OpenRouter/GitHub Copilot/Vercel/Bedrock variants whose entries
# don't yet carry the flag.
if effort == "max" and not (
self._is_claude_4_6_model(model)
or self._is_claude_4_7_model(model)
or self._supports_effort_level(model, "max")
):
# Per-model gating for ``max`` / ``xhigh`` is centralised in
# ``_validate_effort_for_model`` so the chat path and the
# /v1/messages pass-through stay in lock-step when a new model
# tier lands.
gate_error = self._validate_effort_for_model(model, effort)
if gate_error is not None:
raise litellm.exceptions.BadRequestError(
message=(
f"effort='max' is not supported by this model. "
f"Got model: {model}"
),
model=model,
llm_provider=self.custom_llm_provider or "anthropic",
)
# ``xhigh`` is data-driven via ``supports_xhigh_reasoning_effort`` so
# enabling it for a new model is a pure model-map change.
if effort == "xhigh" and not self._supports_effort_level(model, "xhigh"):
raise litellm.exceptions.BadRequestError(
message=(
f"effort='xhigh' is not supported by this model. "
f"Got model: {model}"
),
message=gate_error,
model=model,
llm_provider=self.custom_llm_provider or "anthropic",
)

42
litellm/llms/anthropic/experimental_pass_through/messages/transformation.py
View File

@ -239,39 +239,15 @@ class AnthropicMessagesConfig(BaseAnthropicMessagesConfig):
),
status_code=400,
)
# Per-model gating: ``xhigh`` and ``max`` are only valid on
# specific tiers (Opus 4.6/4.7 for max; data-driven for xhigh).
# The chat completion path enforces this via
# ``_apply_output_config``; mirror it here so /v1/messages
# callers see a clean 400 instead of a provider-side error.
# ``max`` is supported on Claude 4.6 (Opus + Sonnet) and Claude
# 4.7 adaptive-thinking models. Prefer the data-driven
# ``supports_max_reasoning_effort`` flag in
# ``model_prices_and_context_window.json``; family-level checks
# are a fallback for variants whose entries don't yet carry the
# flag.
if mapped_effort == "max" and not (
AnthropicConfig._is_claude_4_6_model(model)
or AnthropicConfig._is_claude_4_7_model(model)
or AnthropicConfig._supports_effort_level(model, "max")
):
raise AnthropicError(
message=(
f"effort='max' is not supported by this model. "
f"Got model: {model}"
),
status_code=400,
)
if mapped_effort == "xhigh" and not AnthropicConfig._supports_effort_level(
model, "xhigh"
):
raise AnthropicError(
message=(
f"effort='xhigh' is not supported by this model. "
f"Got model: {model}"
),
status_code=400,
)
# Per-model gating for ``max`` / ``xhigh`` is centralised in
# ``AnthropicConfig._validate_effort_for_model`` so the chat
# completion path and this /v1/messages pass-through stay in
# lock-step when a new model tier lands.
gate_error = AnthropicConfig._validate_effort_for_model(
model, mapped_effort
)
if gate_error is not None:
raise AnthropicError(message=gate_error, status_code=400)
existing_output_config = optional_params.get("output_config")
if not isinstance(existing_output_config, dict):
existing_output_config = {}

38
tests/test_litellm/llms/anthropic/chat/test_anthropic_chat_transformation.py
View File

@ -2101,6 +2101,44 @@ def test_supports_effort_level_handles_provider_prefixes(model, level, expected)
assert AnthropicConfig._supports_effort_level(model, level) is expected
@pytest.mark.parametrize(
"model,effort,expect_error",
[
# ``max`` accepted on 4.6 / 4.7 (family fallback) and rejected on 4.5.
("claude-opus-4-6", "max", False),
("claude-sonnet-4-6", "max", False),
("claude-opus-4-7", "max", False),
("claude-opus-4-5-20251101", "max", True),
("claude-sonnet-4-5", "max", True),
# ``xhigh`` data-driven; only 4.7 carries the flag in the model map.
("claude-opus-4-7", "xhigh", False),
("claude-opus-4-6", "xhigh", True),
("claude-sonnet-4-6", "xhigh", True),
# Lower efforts and ``None`` always pass the gate.
("claude-opus-4-5-20251101", "high", False),
("claude-haiku-4-5", "low", False),
("claude-opus-4-5-20251101", None, False),
],
)
def test_validate_effort_for_model_centralises_per_model_gating(
model, effort, expect_error
):
"""``_validate_effort_for_model`` is the single source of truth for the
per-model ``max`` / ``xhigh`` gating that ``_apply_output_config`` (chat
completion path) and ``AnthropicMessagesConfig._translate_reasoning_effort_to_anthropic``
(/v1/messages pass-through) both rely on. Both call sites raise their
own provider-appropriate exception, but the gating decision must come
from one place to prevent drift when a new model tier lands.
"""
err = AnthropicConfig._validate_effort_for_model(model, effort)
if expect_error:
assert err is not None
assert effort in err
assert model in err
else:
assert err is None
def test_transform_request_uses_dynamic_max_tokens():
"""
Test that transform_request uses dynamic max_tokens based on model