Skip to content

CubeLoop — a Pythonic, async-native agent harness core designed for high performance, readability, and production-grade persistence.

feat(providers): drop the "minimal" ThinkingLevel

xf gongb8aebc189405
diff --git a/cubepi/providers/anthropic.py b/cubepi/providers/anthropic.py
index 72fe3c1..2b8ab39 100644
--- a/cubepi/providers/anthropic.py
+++ b/cubepi/providers/anthropic.py
@@ -49,7 +49,7 @@ CacheRetention = Literal["short", "long", "none"]
# - reasoning_on_payload writes {"thinking": {"type": "enabled"}}, then
# reasoning_level adds budget_tokens at thinking.budget_tokens.
# - level_budgets mirrors cubepi.providers.base.ThinkingBudgets defaults
-# (minimal=1024, low=2048, medium=8192, high=16384). xhigh clamps to high
+# (low=2048, medium=8192, high=16384). xhigh clamps to high
# to match adjust_max_tokens_for_thinking's "xhigh -> high" mapping.
# "off" is included for completeness but reasoning_level is never written
# on the off-branch.
@@ -65,7 +65,6 @@ _ANTHROPIC_DEFAULT_CAPABILITY = CapabilityDescriptor(
kind="int_budget",
level_budgets={
"off": 0,
- "minimal": 1024,
"low": 2048,
"medium": 8192,
"high": 16384,
diff --git a/cubepi/providers/base.py b/cubepi/providers/base.py
index 7daa800..cbaad54 100644
--- a/cubepi/providers/base.py
+++ b/cubepi/providers/base.py
@@ -21,7 +21,7 @@ from cubepi.types import JsonObject, StructuredValue
BaseModelT = TypeVar("BaseModelT", bound=BaseModel)
-ThinkingLevel = Literal["off", "minimal", "low", "medium", "high", "xhigh"]
+ThinkingLevel = Literal["off", "low", "medium", "high", "xhigh"]
ToolChoice = Literal["auto", "required", "none"] | str
@@ -29,7 +29,6 @@ ToolChoice = Literal["auto", "required", "none"] | str
class ThinkingBudgets(BaseModel):
"""Token budgets for each thinking level."""
- minimal: int = 1024
low: int = 2048
medium: int = 8192
high: int = 16384
diff --git a/cubepi/providers/models.py b/cubepi/providers/models.py
index 16c9b41..3c4b562 100644
--- a/cubepi/providers/models.py
+++ b/cubepi/providers/models.py
@@ -11,7 +11,6 @@ from cubepi.providers.base import Model, ThinkingLevel
# Ordered list of all thinking levels from lowest to highest.
THINKING_LEVELS: list[ThinkingLevel] = [
"off",
- "minimal",
"low",
"medium",
"high",
diff --git a/cubepi/providers/openai_responses.py b/cubepi/providers/openai_responses.py
index 9473f5a..43100a6 100644
--- a/cubepi/providers/openai_responses.py
+++ b/cubepi/providers/openai_responses.py
@@ -38,10 +38,8 @@ from cubepi.providers.capability import (
# Map cubepi ThinkingLevel to OpenAI reasoning.effort values.
# "off" means no reasoning parameter is sent.
-# "minimal" is not a valid OpenAI effort level, so we map it to "low".
_THINKING_TO_EFFORT: dict[ThinkingLevel, str | None] = {
"off": None,
- "minimal": "low",
"low": "low",
"medium": "medium",
"high": "high",
diff --git a/tests/providers/test_models.py b/tests/providers/test_models.py
index 8957a61..389d468 100644
--- a/tests/providers/test_models.py
+++ b/tests/providers/test_models.py
@@ -44,7 +44,6 @@ class TestGetSupportedThinkingLevels:
model = _model(reasoning=True)
assert get_supported_thinking_levels(model) == [
"off",
- "minimal",
"low",
"medium",
"high",
@@ -58,7 +57,7 @@ class TestGetSupportedThinkingLevels:
levels = get_supported_thinking_levels(model)
assert "xhigh" in levels
# All standard levels should still be present
- for lvl in ("off", "minimal", "low", "medium", "high"):
+ for lvl in ("off", "low", "medium", "high"):
assert lvl in levels
def test_xhigh_mapped_to_none_is_excluded(self):
@@ -71,10 +70,9 @@ class TestGetSupportedThinkingLevels:
def test_level_mapped_to_none_is_excluded(self):
model = _model(
reasoning=True,
- thinking_level_map={"minimal": None, "low": None},
+ thinking_level_map={"low": None},
)
levels = get_supported_thinking_levels(model)
- assert "minimal" not in levels
assert "low" not in levels
# Others remain
assert "off" in levels
@@ -84,13 +82,12 @@ class TestGetSupportedThinkingLevels:
def test_empty_map_same_as_no_map(self):
model = _model(reasoning=True, thinking_level_map={})
levels = get_supported_thinking_levels(model)
- assert levels == ["off", "minimal", "low", "medium", "high"]
+ assert levels == ["off", "low", "medium", "high"]
def test_all_levels_disabled_except_off(self):
model = _model(
reasoning=True,
thinking_level_map={
- "minimal": None,
"low": None,
"medium": None,
"high": None,
@@ -126,10 +123,10 @@ class TestClampThinkingLevel:
"""When a level is disabled, search downward first (prefer cheaper)."""
model = _model(
reasoning=True,
- thinking_level_map={"low": None},
+ thinking_level_map={"medium": None},
)
- # "low" disabled -> next down is "minimal"
- assert clamp_thinking_level(model, "low") == "minimal"
+ # "medium" disabled -> next down is "low"
+ assert clamp_thinking_level(model, "medium") == "low"
def test_clamp_down_when_nothing_above(self):
"""When all higher levels are disabled, clamp down."""
@@ -150,7 +147,6 @@ class TestClampThinkingLevel:
model = _model(
reasoning=True,
thinking_level_map={
- "minimal": None,
"low": None,
"medium": None,
"high": None,
@@ -209,7 +205,6 @@ class TestThinkingLevelsOrdering:
def test_correct_order(self):
assert THINKING_LEVELS == [
"off",
- "minimal",
"low",
"medium",
"high",
@@ -217,5 +212,5 @@ class TestThinkingLevelsOrdering:
]
def test_all_levels_present(self):
- expected = {"off", "minimal", "low", "medium", "high", "xhigh"}
+ expected = {"off", "low", "medium", "high", "xhigh"}
assert set(THINKING_LEVELS) == expected
diff --git a/tests/providers/test_thinking_budgets.py b/tests/providers/test_thinking_budgets.py
index 1fa32b1..6809cb1 100644
--- a/tests/providers/test_thinking_budgets.py
+++ b/tests/providers/test_thinking_budgets.py
@@ -4,14 +4,13 @@ from cubepi.providers.base import ThinkingBudgets, adjust_max_tokens_for_thinkin
class TestThinkingBudgets:
def test_defaults(self):
b = ThinkingBudgets()
- assert b.minimal == 1024
assert b.low == 2048
assert b.medium == 8192
assert b.high == 16384
def test_custom_values(self):
- b = ThinkingBudgets(minimal=512, low=1024, medium=4096, high=8192)
- assert b.minimal == 512
+ b = ThinkingBudgets(low=1024, medium=4096, high=8192)
+ assert b.low == 1024
assert b.high == 8192
@@ -25,15 +24,6 @@ class TestAdjustMaxTokensForThinking:
assert max_tokens == 8192
assert budget == 0
- def test_minimal_adds_1024(self):
- max_tokens, budget = adjust_max_tokens_for_thinking(
- base_max_tokens=8192,
- model_max_tokens=128_000,
- reasoning_level="minimal",
- )
- assert budget == 1024
- assert max_tokens == 8192 + 1024
-
def test_low_adds_2048(self):
max_tokens, budget = adjust_max_tokens_for_thinking(
base_max_tokens=8192,
@@ -105,7 +95,7 @@ class TestAdjustMaxTokensForThinking:
assert budget == 0
def test_custom_budgets_override_defaults(self):
- custom = ThinkingBudgets(minimal=256, low=512, medium=2048, high=4096)
+ custom = ThinkingBudgets(low=512, medium=2048, high=4096)
max_tokens, budget = adjust_max_tokens_for_thinking(
base_max_tokens=8192,
model_max_tokens=128_000,
@@ -128,26 +118,27 @@ class TestAdjustMaxTokensForThinking:
def test_exact_model_cap_equals_budget(self):
# Edge case: model cap exactly equals the budget
- # max_tokens = min(8192 + 1024, 1024) = 1024
- # 1024 <= 1024 so budget = max(0, 1024 - 1024) = 0
+ # max_tokens = min(8192 + 2048, 2048) = 2048
+ # 2048 - 2048 = 0 < min_output_tokens(1024), budget = max(0, 2048 - 1024) = 1024
max_tokens, budget = adjust_max_tokens_for_thinking(
base_max_tokens=8192,
- model_max_tokens=1024,
- reasoning_level="minimal",
+ model_max_tokens=2048,
+ reasoning_level="low",
)
- assert max_tokens == 1024
- assert budget == 0
+ assert max_tokens == 2048
+ assert budget == 1024
def test_model_cap_just_above_budget(self):
- # max_tokens = min(8192 + 1024, 1025) = 1025
- # 1025 - 1024 = 1 < min_output_tokens(1024), so budget reduced
+ # base=8192, model=2049, low budget=2048
+ # max_tokens = min(8192 + 2048, 2049) = 2049
+ # 2049 - 2048 = 1 < min_output_tokens(1024), budget = max(0, 2049 - 1024) = 1025
max_tokens, budget = adjust_max_tokens_for_thinking(
base_max_tokens=8192,
- model_max_tokens=1025,
- reasoning_level="minimal",
+ model_max_tokens=2049,
+ reasoning_level="low",
)
- assert max_tokens == 1025
- assert budget == 1 # max(0, 1025 - 1024)
+ assert max_tokens == 2049
+ assert budget == 1025
def test_output_tokens_minimum_guaranteed(self):
# Bug case: base=8192, model=9000, medium budget=8192