Skip to content

The agent that grows with you

test(agent): move closed-level-set cases to a sibling file; cover a rejected enabled level

unsupportedpastelse715a2de0da7
diff --git a/tests/agent/test_error_classifier.py b/tests/agent/test_error_classifier.py
index ca1928a50bd..a7718912624 100644
--- a/tests/agent/test_error_classifier.py
+++ b/tests/agent/test_error_classifier.py
@@ -940,20 +940,15 @@ class TestClassifyApiError:
def test_reasoning_field_rejection_is_reasoning_mandatory(self):
"""A 400 rejecting a reasoning wire control by name — reversed ("reasoning_effort 'none'
unsupported; use ...", #114460), forward ("Unrecognized request argument supplied:
- reasoning_effort"), an enum rejection whose only field name sits in the structured
- 'param' tail (commandcode.ai, #115277), or a closed-set value rejection that names the
- field but blames its value (Alibaba token-plan: "'reasoning_effort' must be one of:
- 'minimal', ...") — takes the drop-the-disable / step-up rung, not the format_error abort;
- a model-id segment (kimi-k2-thinking) stays route gating."""
+ reasoning_effort"), or an enum rejection whose only field name sits in the structured
+ 'param' tail (commandcode.ai, #115277) — takes the drop-the-disable rung, not the
+ format_error abort; a model-id segment (kimi-k2-thinking) stays route gating."""
for msg in (
"Error code: 400 - reasoning_effort 'none' unsupported; use minimal|low|medium|high|xhigh",
"Unrecognized request argument supplied: reasoning_effort",
"Error code: 400 - {'error': {'message': 'Invalid option: expected one of "
"\"low\"|\"medium\"|\"high\"|\"xhigh\"|\"max\"', 'type': 'invalid_request_error', "
"'param': 'reasoning_effort'}}",
- "Error code: 400 - {'error':{'code':'invalid_parameter_error','param':None,"
- "'message':\"'reasoning_effort' must be one of: 'minimal', 'low', 'medium', 'high', "
- "'xhigh', 'max', 'ultra'\",'type':'invalid_request_error'}}",
):
result = classify_api_error(MockAPIError(msg, status_code=400), provider="custom", model="m")
assert result.reason == FailoverReason.reasoning_mandatory, msg
diff --git a/tests/agent/test_error_classifier_reasoning_level.py b/tests/agent/test_error_classifier_reasoning_level.py
new file mode 100644
index 00000000000..266ad910513
--- /dev/null
+++ b/tests/agent/test_error_classifier_reasoning_level.py
@@ -0,0 +1,58 @@
+"""A 400 that names the reasoning field and refuses its *value* against a closed level set.
+
+The classifier routes it to ``reasoning_mandatory`` (not the ``format_error`` abort), and the
+main-loop rung reacts to what was actually sent: a disable steps up to the floor, an enabled level
+the route does not offer (the custom profile's unset default ``medium`` against "off, low, high, or
+max") is dropped for the session so the route default applies.
+"""
+import pytest
+
+from agent.error_classifier import FailoverReason, classify_api_error
+from tests.agent.test_reasoning_disable_recovery import _recover, _WireAgent, _wire_reasoning_config
+
+# Alibaba token-plan / DashScope: the set starts at ``minimal``, so ``none`` is refused.
+_TOKENPLAN_400 = (
+ "Error code: 400 - {'error':{'code':'invalid_parameter_error','param':None,"
+ "'message':\"'reasoning_effort' must be one of: 'minimal', 'low', 'medium', 'high', "
+ "'xhigh', 'max', 'ultra'\",'type':'invalid_request_error'}}"
+)
+# A custom OpenAI-compatible endpoint whose set has no ``medium`` (deepseek-v4.1-flash-fast).
+_OFF_LOW_HIGH_MAX_400 = "HTTP 400: reasoning_effort must be one of off, low, high, or max."
+# Groq gpt-oss: graded levels only.
+_GROQ_400 = (
+ "Error code: 400 - {'error': {'message': \"'reasoning_effort' must be one of 'low', "
+ "'medium', or 'high'\", 'type': 'invalid_request_error'}}"
+)
+
+
+class _APIError(Exception):
+ def __init__(self, message, status_code):
+ super().__init__(message)
+ self.status_code = status_code
+
+
[email protected]("body", [_TOKENPLAN_400, _OFF_LOW_HIGH_MAX_400, _GROQ_400])
+def test_closed_level_set_rejection_is_reasoning_mandatory(body):
+ result = classify_api_error(_APIError(body, 400), provider="custom", model="m")
+ assert result.reason == FailoverReason.reasoning_mandatory
+ assert result.retryable is True and result.should_fallback is False
+
+
+ # The same enum wording about another field is not a reasoning rejection.
+ "Error code: 400 - 'tool_choice' must be one of none, auto, required",
+ # A paired-field constraint, not a closed level set.
+ "Error code: 400 - 'reasoning_effort' must be 'none' when 'enable_thinking' is false",
+])
+def test_other_enum_wording_keeps_its_classification(body):
+ result = classify_api_error(_APIError(body, 400), provider="custom", model="m")
+ assert result.reason != FailoverReason.reasoning_mandatory
+
+
[email protected]("body", [_OFF_LOW_HIGH_MAX_400, _GROQ_400])
+def test_rejected_enabled_level_retries_with_the_route_default(body):
+ agent = _WireAgent({"enabled": True, "effort": "medium"})
+ assert _wire_reasoning_config(agent) == {"enabled": True, "effort": "medium"} # the request that 400ed
+ _recover(agent, body)
+ assert _wire_reasoning_config(agent) is None, "retry must not resend reasoning_effort=medium"
+ assert any("rejects reasoning effort medium" in n for n in agent.notices)