fix(nebius): route effort through canonical clamp_effort — hand-rolled map inverted the ladder

Review finding on salvaged #28253: the hand-rolled mapping sent
ultra -> medium while xhigh -> high (stronger request, weaker wire
value). Declare NEBIUS_EFFORTS in agent/reasoning_effort.py and use
clamp_effort like the zai/kimi/tokenhub call sites; disable detection
stays ahead of the clamp since clamp_effort('none', ...) returns the
floor, not off. Adds a monotonicity regression test.

agent/reasoning_effort.py

@@ -105,6 +105,9 @@ OX_ALPHA_OVERRIDES: dict[str, str] = {"xhigh": "max"}
 #: Tencent TokenHub: low/medium/high.
 TOKENHUB_EFFORTS: tuple[str, ...] = ("low", "medium", "high")
 
+#: Nebius Token Factory: low/medium/high (top-level reasoning_effort knob).
+NEBIUS_EFFORTS: tuple[str, ...] = ("low", "medium", "high")
+
 #: Kimi K3's vendor-documented translation quirks (platform.kimi.ai
 #: thinking-model guide): ``high`` is K3's positional middle AND server
 #: default, so ``medium`` rounds to it rather than down to ``low``; ``xhigh``

plugins/model-providers/nebius-token-factory/__init__.py

@@ -57,12 +57,12 @@ class NebiusTokenFactoryProfile(ProviderProfile):
         effort = str(raw_effort or "medium").strip().lower()
         if enabled is False or effort in {"none", "off", "disabled"}:
             return {}, {}
-        if effort in {"xhigh", "max"}:
-            effort = "high"
-        elif effort == "minimal":
-            effort = "low"
-        elif effort not in {"low", "medium", "high"}:
-            effort = "medium"
+        # Canonical clamp (nearest weaker supported level, never escalate,
+        # monotonic) — the hand-rolled map this replaces inverted the ladder:
+        # ultra fell through to medium while xhigh mapped to high.
+        from agent.reasoning_effort import NEBIUS_EFFORTS, clamp_effort
+
+        effort = clamp_effort(effort, NEBIUS_EFFORTS) or "medium"
 
         return {}, {"reasoning_effort": effort}
 

tests/hermes_cli/test_nebius_token_factory_provider.py

@@ -172,6 +172,28 @@ def test_nebius_reasoning_models_emit_top_level_reasoning_effort():
     assert top_level == {"reasoning_effort": "high"}
 
 
+def test_nebius_effort_clamp_is_monotonic():
+    """Regression: the hand-rolled map sent ultra->medium while xhigh->high,
+    inverting the ladder. The canonical clamp_effort keeps stronger requests
+    at least as strong on the wire (all of xhigh/max/ultra clamp to high)."""
+    from providers import get_provider_profile
+
+    profile = get_provider_profile("nebius-token-factory")
+    assert profile is not None
+
+    def wire(effort):
+        _, top = profile.build_api_kwargs_extras(
+            reasoning_config={"enabled": True, "effort": effort},
+            model="deepseek-ai/DeepSeek-V4-Pro",
+        )
+        return top.get("reasoning_effort")
+
+    assert wire("ultra") == "high"
+    assert wire("max") == "high"
+    assert wire("xhigh") == "high"
+    assert wire("minimal") == "low"
+
+
 def test_nebius_reasoning_defaults_to_medium_for_known_reasoning_model():
     from providers import get_provider_profile