pont: effort -> thinking_token_budget (memes seuils que le patch natif)

anthropic_proxy.py

@@ -329,6 +329,14 @@ def to_openai(body: dict) -> dict:
             out["thinking_token_budget"] = max(1, int(th["budget_tokens"]))
         except (TypeError, ValueError):
             pass
+    # Niveau d'effort (output_config.effort) -> plafond de jetons de raisonnement,
+    # memes seuils que le patch natif (vllm_anthropic_effort_patch.py). Le budget
+    # explicite ci-dessus gagne ; `disabled` aussi.
+    _eff = (body.get("output_config") or {}).get("effort") if isinstance(body.get("output_config"), dict) else None
+    _BUDGET = {"low": 1024, "medium": 4096, "high": 16384, "xhigh": 32768, "max": None}
+    if _eff in _BUDGET and "thinking_token_budget" not in out and "chat_template_kwargs" not in out:
+        if _BUDGET[_eff] is not None:
+            out["thinking_token_budget"] = _BUDGET[_eff]
 
     if body.get("tools"):
         out["tools"] = [{