bugfix: max_tokens hardcoded

libs/agent_framework/build/lib/agent_framework/channels/interruption.py

@@ -58,7 +58,7 @@ async def classify_processing_interruption(
         response = await llm.ainvoke(
             [{"role": "system", "content": prompt}],
             temperature=0,
-            max_tokens=8,
+            fallback_max_tokens=8,
             profile_name=profile_name,
             component_name=profile_name,
             generation_name=f"llm.{profile_name}",

libs/agent_framework/build/lib/agent_framework/guardrails/llm_rails.py

@@ -47,7 +47,7 @@ class LLMGuardrailRail(Guardrail):
                     {"role": "user", "content": prompt},
                 ],
                 temperature=0,
-                max_tokens=600,
+                fallback_max_tokens=600,
                 profile_name=self.profile_name,
                 component_name=self.profile_name,
                 generation_name=f"llm.{self.profile_name}",

libs/agent_framework/build/lib/agent_framework/llm/profile_resolver.py

@@ -2,6 +2,7 @@ from __future__ import annotations
 
 import copy
 import logging
+import os
 import re
 from pathlib import Path
 from typing import Any
@@ -164,6 +165,66 @@ class LLMProfileResolver:
 
         return effective
 
+    def resolve_configured_value(
+        self,
+        profile_name: str | None,
+        key: str,
+        *,
+        env_var: str | None = None,
+        fallback: Any = None,
+    ) -> Any:
+        """Resolve one configuration value with an explicit compatibility fallback.
+
+        Precedence is intentionally aligned with the framework configuration model:
+        1. specific ``llm_profiles.yaml`` profile;
+        2. YAML ``default`` profile;
+        3. explicitly configured general environment variable;
+        4. the caller-provided legacy fallback.
+
+        The last step is important for old inference points that historically had a
+        local constant.  Pydantic's Settings object always has framework defaults, so
+        checking only ``settings.LLM_MAX_TOKENS`` would make an *unset* environment
+        indistinguishable from the global default.
+        """
+        selected_name = self.normalize_profile_name(profile_name)
+        if self.enabled:
+            specific_profile = self._profiles.get(selected_name) or {}
+            default_profile = self._profiles.get("default") or {}
+            if key in specific_profile and specific_profile.get(key) is not None:
+                return specific_profile.get(key)
+            if key in default_profile and default_profile.get(key) is not None:
+                return default_profile.get(key)
+
+        if env_var:
+            raw = os.getenv(env_var)
+            if raw is not None and str(raw).strip() != "":
+                raw = str(raw).strip()
+                try:
+                    if isinstance(fallback, bool):
+                        return raw.lower() in {"1", "true", "yes", "on"}
+                    if isinstance(fallback, int) and not isinstance(fallback, bool):
+                        return int(raw)
+                    if isinstance(fallback, float):
+                        return float(raw)
+                except (TypeError, ValueError):
+                    logger.warning(
+                        "Invalid %s=%r for profile=%s key=%s; using legacy fallback=%r",
+                        env_var, raw, selected_name, key, fallback,
+                    )
+                    return fallback
+                return raw
+
+        # Also honor a value explicitly supplied when Settings was instantiated
+        # programmatically, without treating the class default as configuration.
+        field_name = env_var or ""
+        fields_set = getattr(self.settings, "model_fields_set", set()) or set()
+        if field_name and field_name in fields_set:
+            value = getattr(self.settings, field_name, None)
+            if value is not None:
+                return value
+
+        return fallback
+
     def normalize_profile_name(self, profile_name: str | None) -> str:
         return _canonical_profile_name(profile_name)
 

libs/agent_framework/build/lib/agent_framework/llm/providers.py

@@ -370,6 +370,14 @@ class OCICompatibleOpenAIProvider(LLMProvider):
         profile_name = kwargs.pop("profile_name", None)
         component_name = kwargs.pop("component_name", None) or kwargs.pop("component", None) or profile_name or "default"
         generation_name = kwargs.pop("generation_name", None) or f"llm.{component_name}"
+        fallback_max_tokens = kwargs.pop("fallback_max_tokens", None)
+        if fallback_max_tokens is not None:
+            kwargs["max_tokens"] = self.profile_resolver.resolve_configured_value(
+                profile_name,
+                "max_tokens",
+                env_var="LLM_MAX_TOKENS",
+                fallback=fallback_max_tokens,
+            )
         generation_name, generation_mapping_meta = _normalize_generation_name(self.telemetry, generation_name)
         effective = self.profile_resolver.resolve(profile_name, **kwargs)
         provider = str(effective.get("provider") or self.provider_name)

libs/agent_framework/build/lib/agent_framework/memory/summary_memory.py

@@ -167,7 +167,7 @@ class ConversationSummaryMemory:
             summary = await self.llm.ainvoke([
                 {"role": "system", "content": "Você resume memória conversacional para agentes corporativos."},
                 {"role": "user", "content": prompt},
-            ], max_tokens=max(256, max_summary_chars // 4), temperature=0.1, profile_name="summary_memory", component_name="summary_memory", generation_name="llm.summary_memory")
+            ], fallback_max_tokens=max(256, max_summary_chars // 4), temperature=0.1, profile_name="summary_memory", component_name="summary_memory", generation_name="llm.summary_memory")
             summary = (summary or "").strip()
             if not summary:
                 return self._deterministic_summary(previous_summary=previous_summary, messages=messages, max_chars=max_summary_chars)

libs/agent_framework/build/lib/agent_framework/rag/rag_service.py

@@ -84,7 +84,7 @@ class RagService:
                     {'role': 'user', 'content': prompt},
                 ],
                 temperature=0,
-                max_tokens=300,
+                fallback_max_tokens=300,
                 profile_name=profile_name,
                 component_name=profile_name,
                 generation_name=f"llm.{profile_name}",
@@ -112,7 +112,7 @@ class RagService:
                     {'role': 'user', 'content': prompt},
                 ],
                 temperature=0,
-                max_tokens=max(512, max_chars // 3),
+                fallback_max_tokens=max(512, max_chars // 3),
                 profile_name=profile_name,
                 component_name=profile_name,
                 generation_name=f"llm.{profile_name}",

libs/agent_framework/build/lib/agent_framework/routing/enterprise_router.py

@@ -1045,7 +1045,7 @@ class EnterpriseRouter:
                     {"role": "user", "content": json.dumps(payload, ensure_ascii=False)},
                 ],
                 temperature=0.0,
-                max_tokens=256,
+                fallback_max_tokens=256,
                 profile_name="router",
                 component_name="router.transaction_parameter_relevance",
                 generation_name="llm.transaction_parameter_relevance",
@@ -1328,7 +1328,7 @@ class EnterpriseRouter:
                     {"role": "user", "content": json.dumps(user, ensure_ascii=False)},
                 ],
                 temperature=0.0,
-                max_tokens=512,
+                fallback_max_tokens=512,
                 profile_name="router",
                 component_name="router",
                 generation_name="llm.transaction_intent_shift",
@@ -1674,7 +1674,7 @@ class EnterpriseRouter:
                 {"role": "user", "content": json.dumps(payload, ensure_ascii=False)},
             ],
             temperature=0.0,
-            max_tokens=768,
+            fallback_max_tokens=768,
             profile_name="router",
             component_name="router.multi_intent",
             generation_name="llm.router.multi_intent",
@@ -1730,7 +1730,7 @@ class EnterpriseRouter:
                 {"role": "user", "content": json.dumps(user, ensure_ascii=False)},
             ],
             temperature=0.0,
-            max_tokens=512,
+            fallback_max_tokens=512,
             profile_name="router",
             component_name="router",
             generation_name="llm.router",

libs/agent_framework/build/lib/agent_framework/runtime/agent_runtime.py

@@ -849,7 +849,7 @@ class AgentRuntimeMixin:
                         component_name="mcp_parameter_extraction",
                         generation_name="llm.mcp_parameter_extraction",
                         temperature=0.0,
-                        max_tokens=80,
+                        fallback_max_tokens=80,
                     )
                     raw = self._llm_response_text(response).strip()
                     payload = parse_json_object(raw)

libs/agent_framework/config/llm_profiles.yaml

@@ -15,11 +15,17 @@ profiles:
     temperature: 0
     max_tokens: 80
     timeout_seconds: 5
+  processing_interruption_classifier:
+    provider: oci_openai
+    model: openai.gpt-4.1-mini
+    temperature: 0
+    max_tokens: 8
+    timeout_seconds: 5
   router:
     provider: oci_openai
     model: openai.gpt-4.1
     temperature: 0
-    max_tokens: 500
+    max_tokens: 768
   guardrail:
     provider: oci_openai
     model: openai.gpt-4.1
@@ -72,6 +78,13 @@ profiles:
     provider: oci_openai
     model: openai.gpt-4.1
     temperature: 0.2
+  mcp_parameter_extraction:
+    provider: oci_openai
+    model: openai.gpt-4.1-mini
+    temperature: 0
+    max_tokens: 80
+    timeout_seconds: 5
+
   transaction_parameter_extraction:
     provider: oci_openai
     model: openai.gpt-4.1-mini

libs/agent_framework/llm_profiles.yaml.example

@@ -15,11 +15,18 @@ profiles:
     temperature: 0
     max_tokens: 700
 
+  processing_interruption_classifier:
+    provider: oci_openai
+    model: openai.gpt-4.1-mini
+    temperature: 0
+    max_tokens: 8
+    timeout_seconds: 5
+
   router:
     provider: oci_openai
     model: openai.gpt-4.1
     temperature: 0
-    max_tokens: 500
+    max_tokens: 768
 
   # Safety / evaluation
   guardrail:
@@ -87,3 +94,11 @@ profiles:
     provider: oci_openai
     model: openai.gpt-4.1
     temperature: 0.2
+
+  mcp_parameter_extraction:
+    provider: oci_openai
+    model: openai.gpt-4.1-mini
+    temperature: 0
+    max_tokens: 80
+    timeout_seconds: 5
+

libs/agent_framework/src/agent_framework/channels/interruption.py

@@ -58,7 +58,7 @@ async def classify_processing_interruption(
         response = await llm.ainvoke(
             [{"role": "system", "content": prompt}],
             temperature=0,
-            max_tokens=8,
+            fallback_max_tokens=8,
             profile_name=profile_name,
             component_name=profile_name,
             generation_name=f"llm.{profile_name}",

libs/agent_framework/src/agent_framework/guardrails/llm_rails.py

@@ -49,7 +49,7 @@ class LLMGuardrailRail(Guardrail):
                     {"role": "user", "content": prompt},
                 ],
                 temperature=0,
-                max_tokens=600,
+                fallback_max_tokens=600,
                 profile_name=self.profile_name,
                 component_name=self.profile_name,
                 generation_name=f"llm.{self.profile_name}",

libs/agent_framework/src/agent_framework/llm/profile_resolver.py

@@ -2,6 +2,7 @@ from __future__ import annotations
 
 import copy
 import logging
+import os
 import re
 from pathlib import Path
 from typing import Any
@@ -164,6 +165,66 @@ class LLMProfileResolver:
 
         return effective
 
+    def resolve_configured_value(
+        self,
+        profile_name: str | None,
+        key: str,
+        *,
+        env_var: str | None = None,
+        fallback: Any = None,
+    ) -> Any:
+        """Resolve one configuration value with an explicit compatibility fallback.
+
+        Precedence is intentionally aligned with the framework configuration model:
+        1. specific ``llm_profiles.yaml`` profile;
+        2. YAML ``default`` profile;
+        3. explicitly configured general environment variable;
+        4. the caller-provided legacy fallback.
+
+        The last step is important for old inference points that historically had a
+        local constant.  Pydantic's Settings object always has framework defaults, so
+        checking only ``settings.LLM_MAX_TOKENS`` would make an *unset* environment
+        indistinguishable from the global default.
+        """
+        selected_name = self.normalize_profile_name(profile_name)
+        if self.enabled:
+            specific_profile = self._profiles.get(selected_name) or {}
+            default_profile = self._profiles.get("default") or {}
+            if key in specific_profile and specific_profile.get(key) is not None:
+                return specific_profile.get(key)
+            if key in default_profile and default_profile.get(key) is not None:
+                return default_profile.get(key)
+
+        if env_var:
+            raw = os.getenv(env_var)
+            if raw is not None and str(raw).strip() != "":
+                raw = str(raw).strip()
+                try:
+                    if isinstance(fallback, bool):
+                        return raw.lower() in {"1", "true", "yes", "on"}
+                    if isinstance(fallback, int) and not isinstance(fallback, bool):
+                        return int(raw)
+                    if isinstance(fallback, float):
+                        return float(raw)
+                except (TypeError, ValueError):
+                    logger.warning(
+                        "Invalid %s=%r for profile=%s key=%s; using legacy fallback=%r",
+                        env_var, raw, selected_name, key, fallback,
+                    )
+                    return fallback
+                return raw
+
+        # Also honor a value explicitly supplied when Settings was instantiated
+        # programmatically, without treating the class default as configuration.
+        field_name = env_var or ""
+        fields_set = getattr(self.settings, "model_fields_set", set()) or set()
+        if field_name and field_name in fields_set:
+            value = getattr(self.settings, field_name, None)
+            if value is not None:
+                return value
+
+        return fallback
+
     def normalize_profile_name(self, profile_name: str | None) -> str:
         return _canonical_profile_name(profile_name)
 

libs/agent_framework/src/agent_framework/llm/providers.py

@@ -370,6 +370,14 @@ class OCICompatibleOpenAIProvider(LLMProvider):
         profile_name = kwargs.pop("profile_name", None)
         component_name = kwargs.pop("component_name", None) or kwargs.pop("component", None) or profile_name or "default"
         generation_name = kwargs.pop("generation_name", None) or f"llm.{component_name}"
+        fallback_max_tokens = kwargs.pop("fallback_max_tokens", None)
+        if fallback_max_tokens is not None:
+            kwargs["max_tokens"] = self.profile_resolver.resolve_configured_value(
+                profile_name,
+                "max_tokens",
+                env_var="LLM_MAX_TOKENS",
+                fallback=fallback_max_tokens,
+            )
         generation_name, generation_mapping_meta = _normalize_generation_name(self.telemetry, generation_name)
         effective = self.profile_resolver.resolve(profile_name, **kwargs)
         provider = str(effective.get("provider") or self.provider_name)

libs/agent_framework/src/agent_framework/memory/summary_memory.py

@@ -167,7 +167,7 @@ class ConversationSummaryMemory:
             summary = await self.llm.ainvoke([
                 {"role": "system", "content": "Você resume memória conversacional para agentes corporativos."},
                 {"role": "user", "content": prompt},
-            ], max_tokens=max(256, max_summary_chars // 4), temperature=0.1, profile_name="summary_memory", component_name="summary_memory", generation_name="llm.summary_memory")
+            ], fallback_max_tokens=max(256, max_summary_chars // 4), temperature=0.1, profile_name="summary_memory", component_name="summary_memory", generation_name="llm.summary_memory")
             summary = (summary or "").strip()
             if not summary:
                 return self._deterministic_summary(previous_summary=previous_summary, messages=messages, max_chars=max_summary_chars)

libs/agent_framework/src/agent_framework/rag/rag_service.py

@@ -98,7 +98,7 @@ class RagService:
                     {'role': 'user', 'content': prompt},
                 ],
                 temperature=0,
-                max_tokens=300,
+                fallback_max_tokens=300,
                 profile_name=profile_name,
                 component_name=profile_name,
                 generation_name=f"llm.{profile_name}",
@@ -126,7 +126,7 @@ class RagService:
                     {'role': 'user', 'content': prompt},
                 ],
                 temperature=0,
-                max_tokens=max(512, max_chars // 3),
+                fallback_max_tokens=max(512, max_chars // 3),
                 profile_name=profile_name,
                 component_name=profile_name,
                 generation_name=f"llm.{profile_name}",

libs/agent_framework/src/agent_framework/routing/enterprise_router.py

@@ -1045,7 +1045,7 @@ class EnterpriseRouter:
                     {"role": "user", "content": json.dumps(payload, ensure_ascii=False)},
                 ],
                 temperature=0.0,
-                max_tokens=256,
+                fallback_max_tokens=256,
                 profile_name="router",
                 component_name="router.transaction_parameter_relevance",
                 generation_name="llm.transaction_parameter_relevance",
@@ -1328,7 +1328,7 @@ class EnterpriseRouter:
                     {"role": "user", "content": json.dumps(user, ensure_ascii=False)},
                 ],
                 temperature=0.0,
-                max_tokens=512,
+                fallback_max_tokens=512,
                 profile_name="router",
                 component_name="router",
                 generation_name="llm.transaction_intent_shift",
@@ -1674,7 +1674,7 @@ class EnterpriseRouter:
                 {"role": "user", "content": json.dumps(payload, ensure_ascii=False)},
             ],
             temperature=0.0,
-            max_tokens=768,
+            fallback_max_tokens=768,
             profile_name="router",
             component_name="router.multi_intent",
             generation_name="llm.router.multi_intent",
@@ -1730,7 +1730,7 @@ class EnterpriseRouter:
                 {"role": "user", "content": json.dumps(user, ensure_ascii=False)},
             ],
             temperature=0.0,
-            max_tokens=512,
+            fallback_max_tokens=512,
             profile_name="router",
             component_name="router",
             generation_name="llm.router",

libs/agent_framework/src/agent_framework/runtime/agent_runtime.py

@@ -849,7 +849,7 @@ class AgentRuntimeMixin:
                         component_name="mcp_parameter_extraction",
                         generation_name="llm.mcp_parameter_extraction",
                         temperature=0.0,
-                        max_tokens=80,
+                        fallback_max_tokens=80,
                     )
                     raw = self._llm_response_text(response).strip()
                     payload = parse_json_object(raw)