bugfix: max_tokens hardcoded

MULTI_ITEM_FRAMEWORK_FIX_20260918.md

@@ -0,0 +1,29 @@
+# Multi-item result preservation fix
+
+## Objective
+Preserve the terminal outcome of each item executed by the primary transactional tool. A failure in a later auxiliary/composite step must not erase successful item outcomes or turn an already-completed multi-item operation into a total failure.
+
+## Framework-only change
+No agent/template business logic is required for this fix. The runtime now recognizes an explicit structural contract on the requested tool: a `results` array with two or more entries carrying boolean `success` or `ok` fields.
+
+For workflow-backed tools the authoritative source is preferentially:
+
+`workflow.output[tool_name].results`
+
+The framework derives one of:
+- `SUCCESS`
+- `PARTIAL_SUCCESS`
+- `FAILED`
+
+and publishes a generic `metadata.multi_item` summary with item counts.
+
+When one or more primary items completed successfully but the outer wrapper is `ok=false` because a later auxiliary/post-processing step failed, the normalized result preserves the original failure as `original_ok=false` / `secondary_error` while exposing the requested operation as operationally successful (`ok=true`).
+
+The LLM context also receives an explicit generic instruction that successful per-item terminal evidence cannot be overwritten by a secondary error and that mixed outcomes must be reported individually.
+
+## Regression tests
+Added `tests/test_multi_item_result_normalization.py` covering:
+- 2 success + 1 failure -> `PARTIAL_SUCCESS`;
+- all primary items successful + later secondary failure -> primary `SUCCESS` preserved;
+- single-item behavior unchanged;
+- unrelated nested tool results cannot override failure of the requested primary tool.

Tuning-Performance/Authentication/agent_template_backend_authentication/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

Tuning-Performance/Deterministic_Transactional_Workflow/agent_template_backend/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

Tuning-Performance/External_Guardrails_Judges/agent_template_backend/llm_profiles.yaml

@@ -9,11 +9,18 @@ profiles:
     model: openai.gpt-4.1
     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
   guardrail:
     provider: oci_openai
     model: openai.gpt-4.1

Tuning-Performance/Long_Term_Memory/templates/agent_template_backend/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

Tuning-Performance/Normal/templates/agent_template_backend/llm_profiles.yaml

@@ -9,11 +9,18 @@ profiles:
     model: openai.gpt-4.1
     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
   guardrail:
     provider: oci_openai
     model: openai.gpt-4.1

Tuning-Performance/Normal/templates/agent_template_backend_day_zero/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

Tuning-Performance/Route_Stickness/templates/agent_template_backend/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

Tuning-Performance/Route_Stickness/templates/agent_template_backend_day_zero/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

Tuning-Performance/Transaction_Evidence/agent_template_backend/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

Tuning-Performance/Transaction_Pre_Validation/agent_template_backend/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

docs/FIX_LLM_MAX_TOKENS_CONFIGURATION_20260918.md

@@ -0,0 +1,58 @@
+# Correção — `max_tokens` por `llm_profiles.yaml` / `.env`
+
+## Objetivo
+
+Remover `max_tokens` efetivos fixados nos pontos de inferência do framework sem perder compatibilidade com os budgets anteriores.
+
+A precedência usada pelos pontos corrigidos passa a ser:
+
+1. `max_tokens` do profile selecionado em `llm_profiles.yaml`;
+2. `max_tokens` do profile `default` do mesmo YAML, quando aplicável ao resolver;
+3. `LLM_MAX_TOKENS` configurado no ambiente / `.env`;
+4. valor legado do ponto de inferência, usado somente como fallback de compatibilidade.
+
+O parâmetro interno `fallback_max_tokens` **não é uma nova configuração do usuário**. Ele apenas preserva o valor histórico quando não existe configuração externa.
+
+## Pontos corrigidos
+
+- Enterprise Router:
+  - transaction parameter relevance: fallback legado `256`;
+  - transaction intent shift: fallback legado `512`;
+  - multi-intent: fallback legado `768`;
+  - routing normal: fallback legado `512`.
+- MCP parameter extraction: fallback `80`.
+- Processing interruption classifier: fallback `8`.
+- LLM guardrail: fallback `600`.
+- RAG query rewrite: fallback `300`.
+- RAG context compression: mantém a fórmula histórica `max(512, max_chars // 3)` apenas como fallback.
+- Conversation summary memory: mantém a fórmula histórica `max(256, max_summary_chars // 4)` apenas como fallback.
+
+## Atualização de `llm_profiles.yaml`
+
+Como os quatro pontos do Enterprise Router usam o mesmo profile `router`, o profile foi ajustado para o **maior budget histórico**, conforme solicitado:
+
+```yaml
+router:
+  max_tokens: 768
+```
+
+Também foram garantidos os profiles explícitos:
+
+```yaml
+processing_interruption_classifier:
+  max_tokens: 8
+
+mcp_parameter_extraction:
+  max_tokens: 80
+```
+
+Os profiles já existentes foram mantidos:
+
+- `guardrail: 600`
+- `rag_rewriter: 300`
+- `rag_compressor: 1200`
+- `summary_memory: 1200`
+
+## Compatibilidade
+
+Se `llm_profiles.yaml` não estiver disponível e `LLM_MAX_TOKENS` não estiver definido, cada ponto corrigido continua usando seu budget histórico. Portanto a correção não transforma a ausência de configuração em `2048` apenas por causa do default global de `Settings`.

evals/offline/configs/llm_profiles/llm_profiles.yaml

@@ -9,3 +9,11 @@ profiles:
     model: mock-judge
     temperature: 0
     max_tokens: 300
+
+  mcp_parameter_extraction:
+    provider: oci_openai
+    model: openai.gpt-4.1-mini
+    temperature: 0
+    max_tokens: 80
+    timeout_seconds: 5
+

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)

templates/agent_template_backend/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

templates/agent_template_backend_day_zero/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

tests/test_llm_max_tokens_configuration.py

@@ -0,0 +1,95 @@
+from __future__ import annotations
+
+import os
+import sys
+from pathlib import Path
+from types import SimpleNamespace
+
+import yaml
+
+ROOT = Path(__file__).resolve().parents[1]
+LIB = ROOT / "libs" / "agent_framework" / "src"
+if str(LIB) not in sys.path:
+    sys.path.insert(0, str(LIB))
+
+from agent_framework.llm.profile_resolver import LLMProfileResolver
+
+
+def _settings(path: Path):
+    return SimpleNamespace(
+        LLM_PROFILES_PATH=str(path),
+        LLM_PROVIDER="oci_openai",
+        OCI_GENAI_MODEL="openai.gpt-4.1",
+        LLM_TEMPERATURE=0.2,
+        LLM_MAX_TOKENS=2048,
+        LLM_TIMEOUT_SECONDS=120,
+        OCI_GENAI_BASE_URL="",
+        OCI_GENAI_API_KEY=None,
+        OCI_GENAI_PROJECT_OCID=None,
+        OCI_AUTH_MODE="config_file",
+        OCI_GENAI_ENDPOINT=None,
+        OCI_REGION="",
+        model_fields_set=set(),
+    )
+
+
+def test_profile_max_tokens_wins_over_env_and_legacy_fallback(tmp_path, monkeypatch):
+    profile = tmp_path / "llm_profiles.yaml"
+    profile.write_text(
+        "profiles:\n"
+        "  default:\n"
+        "    provider: oci_openai\n"
+        "  router:\n"
+        "    max_tokens: 768\n",
+        encoding="utf-8",
+    )
+    monkeypatch.setenv("LLM_MAX_TOKENS", "999")
+    resolver = LLMProfileResolver.from_settings(_settings(profile))
+    assert resolver.resolve_configured_value(
+        "router", "max_tokens", env_var="LLM_MAX_TOKENS", fallback=512
+    ) == 768
+
+
+def test_env_max_tokens_wins_when_profiles_file_is_absent(tmp_path, monkeypatch):
+    missing = tmp_path / "missing.yaml"
+    monkeypatch.setenv("LLM_MAX_TOKENS", "901")
+    resolver = LLMProfileResolver.from_settings(_settings(missing))
+    assert resolver.resolve_configured_value(
+        "router", "max_tokens", env_var="LLM_MAX_TOKENS", fallback=512
+    ) == 901
+
+
+def test_legacy_fallback_is_kept_when_yaml_and_env_are_absent(tmp_path, monkeypatch):
+    missing = tmp_path / "missing.yaml"
+    monkeypatch.delenv("LLM_MAX_TOKENS", raising=False)
+    resolver = LLMProfileResolver.from_settings(_settings(missing))
+    assert resolver.resolve_configured_value(
+        "router", "max_tokens", env_var="LLM_MAX_TOKENS", fallback=512
+    ) == 512
+
+
+def test_canonical_profiles_keep_existing_budgets_and_router_uses_largest_value():
+    profile_path = ROOT / "libs" / "agent_framework" / "config" / "llm_profiles.yaml"
+    profiles = yaml.safe_load(profile_path.read_text(encoding="utf-8"))["profiles"]
+    assert profiles["router"]["max_tokens"] == 768
+    assert profiles["processing_interruption_classifier"]["max_tokens"] == 8
+    assert profiles["mcp_parameter_extraction"]["max_tokens"] == 80
+    assert profiles["guardrail"]["max_tokens"] == 600
+    assert profiles["rag_rewriter"]["max_tokens"] == 300
+    assert profiles["rag_compressor"]["max_tokens"] == 1200
+    assert profiles["summary_memory"]["max_tokens"] == 1200
+
+
+def test_inference_points_use_fallback_max_tokens_instead_of_direct_max_tokens():
+    checks = {
+        "libs/agent_framework/src/agent_framework/routing/enterprise_router.py": [256, 512, 768],
+        "libs/agent_framework/src/agent_framework/runtime/agent_runtime.py": [80],
+        "libs/agent_framework/src/agent_framework/channels/interruption.py": [8],
+        "libs/agent_framework/src/agent_framework/guardrails/llm_rails.py": [600],
+        "libs/agent_framework/src/agent_framework/rag/rag_service.py": [300],
+    }
+    for relative, values in checks.items():
+        text = (ROOT / relative).read_text(encoding="utf-8")
+        for value in values:
+            assert f"fallback_max_tokens={value}" in text
+            assert f"max_tokens={value}," not in text.replace(f"fallback_max_tokens={value},", "")