feat_dev(guardrail): externalize labels and tighten censorship logic

config.yaml

@@ -13,6 +13,7 @@ mcp_audience: "https://ap01194-orq-cog-rag-connector-1007577023101.us-central1.r
 
 agent_name: VAia
 agent_model: gemini-2.5-flash
+
 agent_instructions: |
     Eres VAia, el asistente virtual de VA en WhatsApp. VA es la opción digital de Banorte para los jóvenes. Fuiste creado por el equipo de inteligencia artifical de Banorte. Tu rol es resolver dudas sobre educación financiera y los productos/servicios de VA. Hablas como un amigo que sabe de finanzas: siempre vas directo al grano, con calidez y sin rodeos.
 
@@ -50,6 +51,13 @@ agent_instructions: |
 
     El teléfono de centro de contacto de VA es: +52 1 55 5140 5655
 
+# Guardrail config
+guardrail_censored_user_message: "[pregunta mala]"
+guardrail_censored_model_response: "[respuesta de adversidad]"
+guardrail_blocked_label: "[GUARDRAIL_BLOCKED]"
+guardrail_passed_label: "[GUARDRAIL_PASSED]"
+guardrail_error_label: "[GUARDRAIL_ERROR]"
+
 guardrail_instruction: |
     Eres una capa de seguridad y protección de marca para VAia, el asistente virtual de VA en WhatsApp.
     VAia es un asistente de educación financiera y productos/servicios de VA (la opción digital de Banorte para jóvenes)

src/va_agent/config.py

@@ -21,9 +21,16 @@ class AgentSettings(BaseSettings):
 
     # Agent configuration
     agent_name: str
-    agent_instructions: str
-    guardrail_instruction: str
     agent_model: str
+    agent_instructions: str
+
+    # Guardrail configuration
+    guardrail_censored_user_message: str
+    guardrail_censored_model_response: str
+    guardrail_blocked_label: str
+    guardrail_passed_label: str
+    guardrail_error_label: str
+    guardrail_instruction: str
 
     # Firestore configuration
     firestore_db: str

src/va_agent/governance.py

@@ -171,19 +171,19 @@ class GovernancePlugin:
 
             if decision == "unsafe":
                 callback_context.state["guardrail_blocked"] = True
-                callback_context.state["guardrail_message"] = "[GUARDRAIL_BLOCKED]"
+                callback_context.state["guardrail_message"] = settings.guardrail_blocked_label
                 callback_context.state["guardrail_reasoning"] = reasoning
                 return LlmResponse(
                     content=Content(role="model", parts=[Part(text=blocking_response)]),
                     usage_metadata=resp.usage_metadata or None,
                 )
             callback_context.state["guardrail_blocked"] = False
-            callback_context.state["guardrail_message"] = "[GUARDRAIL_PASSED]"
+            callback_context.state["guardrail_message"] = settings.guardrail_passed_label
             callback_context.state["guardrail_reasoning"] = reasoning
 
         except Exception:
             # Fail safe: block with a generic error response and mark the reason
-            callback_context.state["guardrail_message"] = "[GUARDRAIL_ERROR]"
+            callback_context.state["guardrail_message"] = settings.guardrail_error_label
             logger.exception("Guardrail check failed")
             return LlmResponse(
                 content=Content(

src/va_agent/session.py

@@ -25,12 +25,13 @@ from google.cloud.firestore_v1.field_path import FieldPath
 from google.genai.types import Content, Part
 
 from .compaction import SessionCompactor
+from .config import settings
 
 if TYPE_CHECKING:
     from google import genai
     from google.cloud.firestore_v1.async_client import AsyncClient
 
-logger = logging.getLogger("google_adk." + __name__)
+logger = logging.getLogger(__name__)
 
 
 class FirestoreSessionService(BaseSessionService):
@@ -382,7 +383,7 @@ class FirestoreSessionService(BaseSessionService):
         # Determine if we need to censor this event (model response when guardrail blocked)
         should_censor_model = (
             session.state.get("guardrail_blocked", False)
-            and event.author == app_name
+            and event.author != "user"
             and hasattr(event, "content")
             and event.content
             and event.content.parts
@@ -396,34 +397,36 @@ class FirestoreSessionService(BaseSessionService):
 
             # Create a censored version of the model response
             event_to_save = copy.deepcopy(event)
-            event_to_save.content.parts[0].text = "[respuesta de adversidad]"
+            event_to_save.content.parts[0].text = settings.guardrail_censored_model_response
             event_data = event_to_save.model_dump(mode="json", exclude_none=True)
 
             # Also censor the previous user message in Firestore
-            # Find the last user event in the session (skip the current model event we just added)
+            # Find the last user event in the session
-            for i in range(len(session.events) - 2, -1, -1):
+            prev_user_event = next(
-                prev_event = session.events[i]
+                (
-                if (
+                    e
-                    prev_event.author == "user"
+                    for e in reversed(session.events[:-1])
-                    and prev_event.content
+                    if e.author == "user" and e.content and e.content.parts
-                    and prev_event.content.parts
+                ),
-                ):
+                None,
-                    # Update this event in Firestore with censored content
+            )
-                    censored_user_content = Content(
+            if prev_user_event:
-                        role="user", parts=[Part(text="[pregunta mala]")]
+                # Update this event in Firestore with censored content
+                censored_user_content = Content(
+                    role="user",
+                    parts=[Part(text=settings.guardrail_censored_user_message)],
+                )
+                await (
+                    self._events_col(app_name, user_id, session_id)
+                    .document(prev_user_event.id)
+                    .update(
+                        {
+                            "content": censored_user_content.model_dump(
+                                mode="json", exclude_none=True
+                            )
+                        }
                     )
-                    await (
+                )
-                        self._events_col(app_name, user_id, session_id)
-                        .document(prev_event.id)
-                        .update(
-                            {
-                                "content": censored_user_content.model_dump(
-                                    mode="json", exclude_none=True
-                                )
-                            }
-                        )
-                    )
-                    break
         else:
             event_data = event.model_dump(mode="json", exclude_none=True)