iatoolkit 2.56.0__tar.gz → 2.57.0__tar.gz

{iatoolkit-2.56.0/src/iatoolkit.egg-info → iatoolkit-2.57.0}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: iatoolkit
-Version: 2.56.0
+Version: 2.57.0
 Summary: IAToolkit
 Author: Fernando Libedinsky
 License-Expression: MIT

{iatoolkit-2.56.0 → iatoolkit-2.57.0}/src/iatoolkit/__init__.py

@@ -3,7 +3,7 @@
 #
 # IAToolkit is open source software.
 
-__version__ = "2.56.0"
+__version__ = "2.57.0"
 
 # Expose main classes and functions at the top level of the package
 

{iatoolkit-2.56.0 → iatoolkit-2.57.0}/src/iatoolkit/common/routes.py

@@ -5,6 +5,7 @@
 
 from io import BytesIO
 import mimetypes
+import os
 from flask import render_template, redirect, url_for, current_app, abort, send_file
 from flask import jsonify
 from iatoolkit.common.exceptions import IAToolkitException
@@ -59,6 +60,18 @@ def register_views(app):
     # must stay reachable even if the company registry or DB is degraded.
     app.add_url_rule('/', endpoint='liveness', view_func=lambda: ("ok", 200))
 
+    # Browsers and MCP clients (Claude shows it as the connector icon) ask the
+    # host for /favicon.ico regardless of the page.
+    app.add_url_rule(
+        '/favicon.ico',
+        endpoint='favicon',
+        view_func=lambda: send_file(
+            os.path.join(os.path.dirname(os.path.dirname(os.path.abspath(__file__))), 'static', 'images', 'favicon.ico'),
+            mimetype='image/vnd.microsoft.icon',
+            max_age=86400,
+        ),
+    )
+
     # assign root '/' to our new redirect logic
     app.add_url_rule('/home', view_func=RootRedirectView.as_view('root_redirect'))
 

{iatoolkit-2.56.0 → iatoolkit-2.57.0}/src/iatoolkit/infra/llm_providers/gemini_adapter.py

@@ -700,23 +700,34 @@ class GeminiAdapter:
         return {}
 
     def _extract_usage_metadata(self, gemini_response) -> Usage:
-        """Extraer información de uso de tokens de manera segura"""
+        """Extraer información de uso de tokens de manera segura.
+
+        Gemini reports the tokens it bills in four counters. `candidates_token_count`
+        excludes the model's reasoning (`thoughts_token_count`), which Google bills as
+        output, and `prompt_token_count` excludes the results of its built-in tools
+        (`tool_use_prompt_token_count`), billed as input. Reading only the first two
+        priced every thinking model's reasoning at zero.
+        """
         input_tokens = 0
         output_tokens = 0
         total_tokens = 0
 
+        def _count(metadata, name: str) -> int:
+            value = getattr(metadata, name, None)
+            return value if isinstance(value, int) and value > 0 else 0
+
         try:
-            # Verificar si existe usage_metadata
-            if hasattr(gemini_response, 'usage_metadata') and gemini_response.usage_metadata:
-                usage_metadata = gemini_response.usage_metadata
-
-                # Acceder a los atributos directamente, no con .get()
-                if hasattr(usage_metadata, 'prompt_token_count'):
-                    input_tokens = usage_metadata.prompt_token_count
-                if hasattr(usage_metadata, 'candidates_token_count'):
-                    output_tokens = usage_metadata.candidates_token_count
-                if hasattr(usage_metadata, 'total_token_count'):
-                    total_tokens = usage_metadata.total_token_count
+            usage_metadata = getattr(gemini_response, 'usage_metadata', None)
+            if usage_metadata:
+                input_tokens = (
+                    _count(usage_metadata, 'prompt_token_count')
+                    + _count(usage_metadata, 'tool_use_prompt_token_count')
+                )
+                output_tokens = (
+                    _count(usage_metadata, 'candidates_token_count')
+                    + _count(usage_metadata, 'thoughts_token_count')
+                )
+                total_tokens = _count(usage_metadata, 'total_token_count')
 
         except Exception as e:
             logging.warning(f"No se pudo extraer usage_metadata de Gemini: {e}")

{iatoolkit-2.56.0 → iatoolkit-2.57.0}/src/iatoolkit/repositories/llm_query_repo.py

@@ -14,6 +14,7 @@ from iatoolkit.repositories.models import (
 from injector import inject
 from iatoolkit.repositories.database_manager import DatabaseManager
 from sqlalchemy import or_, and_
+from iatoolkit.repositories.models import CONTEXT_INIT_REQUEST_SOURCE
 from typing import List
 
 
@@ -34,10 +35,17 @@ class LLMQueryRepo:
         self.session.commit()
         return query
 
+    @staticmethod
+    def not_context_init():
+        """Filter for rows that are a conversation turn, not a context initialization."""
+        request_source = LLMQuery.stats["request_source"].as_string()
+        return or_(request_source.is_(None), request_source != CONTEXT_INIT_REQUEST_SOURCE)
+
     # get user query history
     def get_history(self, company: Company, user_identifier: str, limit: int = 100) -> list[LLMQuery]:
         return self.session.query(LLMQuery).filter(
             LLMQuery.user_identifier == user_identifier,
+            self.not_context_init(),
         ).filter_by(company_id=company.id).order_by(LLMQuery.created_at.desc()).limit(limit).all()
 
 

{iatoolkit-2.56.0 → iatoolkit-2.57.0}/src/iatoolkit/repositories/models.py

@@ -593,6 +593,12 @@ class MemoryPageSource(Base):
         return {column.key: getattr(self, column.key) for column in class_mapper(self.__class__).columns}
 
 
+#: request_source of the iat_queries row that records a context initialization:
+#: the company context sent once to start a server-side conversation. Its tokens
+#: are billed, but it is not a question, so conversation readers skip it.
+CONTEXT_INIT_REQUEST_SOURCE = "context_init"
+
+
 class LLMQuery(Base):
     """Logs a query made to the LLM, including input, output, and metadata."""
     __tablename__ = 'iat_queries'
@@ -622,6 +628,10 @@ class LLMQuery(Base):
 
     __table_args__ = (
         Index('ix_iat_queries_company_prompt_created', 'company_id', 'prompt_name', 'created_at'),
+        # Executions that did not run as a task (synchronous agent calls). They are a
+        # tiny fraction of the table, which the Agents monitor would otherwise scan whole.
+        Index('ix_iat_queries_company_created_untasked', company_id, created_at.desc(),
+              postgresql_where=task_id.is_(None)),
     )
 
     company = relationship("Company", back_populates="llm_queries")

{iatoolkit-2.56.0 → iatoolkit-2.57.0}/src/iatoolkit/services/history_manager_service.py

@@ -60,7 +60,8 @@ class HistoryManagerService:
             response_id = self.llm_client.set_company_context(
                 company=company,
                 company_base_context=prepared_context,
-                model=model
+                model=model,
+                user_identifier=user_identifier,
             )
             self.session_context.save_last_response_id(company_short_name, user_identifier, response_id, model=model)
             self.session_context.save_initial_response_id(company_short_name, user_identifier, response_id, model=model)