chore: sync main with staging

.env.example

@@ -38,9 +38,9 @@ GOOGLE_GCS_BUCKET=bucket-name
 # ============================================================
 # ==                      LLM settings                      ==
 # ============================================================
-MODEL_URI=google_genai:gemini-3-flash-preview
-MODEL_TEMPERATURE=0.2
-THINKING_LEVEL=low
+OPENAI_API_KEY=openai-api-key
+MODEL_URI=gpt-5.6-luna
+REASONING_EFFORT=medium
 
 # ============================================================
 # ==                   LangSmith settings                   ==

.github/workflows/test-chatbot.yaml

@@ -47,9 +47,9 @@ jobs:
           GOOGLE_GCS_BUCKET: mock-bucket
 
           # Mock LLM configuration
+          OPENAI_API_KEY: mock-openai-api-key
           MODEL_URI: mock-model-uri
-          MODEL_TEMPERATURE: 0.0
-          THINKING_LEVEL: low
+          REASONING_EFFORT: medium
 
           # Mock LangSmith configuration
           LANGSMITH_TRACING: false

alembic/versions/4b3d2fa4a75f_add_thread_language.py

@@ -0,0 +1,32 @@
+"""Add language column to threads table.
+
+Revision ID: 4b3d2fa4a75f
+Revises: f6ce7837e023
+Create Date: 2026-08-04 19:50:02.000000
+"""
+
+from typing import Sequence, Union
+
+import sqlalchemy as sa
+from alembic import op
+
+# revision identifiers, used by Alembic.
+revision: str = "4b3d2fa4a75f"
+down_revision: Union[str, Sequence[str], None] = "f6ce7837e023"
+branch_labels: Union[str, Sequence[str], None] = None
+depends_on: Union[str, Sequence[str], None] = None
+
+
+def upgrade() -> None:
+    """Upgrade schema."""
+    # server_default backfills existing threads with the Portuguese default;
+    # new rows get their value from the application (ThreadPayload.language).
+    op.add_column(
+        "thread",
+        sa.Column("language", sa.String(), nullable=False, server_default="pt"),
+    )
+
+
+def downgrade() -> None:
+    """Downgrade schema."""
+    op.drop_column("thread", "language")

app/agent/context.py

@@ -0,0 +1,20 @@
+from dataclasses import dataclass
+
+from app.i18n import LanguageCode
+
+
+@dataclass
+class AgentContext:
+    """Per-run context for an agent run.
+
+    Read by:
+      - The system-prompt middleware (`language`)
+      - The tools (`language` for localized metadata; `thread_id` and `user_id` for BigQuery job labels).
+
+    `thread_id` is *also* kept in `config["configurable"]` because the langgraph
+    checkpointer keys persistence on it there; on this context it is app metadata.
+    """
+
+    thread_id: str
+    user_id: str
+    language: LanguageCode

app/agent/middleware.py

@@ -0,0 +1,9 @@
+from datetime import date
+
+from langchain.agents.middleware import ModelRequest, dynamic_prompt
+
+
+@dynamic_prompt
+def system_prompt_middleware(request: ModelRequest) -> str:
+    """Render the system prompt template, filling `{current_date}`."""
+    return request.system_message.content.format(current_date=date.today().isoformat())

app/agent/schemas.py

@@ -1,43 +1,6 @@
-from enum import Enum
-
 from pydantic import BaseModel, Field
 
 
-class TemporalGranularity(str, Enum):
-    """Granularity of the data's temporal coverage."""
-
-    DAY = "day"
-    MONTH = "month"
-    YEAR = "year"
-
-
-class TemporalCoverage(BaseModel):
-    """The interval the SQL query actually filtered on in the answer."""
-
-    period_start: str = Field(
-        description=(
-            "Start of the interval filtered by the SQL query (e.g. '2010' for "
-            "`ano = 2010` or `ano BETWEEN 2010 AND 2012`). Format it to match `granularity`: "
-            "YYYY (year), YYYY-MM (month) or YYYY-MM-DD (day) — e.g. '2010', '2010-01', '2010-01-01'. "
-            "May be narrower than the table's full coverage."
-        )
-    )
-    period_end: str = Field(
-        description=(
-            "End of the interval filtered by the SQL query (e.g. '2010' for "
-            "`ano = 2010`; '2012' for `ano BETWEEN 2010 AND 2012`). Format it to match `granularity`: "
-            "YYYY (year), YYYY-MM (month) or YYYY-MM-DD (day) — e.g. '2012', '2012-01', '2012-01-01'. "
-            "May be narrower than the table's full coverage."
-        )
-    )
-    granularity: TemporalGranularity = Field(
-        description=(
-            "Granularity of `period_start`/`period_end`, matching their format: "
-            "YYYY (year), YYYY-MM (month), YYYY-MM-DD (day)."
-        )
-    )
-
-
 class DataSource(BaseModel):
     """A Base dos Dados table the answer draws on or points the user to."""
 
@@ -73,16 +36,11 @@ class StructuredResponse(BaseModel):
             "on clarification turns. Leave empty (None) when no table is relevant."
         ),
     )
-    temporal_coverage: TemporalCoverage | None = Field(
-        default=None,
-        description=(
-            "The interval your SQL query actually filtered. Leave empty (None) when no query "
-            "was run (e.g. a clarification turn) or the answer has no temporal dimension."
-        ),
-    )
-    follow_up_questions: list[str] | None = Field(
+    follow_up_prompts: list[str] | None = Field(
         default=None,
         description=(
-            "3 suggested follow-up questions (in the user's language) to explore the data further."
+            "3 next prompts the user could send you to explore the data further, each written "
+            "in the user's own voice — a message the user types to you, never a question you ask "
+            "the user. In the user's language."
         ),
     )

app/agent/tools/__init__.py

@@ -1,10 +1,7 @@
 from langchain_core.tools import BaseTool
 
 from app.agent.tools.api import get_dataset_details, get_table_details, search_datasets
-from app.agent.tools.bigquery import (
-    decode_table_values,
-    execute_bigquery_sql,
-)
+from app.agent.tools.bigquery import decode_table_values, execute_bigquery_sql
 
 
 class BDToolkit: