verbose live UI + tool-level SSE events + Groq default + regression tests

api/tools/execute_sql.py

@@ -0,0 +1,55 @@
+"""Execute read-only SQL against the financial warehouse.
+
+Defence in depth:
+- Engine-level `search_path` and `statement_timeout` via libpq options in
+  `api.tools.db.get_engine` — no SQL-side SETs.
+- `execution_options(postgresql_readonly=True, postgresql_deferrable=True)`
+  — SQLAlchemy emits `SET TRANSACTION READ ONLY DEFERRABLE` internally; we
+  don't compose a SQL string for it.
+- Hard refusal of any statement that doesn't begin with SELECT or WITH.
+- Row cap on the returned result set.
+- No commit — read-only transactions roll back on context exit.
+
+The SQL passed in IS a string (it's the LLM's output); nothing on the Python
+side concatenates SQL fragments around it.
+"""
+
+from __future__ import annotations
+
+import logging
+
+from sqlalchemy import text
+
+from api import langfuse_client as lf
+from api.tools.db import get_engine
+from api.tools.types import QueryResult
+
+logger = logging.getLogger("nvi.tools.execute_sql")
+
+MAX_ROWS = 1000
+
+
+def execute_sql(sql: str) -> QueryResult:
+    sql_clean = sql.strip().rstrip(";").strip()
+    head = sql_clean.split(None, 1)[0].upper() if sql_clean else ""
+    if head not in {"SELECT", "WITH"}:
+        raise ValueError(f"refusing non-SELECT statement (starts with {head!r})")
+
+    engine = get_engine()
+    with lf.span("execute_sql", input={"sql": sql_clean}) as span:
+        ro_engine = engine.execution_options(
+            postgresql_readonly=True,
+            postgresql_deferrable=True,
+        )
+        with ro_engine.connect() as conn:
+            res = conn.execute(text(sql_clean))
+            columns = list(res.keys())
+            fetched = res.fetchmany(MAX_ROWS + 1)
+
+        truncated = len(fetched) > MAX_ROWS
+        rows = [list(r) for r in fetched[:MAX_ROWS]]
+        result = QueryResult(
+            columns=columns, rows=rows, row_count=len(rows), truncated=truncated
+        )
+        span.update(output={"row_count": result.row_count, "truncated": truncated})
+        return result