Add /{database}/-/editor-schema.json endpoint for SQL editor consumers

Neutral {database, tables: [{name, view, columns: [{name, type}]}]}
shape, gated on view-database + execute-sql with no table-name leak on
403, hidden tables excluded. /-/schema.json was already taken by the
DDL endpoint, hence editor-schema.json. _editor_schema() now maps from
the shared _schema_tables() introspection helper.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
This commit is contained in:
Alex Garcia 2026-07-10 10:56:21 -07:00
commit b9716d4278
5 changed files with 336 additions and 22 deletions

View file

@ -85,6 +85,7 @@ from .views.special import (
JumpView,
InstanceSchemaView,
DatabaseSchemaView,
DatabaseEditorSchemaView,
TableSchemaView,
)
from .views.table import (
@ -2716,6 +2717,10 @@ class Datasette:
DatabaseSchemaView.as_view(self),
r"/(?P<database>[^\/\.]+)/-/schema(\.(?P<format>json|md))?$",
)
add_route(
DatabaseEditorSchemaView.as_view(self),
r"/(?P<database>[^\/\.]+)/-/editor-schema\.json$",
)
add_route(
QueryParametersView.as_view(self),
r"/(?P<database>[^\/\.]+)/-/query/parameters$",

View file

@ -649,6 +649,52 @@ def _column_completion(name, type_):
return completion
async def _schema_tables(datasette, database_name, *, include_hidden=True):
"""
Neutral introspection of a database's tables and views for SQL editors.
Returns an ordered list of dicts, one per table or view::
{"name": str, "view": bool,
"columns": [{"name": str, "type": str}, ...]}
``type`` is the SQLite declared column type (empty string when the column
has no declared type). Regular-table columns come from the internal
``catalog_columns`` catalog; views are absent from that catalog so their
columns are read directly via PRAGMA table_xinfo. Hidden tables (FTS shadow
tables and the like) are excluded unless ``include_hidden`` is True. This is
the shared, serialization-agnostic source for both ``_editor_schema`` (which
maps it to lang-sql Completion objects) and the ``/-/editor-schema.json``
endpoint (which emits it directly).
"""
internal_db = datasette.get_internal_database()
result = await internal_db.execute(
"select table_name, name, type from catalog_columns where database_name = ?",
[database_name],
)
table_columns = {}
for row in result.rows:
table_columns.setdefault(row["table_name"], []).append(
{"name": row["name"], "type": row["type"]}
)
db = datasette.get_database(database_name)
hidden = set() if include_hidden else set(await db.hidden_table_names())
tables = []
for table_name, columns in table_columns.items():
if table_name in hidden:
continue
tables.append({"name": table_name, "view": False, "columns": columns})
# Views are not represented in catalog_columns, so pull their real columns
# directly (PRAGMA table_xinfo works against views too).
for view_name in await db.view_names():
columns = [
{"name": column.name, "type": column.type}
for column in await db.table_column_details(view_name)
]
tables.append({"name": view_name, "view": True, "columns": columns})
return tables
async def _editor_schema(datasette, database_name):
"""
Build a lang-sql SQLNamespace for the CodeMirror SQL editor autocomplete.
@ -659,29 +705,21 @@ async def _editor_schema(datasette, database_name):
container so the popup can label them as views while still completing their
real columns. See @codemirror/lang-sql >= 6.6 SQLNamespace / Completion.
"""
internal_db = datasette.get_internal_database()
result = await internal_db.execute(
"select table_name, name, type from catalog_columns where database_name = ?",
[database_name],
)
schema = {}
for row in result.rows:
schema.setdefault(row["table_name"], []).append(
_column_completion(row["name"], row["type"])
)
# Views are not represented in catalog_columns, so pull their real columns
# directly (PRAGMA table_xinfo works against views too).
db = datasette.get_database(database_name)
for view_name in await db.view_names():
columns = await db.table_column_details(view_name)
schema[view_name] = {
"self": {
"label": view_name,
"type": "class",
"detail": "view",
},
"children": [
_column_completion(column.name, column.type) for column in columns
],
}
for table in await _schema_tables(datasette, database_name, include_hidden=True):
completions = [
_column_completion(column["name"], column["type"])
for column in table["columns"]
]
if table["view"]:
schema[table["name"]] = {
"self": {
"label": table["name"],
"type": "class",
"detail": "view",
},
"children": completions,
}
else:
schema[table["name"]] = completions
return schema

View file

@ -1345,6 +1345,59 @@ class DatabaseSchemaView(SchemaBaseView):
return await self.format_html_response(request, schemas)
class DatabaseEditorSchemaView(BaseView):
"""
JSON introspection of a database's tables, views and columns shaped for SQL
editor autocomplete consumers (the CodeMirror ``<datasette-sql-editor>``
component and external clients such as datasette-paper).
Distinct from :class:`DatabaseSchemaView` (``/<db>/-/schema.json``), which
returns the raw DDL as a SQL string gated on ``view-database`` alone. This
endpoint returns a neutral structured payload and is gated on both
``view-database`` and ``execute-sql`` — the same permissions as the inline
editor schema handed to the SQL query page.
"""
name = "database_editor_schema"
has_json_alternate = False
async def get(self, request):
from .query_helpers import _schema_tables
database_name = request.url_vars["database"]
# view-database is checked first so actors without it cannot
# distinguish an existing database from a missing one, and a denied
# request only ever leaks the permission action name, never table names.
await self.ds.ensure_permission(
action="view-database",
resource=DatabaseResource(database=database_name),
actor=request.actor,
)
if database_name not in self.ds.databases:
headers = {}
if self.ds.cors:
add_cors_headers(headers)
return Response.json(
error_body("Database not found", 404), status=404, headers=headers
)
await self.ds.ensure_permission(
action="execute-sql",
resource=DatabaseResource(database=database_name),
actor=request.actor,
)
await self.ds.refresh_schemas()
tables = await _schema_tables(self.ds, database_name, include_hidden=False)
headers = {}
if self.ds.cors:
add_cors_headers(headers)
return Response.json(
{"database": database_name, "tables": tables}, headers=headers
)
class TableSchemaView(SchemaBaseView):
"""
Displays schema for a specific table.

View file

@ -152,6 +152,60 @@ Values for named SQL parameters can be provided as additional query string param
The response uses the same default representation described above.
.. _json_api_editor_schema:
Schema for SQL editors
----------------------
The ``/-/editor-schema.json`` endpoint returns a machine-readable description of
a database's tables, views and columns, shaped for SQL editor autocomplete. It
powers Datasette's own CodeMirror SQL editor and is available for external
consumers such as embeddable editor components.
::
GET /<database>/-/editor-schema.json
Access requires both the :ref:`actions_view_database` and
:ref:`actions_execute_sql` permissions for the database - the same gate as the
inline editor schema on the SQL query page. A request that fails either check
receives a ``403`` JSON error that does not reveal any table or column names.
The response is a neutral structure - a ``database`` name and a list of
``tables``, each with a ``view`` flag (``true`` for SQL views) and a list of
``columns`` carrying the SQLite declared ``type`` (an empty string when the
column has no declared type):
.. code-block:: json
{
"database": "fixtures",
"tables": [
{
"name": "facetable",
"view": false,
"columns": [
{"name": "pk", "type": "INTEGER"},
{"name": "state", "type": "TEXT"}
]
},
{
"name": "paginated_view",
"view": true,
"columns": [
{"name": "content", "type": "TEXT"}
]
}
]
}
Hidden tables - such as the shadow tables that back SQLite full-text search -
are excluded from the response.
This endpoint is distinct from the :ref:`database schema endpoint <DatabaseSchemaView>`
at ``/<database>/-/schema.json``, which returns the raw ``CREATE`` statements as
a SQL string.
.. _json_api_shapes:
Different shapes

View file

@ -1,3 +1,5 @@
import json
import pytest
import pytest_asyncio
from datasette.app import Datasette
@ -245,3 +247,165 @@ async def test_table_not_exists(schema_ds):
response = await schema_ds.client.get("/schema_public_db/nonexistent/-/schema.md")
assert response.status_code == 404
assert "not found" in response.text.lower()
# ---------------------------------------------------------------------------
# /<database>/-/editor-schema.json — neutral structured schema for SQL editors
# ---------------------------------------------------------------------------
@pytest_asyncio.fixture(scope="module")
async def editor_schema_ds():
"""Datasette instance exercising the editor-schema endpoint.
- public db: tables + a view + an FTS table (hidden shadow tables)
- private db: gated behind view-database (allow root only)
- noexec db: view-database allowed for anyone, execute-sql denied
"""
ds = Datasette(
config={
"databases": {
"editor_private_db": {"allow": {"id": "root"}},
"editor_noexec_db": {
# Everyone may view the database, but nobody may run SQL
"allow_sql": {"id": "root"},
},
}
}
)
public_db = ds.add_memory_database("editor_public_db")
await public_db.execute_write(
"CREATE TABLE users (id INTEGER PRIMARY KEY, name TEXT)"
)
await public_db.execute_write(
"CREATE TABLE posts (id INTEGER PRIMARY KEY, title TEXT, body TEXT)"
)
await public_db.execute_write(
"CREATE VIEW recent_posts AS SELECT id, title FROM posts ORDER BY id DESC"
)
# An FTS table produces hidden shadow tables (users_fts_data, etc.)
await public_db.execute_write(
"CREATE VIRTUAL TABLE users_fts USING fts5(name, content=users)"
)
private_db = ds.add_memory_database("editor_private_db")
await private_db.execute_write(
"CREATE TABLE secret_data (id INTEGER PRIMARY KEY, value TEXT)"
)
noexec_db = ds.add_memory_database("editor_noexec_db")
await noexec_db.execute_write(
"CREATE TABLE locked (id INTEGER PRIMARY KEY, value TEXT)"
)
await ds.invoke_startup()
await ds.refresh_schemas()
return ds
@pytest.mark.asyncio
async def test_editor_schema_allowed_shape(editor_schema_ds):
"""Authorized fetch returns tables, columns, types and views in the
documented neutral shape."""
response = await editor_schema_ds.client.get(
"/editor_public_db/-/editor-schema.json"
)
assert response.status_code == 200
data = response.json()
assert data["database"] == "editor_public_db"
assert isinstance(data["tables"], list)
by_name = {t["name"]: t for t in data["tables"]}
# Regular table with columns + declared types
users = by_name["users"]
assert users["view"] is False
assert users["columns"] == [
{"name": "id", "type": "INTEGER"},
{"name": "name", "type": "TEXT"},
]
posts = by_name["posts"]
assert posts["view"] is False
assert {c["name"] for c in posts["columns"]} == {"id", "title", "body"}
# View is flagged and carries its real columns
view = by_name["recent_posts"]
assert view["view"] is True
assert [c["name"] for c in view["columns"]] == ["id", "title"]
# Whole payload is JSON-serializable and every entry matches the shape
for table in data["tables"]:
assert set(table) == {"name", "view", "columns"}
for column in table["columns"]:
assert set(column) == {"name", "type"}
@pytest.mark.asyncio
async def test_editor_schema_excludes_hidden_tables(editor_schema_ds):
"""FTS shadow tables (hidden_table_names) must not appear."""
response = await editor_schema_ds.client.get(
"/editor_public_db/-/editor-schema.json"
)
assert response.status_code == 200
names = {t["name"] for t in response.json()["tables"]}
assert not any("_fts_" in name or name.endswith("_fts") for name in names), names
# Sanity: the visible objects are still there
assert {"users", "posts", "recent_posts"} <= names
@pytest.mark.asyncio
async def test_editor_schema_denied_view_database_403_no_leak(editor_schema_ds):
"""Anonymous user lacking view-database gets a 403 that leaks no names."""
response = await editor_schema_ds.client.get(
"/editor_private_db/-/editor-schema.json"
)
assert response.status_code == 403
body = response.text
assert "secret_data" not in body
data = response.json()
assert data["ok"] is False
assert "secret_data" not in json.dumps(data)
# The permitted actor can read it
response = await editor_schema_ds.client.get(
"/editor_private_db/-/editor-schema.json", actor={"id": "root"}
)
assert response.status_code == 200
names = {t["name"] for t in response.json()["tables"]}
assert "secret_data" in names
@pytest.mark.asyncio
async def test_editor_schema_denied_execute_sql_403_no_leak(editor_schema_ds):
"""A viewer who lacks execute-sql gets a 403 with no schema data."""
# Anonymous user may view editor_noexec_db but not run SQL against it
response = await editor_schema_ds.client.get(
"/editor_noexec_db/-/editor-schema.json"
)
assert response.status_code == 403
data = response.json()
assert data["ok"] is False
assert "tables" not in data
assert "locked" not in json.dumps(data)
# The actor granted execute-sql can read the schema
response = await editor_schema_ds.client.get(
"/editor_noexec_db/-/editor-schema.json", actor={"id": "root"}
)
assert response.status_code == 200
names = {t["name"] for t in response.json()["tables"]}
assert "locked" in names
@pytest.mark.asyncio
async def test_editor_schema_database_not_found(editor_schema_ds):
"""A non-existent database returns a 404 JSON error."""
response = await editor_schema_ds.client.get(
"/nonexistent_db/-/editor-schema.json"
)
assert response.status_code == 404
data = response.json()
assert data["ok"] is False
assert "not found" in data["error"].lower()