libredb/libredb-studio

Follow-up to #1436, as promised in that PR. #1436 stops the table statistics from failing on CockroachDB: the size builtins take relid now, so the read no longer dies on pg_size_pretty(): unknown signature: pg_table_size(string). What it cannot fix is that the panel is then empty, and that has a different cause. Measured (CockroachDB CCL v26.3.2, cockroachdb/cockroach:latest, single node, 2026-10-04) CREATE TABLE probe_t (id int primary key, note varchar); INSERT INTO probe_t VALUES (1, 'x'); SELECT count(*) FROM pg_class c JOIN pg_namespace n ON n.oid = c.relnamespace WHERE n.nspname = 'public' AND c.relkind = 'r'; -- 1 SELECT count(*) FROM pg_stat_user_tables; -- 0 SELECT count(*) FROM pg_stat_all_tables; -- 0 SELECT pg_table_size('public.probe_t'::regclass); -- NULL So pg_class knows the user's table, both pg_stat_* views are empty, and pg_table_size() exists but answers NULL for a table it has. getTableStats() reads FROM pg_stat_user_tables, so on CockroachDB it

Summary On a MongoDB collection whose documents hold a subdocument, Generate Test Data builds each insertMany document from the flat inferred column list. That list names the subdocument and its children side by side (address, address.city, address.geo.lat), so every generated document gets "address": "value_0" and a separate top-level key literally named "address.city". MongoDB 5.0 and later accept field names containing a dot, so the insert succeeds and writes documents whose shape matches nothing in the collection: the subdocument becomes a string, and the dotted keys are unreachable through ordinary dot-notation queries. Measured (MongoDB 8.2.12, mongo:8.2, E2E pass 2026-10-03/04) Collection documents shaped { name, address: { city, geo: { lat, lng } } }; right-click > Generate Test Data, Run. The inserted documents read back as: { _id: ..., name: "...", address: "value_0", "address.city": "...", "address.geo": "value_0", "address.geo.lat": "..." } db.c.find({"address.city": {$exi

Summary User menu > MCP opens /settings/mcp, a full page with the status and token controls and nothing else: no header link, no back button. The only way back to the editor is the browser's Back button or typing the address. The sibling page /settings/authenticator has a "Back to the editor" link at the top. Measured (E2E pass 2026-10-03/04, production build, engine independent) Signed in as admin and as user, User menu > MCP: the page shows the MCP heading, status and token controls; no link leads back. Where it lives (main f9a17e8) src/app/settings/mcp/page.tsx: renders <McpSettings /> alone. src/components/mcp/McpSettings.tsx:102-110 (and the failure state at :94-97): the <main> starts with the heading. src/app/settings/authenticator/page.tsx:20-26: the "Back to the editor" link to copy. What to do Add the same "Back to the editor" link (href="/", ArrowLeft icon, same classes) at the top of /settings/mcp, in both the loaded and the failed state. Putting it in the page rather tha

Summary docs/MCP.md says only run_read_query needs a least-privilege principal on PostgreSQL and SQL Server, and that "inspect_schema has no such requirement". run_read_query's own refusal adds "inspect_schema works on every engine". In fact inspect_schema acquires its provider under the agent-operations execution profile, which opens with readOnly: true, and the PostgreSQL provider verifies the role at connect under that context. So on a seed connecting as postgres (or any superuser), inspect_schema fails too. Its refusal is also wrapped in the "untrusted data read from a database" notice and the "The database refused or failed the call:" prefix, although the text is Studio's own, and run_read_query's refusal ends in a double period. Measured (PostgreSQL 18.6, postgres:latest, MCP on a second instance of the same build, E2E pass 2026-10-03/04) Seed connection as postgres with mcp: true. inspect_schema {"table":"order_items"} answers isError: true: <untrusted-data notice> The database

Summary When the model does not answer within the turn limit, the investigation loop ends the run as failed with stop reason model-timeout, but with no failure reason. The answer card's failure note reads only failureReason, so it shows the unclassified text: "This run ended as failed and its own record names no reason for it. The server log is where the cause is." The record does name a reason, and the timeline already has the right sentence for it ("The model did not answer in time. Starting the run again is reasonable."). The card contradicts the run's own record and sends the user to the server log for nothing. Measured (E2E pass 2026-10-03/04, SQLite 3.50.4 Sample (Employees) and PostgreSQL 18.6, a slow local OpenAI-compatible model) Agent rail, Plan, any objective; the model does not answer within the turn limit. The card reads: Run failed This run ended as failed and its own record names no reason for it. The server log is where the cause is. The run's run-finished event carrie