* fix(windows): keep the 8 MiB main-thread stack on legacy builds
Windows 7 / Server 2012 R2 users on 0.6.58 report the WER dialog "DBX 已停止
工作" with APPCRASH signature dbx.exe / faulting module dbx.exe / exception
c00000fd (STATUS_STACK_OVERFLOW) as soon as they click a database connection
(MySQL, Oracle 11 and Dameng 8 all reproduce), while 0.6.55 is fine.
Root cause: .cargo/config.toml raises the main-thread stack to 8 MiB with
-C link-arg=/STACK:8388608, but Cargo ignores target rustflags whenever a
RUSTFLAGS env var is set. The Win7/Server 2012 R2 jobs set RUSTFLAGS for the
static CRT, so the legacy PE kept the 1 MiB linker default while the standard
Windows build kept 8 MiB. The 0.6.56 connect-path rewrite made the DB query
stack deeper, and the legacy 1 MiB stack overflows on connect.
- repeat -C link-arg=/STACK:8388608 in every workflow RUSTFLAGS that links a
Windows PE (release Win7 app + dbx-web, CI Win7/standard/duckdb jobs, SQL
Server 2000 Win7 test build)
- add the same flag to the legacy target table in .cargo/config.toml for local
builds
- fail the Win7 PE audit when SizeOfStackReserve drops below 8 MiB, so this
cannot silently regress again
- assert the flag stays in the workflows from ci-workflow.test.mjs
* chore(clippy): indent the MariaDB KB doc continuation
main's 7958f7ad9 added an unindented continuation line after a doc list
item, so rust-fmt-clippy now fails on main and on every PR with
clippy::doc_lazy_continuation under -D warnings. Indent the line the way
the lint documents so the build can go green.
---------
Co-authored-by: DBX Agent <agent@dbx.local>
* fix(tabs): distinguish the active two-row group
* fix(tabs): apply group focus styling to single-row headers
* fix(tabs): retain faint group rails and outline collapsed counts
* fix(tabs): soften collapsed group count outline
---------
Co-authored-by: zipg <4047349+zipg@users.noreply.github.com>
The v0.6.56 switch (#13302 / #13376) defaulted off and gated connection
loader expansion, which turned into the "connection nodes cannot be
expanded" regression (#13536, #13579). Making the switch default on
(#13597) did not solve the underlying UX mismatch: clicking a connection
node both connects and expands, so an off switch still expanded on click
and looked broken.
Restore the long-standing behavior: connection loaders always expand
after loading again. Keep the #13557 manual-toggle expansion fix so the
first click always expands reliably. Remove the setting from the editor
settings store, settings dialog, settings search, draft/transfer
allowlists and all locales; existing persisted values are ignored by the
normalizer.
The #13302 performance concern is preserved as a possible future
optimization, but not as a default-off behavior switch.
Co-authored-by: zipg <4047349+zipg@users.noreply.github.com>
MQTT was reported separately: its connection node uses its own loadMqttTopics loader, but shares the same toggle path, so it has the same guarantee as redis/mongo/etcd/milvus/es. Add MQTT plus mq, nacos, dynamodb, zookeeper, consul and qdrant cases; reverting the #13557 fix fails all 12 new cases.
Co-authored-by: zipg <4047349+zipg@users.noreply.github.com>
v0.6.56 turned the long-standing auto-expand behavior into a default-off switch (#13302/#13376). That changed the default UX and the default-off flip also exposed the manual expansion regression #13557. Restore the old default while keeping the opt-out switch: users with many schemas can still disable it.
Existing blobs from the default-off builds are migrated once via sidebarExpandTreeOnConnectMigrationVersion; an explicit toggle afterwards is honored.
Co-authored-by: zipg <4047349+zipg@users.noreply.github.com>
The #13536 regression affected Redis, MongoDB, etcd, Milvus and Elasticsearch as well, since all of them share the connection toggle path. Lock the cross-type behavior in the expansion suite so a future loader gating change cannot regress it silently.
Co-authored-by: zipg <4047349+zipg@users.noreply.github.com>
`sidebarExpandTreeOnConnect` was registered on the EditorSettings interface, in
DEFAULT_EDITOR_SETTINGS, in the normalizeEditorSettings whitelist and in the
settings-dialog draft, but never in `applyEditorSettingsPatch()`. That function
assigns every field explicitly, so the key was dropped on every apply:
- the store value never changed, so the toggle had no effect at all;
- the persisted snapshot kept the old value;
- the dialog rebuilt its base from the unchanged store, so the draft stayed
dirty forever and "apply and close" kept reporting unsaved changes.
Add the missing branch and cover it with two regression tests: one asserts the
toggle reaches the store and disk, the other asserts the dialog dirty state
clears once the toggle is applied.
v0.6.56 introduced prefetchCompletionTablesIfEnabled() with a body that
called itself instead of completionMetadata.prefetchCompletionTables().
Query tabs pass prefetchCompletion=true and the helper runs synchronously
during setup, so every query tab mount (including opening a .sql file)
recursed until the stack overflow aborted setup, leaving a blank,
unresponsive editor.
Call the composable's prefetch again behind the same query-tab guard and
cover the mount path with a regression test.
Co-authored-by: zipg <4047349+zipg@users.noreply.github.com>
OSCAR rejects statements that combine its ROWNUM pseudo-column with a LIMIT. The JDBC agent caps table-data statements with Statement.setMaxRows(), which the OSCAR driver forwards as a row limit, so later ROWNUM-wrapped pages collided with that cap ("rownum and limit cannot be mixed").
Use LIMIT/OFFSET for OSCAR table previews, which the driver supports directly, and keep ROWNUM for bounded reads that do not set the driver cap. Preserve the explicit column projection when switching pagination forms.
Closes#13515
lxk955