Skip to content

Issue sprint: theme editor, motion and layout, tear-off tabs, Informix row history, test data generator - #173

Merged
e-omunoz merged 89 commits into
mainfrom
feat/integration-2.4.0
Oct 13, 2026
Merged

e-omunoz merged 89 commits into
mainfrom
feat/integration-2.4.0

Conversation

@e-omunoz

@e-omunoz e-omunoz commented Oct 13, 2026 •

Copy link
Copy Markdown
Owner

Integration of the 2.4.0 issue sprint: three tested branches merged with --no-ff onto main, plus two integration fixes.

Branch Issues Tested at
issue/theme-motion-layout #107, #110, #108, #111 6de2dc4 (desktop slot 1)
issue/informix-row-history #123 99bca1e (desktop slot 2)
issue/test-data-generator #172 f908d7e (desktop slot 3)

Integration (this branch)

Checks on the integrated head

  • Fast checks (WSL): every dev/*-check.ts passes; tsc --noEmit -p . and -p installer pass; npm run build passes.

  • bash dev/wsl/engines.sh test at 7ca9cf6 printed ALL ENGINES PASS (exit 0). Results by engine:

    Engine Result Modules
    PostgreSQL ✅ postgres 17
    MySQL ✅ mysql 16
    MariaDB ✅ mysql 16, engine_tests 6, mcp 2, probe 1
    SQL Server ✅ mssql 12
    Informix DRDA ✅ engine_tests, including testdata_informix
    Informix JDBC ✅ engine_tests, including informix_jdbc_row_history and testdata_informix_jdbc
    SQLite ✅ sqlite 10
    ODBC ✅ odbc 3, odbc_driver 6

    Other modules: rowhistory 15, windows 7 and ssh 4. Zero failures; 1 ignored in sheets, as before.
    bc294c5 changes only a dev check.

  • dev\check-all.ps1 at 7ca9cf6 (Windows, debug desktop app): every step passed except these four.

    • motion, focus and drag and drop: the Windows path bug above. Fixed in bc294c5, and it passes on Windows now.
    • menus, settings and theme editor (preview): the vite server already running on port 1420 served a blank page. It had been started the day before and did not pick up the merged modules. Against a fresh vite on the merged tree, layout-check.mjs passed 31/31, with no page errors.
    • restore (prepare/verify): Execution context was destroyed while the app was starting. That is a start-up reload; verify failed only because prepare had not run. I ran restart → prepare → restart → verify again twice and it passed both times (5/5 ok).
  • The e2e steps passed, including test data generator, generated scripts, data/schema comparison, keyboard shortcuts (desktop) and Gib companion.


#107 Theme editor: full visual customisation

What changed: Ajustes › General › Editor de tema (also the palette entry and the pencil on a custom card) builds a custom theme from any built-in one:

  • every token group, plus fonts, radius, row height, shadows and animation speed, with a live preview of the real UI;
  • a WCAG AA contrast checker;
  • new/duplicate/rename/delete/save/Guardar y usar/discard;
  • .celer-theme.json export and import (hostile, garbage or newer files are refused);
  • custom themes in Seguir al sistema; tear-off windows follow.

Tests (desktop slot 1; no connection involved):

  • ✅ All 10 token groups are editable and change the real UI live.
  • ✅ Fonts, radius, density, shadows and speed reach every component.
  • ✅ Contrast warnings appear below 4.5:1.
  • ✅ The theme life cycle works with no DataCloneError.
  • ✅ An exported theme imported on a fresh data folder is identical (107 properties).
  • ✅ Hostile files are refused.
  • ✅ System light/dark switching works with a custom theme.
  • ✅ Tear-offs follow the saved theme. They do not follow the unsaved draft, which is acceptable.
  • ✅ themes-check and contrast-check pass.

#110 Global animations and styles

What changed:

  • Motion tokens everywhere, honouring Animaciones and reduced motion.
  • Exit animations for dialogs, menus, palette, toasts, panels and tabs.
  • Consistent controls, focus rings and badges.
  • F6 / Shift+F6 cycle the panels.
  • One drag-and-drop look.
  • Shortcut hints in menus.

Tests:

  • ✅ motion-check passes; reduced motion gives 0.06 s in the main window and tear-offs.
  • ✅ F6 order is right, it does nothing behind overlays, and closing a panel hands focus on.
  • ✅ Menu hints: shortcuts and muted notes.
  • ✅ Screenshots of all 7 built-in themes and a custom theme.
  • Result-column drag on every engine: PG ✅ MySQL ✅ MariaDB ✅ MSSQL ✅ IFX-DRDA ✅ IFX-JDBC ✅ SQLite ✅ ODBC ✅
  • ⚠️ Transitions were checked in code and CSS, not captured frame by frame.

#108 Tear-off windows: make dragging a tab out obvious

What changed:

  • A ghost card follows the pointer across Celer windows.
  • Outside every window it becomes a dashed outline of the new window, with «Soltar para abrir en una ventana nueva».
  • Insertion line over another tab bar; Esc cancels.
  • Abrir en ventana nueva in the tab menu, plus a one-time hint.

Tests:

  • ✅ Real mouse drags (SendInput): the card shows inside, over another window and outside.
  • ✅ The outline matches the new window: 580,482 at 1280×820.
  • ✅ The window is clamped on screen at the corner.
  • ✅ Esc and coming back over a window clear the outline.
  • ✅ Dropping into another tab bar works.
  • ✅ windows-check passes; cargo windows:: 7 ok.
  • Console with an open manual transaction moved to a new window keeps its session, its uncommitted row and its column order: PG ✅ MySQL ✅ MariaDB ✅ MSSQL ✅ IFX-DRDA ✅ IFX-JDBC ✅ SQLite ✅ ODBC ✅
  • ⚠️ Not tested: 150 % scaling and a second monitor. This machine has a single 100 % monitor; that path is covered only by unit checks.

#111 Reorganise overgrown parts (nothing removed)

What changed:

  • No explorer button in the top bar; the stripe and Alt+1 toggle the explorer.
  • The right panel header shows only its title.
  • One Importar conexiones ›.
  • Table menu: open/structure/Ver DDL/diagram on top, then Generar SQL › and Comparar ›.
  • Cell menu: Copiar como › and Filtrar ›.
  • Connection dialog: one Entorno row.
  • Ajustes grouped General/Editor/Conexiones/IA y MCP.
  • Keyboard-operable submenus.

Tests:

  • ✅ layout-check 31/31 (also on the integrated head).
  • ✅ Tour 7/7.
  • ✅ Every settings section reachable.
  • ✅ Diff review: no entry removed, and the split button was not done.
  • Table menu order, Ver DDL › Abrir en consola on the same connection, Generar SQL › with dialect-correct SQL, and all 9 copy formats plus Filtrar › Por este valor: PG ✅ MySQL ✅ MariaDB ✅ MSSQL ✅ IFX-DRDA ✅ IFX-JDBC ✅ SQLite ✅ ODBC ✅

#123 Informix row history from the logical logs

What changed: Historial de la fila in a table viewer's row menu on Informix over JDBC reads the row's changes from the logical logs through the CDC API, newest first:

  • each change shows commit time, transaction, user and operation, with before → after values, Ver la fila en este momento and Todas las columnas;
  • every missing precondition is stated as «No se puede reconstruir: …»;
  • full row logging is turned on only for the read and then off again, and a failed clean-up is reported and repaired on the next read;
  • partial histories are labelled with the reason.

Tests (desktop slot 2):

  • IFX-JDBC ✅ history with grouped transactions and unicode.
  • IFX-JDBC ✅ every unavailable reason, each checked in the app.
  • IFX-JDBC ✅ full row logging is off again after a normal, stopped, killed or stalled read (sysptnhdr flag checked).
  • IFX-JDBC ✅ every partial-history reason.
  • IFX-DRDA ✅ «usa DRDA» reason.
  • IFX Client SDK ⚠️ backend only (invoke); the SDK is not installed here.
  • Disabled «No disponible en …» in table and query menus: PG ✅ MySQL ✅ MariaDB ✅ MSSQL ✅ SQLite ✅ ODBC ✅. An Informix query result says to open the table instead.
  • ✅ informix_jdbc_row_history, rowhistory 15 and rowhistory-check pass.

Known non-blocking nits (follow-up):

  • If the current row cannot be read for the final comparison (e.g. -245, lock held), the dialog still says «Faltan cambios…».
  • A server that is down at read start shows the raw INFORMIX_GUIDE: code prefix in the error box.
  • Small wording and layout nits on the log-range line and the Ver la fila button.

#172 Test data generator

What changed: Generar datos de prueba… opens from the table menu (explorer and tab) and the palette.

  • It infers a generator per column from the type and Spanish/English names; all 41 generator kinds can be chosen.
  • It respects NOT NULL, CHECK, enums, unique keys (also against existing rows) and foreign keys.
  • The same seed gives the same rows.
  • Rows go in through the import pipeline in one transaction, with progress and a Cancel that rolls back. Read-only connections refuse; production connections ask first.
  • Output can also be a script, CSV or JSON.
  • Also: SQLite only treats the real rowid alias as server-filled, and the import writes 0x… into binary columns as bytes.

Tests (desktop slot 3):

  • Dialog and inference, 10,000 valid rows, FK and uniqueness checked by SQL, empty parent offer, read-only/production/Cancel/console transaction left untouched, and script/CSV/JSON identical to the insert: PG ✅ MySQL ✅ MariaDB ✅ MSSQL ✅ IFX-DRDA ✅ IFX-JDBC ✅ SQLite ✅ ODBC ✅
  • On Informix, BYTE/BLOB columns get NULL because the engine takes no literals there (documented).
  • ✅ 100,000 rows into PostgreSQL in 9.0 s; Cancel during a 300k run stops in about 100 ms.
  • ✅ testdata-check passes; cargo testdata_* passes on every engine; the browser demo works.
  • ⚠️ After integration the explorer entry moved out of the list of Generar … entries to sit next to Importar datos…. Same label, same handler. The desktop e2e test data generator step (PostgreSQL) passed at 7ca9cf6. Every engine still has the tab-menu and palette paths the tester walked.

Closes #107
Closes #110
Closes #108
Closes #111
Closes #123
Closes #172

Issues left out: none.

Review

Macro review 2.4 (/macro-review full, all areas, focused on the 2.4 features; tracking issue #171). The full report is in docs/review/REPORT.md.

41 issues filed (1 high, 13 medium, 27 low). All 41 are fixed on this branch: none deferred, none left open. The final verification is green, with no failures and no regressions.

Issue Severity Area Title Status
#174 high security Test data: MySQL/MariaDB catalog queries escape only quotes, so a table name with a backslash runs injected SQL when the dialog opens fixed in 48be1ed
#175 medium ui-visual Theme editor: Contraste panel shrinks to a sliver at 1024x640 and 1366x768 at 125 % fixed in e5bcb73
#176 medium ui-visual Built-in themes fail the editor's CONTRAST_PAIRS; Fjord's --danger is 2.46:1 on dialogs fixed in 321fc4a
#177 medium ui-functional Test data «Insertar» (and CSV/Excel import) reloads open tabs of the table, discarding unsaved edits fixed in 73f9106
#178 medium drivers-open MySQL 8: DEFAULT CURRENT_TIMESTAMP columns (DEFAULT_GENERATED) reported as identity fixed in aedb3d1
#179 medium drivers-open Test data on MariaDB JSON columns: plain text rejected by the implicit json_valid CHECK fixed in 8c37328
#180 medium drivers-enterprise / connections / core-rust Row history bypasses the connection's SSH tunnel and startup script fixed in 23699b6, 025c421
#181 medium drivers-enterprise Test data over ODBC to MariaDB: out-of-list ENUM values, downgraded error, success reported with 0 rows fixed in d8322d9
#182 medium drivers-enterprise Test data on Informix: multi-field INTERVAL columns get '1' (-1263) fixed in 3416b13
#183 medium drivers-enterprise Test data on SQL Server: CHECK with N'...' literals not parsed fixed in 3f9616d
#184 medium frontend-state Test data dialog: Sets of up to 200,000 values rebuilt through the store proxy on each change (freezes, ~300 MB kept) fixed in 99565cd
#185 medium frontend-state Test data JSON export of 1M wide rows fails with «Invalid string length» fixed in 4a1b881
#186 medium frontend-state Scrim mousedown or Esc during a test data insert cancels and rolls back with no confirmation (import too) fixed in 0b4b4fe
#187 medium release STATUS.md and docs/ROADMAP.md still describe the 2.3 sprint fixed in 60af03e
#188 low ui-visual Row history .rh-op chips and theme editor level chips below 4.5:1 on their tint fixed in 67fa6fe
#189 low ui-visual Test data: pattern option always reads «Patrón ([A-Z]{3}-\d{4})» fixed in 3ed8a89
#190 low ui-visual Test data dialog: generator labels and preview cells cut off, no tooltip fixed in ba3454d
#191 low ui-visual Row history: failure to turn full row logging back off shown only as a muted note fixed in 9a092da
#192 low ui-visual SQLite: INTEGER PRIMARY KEY reported as nullable («NULL» badge on the key) fixed in 7af539f
#193 low ui-functional Workspace shortcuts (Ctrl+W, Ctrl+Tab, F5…) fire behind modal dialogs fixed in 740db2c
#194 low ui-functional Test data on Informix: all-default tables generate INSERT … DEFAULT VALUES (syntax error) fixed in 7957ce4
#195 low ui-functional Test data on SQL Server/Informix: existing NULL in a nullable UNIQUE column not seeded fixed in 547e20f
#196 low drivers-open Test data: a sequence on a unique integer key can run past the type's range without warning fixed in 742b481
#197 low drivers-open Test data and import on MyISAM/Aria: message says nothing was inserted though rows stay fixed in 670e5b2
#198 low drivers-open Test data on MySQL/MariaDB BIT(n>1) and spatial columns: rejected values fixed in 639d9f5
#199 low drivers-open Test data on PostgreSQL domains ignores base type modifiers, NOT NULL and CHECK fixed in 9367968
#200 low drivers-open CHECK reader misses negative bounds as PostgreSQL and MySQL 8 store them fixed in 9606590
#201 low drivers-open SQLite: INTEGER PRIMARY KEY DESC treated as the rowid alias fixed in e16becb
#202 low drivers-enterprise Row history ledger keyed by DBSERVERNAME and table only (cross-server confusion) fixed in d26d2fd
#203 low drivers-enterprise Test data on an unrecognised ODBC engine loads the whole referenced table fixed in bbc5f8d
#204 low core-rust Test data CSV export uses ',' and no BOM, unlike Celer's CSV export fixed in f0f2382
#205 low core-rust Overlapping row-history reads of one table turn full row logging off under each other fixed in 7917f2f
#206 low core-rust Engine gate flaky: informix_jdbc_row_history CDC retry and its docker restart fixed in 81a317b
#207 low core-rust write_text_file is a sync command on the main thread; large exports freeze every window fixed in cbdd2f3
#208 low frontend-state «Generar antes datos de …» resets the child table's settings fixed in 503d401
#209 low frontend-state Tab drag sends the console's full SQL as the card preview to every window fixed in a62acca
#210 low security Test data script: a newline in the table name ends the «--» header comment fixed in 4f2c874
#211 low security Row history on a read-only Informix connection turns full row logging off on a left-over ledger entry fixed in 3808d53
#212 low release STATUS test inventory stale; check-all.ps1 skips 15 .ts checks and 11 .mjs scripts fixed in d8c9eb1
#213 low release READMEs omit 2.4's features and the Alt+1 / F6 shortcuts fixed in ce8a68e
#214 low release docs/media screenshots show the removed explorer button and lack the Variables button fixed in f3474a5

#180 was filed by three auditors (drivers-enterprise, connections, core-rust) and is counted once. It was fixed in two parts: the SSH tunnel (23699b6) and the startup script (025c421).

Not fixed, and outside these issues: with the Teal or Green accent, accent-coloured text stays below 4.5:1 in the light themes.

What the review did not check (the full list is in the report):

  • The desktop app (WebView2) was not driven for most fixes. Those fixes rest on engine tests, logic checks and the browser preview.
  • The tear-off outline (Tear-off windows: make it obvious that a tab can be dragged out into its own window #108) needs a real OS drag, so it was reviewed in code only.
  • No live Informix CDC events were read: other agents restarted celer-ifx and left it quiescent.
  • Generic ODBC was tested only against PostgreSQL and MariaDB. ODBC to SQL Server or Informix was not tested, and neither were Windows DSNs.
  • macOS was not checked.

Final engine matrix (whole branch)

These results come from the latest bash dev/wsl/engines.sh test run (the WSL side of dev\wsl.ps1 test). It ran on the branch's final code, 13 Oct 09:09. The commits after it change only docs, screenshots, CHANGELOG and dev/check-all.ps1. First run: 281 passed, 0 failed, 1 ignored (sheets). MariaDB rerun: 29 passed. The run printed ALL ENGINES PASS, exit 0.

Engine Result Coverage
PostgreSQL 17 ✅ postgres 17; testdata_postgres, testdata_postgres_100k, testdata_pg_domains and the cross-engine testdata_* tests
MySQL 8.4 ✅ mysql 16; testdata_mysql, testdata_mysql_types, testdata_mysql_non_transactional, mysql_default_generated_is_not_identity
MariaDB 11.4 ✅ mysql 16, engine_tests 10, mcp 2, probe 1
SQL Server 2022 ✅ mssql 12; testdata_mssql
Informix DRDA ✅ engine_tests: informix_engine, testdata_informix
Informix JDBC ✅ engine_tests: informix_jdbc_engine, informix_jdbc_reconnects, informix_jdbc_row_history, testdata_informix_jdbc
SQLite ✅ sqlite 10 (incl. identity_is_only_the_rowid_alias); testdata_sqlite, testdata_sqlite_int_keys
ODBC ✅ odbc 3, odbc_driver 6; testdata_odbc (psqlODBC → PostgreSQL), testdata_odbc_mysql (MariaDB Connector/ODBC → MariaDB)
SSH tunnels ✅ ssh 4; ssh_tunnel_row_history

Other modules: engine_tests 62, rowhistory 17, windows 7, report 5, write_text_tests 1 and data_dir_tests 1. All pass.
Fast checks pass on every area: all 35 dev/*-check.ts, tsc -p ., tsc -p installer and npm run build.
GitHub Actions on this head: ci.yml, engines.yml (SQL Server + Informix) and a manual release-desktop.yml packages-only run (Windows/macOS/Linux, celer-release artifact, nothing published) have been triggered.

Review issues closed by this PR

Closes #174
Closes #175
Closes #176
Closes #177
Closes #178
Closes #179
Closes #180
Closes #181
Closes #182
Closes #183
Closes #184
Closes #185
Closes #186
Closes #187
Closes #188
Closes #189
Closes #190
Closes #191
Closes #192
Closes #193
Closes #194
Closes #195
Closes #196
Closes #197
Closes #198
Closes #199
Closes #200
Closes #201
Closes #202
Closes #203
Closes #204
Closes #205
Closes #206
Closes #207
Closes #208
Closes #209
Closes #210
Closes #211
Closes #212
Closes #213
Closes #214

🤖 Generated with Claude Code

e-omunoz and others added 30 commits October 13, 2026 00:16
…h live preview, contrast checks and JSON export/import (#107)

Every token is editable by group (surfaces, text, borders, accent, states, grid,
editor syntax, explorer, Gib), plus fonts, radius, row height, shadows and
animation speed. Custom themes sit next to the built-in ones, can be chosen for
the system's light and dark modes, and tear-off windows follow them through the
shared settings. Hard-coded corner radii now use the radius tokens so the
radius setting reaches every component.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…ons, shared component states, F6 between panels and one drag-and-drop look (#110)

Every transition and animation takes --motion-fast/normal/slow (scaled by the
theme's speed, cut by reduced motion); dev/motion-check.ts fails on a stray
duration. Panels, tabs, toasts, dialogs, menus and the palette animate in and
out, buttons share hover and press, text fields one focus ring, menus and
popovers one surface, badges and empty states one style. F6 / Shift+F6 move
the focus between panels, context menus show shortcuts from the keymap in one
notation, and tabs, result columns, explorer and library share the drag
tokens and ghost card.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…he new window's outline, and tab bars show where it lands (#108)

The drag image is a card with the tab's title, connection and first lines.
Outside every Celer window the core shows a transparent, click-through outline
window (public/drag-ghost.html) at the very spot the new window will open, with
«Soltar para abrir en una ventana nueva»; over a window it hides and that
window's tab bar draws an insertion line (also for reordering within a bar).
The tab menu says «Abrir en ventana nueva», the tooltip mentions dragging, a
one-time hint (Gib or a notice) explains it, and a window opened by a drop
grows into place.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
… or the guide is on top (#110)

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
… which threw DataCloneError on store proxies and kept saved themes from opening (#107)

Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
…speed (#110)

Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
… image no longer shows over it, and a dropped window is kept on the screen (#108)

Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
…eler window too (#108)

Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
…y precondition and partial reason said (#123)

Brings wip/informix-row-history-2.3 up to main as one change and fixes what Sprint 2.3's testing found:
- A transaction is named by the log position of its commit; the records' transaction number is a reused slot.
- A read stopped or cut short shows how far it got and never says «sin cambios»; a connection lost mid-read
  (killed session, server restart) keeps what was read, labelled partial, instead of failing the dialog.
- No-changes text no longer infers the row's earlier values.
- Fixed-length-only tables: the full-row-logging offer says earlier updates will not appear, and a history read
  with it turned on just for the read is labelled partial; tables with a VARCHAR are told their changes read whole.
- DRDA and Client SDK are preconditions of the connection: the entry stays enabled on Informix and the dialog says
  «No se puede reconstruir: la conexión usa DRDA…». Other engines keep it disabled with «No disponible en …».
- The row menu entry follows main's greyed-out convention (hint and tooltip); no shared menu component changes.
- CHANGELOG entry under [Unreleased].

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…will open, kept on its monitor, and Esc leaves the tab where it was (#108)

Sprint 2.3 failed here: dragged into the lower half of a large screen, the fixed
1280x820 outline hung off the bottom with its label, while the dropped window
opened clamped somewhere else. Now one frame (drop_frame in windows.rs) places
both: the full window's size at the scale of the monitor under the pointer, its
corner at the drop offset, kept inside that monitor's work area. The outline
resizes when it crosses to a monitor with another scale, and a dropped window
is sized and placed again after landing there.

The label sits at the top of the outline, next to the pointer. On Windows the
core watches the mouse button and Esc while the system runs the drag: the
outline goes the moment the button is let go or Esc is pressed, the window
opens where the button was let go (not where the pointer is a moment later),
and Esc no longer opens a window.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…rir en ventana nueva» and describe the outline (#108)

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
… the tail of the current log with one log switch, as a DBA would, instead of failing on the server's state (#123)

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…n Entorno row and grouped settings, as approved (#111)

- Context menus take submenus (MenuItem.submenu): arrows, → / Intro open one with
  its first item active, ← / Esc go back, the mouse opens one by resting on its
  row; a submenu sits beside its row and flips or moves to stay in the window
  (src/menu.ts, dev/menu-check.ts).
- The top bar loses its explorer button (the stripe and Alt+1 stay). The right
  panel's header keeps only Valor / Registro, or the title of the tool open
  (Historial, Biblioteca, Variables, Asistente IA), which open from the stripe
  (Variables gets its button there) and their shortcuts.
- One «Importar conexiones ›» (fichero de Celer · DBeaver · DbVisualizer) in
  Nuevo and in the explorer's options menu; the background menu keeps only the
  order. Palette commands for each import and for exporting every connection.
- Tables and views: open, structure, «Ver DDL» and diagram on top, then
  «Generar SQL ›» (eight entries for a table, four for a view) and
  «Comparar ›»; «DDL en consola» is the DDL tab's «Abrir en consola», which now
  opens on the table's database.
- Cell menu: Copiar, Copiar con cabeceras, «Copiar como ›» (CSV, JSON,
  Markdown, XML, INSERT, IN, WHERE) and «Filtrar ›», each running what its
  entry ran before.
- Connection dialog: environment, Solo lectura and Color framed as one Entorno
  row below the main block.
- The palette's «Editor de tema…» opens its section in Ajustes.
- The welcome tour, Gib's tip, the e2e helpers (menu paths through submenus)
  and dev/layout-check.mjs (the whole walk on the browser preview) follow.

The split button Siguiente página / Cargar todo was not done, as decided.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
… ring, and the grouped settings nav fits without scrolling (#110, #111)

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…enus, the reorganised menus and settings, and the tear-off outline (#107, #108, #110, #111)

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…ont or an out-of-range number is refused with what is wrong instead of being imported silently cleaned (#107)

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…unsaved, so leaving, importing or switching asks only when there is something to lose (#107)

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…he open results, so text no longer spills into the next column until the query is rerun; hand-resized columns keep their width (#107)

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…showed: its size is measured and asked again (the first resize of a new undecorated window added a title bar, 30 px taller) and its corner moved by the invisible shadow frame (8 px off), so a drop near the bottom no longer leaves it under the taskbar (#108)

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…ion Windows shows the move cursor instead of its no-drop sign on top of «Soltar para abrir en una ventana nueva» (#108)

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…e he is there, instead of in whichever window hosts him (#108)

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…n use, the one a click changes, not the global one (#107)

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…er window (#108)

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
… animation, so «Esc closes the menu» no longer fails on a cold run (#111), and the layout check sees an untouched new theme leave the editor without asking (#107)

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…pop-in still scales it down, so one opened at the window's right edge no longer overflows it (#110)

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…sing the start-up guide, so on a cold dev server the guide no longer opens mid-check and takes the Esc meant to close the menu (#111)

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…its tab (the offset is scaled to the monitor), and moved onto a monitor with another scale it is placed before it is sized and measured again, so the system's rescale no longer leaves it a different size from the window that opens (#108)

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…alid rows on every engine (#172)

A dialog from the table's menu (explorer and tab) and the palette proposes a generator per column from its type
and its Spanish or English name, reads the table's unique keys, CHECK lists and ranges, native enums, identity
columns and foreign keys (with a sample of the referenced keys) per dialect, and generates reproducible rows
(seed) with a 20-row preview. The rows go in through the import's pipeline (now a shared insertInTransaction: own
session, one transaction, progress, rows/s, Cancelar rolls back; read-only refuses, production asks), or to a
script in a new console (a file past 10,000 rows), or to CSV / JSON. An empty referenced table is offered to be
filled first. Import: 0x… values go into binary columns as bytes.

Pure logic in src/testData*.ts (dev/testdata-check.ts); per-dialect SQL checked live on PostgreSQL, MySQL,
MariaDB, SQL Server, Informix (DRDA and JDBC), SQLite and ODBC by engine_tests.rs testdata_* (10,000 rows, count,
FK integrity, uniqueness, lists and ranges, a row read back, rollback, another session's open transaction), plus
100,000 rows into PostgreSQL; dev/testdata-e2e-check.mjs drives the dialog in the desktop app.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…ancelled (#172)

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…site primary keys get a value from the test data generator, the grid and the import (#172)

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
e-omunoz and others added 16 commits October 13, 2026 08:01
… ones) runs the connection's startup script first, as its sessions do, and a read-only connection refuses a script that writes; engine test on Informix JDBC with a broken and a good script (#180)

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…r instead of a bare comma without BOM, so accented test data opens correctly in Excel (#204)

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
… saving a test-data script or export of tens of MB no longer freezes every window (#207)

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…gging Celer turned on, and only the last of them turns it off, instead of the first to end turning it off under the others; engine test with two overlapping reads on Informix JDBC (#205)

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…a read whose newest log part the CDC API did not have ready yet, and the large-schema keys bound is timed again when the first read hits a catalog cold from the container restart, so the Informix gate no longer fails intermittently (#206)

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
… existing values of unique columns are indexed once per model, so editing the dialog's settings on a table with 200,000 rows no longer freezes the window, and closing the dialog releases them (#184)

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…ough an append-capable write_text_file and a new finish_text_file, so a 1,000,000-row JSON of a wide table no longer fails with «Invalid string length», and a cancelled run leaves no partial file (#185)

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…g inserted no longer cancels the run and rolls back the transaction; only the footer's Cancelar / Detener stops it, as in the export dialog (#186)

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
… back as the user left it (rows, seed, language, output, view and each still-existing column's generator) instead of resetting it (#208)

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…y that to the other windows, instead of a console's full SQL text through the core (#209)

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…eme editor, tear-off outline, motion, reorganised menus, Informix row history, test data generator) instead of listing #107/#108/#110/#111/#123 as held back (#187)

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…, plus the editor, result compare, typed import and windows e2e scripts, and STATUS states the current test counts (282 / 35 / 28) (#212)

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…r, Informix row history and tear-off outline, and list Alt+1 and F6 / Shift+F6 among the shortcuts (#213)

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…uild, so they show the top bar without the explorer button and the stripe with its Variables button (#214)

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
… gets its Fixed entry, since it changes the structure view, schema comparison and diagram as well as the test data dialog (review follow-up)

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
@e-omunoz

Copy link
Copy Markdown
Owner Author

Release gate: not passed

Checked on head a29118ce036ce9885c9f6d630c062aa80226cae5. Gate 5 fails. The Engines (SQL Server, Informix) run is red on this SHA, with the same failure on both attempts. Every other gate passes. Nothing was merged or released.

# Gate Result Evidence
1 Acceptance criteria on every engine ✅ Tester rounds pass for #107, #108, #110, #111, #123 and #172 on every applicable engine. The one criterion the tester left untested was #108 at 150 % scaling, so I checked it here with the release portable of this SHA. I set the display to 150 % (the work area went to 1368 px) and did a real mouse drag of a tab outside the window. The outline was the window's size (about 1267×807 logical; the window is 1280×820) and stayed on screen, with «Soltar para abrir en una ventana nueva»; Esc cancelled. A real drop at the bottom-left corner opened win-2 at 0,92 at 1280×820 logical, which fits inside the 912 px work area, in the same place as the outline. Then I set the scaling back to 100 %. There is no second monitor on this machine, so that part was not tested (the criterion accepts that). Screenshots: review-out/gate/smoke-a29118c/shots-150/.
2 Macro review ran, final verification green ✅ verify.green: true, with no failures or regressions.
3 No open review issue with sev:critical/high ✅ The only high issue is #174. Its fix is 48be1ed, which I read: MySQL/MariaDB catalog literals now double backslashes, and a catalog query the splitter reads as more than one statement is not sent. The PR has Closes #174. #175–#214 are medium/low; every one has a fix commit and a Closes line.
4 Local checks on the head SHA ✅ dev/wsl/engines.sh test printed ALL ENGINES PASS (EXIT 0): PG, MySQL 8.4, MariaDB 11.4, MSSQL, IFX-DRDA, IFX-JDBC (incl. informix_jdbc_row_history), SQLite, ODBC, SSH; 281 lib tests plus engine_tests 62 and MariaDB. tsc -p installer and npm run build were OK. dev\check-all.ps1 gave EXIT 0 with all 60 steps ok (typescript, all 35 logic checks, layout, rust unit tests, e2e, test data e2e, several windows, restore…). Logs: review-out/gate/engines-a29118c.log, check-all-a29118c.log.
5 GitHub Actions green on the head SHA ❌ CI 38361257456 ✅ · CI (PR) 38361202530 ✅ · Release 38361264764 ✅ · Engines 38361260998 ❌, on attempt 1 and on the rerun (attempt 2). Both attempts fail engine_tests::informix_jdbc_row_history at src/engine_tests.rs:3265: ask_whole("rh_t", &id2) returns no events where it expects ["insert", "delete"]. The log-switch retry for #206 does not cover this case, because the history comes back empty without the «se detuvo» / «todavía no están disponibles» reasons it retries on. The test passes on the local WSL Informix but fails every time on the CI server, so #206 is not fixed for CI. Someone needs to find out whether the empty history is the test's timing or the feature returning an empty, non-partial history for a deleted row. A deleted row's history coming back empty without saying it is partial would break the #123 rule that "a read cut short never says the row had no changes".
6 Smoke test of the built packages ✅ sha256sum -c SHA256SUMS.txt: all 6 files OK. I started Celer-Portable-Windows.exe with --data-dir D:\celer-gate-a29\data. It downloaded the IBM CLI and Informix JDBC drivers through its own commands, then opened a console and ran a query on PG ✅ MySQL ✅ MariaDB ✅ MSSQL ✅ IFX-DRDA ✅ IFX-JDBC ✅ SQLite ✅ ODBC (SQL Server DSN) ✅, each checked both in the UI and through a core session (smoke.json, one screenshot per engine). The installer ran with --silent --sandbox D:\celer-gate-a29\sbx. It created the Celer-sandbox uninstall key, Celer.sandbox/.celersandbox and shortcuts carrying --data-dir. The installed app started and ran queries on PG, MSSQL and SQLite. The sandbox uninstall removed the key, the association and the shortcuts. The real install (key Celer) and %APPDATA%\es.celer.app were left alone (last write 10/10). Screenshots: review-out/gate/smoke-a29118c/.
7 Mergeable, CHANGELOG, docs ✅ mergeable: MERGEABLE (state UNSTABLE only because of the red Engines check). [Unreleased] covers #107/#108/#110/#111/#123/#172 and every user-visible review fix. GUIA has «Datos de prueba», «Historial de la fila» and «Editor de tema»; both READMEs cover the 2.4 features.
8 No open product decision on a merged issue ✅ #123's only decision is settled in the thread: Informix only. No issue the PR closes carries a decision label.

What is needed to pass

  1. Make informix_jdbc_row_history pass on the CI Informix. First find out why the deleted row id2 comes back with no events and no partial reason. If the feature can return an empty history without marking it partial, that is a defect in Informix: see what a row looked like at a past moment, from the logical logs (feasibility first) #123 and has to be fixed in the code, not only in the test.
  2. Run engines.yml again on the new head and get it green. Then this gate can be re-run; the other gates passed on this SHA.

Version bump once it passes: minor (features #107, #108, #110, #111, #123, #172).

e-omunoz and others added 2 commits October 13, 2026 11:27
…ed and no longer remembered for the session, and the engine test stops racing them (#123, #206)

About 90 s after Informix starts, the sysadmin scheduler's start-up sensors (mon_sysenv, mon_config_startup)
write a burst of transactions to the log, and the CDC reader stops where it meets them for minutes. The engine
test restarted the container first, so on the slower CI runner its reads ran into that window: every later
history was partial, ask_whole's retries jumped to new logs the reader did not deliver, a cdc_deactivatesess
went unanswered, and since the test reused one connection for every read the next cdc_opensess was refused
(-83715). That history was "unavailable" with its reason, but the assertion only printed its empty op list.

- rowhistory: the stop message names both causes and says to read again later; a wall found by one read is
  trusted for five minutes only (the reader was measured to read past it again).
- engine test: waits for the start-up sensors, reads a short probe history and restarts the container only when
  the reader does not give it whole (no restart of a shared server otherwise), reads each history on its own
  connection as the app does, refuses an unavailable history with its reason, prints the history on failure, and
  runs the two simultaneous reads last.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…tart-up tasks; the engine test no longer restarts the shared server (#123, #206)

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
@e-omunoz

Copy link
Copy Markdown
Owner Author

Release gate: not passed

Checked on head 637c34eb4cb35bbc384e06ba4e84e4ffa3a0c75a. Gate 1 fails. While I was checking #123 in the desktop app, I found a new row-history defect. Gate 5 now passes: Engines is green on this SHA. Every other gate passes. Nothing was merged or released.

# Gate Result Evidence
1 Acceptance criteria on every engine ❌ The code changed since a29118c is only rowhistory.rs, engine_tests.rs, CHANGELOG and GUIA, so the tester and gate evidence for #107, #108, #110, #111 and #172 still stands. For #123 I used slot 1 (run-desktop.ps1 -Slot 1 -Ref origin/feat/integration-2.4.0, worktree at 637c34e) with the Informix JDBC connection, and read the history of a table row from the row menu. What works: a fresh table gave a complete history: insert and 2 updates, not partial, full row logging switched on only for the read and then off again. A read made while another session was writing the log stopped and says «Historial parcial» with «todavía no están disponibles». The next read met that stop and says partial with the new wording: «…se detuvo en la posición 602:0x9c21c del log 602… el servidor no entrega nada más de él por ahora. Pasa tras una lectura que llegó al final del log mientras se escribía … y en los minutos después de arrancar Informix … vuelve a leer más tarde.» It listed 10 of the 11 changes. Screenshots: review-out/gate-637c34e/rh/rh-1-complete.png, rh-loop-13.png. New defect (not filed yet, I suggest sev:medium): after Informix fast recovery replays a table's CREATE TABLE, the row history drops that table's real changes. Fast recovery runs after an unclean stop that comes before the next checkpoint. Recovery stamps sysmaster:sysptnhdr.created with the time of the replay. rowhistory.rs uses that column as the table's creation time (Builder::created, around line 1337) and discards every earlier commit with the note «eran de una tabla borrada que tenía el mismo número de partición». Repro: create a table, insert, update, docker kill -s KILL celer-ifx before a checkpoint, then start it. created moved from 20:33:37 to 20:34:06, the recovery time. The app then showed «Historial parcial: Ningún cambio de esta fila…», «Se descartaron 2 cambios…» and «Sin cambios de esta fila en los logs leídos», although the insert and the update are in log 600 (rh-3-recovered-table.png). A clean restart does not do this, and neither does switching full row logging on or off. This breaks #123's criteria that each partial history gives «the right reason» and that the dialog lists the row's changes, because the reason it gives is false.
2 Macro review ran, final verification green ✅ verify.green: true from the macro review, as before. The only change since then is fix/rowhistory-ci, which is covered by gates 4 and 5.
3 No open review issue with sev:critical/high ✅ The only open high issue is #174. Its fix, 48be1ed, is an ancestor of this SHA: MySQL/MariaDB catalog literals go through sqlLiteral(…, "mysql"), which doubles backslashes, and a catalog query the splitter reads as more than one statement is not sent. The PR has Closes #174. #175–#214 are medium/low and the PR closes all of them.
4 Local checks on the head SHA ✅ In the WSL scratch worktree at 637c34e, bash dev/wsl/engines.sh test printed ALL ENGINES PASS. That run covers all 35 dev/*-check.ts and tsc -p .. It ran 281 lib tests (1 ignored) on PG, MySQL 8.4, MSSQL, IFX-DRDA, IFX-JDBC (incl. informix_jdbc_row_history), SQLite, ODBC and SSH, plus 29 on MariaDB 11.4. tsc -p installer was OK and npm run build was OK. On Windows, the checkout C:\Users\oscar\Projects\celer is at 637c34e and clean, and dev\check-all.ps1 (JAVA_HOME jdk-21) ran in one full run with all 60 steps ok, EXIT 0. Logs: review-out/gate/engines-637c34e.log, check-all-637c34e.log.
5 GitHub Actions green on the head SHA ✅ CI (PR) 38378977073 ✅ · CI 38379042234 ✅ · Engines 38379046419 ✅ · Release 38379051870 ✅. All four ran on this SHA.
6 Smoke test of the built packages ✅ The artifacts of Release 38379051870 (headSha 637c34e) pass sha256sum -c SHA256SUMS.txt for all 6 files. I started Celer-Portable-Windows.exe --data-dir D:\celer-gate-637\data, with connections from devdata-connections.mjs and the drivers copied in. It connected and ran a query in a console and in a core session on PG ✅ MySQL ✅ MariaDB ✅ MSSQL ✅ IFX-DRDA ✅ IFX-JDBC ✅ SQLite ✅ ODBC (SQL Server) ✅. I then ran the installer with CELER_SETUP_SANDBOX=D:\celer-gate-637\sbx (empty) and --silent --desktop --start-menu --associate-sql --launch. It created the key Celer-sandbox, .celersandbox→Celer.sandbox and shortcuts carrying --data-dir …\sbx\data. The app started on --data-dir D:\celer-gate-637\sbx\data, and when relaunched from its shortcut it ran queries on all 8 engines. Uninstalling with the registered QuietUninstallString (--uninstall --silent --sandbox …, no env var) exited 0 and removed the key, the association and the shortcuts. Before and after snapshots of the real install are identical: key Celer 1.3.1, shortcuts, no .sql association, %APPDATA%\es.celer.app (newest file 10/10), %LOCALAPPDATA%\es.celer.app, and the hash of Programs\Celer\celer.exe. Screenshots: review-out/gate-637c34e/smoke/.
7 Mergeable, CHANGELOG, docs ✅ mergeable: MERGEABLE. [Unreleased] has the new #206 entry and covers #107/#108/#110/#111/#123/#172 and the review fixes. GUIA covers «Datos de prueba», «Editor de tema» and «Historial de la fila», including the new start-up case. README.md and README.es.md cover the test data generator, the theme editor and the row history, and the shortcut tables have F6/Shift+F6 and Alt+1. STATUS.md is on 2.4, and its count of 282 lib tests and 35 checks matches the code. SECURITY.md contradicts none of these features: the files they save fall under «the files you open, save or export».
8 No open product decision on a merged issue ✅ #123: «Decision taken (owner delegated, 2026-10-13): Informix only». #111: «Decisión (el propietario delegó el criterio, 2026-10-12)», approved except the split button, which was not built. No closed issue has a decision label.

What is needed to pass

  1. Settle the fast-recovery defect in Informix: see what a row looked like at a past moment, from the logical logs (feasibility first) #123's row history. Either fix it, by not trusting sysptnhdr.created once fast recovery may have rewritten it, or by not calling those commits "from a dropped table" unless that is shown. Or file it as a review issue (I suggest sev:medium) and defer it with a written reason, if the owner accepts Informix: see what a row looked like at a past moment, from the logical logs (feasibility first) #123 shipping with it.
  2. Then re-run the gate on the new head. Gates 2–8 passed on this SHA.

Version bump once it passes: minor.

e-omunoz and others added 2 commits October 13, 2026 23:39
… keeps its own changes instead of having them dropped as a dropped table's (#123)

After an unclean stop before the checkpoint that followed a CREATE TABLE, fast recovery replays it on start and
stamps sysptnhdr.created with the replay time. The history took that as the creation time, skipped the logs before
it and discarded the row's real insert and updates as «de una tabla borrada que tenía el mismo número de partición».
A creation time between the server's start (sh_boottime) and its on-line mode (the Startup checkpoint in
syscheckpoint, else the first internal session) is now taken as the replay's: nothing is skipped or discarded by
time, the changes are shown, and the history is partial with the true reason. A part of the log the CDC API then
refuses (-83800 on the first insert, seen once) is said as a gap, and no gap lets the history claim where the row
began. The engine test reproduces it with docker kill, only with CELER_INFORMIX_KILLABLE=1.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…on time was rewritten by Informix fast recovery, with the true partial reason (#123)

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
@e-omunoz

Copy link
Copy Markdown
Owner Author

Release gate: not passed

Checked on head 70b8987d959d4e1b57194d93350d2b1ebb7aa0da. Gate 1 fails. The fast-recovery fix (5b8fb74) removed the false «tabla borrada» reason, but it causes a new error. On a recovered table, the row history now lists changes from earlier, dropped tables that had the same partition number as if they were this row's changes. The fix's own engine test fails 5 times out of 7. Gate 7 has two small docs gaps. Every other gate passes. Nothing was merged or released.

# Gate Result Evidence
1 Acceptance criteria on every engine ❌ The only code change since 637c34e is rowhistory.rs and engine_tests.rs. The earlier evidence for #107, #108, #110, #111 and #172 stands. #123 on my own killable container (celer-gate-ifx, informix-developer-database, ports 19088/19089, celer + syscdcv1 created as in engines.yml). I ran informix_jdbc_row_history_after_fast_recovery with CELER_INFORMIX_KILLABLE=1 CELER_INFORMIX_CONTAINER=celer-gate-ifx 7 times. In every run fast recovery rewrote sysptnhdr.created (e.g. 1791927826 → 1791927838). Runs 1 and 2 passed: insert «antes» and update «despues» were shown, marked partial with the true reason («…su recuperación rápida rehízo la tabla… Los 2 cambios…»), and no «tabla borrada». Runs 3–7 failed at assert_eq!(ops, ["insert", "update"]). The table got a partition number that an earlier, dropped table had used. Replay mode reads every log with cutoff 0, so that table's changes still in the logs were listed as this row's changes: Los 4 / 6 cambios (the dropped rh_rec from earlier runs), and Los 37 / 39 / 41 cambios (records with v='' and truncates at 21:42:45–56, from the server's set-up 4 minutes before the table existed). Desktop app: slot 1 (run-desktop.ps1 -Slot 1 -Ref origin/feat/integration-2.4.0, worktree at 70b8987), JDBC connection to that container. I created gate_rh_app after onmode -c, inserted 1 'antes', updated it to 'despues', killed the server and started it again (created 1791928695 → 1791928706, boot 1791928699). «Historial de la fila» shows «Historial parcial: …recuperación rápida… Los 8 cambios de esta fila anteriores a esa hora se muestran como de esta tabla…». It lists 4 «Alta» for id 1 with no delete in between: 23:45:20, 23:49:09 and 23:51:59, all from dropped tables, and 23:58:15, the real one. That is 6 changes this row never had, under the header «sin suponer nada». Screenshot: review-out/gate-70b8987/rh/rh-app-recovered-1.png. This breaks #123's rule «no guessing, no invented values» and the criterion that the dialog lists that row's changes. The caveat in the partial box does not make the listed changes true. It is not the deferred #215 case either: #215 covers a second restart and a drop + re-create in the same second, but here the dropped tables are minutes older. Normal read on the shared engine (Informix JDBC, celerdemo.gate_rh_70b, dropped afterwards): complete, not partial, with insert + 2 updates. Full row logging was switched on only for the read and then off again (rh/shared-step3.png). Test logs: review-out/gate/fastrec-70b8987-run1..7.log (WSL checkout).
2 Macro review ran, final verification green ✅ verify.green: true from the macro review, as before. The only change since then is fix/rowhistory-recovery, which gates 1, 4 and 5 cover.
3 No open review issue with sev:critical/high ✅ The only open high issue is #174. Its fix, 48be1ed, is an ancestor of this SHA, and the PR has Closes #174. #215 and #216 are sev:low and each has a written Deferred reason (owner delegated). All other open review issues are medium/low and closed by this PR.
4 Local checks on the head SHA ✅ In the WSL scratch worktree at 70b8987, bash dev/wsl/engines.sh test printed ALL ENGINES PASS and exited 0. That run covers all 35 dev/*-check.ts and tsc -p .. It ran 284 lib tests (283 passed, 1 ignored), including engine_tests 64 ok and rowhistory 18 ok, on PG, MySQL 8.4, MSSQL, IFX-DRDA, IFX-JDBC, SQLite, ODBC and SSH, plus 29 on MariaDB 11.4. informix_jdbc_row_history passed. celer-ifx was not restarted: it had been up 36 minutes when the run ended, so the #216 path was not hit. tsc -p installer was OK and npm run build was OK. On Windows, the checkout C:\Users\oscar\Projects\celer is at 70b8987 and clean, and dev\check-all.ps1 (JAVA_HOME jdk-21) ran in one full run with all 58 steps ok, EXIT 0. Logs: review-out/gate/engines-70b8987.log, check-all-70b8987.log, npm-build-70b8987.log, tsc-installer-70b8987.log (WSL checkout).
5 GitHub Actions green on the head SHA ✅ CI (pull_request) 38472415384 ✅ · Engines 38472412831 ✅ · Release 38472417148 ✅. All three ran on headSha 70b8987. CI runs on push only for main.
6 Smoke test of the built packages ✅ The artifacts of Release 38472417148 pass sha256sum -c SHA256SUMS.txt for all 6 files, and the Windows exes are byte-identical to the celer-windows job's. I started Celer-Portable-Windows.exe --data-dir D:\celer-gate-70b\data, with connections from devdata-connections.mjs and D:\celer-devdata\drivers copied in. It connected and ran a query in a console and in a core session on PG ✅ MySQL ✅ MariaDB ✅ MSSQL ✅ IFX-DRDA ✅ IFX-JDBC ✅ SQLite ✅ ODBC (SQL Server) ✅. I then ran the installer with CELER_SETUP_SANDBOX=D:\celer-gate-70b\sbx (empty) and --silent --desktop --start-menu --associate-sql --launch. It exited 0 and created the key Celer-sandbox, .celersandbox→Celer.sandbox and shortcuts carrying --data-dir …\sbx\data. The app started on that folder. Relaunched from its desktop shortcut, it ran queries on all 8 engines. Uninstalling with the registered QuietUninstallString (--uninstall --silent --sandbox …, env var removed) exited 0 and removed the key, the classes and the shortcuts. Only uninstall.exe was left, which is by design (setup.rs:768). Before and after snapshots of the real install are identical: key Celer 1.3.1, shortcuts, no .sql association, %APPDATA%\es.celer.app (only file metadata read), %LOCALAPPDATA%\es.celer.app, and the hash of Programs\Celer\celer.exe. Evidence: review-out/gate-70b8987/smoke/ (shots, shots-installed, before/after.json).
7 Mergeable, CHANGELOG, docs ❌ (minor) mergeable: MERGEABLE. [Unreleased] has the new fast-recovery entry, and DRIVERS.md describes it. Gaps: (a) docs/GUIA.md lists every reason a row history is shown as partial (lines 313–319) but not the new one, the recovery after an unclean stop. (b) STATUS.md says cargo test --lib has 282 tests, but this SHA has 284. README.md, README.es.md and SECURITY.md are unchanged and still match.
8 No open product decision on a merged issue ✅ #123: «Decision taken (owner delegated, 2026-10-13): Informix only». #111: «Decisión (el propietario delegó el criterio, 2026-10-12)». #215 and #216 each have a written deferral. No decision_needed label exists.

What is needed to pass

  1. Informix: see what a row looked like at a past moment, from the logical logs (feasibility first) #123, the recovered-table history. A recovered table's history must not list a dropped table's changes as this row's. A hint: fast recovery replays only from the last checkpoint before the crash, so the table's real creation is after that checkpoint. Use that checkpoint (or the log position where recovery started) as the lower bound instead of 0. Where that cannot be told, show the changes before the replay as unattributable rather than as the row's. At the least, an «Alta» of a key that is still live should end what is attributed to this table. File it as a review issue (I suggest sev:medium, area:drivers-enterprise). Then either fix it, or defer it with a written reason and relax the engine test to match. The test as written fails 5 out of 7 times.
  2. Docs. Add the fast-recovery partial reason to GUIA's list, and update the STATUS test count to 284.
  3. Re-run the gate on the new head. Gates 2–6 and 8 passed on this SHA.

Version bump once it passes: minor.

e-omunoz and others added 2 commits October 14, 2026 00:45
…n time, changes before it are left out with the true reason instead of a dropped table's being listed as the row's (#123)

5b8fb74 read every log from the start when sysptnhdr.created fell between the server's start and its on-line mode,
and listed every earlier commit as the table's: a dropped table that had the same partnum (even the server's set-up
records) showed up as the row's own inserts and updates. The read now keeps the normal cutoff at that creation time
and the history says why it is partial: «El servidor reescribió la hora de creación de esta tabla al recuperarse de
una parada brusca (…): los cambios anteriores a esa hora no se pueden atribuir a esta tabla con seguridad y no se
muestran.» It never calls them a dropped table's, never says the row had no changes or where it began; the dialog's
empty state says no change was read since that time (new History.shownFrom). Changes after the recovery are shown.
The detection, the -83800 gap and the start-up log note are kept. The engine test creates and drops tables of the
same name first (their partnum is reused), kills the server and checks that nothing before the recovery is listed and
that an update after the restart is.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…y shows nothing it cannot attribute and says why (#123)

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
@e-omunoz

Copy link
Copy Markdown
Owner Author

Release gate: passed

Checked on head eed73c0e59352211d6e882f002da50f7dfc3c2fb. All 8 gates pass. The two failures from 70b8987 are fixed. Gate 1: after a fast recovery, the row history no longer shows any dropped table's change as the row's. Gate 7: the GUIA list of partial reasons and the STATUS test count are now correct. Nothing was merged or released by this check.

# Gate Result Evidence
1 Acceptance criteria on every engine ✅ Since 70b8987 the only change is fix/rowhistory-recovery-2 (131a74b): rowhistory.rs, rowHistoryView.ts, types.ts, demo.ts, rowhistory-check.ts, engine_tests.rs, plus docs. The earlier evidence for #107, #108, #110, #111, #172 and #174–#214 still stands, because nothing outside row history changed. #123 on my own killable container (celer-gate-ifx, informix-developer-database, ports 19088/19089, named volumes, celer + celerdemo + syscdcv1 as in engines.yml): I ran informix_jdbc_row_history_after_fast_recovery with CELER_INFORMIX_KILLABLE=1 CELER_INFORMIX_CONTAINER=celer-gate-ifx 6 times, and all 6 passed. In every run the three dropped rh_rec tables had the same partnum as the new table (1049188 or 1049178), and fast recovery rewrote created (e.g. 1791931746 → 1791931757). After the restart the read showed no events, with one partial reason: «El servidor reescribió la hora de creación … no se muestran». An UPDATE made after the restart was then shown alone (update … 'tras'). There was no «tabla borrada», «Se descartaron», «Ningún cambio» or «ya existía». Desktop app, slot 1 (run-desktop.ps1 -Slot 1 -Ref origin/feat/integration-2.4.0, worktree at eed73c0), JDBC connection «Gate IFX JDBC» to that container. I reproduced last gate's failure case: gate_rh_eedapp was created and dropped 3 times with id 1 rows, all 3 on partnum 1049188. Then I ran onmode -c, created the table again (same partnum), inserted 'antes', updated it to 'despues', ran docker kill -s KILL and started the server again (created 1791932334 → 1791932343, boot 1791932336). «Historial de la fila» showed no change at all: none from before the recovery and none from the dropped tables. Its partial box held only «El servidor reescribió la hora de creación de esta tabla al recuperarse de una parada brusca (2026-10-13 22:59:03 UTC): los cambios anteriores a esa hora no se pueden atribuir a esta tabla con seguridad y no se muestran.» The empty state read «No se leyó ningún cambio de esta fila posterior al 14/10/26, 0:59:03; los anteriores no se muestran (ver arriba).» After an UPDATE … 'tras', «Volver a leer» listed exactly one «Modificación» (despues → tras) with the same partial reason. Screenshots: review-out/gate-eed73c0/rh/rh-recovered.png, rh-recovered-after-update.png. Normal read on the shared Informix JDBC (celerdemo.gate_rh_eed, dropped afterwards): complete, with no partial box, showing the insert and 2 updates. Full row logging was on only during the read and was off again before the drop (rh-shared.png). The extra connection, its credential and its tabs were removed: connections.json is identical to before, and cmdkey has no entry for it.
2 Macro review ran, final verification green ✅ verify.green: true from the macro review. Gates 1, 4 and 5 cover the two row-history fixes since then.
3 No open review issue with sev:critical/high ✅ The only open high issue is #174. Its fix, 48be1ed, is an ancestor of this SHA, and the PR has Closes #174. #215 and #216 are sev:low and each has a written Deferred reason (owner delegated). Every other open review issue is medium or low, and the PR closes it.
4 Local checks on the head SHA ✅ In a WSL scratch worktree at eed73c0, bash dev/wsl/engines.sh test printed ALL ENGINES PASS and exited 0. It ran 284 lib tests (283 passed, 1 ignored), including engine_tests 64 ok and rowhistory 18 ok, on PG, MySQL 8.4, MSSQL, IFX-DRDA, IFX-JDBC, SQLite, ODBC and SSH, plus 29 on MariaDB 11.4. celer-ifx was not restarted (up 2 h afterwards). All 35 dev/*-check.ts passed, tsc -p . OK, tsc -p installer OK, npm run build OK. On Windows, the checkout C:\Users\oscar\Projects\celer is at eed73c0 and clean. dev\check-all.ps1 (JAVA_HOME jdk-21) ran in one full run with 58 steps ok, EXIT 0, including cargo 283 passed / 1 ignored. Logs: review-out/gate/engines-eed73c0.log, check-all-eed73c0.log, npm-build-eed73c0.log, tsc-installer-eed73c0.log, eed/fastrec-run1..6.log (WSL checkout).
5 GitHub Actions green on the head SHA ✅ CI (pull_request) 38479995561 ✅ · Engines 38479995025 ✅ · Release 38479998129 ✅ (macOS, Windows and Linux packages, release job). All three ran on headSha eed73c0.
6 Smoke test of the built packages ✅ The artifacts of Release 38479998129 pass sha256sum -c SHA256SUMS.txt for all 6 files, and the Windows exes are byte-identical to the celer-windows job's. I started Celer-Portable-Windows.exe --data-dir D:\celer-gate-eed\data, with connections from devdata-connections.mjs and D:\celer-devdata\drivers copied in. It connected and ran a query in a console and in a core session on PG ✅ MySQL ✅ MariaDB ✅ MSSQL ✅ IFX-DRDA ✅ IFX-JDBC ✅ SQLite ✅ ODBC (SQL Server) ✅. I then ran the installer with CELER_SETUP_SANDBOX=D:\celer-gate-eed\sbx (empty) and --silent --desktop --start-menu --associate-sql --launch. It exited 0 and created the key Celer-sandbox («Celer (prueba del instalador)» 2.3.0), .celersandbox → Celer.sandbox and shortcuts carrying --data-dir …\sbx\data. The app started on that folder. Relaunched from its desktop shortcut, it ran queries on all 8 engines. Uninstalling with the registered QuietUninstallString (--uninstall --silent --sandbox …, env var removed) exited 0 and removed the key, the classes and the shortcuts. Only uninstall.exe was left, which is by design. Before and after snapshots of the real install are identical: key Celer 1.3.1, shortcuts, no .sql association, %APPDATA%\es.celer.app (file metadata only), %LOCALAPPDATA%\es.celer.app, and the hash of Programs\Celer\celer.exe. Evidence: review-out/gate-eed73c0/smoke/ (shots, shots-installed, before/after.json, install/postinst/uninstall.txt).
7 Mergeable, CHANGELOG, docs ✅ mergeable: MERGEABLE, mergeStateStatus: CLEAN, and there are no conflicts with main. In [Unreleased], the fast-recovery entry now describes the conservative behaviour. docs/GUIA.md lists the new partial reason (changes before the rewritten time are not shown, later ones are). docs/DRIVERS.md matches the code. STATUS.md gives 284 tests (cargo test --lib on this SHA: 284), 35 .ts checks and 28 .mjs scripts, and all three counts match. README.md, README.es.md and SECURITY.md are unchanged since the last gate and still match.
8 No open product decision on a merged issue ✅ #123: «Decision taken (owner delegated, 2026-10-13): Informix only». #111: «Decisión (el propietario delegó el criterio, 2026-10-12)». #215 and #216 each have a written deferral. No decision_needed label exists.

Observation (not blocking, older than this fix): on a table whose read starts after older logs that are still on disk, the range line says «Lo anterior ya no está en los logs». The note just below gives the real reason («La lectura empieza en el log 6: los anteriores se llenaron antes…»). This line also shows for normal tables and was in the earlier passing gates. It could be reworded in a later low-severity issue.

Version bump: minor (2.4.0 has features).

@e-omunoz
e-omunoz marked this pull request as ready for review October 13, 2026 23:16
@e-omunoz
e-omunoz merged commit c784dbd into main Oct 13, 2026
6 checks passed
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment