fix(ci): point ui-preview planner at the add-row menu's aria-label

2026-10-19 16:16:45 +00:00 · 2026-10-09 13:29:41 +08:00
parent 2f56bc7802
commit b3acccac7c
1 changed files with 1 additions and 1 deletions
@@ -225,7 +225,7 @@ async function codeContext(diff, src) {

 const GUIDE = `You plan a short screen recording of dbx, a database GUI (shown here in a browser at ${VIEW.width}x${VIEW.height}). It runs for real on a sandbox machine: the web build of the app against a dbx-web backend with password protection off, and made-up data beside it (seed.mjs) so the pages that need data have some. That data: two saved SQLite connections, "商店 Shop (SQLite)" (tables customers, products, orders, audit_log, a view v_daily_revenue, foreign keys and a trigger) and "分析 Analytics (SQLite)" (events with 5000 rows — the grid pages — and daily_stats with 90 days); a small sample.csv (name,category,price,stock) for import dialogs; a saved-SQL folder "常用查询" with three files; and query history over the last three days across both connections, real executions mixed with seeded entries, a few failures among them. The UI language is ${LOCALE}. What it doesn't have: any server database (MySQL, PostgreSQL, Redis, MongoDB… none are connected — connection dialogs can be opened but not completed), no AI provider configured, no plugins or JDBC drivers installed, no SSH tunnels.

-dbx is a single-page app with no URL routing: every recording starts from the page as it opens. The left sidebar is a tree that opens one level at a time: a click on a connection lists its databases (SQLite shows one, "main 默认库"), a click on that database node lists its tables, and a click on a table opens its data grid — the query editor, table data, structure, ER diagram are that connection's tabs. A double-click on a connection instead opens a database-browser tab, so working down the tree is how a table is reached; the outline you are given shows the app at rest, connections collapsed, so a node the plan needs that isn't in it has to be expanded into view first, not looked for on other pages. The search box above the tree (placeholder 搜索...) can reveal a node without that clicking, but while it holds text it filters the tree — a blank tree with text in the box means no match, and its text survives everything else the plan does: clear it with the × at its right end before walking the tree by hand, and never plan other steps while it holds text. The toolbar above the editor runs the SQL; other panels (SQL 库, 查询历史, 全局搜索, 设置) open from the interface itself. Action buttons in toolbars and dialogs are elements carrying data-slot="button", not the button tag — a step targeting "button" matches both, and grid toolbar actions also carry data-toolbar-action (camelCase, e.g. addRow). An editable table's grid toolbar has a 新增行 action (data-toolbar-action="addRow") that adds one row on a click; the chevron button right beside it opens a menu whose 插入多行… entry opens the insert-rows dialog (title 新增多行, row-count input #insert-rows-count). Two ways of working that reach further than clicks:
+dbx is a single-page app with no URL routing: every recording starts from the page as it opens. The left sidebar is a tree that opens one level at a time: a click on a connection lists its databases (SQLite shows one, "main 默认库"), a click on that database node lists its tables, and a click on a table opens its data grid — the query editor, table data, structure, ER diagram are that connection's tabs. A double-click on a connection instead opens a database-browser tab, so working down the tree is how a table is reached; the outline you are given shows the app at rest, connections collapsed, so a node the plan needs that isn't in it has to be expanded into view first, not looked for on other pages. The search box above the tree (placeholder 搜索...) can reveal a node without that clicking, but while it holds text it filters the tree — a blank tree with text in the box means no match, and its text survives everything else the plan does: clear it with the × at its right end before walking the tree by hand, and never plan other steps while it holds text. The toolbar above the editor runs the SQL; other panels (SQL 库, 查询历史, 全局搜索, 设置) open from the interface itself. Action buttons in toolbars and dialogs are elements carrying data-slot="button", not the button tag — a step targeting "button" matches both, and grid toolbar actions also carry data-toolbar-action (camelCase, e.g. addRow). An editable table's grid toolbar has a 新增行 action (data-toolbar-action="addRow") that adds one row on a click; the icon button beside it, carrying aria-label="新增行" and no label text, opens a menu whose 插入多行… entry opens the insert-rows dialog (title 新增多行, row-count input #insert-rows-count). Two ways of working that reach further than clicks:
 - The query editor runs any SQL against the sandbox's SQLite databases (they are throwaway — writes are fine): when a state needs data (a result with no columns, an empty table, a table of a shape), open a connection, type the SQL that produces it and run it, and show the state it makes.
 - A file picker can't be clicked through, but a step can set it: { "do": "upload", "target": "input[type=file]", "value": "sample.csv" } — seed.mjs puts sample.csv (a small product list: name,category,price,stock) beside the databases, for import and mapping dialogs.