Make the engine's own checks fire, and version core 0.7.0 - #28
Merged
Conversation
Every catalog and semantic check returns "nothing found" on anything it cannot parse, so a check that stops reading a statement is indistinguishable from one that found nothing, and no suite goes red. Five were dead on Oracle: a top-N question makes the model write FETCH FIRST n ROWS ONLY, which the parser these checks use cannot read, so the column floor, the table floor, the fan-out floor, the ungrouped-aggregate lint and the ambiguous-column floor all went quiet, and a query selecting a column no table has reached the database. The MongoDB no-op check had the same hole for shell-form JSON, which is what a small model writes. Each now reads the statement the way the guard does, and a new test gives every check a query it must flag on every dialect. Removing the fix fails fifteen of them, so one going quiet again fails the suite rather than shipping. A correction now names where the column actually lives. Told only that a column does not exist, a model renames the alias and fails the same way; told which table has it and the join that reaches it, it fixes the query. Measured on a 7B model against Oracle: none of three recovered before, three of three after. Structure questions are answered with SQL written here rather than guessed. "How many rows are in each table", "which tables have no primary key" and "what tables are in this database" read each engine's own catalog with every name quoted, because a model that has never seen information_schema invents columns on it. A relationship question is answered from the foreign keys instead of returning rows of a join, on documents as well as SQL, and a question about nothing in the database is declined in one sentence. Oracle no longer refuses the one thing a small model cannot stop doing: a plain trailing LIMIT n becomes FETCH FIRST n ROWS ONLY, which is the same query, while the forms with no equivalent are still refused. An account holding grants on another owner's tables can now name that schema instead of seeing an empty database, and CURRVAL is allowed, since it reports the current value without advancing anything. Database error text is redacted before it reaches a model, a scoped caller can no longer tell an existing connection from a missing one by the status code, a composite foreign key keeps its column order, and values from the DuckDB drivers arrive as values rather than as their storage.
Clearing the panel during a question left the button reading "Cancel" with nothing to cancel, and refused the next question because the panel still believed it was busy. The result store measured a result by sampling its first rows and pricing any object at a flat 32 bytes, so a result holding large JSON or BLOB cells was counted as if every cell were tiny and the store could hold far more memory than its budget allowed. It now samples across the result and measures a large value up to a bound, which costs less than sizing a megabyte cell exactly. The extension also carries the core changes it imports: structure questions answered from the catalog, relationship questions answered in prose, the Oracle LIMIT translation, and error text redacted before it reaches a model. An unused settings interface goes with it, since nothing has ever referenced it.
Swing reads a string that opens with <html> as markup, and a result cell, a schema tree label and a chart tooltip all come from the database. A value like "<html><img src=http://...>" in a table the plugin displayed would fetch that URL from inside the IDE. All three now render as text. The "suggest a fix" prompt carried raw driver error text, which quotes the offending row, so cell values reached the model. They are redacted first, the same way the TypeScript engine does it. A whole number is shown without a decimal the database never had: an INTEGER column and every MongoDB integer rendered as "1.0" in the grid, the copy buffer and exported CSV. One formatter now serves all of them, and INTEGER travels as a number so it sorts and charts as one. The plugin also gains the engine work from this cut: structure questions answered from the catalog, relationship questions answered in prose on documents as well as SQL, the fan-out floor that reports a SUM inflated by a join, and the Oracle LIMIT translation. A failed schema read is no longer cached as an empty database for five minutes, a transient rate limit is reported as one rather than as a billing problem, and a mid-stream provider error is surfaced instead of being returned as a truncated answer.
core and oracle take a minor: core gains public routing predicates and the mongo pipeline helpers, oracle a schema option for an account that reads another owner's tables. server, react, postgres and duckdb take a patch.
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Four commits: the engine and connector work, the two hand-versioned surfaces, and the version bumps.
Why this cut exists
Five catalog and semantic checks were silently dead on Oracle. Every one of them fails open by
design, so a check that stops parsing a statement is indistinguishable from a check that found
nothing, and no suite goes red. A top-N question makes the model write
FETCH FIRST n ROWS ONLY,which the parser those checks use cannot read, so the column floor, the table floor, the fan-out
floor, the ungrouped-aggregate lint and the ambiguous-column floor all went quiet, and a query
selecting a column no table has reached the database. The MongoDB no-op check had the same hole for
shell-form JSON, which is what a small model writes.
checks-alive.test.tsandChecksAliveTest.ktnow give every check a query it must flag on everydialect. Reverting the fix fails fifteen of them.
Versions
@asksql/core@asksql/oracle@asksql/server@asksql/react@asksql/postgres@asksql/duckdbBehaviour a user can see change
LIMIT nis translated toFETCH FIRST n ROWS ONLYinstead of beingrefused. Offset and placeholder forms are still refused.
returning documents.
1, not1.0, in the grid, thecopy buffer and exported CSV.
Verification
pnpm verifyNot verified: no human has driven the JetBrains or VS Code UIs, and the Anthropic, OpenAI, Google
and Azure providers were not exercised.