Quote identifiers a database will not read back as itself

.changeset/identifier-and-semantic-floors.md

@@ -0,0 +1,39 @@
+---
+'@asksql/core': patch
+---
+
+Quote the identifiers a database would not read back as itself. A mixed-case Postgres schema failed
+every query, because an unquoted name folds to lower case and resolves to nothing; Oracle folds the
+other way, and MySQL on Linux compares table names case-sensitively. Table and column names are now
+quoted from the catalog before the query is validated. A name that is already correct is left
+untouched, so MySQL output is unchanged, and a name spelled two ways across the catalog is skipped
+rather than guessed. A table named like a parser keyword, such as `order` or `Nulls`, also works now:
+the bare form could not be parsed at all and the question failed after three attempts.
+
+Reserved words now come from each database itself, through `pg_get_keywords()`,
+`information_schema.KEYWORDS`, `V$RESERVED_WORDS` and `duckdb_keywords()`, rather than one shared list
+that applied MySQL's rules to Postgres and missed most of MySQL's own: MySQL reserves 262 words where
+the shared list had about a hundred. Regenerate with `node tools/generate-sql-keywords.mjs`.
+
+Quoting knows where a word is syntax rather than a name, so `CAST(x AS DATE)` and
+`EXTRACT(MONTH FROM d)` are left alone, and a CTE is still recognised once its name is quoted.
+
+When a database rejects a name, the corrected query is derived from the catalog rather than from a
+second model call, and the table repair names the closest match it already knew.
+
+Tell the model which database and schema it is connected to. Without that it wrote
+`table_schema = 'your_database_name'` against `information_schema` and returned nothing at all, which
+reads as an empty database rather than an error. Structure questions also get a correct catalog query
+for the engine to build on.
+
+Say what is actually wrong when a statement will not parse. An apostrophe inside a value, as in
+`'O'Brien'`, only produced "could not parse", so the model returned the same statement until it ran
+out of attempts; it is now told to double the quote.
+
+Reject `AVG(SUM(x))` before it reaches the database. Nested aggregates are invalid everywhere, so the
+query is repaired rather than run and failed.
+
+Fix two false alarms. A `UNION ALL` of per-table counts was blocked as a hallucinated column, because
+each branch's columns were judged against every branch's tables; per-table row counts now work. Index
+columns arrived already quoted from introspection and were quoted a second time, so every prompt
+carried `"""ColumnName"""`.

packages/browser-extension/STORE-CERTIFICATION-NOTES.md

@@ -0,0 +1,50 @@
+# Notes for Certification
+
+Paste the section below into **Submission Options > Notes for Certification** when resubmitting to the
+Microsoft Edge Add-ons store. It answers policy 1.3.1 (Product is Testable), which the 08/13/2026
+review flagged. Product ID: 248cd48a-7dbe-4cfd-8ec0-df1e07231acd
+
+---
+
+Product ID: 248cd48a-7dbe-4cfd-8ec0-df1e07231acd
+
+**Why no test account credentials are provided**
+
+AskSQL has no accounts, no sign-in, and no server of our own. Nothing is hosted by us, so there is no
+credential we could issue. The extension stores its settings locally and talks only to two things the
+user chooses: their own data files, and their own AI model provider.
+
+Because of that, testing needs no credentials from us. It needs a model provider and a data file, and
+both can be supplied at no cost in a few minutes.
+
+**Fastest way to test, with no API key and no account (about 5 minutes)**
+
+1. Install Ollama from https://ollama.com (free, no account required) and run:
+   `ollama pull qwen2.5-coder:7b`
+2. Start Ollama with `OLLAMA_ORIGINS=* ollama serve` so it serves on http://127.0.0.1:11434.
+   The variable matters: fetching the model list works without it, but asking a question fails with
+   403, because Ollama rejects the extension's origin on POST requests.
+3. Open the extension's Options page, choose provider **Ollama**, click **Fetch models**, pick the
+   model, and click **Test provider**. It should report success.
+4. Add a connection: click **Add connection**, choose **Data files**, and select any CSV or Excel
+   file. Any small spreadsheet works; no database server is needed.
+5. Open the side panel and ask a question about the file, for example "how many rows are there?" or
+   "show me the first 10 records".
+
+**Alternative, if you prefer a hosted provider**
+
+Any OpenAI, Anthropic or Groq API key works. Enter it in the Options page under the matching provider
+and follow steps 3 to 5 above. We cannot include one of our keys in this submission, because the key
+would be visible to anyone who reads the listing and would be billed to us.
+
+**What the extension sends where**
+
+Questions and database schema go only to the provider the user configures, over a connection they
+control. Data files are read in the browser and never uploaded to us. The extension has no analytics
+and no backend. Generated SQL is read-only and is checked before it runs, so a query cannot modify
+the user's data.
+
+**If anything blocks the review**
+
+Please include the Product ID in any reply and we will respond quickly with whatever else is helpful,
+including a recorded walkthrough if that is easier than running it locally.

packages/core/README.md

@@ -97,6 +97,15 @@ Beyond ask -> approve -> run, all optional:
   column (a common small-model slip), it is handed the real column list and re-asked, so the
   fix happens before the database ever sees the query. The schema is also auto-shrunk and
   retried once on context overflow.
+- **Identifier quoting** - names a database would not read back as themselves are quoted from the
+  catalog before the query is validated, following each engine's own rule: Postgres folds unquoted
+  names down, Oracle folds them up, MySQL on Linux compares table names case-sensitively, and every
+  engine has reserved words. A mixed-case schema therefore works without the model having to
+  remember quotes, and names that are already correct are left alone. If a database still rejects a
+  name, the corrected query comes from the catalog rather than a second model call.
+- **Semantic floors** - a query that would be rejected or would answer the wrong question is
+  repaired before it runs: an aggregate beside a bare column with no `GROUP BY`, an aggregate nested
+  inside another (`AVG(SUM(x))`), and a one-to-many join that inflates a `SUM`.
 - **Follow-up context** - prior turns are threaded into the prompt so "now break that down by
   month" works.
 - **Query history** - `config.history` records every attempt (status, duration), backed by an

packages/core/src/catalog.ts

@@ -5,6 +5,7 @@
  */
 
 import type { PrunerSettings, SchemaCatalog, TableInfo } from './types.js';
+import { reservedWordsFor } from './sql-keywords.js';
 import { VALUE_SAMPLE_MAX_DISTINCT } from './types.js';
 import { dialectFor } from './dialects.js';
 
@@ -35,33 +36,22 @@ function sanitizeComment(comment: string | null | undefined): string | null {
 /** A name that can be written without quotes; anything else is rendered quoted. */
 const PLAIN_IDENTIFIER_RE = /^[A-Za-z_][A-Za-z0-9_]*$/u;
 
-/** Words an engine will not accept as a bare identifier; not exhaustive across every dialect. */
-const RESERVED_WORDS: ReadonlySet<string> = new Set(
-  (
-    'select from where group by order having limit offset union all distinct join inner outer left ' +
-    'right full cross natural on using as into insert update delete set values create drop alter ' +
-    'table column view index key primary foreign unique constraint references default check null ' +
-    'not and or in is like between case when then else end exists any some cast collate with ' +
-    'recursive returning window over partition range rows current session system user grant revoke ' +
-    'to begin commit rollback transaction lock database schema trigger procedure function ' +
-    'desc asc date time timestamp interval level size type comment position language'
-  ).split(' '),
-);
-
 /**
  * True when the engine would not read the bare name back as itself; an unquoted identifier folds
  * case - PostgreSQL to lower, Oracle to upper.
  */
-function needsQuoting(name: string, engine: string): boolean {
+export function needsQuoting(name: string, engine: string): boolean {
   if (!PLAIN_IDENTIFIER_RE.test(name)) return true;
-  if (RESERVED_WORDS.has(name.toLowerCase())) return true;
+  if (reservedWordsFor(engine).has(name.toLowerCase())) return true;
   if (engine === 'oracle') return name !== name.toUpperCase();
   // MySQL, SQLite and DuckDB match identifiers case-insensitively, so folding cannot lose a name.
   if (engine === 'mysql' || engine === 'sqlite' || engine === 'duckdb') return false;
   return name !== name.toLowerCase();
 }
 
-function promptIdentifier(name: string, quote: string, engine: string): string {
+function promptIdentifier(raw: string, quote: string, engine: string): string {
+  // Index columns arrive already quoted from introspection; quoting twice escapes the quotes into the name.
+  const name = raw.length > 1 && raw.startsWith(quote) && raw.endsWith(quote) ? raw.slice(1, -1) : raw;
   if (!needsQuoting(name, engine)) return name;
   // Doubling is how every supported engine escapes its own quote character inside an identifier.
   return `${quote}${name.split(quote).join(quote + quote)}${quote}`;