Describe what a column's type leaves out, and say what the provider said

packages/browser-extension/CHANGELOG.md

@@ -4,6 +4,26 @@ All notable changes to the browser extension are documented here. Versions match
 `manifest.json`. The format follows [Keep a Changelog](https://keepachangelog.com/en/1.1.0/) and
 the project follows [Semantic Versioning](https://semver.org/spec/v2.0.0.html).
 
+## [0.4.0] - 2026-08-20
+
+Carries the engine changes the extension bundles, plus its own model picker.
+
+### Added
+- The schema now states what a column's type cannot: the unit of an integer timestamp, and the shape of
+  a JSON column, including whether it holds objects or a list. Comparing epoch milliseconds against
+  epoch seconds matches every row and raises no error, and a guessed JSON key matches none, so both
+  produced confident wrong answers. Supported on SQLite, PostgreSQL, MySQL, DuckDB and Oracle.
+- A filter that compares a column against a value the column does not hold is now reported, rather than
+  answering zero as though nothing matched the question.
+
+### Fixed
+- Model listing reports the provider's own error instead of coming back empty, so a rejected key, a rate
+  limit and an outage are no longer indistinguishable.
+- Models that cannot answer a question are no longer offered in the picker.
+- Output from reasoning models no longer appears in answers, explanations or the live token stream.
+- A hosted provider configured with a local base URL is refused instead of being sent the API key.
+- Uploaded CSV and Parquet files now get the same column descriptions as a database connection.
+
 ## [0.3.2] - 2026-08-18
 
 Carries an engine fix the extension bundles. Nothing in the extension's own code changed.

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

@@ -1,50 +1,89 @@
 # 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
+Paste the block below into **Submission Options > Notes for Certification** every time the extension is
+submitted or resubmitted. This field is **private to the review team** and is not shown on the public
+listing, so a temporary API key can safely go in it.
+
+The 08/13/2026 and 08/18/2026 reviews both flagged policy 1.3.1 (Product is Testable) with identical
+wording. The cause was not the wording of these notes: the field reaching Microsoft did not contain
+them. The release workflow can send them automatically, but its Edge upload step is gated behind
+`EDGE_PUBLISH_ENABLED` and has never run, so every submission so far has been manual. If the field is
+filled in by hand, it must be filled in with this.
+
+## Before submitting
+
+1. Create a **free Groq API key** at <https://console.groq.com> (no card required) and paste it into
+   the block below where it says `PASTE_KEY_HERE`. Groq's free tier is rate limited and costs nothing.
+2. Note the date you created it. **Revoke it once the review completes** - it exists only for the
+   reviewer.
+3. Do not reuse a key that has billing attached, and never commit a real key to this file.
+4. Confirm the model named below still exists: `curl https://api.groq.com/openai/v1/models -H "Authorization: Bearer $KEY"`.
+   Providers retire models without notice - Groq removed every Llama chat model, and the name previously
+   printed here answered 404, which would read to a reviewer as a product that does not work.
 
 ---
 
 Product ID: 248cd48a-7dbe-4cfd-8ec0-df1e07231acd
 
-**Why no test account credentials are provided**
+**Test credentials**
+
+AskSQL has no accounts and no sign-in, so there is no account to issue. What it does need is an AI
+provider, which the user supplies. So that the review does not depend on you creating one, here is a
+temporary key we created for this submission and will revoke afterwards:
+
+    Provider: Groq
+    API key:  PASTE_KEY_HERE
+
+Nothing else is needed. There is no database to connect to for this test, no server of ours, and
+nothing to install.
+
+**Test it in about two minutes**
+
+1. Save these six lines as `sales.csv` anywhere on the machine:
+
+        id,customer,region,amount
+        1,Ada,EU,1200.50
+        2,Grace,NA,980.00
+        3,Kat,NA,1500.25
+        4,Ada,EU,300.00
+        5,Linus,APAC,75.99
+
+2. Click the AskSQL toolbar icon to open the side panel, then open **Settings**.
+3. Under **AI provider**, choose **Groq**, paste the key above, click **Fetch models**, pick
+   `openai/gpt-oss-20b`, and click **Test provider**. It reports success.
+4. Under **Connections**, click **Add connection**, choose **Data files**, and select `sales.csv`.
+   The file is read inside the browser into DuckDB-WASM; nothing is uploaded.
+5. In the side panel, ask: **"What is the total amount per region?"**
+   You should see the SQL it wrote, and a result of three rows: EU 1500.50, NA 2480.25, APAC 75.99.
 
-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.
+Asking "delete all rows" is a good second test: the extension refuses it, because the generated SQL is
+checked and only read-only statements are allowed to run.
 
-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.
+**If you prefer to use no key at all**
 
-**Fastest way to test, with no API key and no account (about 5 minutes)**
+Steps 1, 2 and 4 work with no provider configured and no network access: the file loads, the tables
+and columns are listed, and the UI is fully exercised. Only step 5, which needs a model, requires the
+key. A local model also works: install Ollama from ollama.com, run `ollama pull qwen2.5-coder:7b`, and
+choose provider **Ollama** with base URL `http://localhost:11434/v1` and no key.
 
-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".
+**Permissions**
 
-**Alternative, if you prefer a hosted provider**
+Host permissions are optional and requested per site, only when the tester configures an AI endpoint or
+an AskSQL server at that address. They are never requested up front.
 
-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.
+`declarativeNetRequestWithHostAccess` is used for exactly one purpose: removing the `Origin` header
+from requests to the AI endpoint the user configured. Local AI servers such as Ollama and LM Studio
+reject a `chrome-extension://` origin by default, which would otherwise make the extension unusable
+with a local model. The rule removes a header. It never adds, forges, blocks or redirects anything, and
+never applies to any other address. PRIVACY.md in the package documents this.
 
-**What the extension sends where**
+**Data handling**
 
-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.
+No analytics, no telemetry, and no server operated by us. Only the schema - table and column names -
+and the question the user typed are sent to the AI endpoint the user chose. Row data and query results
+are never sent. Data files are read in the browser and never uploaded.
 
 **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.
+Please include the Product ID in any reply. We will respond quickly with whatever helps, including a
+recorded walkthrough if that is easier than running it.

packages/browser-extension/manifest.json

@@ -2,7 +2,7 @@
   "manifest_version": 3,
   "name": "AskSQL",
   "short_name": "AskSQL",
-  "version": "0.3.2",
+  "version": "0.4.0",
   "description": "Ask your database questions in plain language. Read-only by design, zero telemetry. Query files in-browser or your AskSQL server.",
   "minimum_chrome_version": "116",
   "key": "MIIBIjANBgkqhkiG9w0BAQEFAAOCAQ8AMIIBCgKCAQEAk17oLrOl3oNPAgteUmUQDhUJJo+lxrQDt7baBBxmyQ3w3nE1g7IOZVGtq6q+gItzvs8aJAe+dHKanZynRWJfeK8zqCiBNDzyMqnRpAMOOtvKlJJOTqe9N9Lj/RJgxRgfylPnYJoPNXMAPd2y4ceeJX6yj0GpTPiS9AIfTO6p6WXB9YGWl5gdwjhGuYEAB+S7vD+M8yaCCn8C1b2+ne0yboTf/3nUyh6XysrAPafNCMlUbU1qBjAQGNJgXdsbRtWfaIbKorKhN4tWgsc3SsXZBdIxXsz/oL3EMzspjWjTEIrf/n5p1q3kNJ+SQ40SunWXdjDY4q/LnPnGmQFCUona7wIDAQAB",
@@ -29,8 +29,16 @@
     "default_path": "sidepanel/index.html"
   },
   "options_page": "options/index.html",
-  "permissions": ["sidePanel", "storage", "contextMenus", "declarativeNetRequestWithHostAccess"],
-  "optional_host_permissions": ["http://*/*", "https://*/*"],
+  "permissions": [
+    "sidePanel",
+    "storage",
+    "contextMenus",
+    "declarativeNetRequestWithHostAccess"
+  ],
+  "optional_host_permissions": [
+    "http://*/*",
+    "https://*/*"
+  ],
   "content_security_policy": {
     "extension_pages": "script-src 'self' 'wasm-unsafe-eval'; object-src 'self'"
   }

packages/browser-extension/package.json

@@ -2,7 +2,7 @@
   "name": "asksql-browser-extension",
   "private": true,
   "type": "module",
-  "version": "0.3.2",
+  "version": "0.4.0",
   "description": "AskSQL for Chromium browsers (Edge + Chrome): zero-backend DuckDB-WASM file chat and @asksql/server sidecar mode in a Manifest V3 side panel.",
   "scripts": {
     "fetch-duckdb-extensions": "node scripts/fetch-duckdb-extensions.mjs",

packages/browser-extension/src/listModels.ts

@@ -11,7 +11,15 @@ import { ensureProviderOriginAccess } from './providerAccess.js';
 const LISTABLE_HOSTED: ReadonlySet<ProviderName> = new Set(['openai', 'groq', 'nvidia']);
 const MODEL_LOOKUP_TIMEOUT_MS = 10_000;
 
-const isNotChatModel = (name: string): boolean => /embed|embedding|rerank|retriever|[-/]parse$|\bocr\b/i.test(name);
+/**
+ * Listed beside chat models but rejected by /chat/completions. Groq is why speech, TTS and classifier
+ * models are here: of its 13 entries only 7 can chat, and the list is alphabetical, so a broken one sits
+ * where the user picks. `\bguard\b` and not `guard`, because gpt-oss-safeguard DOES chat.
+ */
+const isNotChatModel = (name: string): boolean =>
+  /embed|embedding|rerank|retriever|[-/]parse$|\bocr\b|whisper|\btts\b|speech|transcribe|orpheus|\bguard\b|moderation/i.test(
+    name,
+  );
 
 /** The endpoint we can list models from, if any (anthropic/google/azure have no such listing). */
 export function listableBaseUrl(provider: ProviderName, configuredBaseURL: string | undefined): string | undefined {

packages/browser-extension/src/options/main.tsx

@@ -190,7 +190,7 @@ function ProviderSection({
           id="model"
           type="text"
           value={provider.model}
-          placeholder="e.g. llama-3.3-70b-versatile"
+          placeholder="click Fetch models, or type a model id"
           onChange={(e) => edit({ ...provider, model: e.target.value })}
         />
         {canListModels && (

packages/browser-extension/test/listModels.test.ts

@@ -65,6 +65,42 @@ describe('fetchProviderModels', () => {
     expect(await fetchProviderModels('nvidia', undefined, 'nvapi-1')).toEqual(['meta/llama-3.3-70b-instruct']);
   });
 
+  // Groq's real catalogue as of 2026-08-19, verified model by model against /chat/completions: six of
+  // the thirteen reject a chat request ("does not support chat completions", "text classification models
+  // do not support streaming", "requires terms acceptance"). The list is alphabetical, so two broken ones
+  // sat at positions 2 and 3 - where a user picks - and the query failed after the model list looked fine.
+  it('offers only the Groq models that can actually answer a query', async () => {
+    globalThis.fetch = vi.fn(async () =>
+      jsonResponse(200, {
+        data: [
+          { id: 'allam-2-7b' },
+          { id: 'canopylabs/orpheus-arabic-saudi' },
+          { id: 'canopylabs/orpheus-v1-english' },
+          { id: 'groq/compound' },
+          { id: 'groq/compound-mini' },
+          { id: 'meta-llama/llama-prompt-guard-2-22m' },
+          { id: 'meta-llama/llama-prompt-guard-2-86m' },
+          { id: 'openai/gpt-oss-120b' },
+          { id: 'openai/gpt-oss-20b' },
+          { id: 'openai/gpt-oss-safeguard-20b' },
+          { id: 'qwen/qwen3.6-27b' },
+          { id: 'whisper-large-v3' },
+          { id: 'whisper-large-v3-turbo' },
+        ],
+      }),
+    ) as typeof fetch;
+    // gpt-oss-safeguard is kept on purpose: it contains "guard" but chats, so the test is \bguard\b.
+    expect(await fetchProviderModels('groq', undefined, 'gsk-test')).toEqual([
+      'allam-2-7b',
+      'groq/compound',
+      'groq/compound-mini',
+      'openai/gpt-oss-120b',
+      'openai/gpt-oss-20b',
+      'openai/gpt-oss-safeguard-20b',
+      'qwen/qwen3.6-27b',
+    ]);
+  });
+
   it('lists ollama models, filtering out embedding models', async () => {
     globalThis.fetch = vi.fn(async () =>
       jsonResponse(200, { models: [{ name: 'llama3.2' }, { name: 'nomic-embed-text' }] }),

packages/core/CHANGELOG.md

@@ -1,5 +1,21 @@
 # @asksql/core
 
+## 0.9.0
+
+### Minor Changes
+
+- Describe what a column's type leaves out, and fix a set of AI provider issues.
+
+  The schema now states two things a column type cannot: the unit of an integer timestamp, and the shape
+  of a JSON column. Comparing epoch milliseconds against epoch seconds matches every row and raises no
+  error, and a guessed JSON key matches none, so both produced confident wrong answers. Every engine emits
+  the hint in its own syntax, from one shared implementation.
+
+  Also fixed: model listing now reports the provider's own error instead of returning an empty list;
+  models that cannot answer a question are no longer offered; a hosted provider configured with a local
+  base URL is refused rather than sent the API key; and reasoning-model output no longer appears in
+  answers, explanations, or the token stream.
+
 ## 0.8.1
 
 ### Patch Changes

packages/core/README.md

@@ -24,7 +24,7 @@ import { PostgresConnector } from '@asksql/postgres';
 
 const model = await resolveModel({
   provider: 'groq',                       // openai | anthropic | google | azure | groq | nvidia | ollama | openai-compatible
-  model: 'llama-3.3-70b-versatile',
+  model: 'openai/gpt-oss-20b', // an example: use whatever your provider lists at /models
   apiKey: process.env.GROQ_API_KEY,
 });
 
@@ -113,10 +113,14 @@ Beyond ask -> approve -> run, all optional:
 - **Schema pruning + token budget** - large catalogs are pruned to the most relevant tables
   under a token budget (`config.pruner`) before prompting.
 - **Privacy by default** - only the schema is ever sent. `allowDataInPrompt` (default off) is the
-  opt-in for sampled cell values; with it off they are stripped at the single exit from the catalog,
-  so a connector that samples cannot leak them into any prompt - the first prompt, a repair,
-  `explain`, or `explainSchema`. The MongoDB engine takes the same option, gating the values its
-  document sampling infers. Declared enum labels come from the schema and are kept either way.
+  opt-in for cell values, and it now gates three channels, not one: sampled column values, stripped
+  from the catalog so a connector that samples cannot leak them into any prompt (the first prompt, a
+  repair, `explain`, or `explainSchema`); the key NAMES inside a JSON column, where the default states
+  how many recur but not which, because a map with a stable key set is structurally identical to a
+  record; and the distinct values named in a coded-column repair, where the default attaches a caveat
+  for the reader instead. The MongoDB engine takes the same option for the values its document
+  sampling infers. Declared enum labels come from the schema and are kept either way. Query results
+  are never sent on any path.
 
 Prompts, model sampling, guard policy, and grounding (glossary, few-shots) are configurable
 without forking; see

packages/core/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@asksql/core",
-  "version": "0.8.1",
+  "version": "0.9.0",
   "description": "AskSQL engine: schema catalog, AST SQL guard, prompt pipeline, LLM orchestration. Zero database drivers.",
   "type": "module",
   "main": "./dist/index.js",

packages/core/src/catalog.ts

@@ -21,7 +21,7 @@ const FK_CLOSURE_HOPS = 2;
 const VALUE_SAMPLE_CAP = 80;
 
 /** A sampled/enum value rendered into the schema: `|` is replaced, whitespace flattened, length capped. */
-function sanitizeValue(v: string): string {
+export function sanitizeValue(v: string): string {
   const flat = v.replace(/\s+/gu, ' ').trim().replace(/\|/gu, '/');
   return flat.length > VALUE_SAMPLE_CAP ? flat.slice(0, VALUE_SAMPLE_CAP) : flat;
 }