JetBrains chat reads as a conversation (0.5.0)

.github/workflows/jetbrains-ci.yml

@@ -89,6 +89,9 @@ jobs:
           distribution: temurin
           java-version: 21
       - uses: gradle/actions/setup-gradle@v6
+        with:
+          # One cache writer: two filled the repo quota and evicted the pnpm caches.
+          cache-read-only: true
       - run: ./gradlew test -PintegrationTests=true --stacktrace
         env:
           # Rootful Docker on the runner; skip the rootless strategy, whose JNA collides with the IntelliJ Platform's bundled jnidispatch.

packages/jetbrains/CHANGELOG.md

@@ -5,6 +5,43 @@ Format loosely follows [Keep a Changelog](https://keepachangelog.com/en/1.1.0/).
 
 ## [Unreleased]
 
+## [0.5.0] - 2026-08-09
+
+### Fixed
+- A two-line answer is no longer clipped to one, with the note below it painting over the rest. The
+  text pane reported its height measured at its unwrapped width rather than the width it was given.
+- The **Explain** button comes back after a failure. It disabled itself on click and, if the model
+  call failed, stayed disabled for the rest of that turn.
+- A MongoDB turn no longer says "Writing SQL" or "Finding relevant tables" while it builds a pipeline.
+- Content appended to a turn already on screen is scrolled into view, so an approval bar below the
+  fold is reachable. Scrolling only follows when you are already at the bottom, so it will not pull
+  you away from an earlier turn you are reading.
+- A backticked placeholder such as `?`, a date, or `:param` is no longer reported as a name missing
+  from your schema.
+- A proposed write showed three separate lines saying nothing had run. It shows one.
+- The copy control reports failure instead of showing a tick when the clipboard write is refused.
+
+### Added
+- **A copy control on every answer**, not only on result grids: explanations, schema answers, and
+  errors each copy the model's own text rather than the rendered markup.
+- **A query inside a prose answer renders as a real code block** with syntax highlighting and its own
+  Copy, instead of running together with the sentences around it.
+- **The transcript reads as a conversation.** Your question sits on the right in a tinted bubble,
+  AskSQL answers on the left, and a rule separates one turn from the next.
+- Progress is shown for the states that used to sit blank: preparing a turn, answering from the
+  schema, and correcting a query the database rejected.
+- A MongoDB pipeline's collection name is selectable and travels with Copy, as a
+  `db.getCollection(...).aggregate(...)` call that pastes straight into mongosh.
+- Progress is announced to screen readers where the IDE runtime supports it.
+
+### Changed
+- Connections are released when the project's services are disposed rather than through an
+  experimental platform listener, so cleanup also covers disabling or unloading the plugin. The
+  listener also woke both connection registries on every project close, including in projects where
+  AskSQL was never opened.
+- The guard detects `LIMIT ALL` through JSqlParser's supported API. The two methods it used before
+  are deprecated, and the guard failing to compile on a future parser release is not a small problem.
+
 ## [0.4.2] - 2026-08-09
 
 ### Fixed

packages/jetbrains/README.md

@@ -16,9 +16,13 @@ plain language, review the generated SQL, approve it, and get results.
 
 ![Schema tree browsing a connected MySQL database above the chat panel, with sample questions to get started](https://github.com/rahulmahadik/AskSQL/raw/HEAD/packages/jetbrains/images/schema-and-chat.png)
 
-![AskSQL settings configured against a local Ollama model, no API key needed](https://github.com/rahulmahadik/AskSQL/raw/HEAD/packages/jetbrains/images/settings-ollama.png)
+![AskSQL settings configured against a local Ollama model, no API key needed](https://github.com/rahulmahadik/AskSQL/raw/HEAD/packages/jetbrains/images/settings.png)
 
-![The AI provider dropdown in settings, listing OpenAI, Anthropic, Gemini, Groq, Ollama and the rest](https://github.com/rahulmahadik/AskSQL/raw/HEAD/packages/jetbrains/images/settings-providers.png)
+![The AI provider dropdown in settings, listing OpenAI, Anthropic, Google, Groq, Ollama, an OpenAI-compatible endpoint, LM Studio and NVIDIA](https://github.com/rahulmahadik/AskSQL/raw/HEAD/packages/jetbrains/images/settings-ai-providers.png)
+
+![A query result drawn as a bar chart, with a Table toggle beside Export CSV, Copy, Open in Editor and Explain](https://github.com/rahulmahadik/AskSQL/raw/HEAD/packages/jetbrains/images/show-chart.png)
+
+![Adding a database connection from the AskSQL settings page](https://github.com/rahulmahadik/AskSQL/raw/HEAD/packages/jetbrains/images/settings-add-db-connection.png)
 
 ## Getting started (using the plugin)
 
@@ -28,7 +32,7 @@ plain language, review the generated SQL, approve it, and get results.
    `build/distributions/asksql-jetbrains-<version>.zip`, which installs via
    **Plugins → ⚙ → Install Plugin from Disk...**.)
 2. Restart the IDE when prompted.
-3. Open the **AskSQL** tool window (usually a tab on the right/bottom edge).
+3. Open the **AskSQL** tool window (a tab on the right edge).
 4. **Add a connection**: on the empty-state screen, click **Add Connection** (Postgres,
    MySQL, SQLite, DuckDB, Oracle, or MongoDB), or click **Try sample data** for a
    ready-made SQLite database with no setup, good for a first look.
@@ -131,8 +135,11 @@ These hold on every engine (the plan doc carries the full list):
   unless **Send sample column values to the model** is turned on in Settings. It defaults to off,
   so out of the box the model sees field names, types and presence percentages only. Full row
   data (arbitrary query results) is never sent on any engine.
-- Chat history and query results are **in-memory only**; nothing is written to disk
-  except settings, and secrets live only in the OS keychain via PasswordSafe.
+- Chat history and query results are **in-memory only**, and secrets live only in the OS
+  keychain via PasswordSafe. Apart from settings, the only things the plugin writes to disk
+  are under the IDE's system directory and are ones you asked for: the JDBC driver jars it
+  downloads on demand (`asksql/drivers`), the DuckDB database built when you load data files
+  (`asksql/uploads`), the sample SQLite database (`asksql/sample`), and a CSV you export.
 - Zero telemetry.
 
 ## Requirements
@@ -142,7 +149,8 @@ These hold on every engine (the plan doc carries the full list):
 - Docker, only if you want to run the Testcontainers-backed integration tests
   (`./gradlew test -PintegrationTests=true`). Everything else (build, unit tests,
   `runIde`) needs no Docker.
-- Node.js 18+, only for `./gradlew parityVectors` (see below). Never required to build
+- Node.js 20+ (the monorepo's `engines` floor; CI uses 22), only for
+  `./gradlew parityVectors` (see below), which builds `@asksql/core` from this repo. Never required to build
   or run the plugin itself.
 
 ## Compatibility
@@ -157,10 +165,17 @@ Every release is run through the JetBrains **Plugin Verifier**: 15 verifications
 
 | IDE | Versions verified |
 | --- | --- |
-| IntelliJ IDEA Community | 2024.2 (floor), 2024.3, 2025.1, 2025.2, 2025.3 |
-| IntelliJ IDEA Ultimate, PyCharm Professional, WebStorm, PhpStorm, GoLand, Rider, CLion, RubyMine, RustRover | 2026.1 (latest stable) |
+| IntelliJ IDEA Community | 2024.2 (floor), 2024.3, 2025.1 (251.29188.72), 2025.2.6.2 (252.28539.54), 2025.3 (253.28294.334) |
+| IntelliJ IDEA Ultimate, PyCharm Professional, WebStorm, PhpStorm, GoLand, Rider, CLion, RubyMine, RustRover | 2026.1.4 (latest stable) |
 | Android Studio | whichever build is installed locally |
 
+The IC entries from 2025.1 on are pinned by build number because ideaIC publishes both build-number
+and marketing-version artifacts; the other IDEs publish marketing versions only.
+
+`untilBuild` is deliberately open-ended (`null`), so the plugin stays installable on IntelliJ
+majors that did not exist when it was published. The Plugin Verifier is the safety net: a
+verification failure blocks the release.
+
 2024.1 is out of reach: `com.intellij.util.net.JdkProxyProvider`, which routes model calls through
 the IDE's proxy settings, arrives in 242.
 

packages/jetbrains/build.gradle.kts

@@ -223,6 +223,11 @@ tasks {
             }
         }
         systemProperty("idea.force.use.core.classloader", "true")
+        // Painting Swing to a PNG needs real font metrics and a window peer: opt in with -PrenderUi=true.
+        if (providers.gradleProperty("renderUi").orNull == "true") {
+            systemProperty("java.awt.headless", "false")
+            systemProperty("renderUi", "true")
+        }
         maxHeapSize = "2g"
     }
 

packages/jetbrains/gradle.properties

@@ -4,7 +4,7 @@
 
 pluginGroup = com.rahulmahadik.asksql
 pluginName = AskSQL
-pluginVersion = 0.4.2
+pluginVersion = 0.5.0
 
 # IntelliJ Platform target used to COMPILE and RUN the sandbox. Broad
 # compatibility is governed by pluginSinceBuild/pluginUntilBuild in

packages/jetbrains/src/main/kotlin/com/rahulmahadik/asksql/ide/db/ConnectionRegistry.kt

@@ -1,5 +1,6 @@
 package com.rahulmahadik.asksql.ide.db
 
+import com.intellij.openapi.Disposable
 import com.intellij.openapi.components.Service
 import com.intellij.openapi.diagnostic.logger
 import com.intellij.openapi.project.Project
@@ -17,7 +18,7 @@ import java.util.concurrent.atomic.AtomicInteger
  * in-flight [withConnection] still uses: each [Slot] tracks a lease count and closes only once the last lease ends.
  */
 @Service(Service.Level.PROJECT)
-class ConnectionRegistry(private val project: Project, private val scope: CoroutineScope) {
+class ConnectionRegistry(private val project: Project, private val scope: CoroutineScope) : Disposable {
 
     private val log = logger<ConnectionRegistry>()
 
@@ -140,7 +141,11 @@ class ConnectionRegistry(private val project: Project, private val scope: Corout
         }
     }
 
-    /** Called by [com.rahulmahadik.asksql.ide.AskSqlProjectCloseListener] to release every connection deterministically before the scope is torn down. */
+    /** [closeNowOrCancel] is synchronous, so it is safe at dispose time. */
+    override fun dispose() {
+        closeAll()
+    }
+
     fun closeAll() {
         slots.keys.toList().forEach { id ->
             slots.remove(id)?.let {

packages/jetbrains/src/main/kotlin/com/rahulmahadik/asksql/ide/db/MongoClientRegistry.kt

@@ -1,5 +1,6 @@
 package com.rahulmahadik.asksql.ide.db
 
+import com.intellij.openapi.Disposable
 import com.intellij.openapi.components.Service
 import com.intellij.openapi.diagnostic.logger
 import com.intellij.openapi.project.Project
@@ -13,7 +14,7 @@ import java.util.concurrent.atomic.AtomicInteger
 
 /** Owns the lifecycle of every configured [MongoClient] for a project. */
 @Service(Service.Level.PROJECT)
-class MongoClientRegistry(private val project: Project, private val scope: CoroutineScope) {
+class MongoClientRegistry(private val project: Project, private val scope: CoroutineScope) : Disposable {
 
     private val log = logger<MongoClientRegistry>()
 
@@ -101,7 +102,11 @@ class MongoClientRegistry(private val project: Project, private val scope: Corou
         }
     }
 
-    /** Called by [com.rahulmahadik.asksql.ide.AskSqlProjectCloseListener] to release every client deterministically before the scope is torn down. */
+    /** [closeNowOrCancel] is synchronous, so it is safe at dispose time. */
+    override fun dispose() {
+        closeAll()
+    }
+
     fun closeAll() {
         slots.keys.toList().forEach { id ->
             slots.remove(id)?.let {

packages/jetbrains/src/main/kotlin/com/rahulmahadik/asksql/ide/engine/Grounding.kt

@@ -90,7 +90,15 @@ object Grounding {
         return (aliases + ctes).toList()
     }
 
-    private val PROSE_IDENTIFIER_RE = Regex("""`([^`\s]+)`|"([\w.]+)"|\b([a-z][a-z0-9]*(?:_[a-z0-9]+)+)\b""", RegexOption.IGNORE_CASE)
+    // Java's \s is ASCII-only where JavaScript's is not; these are the extras JS matches.
+    // UNICODE_CHARACTER_CLASS would widen \w below and diverge the other way.
+    private val PROSE_IDENTIFIER_RE = Regex(
+        """`([^`\s\u00a0\u1680\u2000-\u200a\u2028\u2029\u202f\u205f\u3000\ufeff]+)`|"([\w.]+)"|\b([a-z][a-z0-9]*(?:_[a-z0-9]+)+)\b""",
+        RegexOption.IGNORE_CASE,
+    )
+
+    /** An identifier, optionally schema-qualified. Placeholders, literals and operators do not match. */
+    private val IDENTIFIER_SHAPE = Regex("""[a-z_][a-z0-9_$-]*(?:\.[a-z_][a-z0-9_$-]*)*""", RegexOption.IGNORE_CASE)
 
     /**
      * Identifier-shaped names in a prose answer absent from the catalog - the grounding floor for
@@ -141,6 +149,8 @@ object Grounding {
         val scanned = if (documentStyle) answer.replace(Regex("```[\\s\\S]*?```"), " ") else answer
         for (m in PROSE_IDENTIFIER_RE.findAll(scanned)) {
             if (documentStyle && m.groupValues[2].isNotEmpty()) continue // "shipped" is a value
+            // Backticks wrap anything, so a placeholder or a literal can arrive here.
+            if (m.groupValues[1].isNotEmpty() && !IDENTIFIER_SHAPE.matches(m.groupValues[1])) continue
             val raw = (m.groupValues[1].ifEmpty { m.groupValues[2] }.ifEmpty { m.groupValues[3] }).lowercase()
             if (raw.startsWith("$")) continue // $lookup / $group are operators
             // Backticked SQL vocabulary is not a name claim; a call with parentheses is a function.

packages/jetbrains/src/main/kotlin/com/rahulmahadik/asksql/ide/guard/SqlGuard.kt

@@ -6,6 +6,7 @@ import com.rahulmahadik.asksql.ide.model.GuardPolicy
 import com.rahulmahadik.asksql.ide.model.GuardVerdict
 import com.rahulmahadik.asksql.ide.model.LimitStyle
 import net.sf.jsqlparser.JSQLParserException
+import net.sf.jsqlparser.expression.AllValue
 import net.sf.jsqlparser.expression.Expression
 import net.sf.jsqlparser.expression.ExpressionVisitorAdapter
 import net.sf.jsqlparser.expression.Function
@@ -535,13 +536,12 @@ object SqlGuard {
             return LimitStatus.Ok
         }
         val limit = target?.limit ?: return LimitStatus.None
-        // `LIMIT ALL` parses to a Limit with no row count. Reading that as "no limit present" left
-        // the statement uncapped, since the clause is already there for an append to bind to.
-        if (limit.isLimitAll) {
-            return LimitStatus.Unbounded {
-                limit.isLimitAll = false
-                limit.rowCount = LongValue(maxRows.toLong())
-            }
+        // `LIMIT ALL` parses to a Limit whose row count is an AllValue, which bounds nothing. Reading
+        // that as "no limit present" left the statement uncapped, since the clause is already there
+        // for an append to bind to. Overwriting the row count is what clears it (the deprecated
+        // setLimitAll(false) was a no-op).
+        if (limit.rowCount is AllValue) {
+            return LimitStatus.Unbounded { limit.rowCount = LongValue(maxRows.toLong()) }
         }
         val rowCount: Expression = limit.rowCount ?: return LimitStatus.None
         val value = (rowCount as? LongValue)?.value ?: return LimitStatus.NonLiteral