ADFA-5403 | Fix LlmInferenceService resolution for Voice-to-Code by jatezzz · Pull Request #93 · appdevforall/plugin-examples

jatezzz · 2026-09-08T19:43:52Z

Description

Updated SpeechToTextPlugin to resolve the LlmInferenceService from the correct global registry. Previously, the plugin failed to find the service in the local context, causing Voice-to-Code to insert raw transcribed text instead of generated code. The service is now resolved from SharedServices first, with the local context.services used as a fallback. Resolution is dynamically retried on every use to account for parallel loading delays when the AI Core activates.

Details

Added resolveLlmService() to cache successful lookups and retry upon failure.
Bound the code generation request to a 60-second timeout (GENERATION_TIMEOUT_SECONDS) to prevent UI freezes.
Implemented stripCodeFences() to remove markdown from the generated output so the editor receives raw code.
Migrated logging from Android Log to the plugin's context.logger.

Screen_Recording_20260903_155612_Code.on.the.Go.mp4

Ticket

ADFA-5403
Parent: ADFA-5402

…(ADFA-5403) context.services is a per-plugin registry that never holds it, so Voice-to-Code always inserted the raw transcript. Resolve per use, bound generation with a timeout, strip markdown fences, and target the open file's language, not Kotlin.

claude

Claude Code Review

This repository is configured for manual code reviews. Comment @claude review for a one-time review, or @claude review always to subscribe this PR to a review on every future push.

_{Tip: disable this comment in your organization's Code Review settings.}

hal-eisen-adfa

Automated review (Claude Code), high and medium findings only. Low-severity notes (duplicate fence stripper, LinkageError on SharedServices, no unit tests) are left out.

F09 (Medium) — line 52, outside the diff, so no inline comment is possible.
private var llmService: LlmInferenceService? = null is a plain non-volatile var, but this PR turns it into a lazily-populated cache. activate() writes it on the main thread; handleTranscript reads it from scope.launch on Dispatchers.IO. There is no happens-before edge, so the IO thread can keep seeing null and re-resolve on every transcript. recordingState in this same class is already @Volatile.

Resolve the LLM service through getPluginService as well, and drop the cached reference when AI Core unloads. Tune the generation request (system prompt, temperature, maxTokens) instead of scraping fences off the reply, size the timeout to that token budget, and await the future cancellably so plugin teardown unwinds it. Rewrite the fence stripper to handle a lead-in line, an unclosed fence and a one-line fenced reply. Guard logging against an uninitialized context during teardown.

jatezzz · 2026-09-09T13:58:38Z

@hal-eisen-adfa F09: fixed. llmService is now @Volatile, like recordingState in the same class, so the IO threads in scope see the value activate() publishes.
dispose() and the new AI Core lifecycle listener (F03) are the only other writers.

hal-eisen-adfa

Second automated review pass (Claude Code, xhigh). All thirteen findings from the first pass are addressed in code and the plugin builds clean; thanks. Five new issues below, four of them in the code this round introduced.

F16 (the Speech-to-Text max_ide_version 26.30 vs ai-core min_ide_version 26.35 gap) is dropped — we do not enforce those ranges.

Restore IDLE state and the raw-transcript fallback when AI Core cancels the shared future, release listener/scope/service in deactivate(), and stop the fence stripper from inserting prose (F17) or eating real code (F18, F20).

hal-eisen-adfa

Third automated review pass (Claude Code, high). The core of the PR is correct. Six new findings, two of them in the code the second pass added.

hal-eisen-adfa · 2026-09-11T02:40:58Z

    override fun activate(): Boolean {
-        Log.i(TAG, "SpeechToTextPlugin activating...")
+        // deactivate() cancels the scope, so a re-enabled plugin needs a fresh one.
+        if (!scope.isActive) {


F21 (Medium) — activate() rebuilds the scope but never resets recordingState.

The field is not part of the scope, so it keeps the value it held when the plugin was disabled. Disable the plugin while a generation runs — the F08 timeout leaves a 132 s window — and re-enable it: the same instance still reads PROCESSING, currentIconRes() returns ic_processing, and the toolbar shows the spinner permanently.

No path can clear it. The coroutine whose finally posts setState(IDLE) was cancelled by teardown(), and startVoiceCapture() does not reset the state before it sets RECORDING.

Add recordingState = RecordingState.IDLE beside the scope rebuild.

hal-eisen-adfa · 2026-09-11T02:40:58Z

+            } finally {
+                // Posted, not dispatched: a cancelled coroutine can no longer suspend, and the
+                // toolbar must leave the spinner even then.
+                runOnMain { setState(RecordingState.IDLE) }


F22 (Medium) — no re-entrancy guard between a running generation and a new capture.

isEnabledProvider = { hasOpenFile() } (line 219) ignores recordingState, and startVoiceCapture() (line 262) does not reject a non-IDLE state. So the mic button stays live while the spinner is up.

Tap the mic, speak, then tap it again during generation #1:

Recognition feat/ADFA-3603 ndk plugin #2 starts and sets RECORDING.

Generation chore: bump plugins to Kotlin 2.3.0 and migrate to compilerOptions DSL #1 completes and insertCodeAtCursor writes its snippet at the cursor in the middle of capture feat/ADFA-3603 ndk plugin #2.

This finally then forces IDLE, so the toolbar drops to the mic icon while capture feat/ADFA-3603 ndk plugin #2 is still listening.

The user gets two insertions and a toolbar that does not match the state. Guard startVoiceCapture() on recordingState == RecordingState.IDLE, and reset here only when the state is still PROCESSING.

hal-eisen-adfa · 2026-09-11T02:40:58Z

+     * @param lines the trimmed reply, split into lines
+     * @return the reply from its first code line on, or empty when it holds none
+     */
+    private fun dropLeadingProse(lines: List<String>): String {


F23 (Low) — dropLeadingProse drops leading prose only, but the KDoc promises more.

The doc says "a lead-in or a refusal is never written into the open file". The scan stops at the first code line and keeps everything after it, so trailing prose survives. For

Here is the code: val x = 1 This adds one.

the return value is val x = 1\nThis adds one., and the English sentence goes into the .kt file.

SuggestionProvider.sanitizeCompletion avoids this because it keeps the first code line only. Either filter prose over the whole reply or narrow the KDoc claim.

hal-eisen-adfa · 2026-09-11T02:40:58Z

+        // With code behind it the tag may be code itself, so require the tag we asked for and a
+        // remainder that starts a name - `c = a + b` and `bash -c "..."` are code, not info.
+        val startsName = rest.first().isLetter() || rest.first() == '_' || rest.first() == '@'
+        return if (startsName && tag.equals(language, ignoreCase = true)) rest else fenceLine


F24 (Low) — the F20 fix compares a fence tag against a host editor id, so it rarely matches.

language comes from IdeEditorService.getCurrentLanguageId(), which returns the host's editor id (kt, java). The model writes its own info string (kotlin, java). The two agree only by chance.

For a one-line reply ```kotlin println("hi")``` in a Kotlin file, tag is kotlin and language is kt, so equals fails and the whole line is returned. The literal word kotlin is inserted into the user's file — the defect F20 asked you to prevent, with a new trigger.

The multi-line case is safe, because rest.isEmpty() returns "" whatever the language. Map the editor id to its tag aliases, or accept any LANGUAGE_TAGS member here as the sibling does and keep the startsName test as the guard.

hal-eisen-adfa · 2026-09-11T02:40:58Z

+private suspend fun <T> CompletableFuture<T>.await(): T =
+    suspendCancellableCoroutine { cont ->
+        whenComplete { value, error ->
+            if (error == null) cont.resume(value) else cont.resumeWithException(error)


F25 (Low) — a null future result is reported as a timeout.

generateCompletion returns a platform-typed CompletableFuture<LlmResponse!>, and cont.resume(value) passes a null through unguarded. withTimeoutOrNull also yields null on a real timeout, so both states reach the same if (response == null) at line 405 and log "Code generation timed out after 132s".

A backend that breaks its contract therefore reads as a slow model, and the 142 s budget looks wrong when it is not. Separate the two — check the future result inside the withTimeoutOrNull block.

hal-eisen-adfa · 2026-09-11T02:40:58Z

+     * would still write into the user's file, host callbacks to a dead instance, and the
+     * cached router that pins AI Core's ClassLoader. Idempotent - dispose() follows deactivate().
+     */
+    private fun teardown() {


F26 (Low) — teardown() keeps the host UI service, which is then used after deactivate() returned.

It clears llmService but leaves editorService and uiService set. The finally in handleTranscript posts runOnMain { setState(IDLE) }, and setState calls uiService?.refreshToolbarActions().

Disable the plugin during a generation: scope.cancel() runs the finally, the post lands after deactivate() returned, and the call reaches into the host toolbar from a disabled plugin. runOnMain has no try/catch, so a host that rejects the call crashes the main thread.

Null editorService and uiService in teardown() beside llmService.

claude Bot reviewed Sep 8, 2026

View reviewed changes

jatezzz requested review from a team, Daniel-ADFA, dara-abijo-adfa and itsaky-adfa September 8, 2026 19:53

hal-eisen-adfa reviewed Sep 9, 2026

View reviewed changes

jatezzz requested a review from hal-eisen-adfa September 9, 2026 14:05

Merge branch 'main' into fix/ADFA-5403-resolve-llm-from-shared-services

c739277

hal-eisen-adfa reviewed Sep 10, 2026

View reviewed changes

hal-eisen-adfa mentioned this pull request Sep 10, 2026

ADFA-5404 | Recover from missing offline language packs #95

Open

jatezzz requested a review from hal-eisen-adfa September 10, 2026 13:45

Merge branch 'main' into fix/ADFA-5403-resolve-llm-from-shared-services

6bdd72a

hal-eisen-adfa reviewed Sep 11, 2026

View reviewed changes

Uh oh!

Uh oh!

Conversation

jatezzz commented Sep 8, 2026 • edited by atlassian Bot Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Description

Details

Ticket

Uh oh!

claude Bot left a comment

Choose a reason for hiding this comment

Claude Code Review

Uh oh!

hal-eisen-adfa left a comment

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

jatezzz commented Sep 9, 2026

Uh oh!

hal-eisen-adfa left a comment

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

hal-eisen-adfa left a comment

Choose a reason for hiding this comment

Uh oh!

hal-eisen-adfa Sep 11, 2026

Choose a reason for hiding this comment

Uh oh!

hal-eisen-adfa Sep 11, 2026

Choose a reason for hiding this comment

Uh oh!

hal-eisen-adfa Sep 11, 2026

Choose a reason for hiding this comment

Uh oh!

hal-eisen-adfa Sep 11, 2026

Choose a reason for hiding this comment

Uh oh!

hal-eisen-adfa Sep 11, 2026

Choose a reason for hiding this comment

Uh oh!

hal-eisen-adfa Sep 11, 2026

Choose a reason for hiding this comment

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

jatezzz commented Sep 8, 2026 •

edited by atlassian Bot

Loading