ADFA-5403 | Fix LlmInferenceService resolution for Voice-to-Code

[jatezzz]

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, then from AI Core's per-plugin export via context.getPluginService(), with the local context.services as the last 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 timeout derived from the token cap (GENERATION_TIMEOUT_SECONDS = MAX_GENERATION_TOKENS / SLOWEST_TOKENS_PER_SECOND + PROMPT_EVAL_SECONDS = 512 / 5 + 30 = 132 s) to prevent UI freezes.
• Implemented stripCodeFences() to remove markdown from the generated output so the editor receives raw code.
• Dropped the cached service when AI Core is deactivated or uninstalled, via a PluginLifecycleListener, so a disabled plugin no longer pins its ClassLoader.
• Raised the manifest's IDE window to 26.29-26.99: SharedServices, PluginContext.getPluginService() and addPluginLifecycleListener() all first ship in 26.29, and the previous 26.30 ceiling excluded every host AI Core runs on.
• 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

[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.

[jatezzz]

@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.

[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.

[itsaky-adfa]

Fourth review pass (/pr-review high). Two IMPORTANT findings; the rest are advisory.

IMPORTANT

• SpeechToTextPlugin.kt:536 - an unlisted fence tag (dart, kts, rust, ...) is inserted as a line of code
• SpeechToTextPlugin.kt:144 - min_ide_version 26.17 predates the three host APIs this PR adds, which shipped in 26.29

MINOR

• SpeechToTextPlugin.kt:678 - max_ide_version 26.30 excludes every host AI Core runs on
• SpeechToTextPlugin.kt:700 - the description's 60 s timeout is 132 s in the code
• SpeechToTextPlugin.kt:521 - isProse deletes real content lines, not just lead-ins
• SpeechToTextPlugin.kt:157 - resolveLlmService can write a service back after teardown() cleared it

NITPICK - 1 inline, not listed

Stack

This is the bottom layer: main <- #93 (c1e9b9b) <- #95 (000cc97). Comments are anchored to #93's head, but every claim above was verified against the tip 000cc97, since the stack merges atomically and the tip is the state that reaches main.

One finding was dropped for that reason, and it is worth naming. The new recordingState != RecordingState.IDLE guard in startVoiceCapture removes main's only recovery from a wedged SpeechRecognizer: on main, a second tap ran destroyRecognizer() and started fresh, so a recognizer that never delivers a terminal callback left the toolbar stuck on the waves icon with every tap answered by stt_busy. #95 repairs it in the same condition by adding elapsed < STALE_CAPTURE_MS (SpeechToTextPlugin.kt:318 at the tip), freeing the button after 147 s. It cannot reach main on its own, so it is not posted - but this layer does not stand alone on that point.

Re-check of the earlier rounds

All 26 prior findings (F01-F26) are fixed at c1e9b9b. Verified in the code, not from the replies:

• F01/F02/F18 fence parsing - I re-implemented stripCodeFences and ran the original repros: ```println("hi")``` yields println("hi"), an unclosed fence keeps everything that arrived, and code on the opening line survives a multi-line block.
• F20/F24 stripLanguageInfo - ```c = a + b``` and ```bash -c "..." are left intact, so the over-deletion is gone. The allowlist that replaced thelanguage` comparison is what the first IMPORTANT above is about - a new hole in the same function, not a regression of F20.
• F03/F19 teardown - deactivate() and dispose() both route through one idempotent teardown() that removes the listener, clears the three services and cancels the scope.
• F04/F14 cancellation - the CancellationException catch rethrows only when the job is genuinely cancelled, and the finally posts setState(IDLE) through runOnMain rather than suspending.
• F05 language is a required parameter read on the main thread. F06/F07 await() replaces the blocking get, and the cancel(true) claim is gone from the comment. F08 the timeout is derived from maxTokens (see the MINOR above). F09 llmService is @Volatile. F10 logger is guarded by ::context.isInitialized. F12 systemPrompt/temperature/maxTokens are set on the request. F13 the getPluginService rung is present. F17/F23 dropProse filters the whole reply (see the MINOR above). F21/F22 state reset on activate, plus the busy guard. F25 completedWithoutResponse separates a null response from a timeout. F26 editorService/uiService are cleared in teardown.

Build

../../gradlew assemblePlugin at c1e9b9b succeeds in a clean worktree and produces speech-to-text.cgp, with no Kotlin warnings. Per CLAUDE.md that is not verification: both IMPORTANT findings are device-visible and neither shows up in a build.

Rules applied

This repo has no REVIEW.md, CONTRIBUTING.md or PR template, and the cogo-plugin-review skill states no approve/request-changes rule, so CLAUDE.md is the only governing document and the default verdict rule applied.

[itsaky-adfa]

Requesting changes on the two IMPORTANT findings from the review above; the four MINOR ones and the nitpick do not block.

1. SpeechToTextPlugin.kt:536 - gate the LANGUAGE_TAGS allowlist on the inline ```tag code``` form. On a plain opening fence line the remainder is an info string by definition, so drop it whatever the language, and dart, kts, rust, html, yaml and cmake stop being written into the file as code.
2. plugins/Speech-to-Text/src/main/AndroidManifest.xml:34 - raise plugin.min_ide_version to at least 26.29, where SharedServices, getPluginService and addPluginLifecycleListener first shipped. 26.35 is the better value, since that is the lowest host AI Core supports.

Both are device-visible and neither shows up in assemblePlugin, which passes at c1e9b9b. Worth exercising the fence path on a non-Kotlin file (a .dart or build.gradle.kts) on a device before the next round, per the verification note in CLAUDE.md.

[itsaky-adfa]

Fourth pass, against head 96067fb.

Stack. This is the lower layer of a two-PR stack: main <- #93 <- #95 (ADFA-5404). Every claim below was checked against the stack tip e2df93b rather than this PR's own head, because the stack merges atomically. #95 touches this same file but not one line of stripCodeFences, dropProse, isProse, stripLanguageInfo, resolveLlmService or forgetLlmService, so nothing below is already repaired higher up.

My CHANGES_REQUESTED of 2026-09-14 should be lifted - both findings it named are fixed. Re-check of all seven prior findings, read at the tip rather than taken from the replies:

Prior finding	State at head
IMPORTANT stripCodeFences - a tag outside LANGUAGE_TAGS inserted as code	Fixed. The !closesInline && !opener.contains(' ') gate drops a lone opener whatever it names. Re-ran my implementation: dart, kts, rust, html, yaml, cmake and the unlisted jsx all now yield only the body.
IMPORTANT min_ide_version 26.17	Fixed - 26.29, where SharedServices, getPluginService and addPluginLifecycleListener first shipped.
MINOR max_ide_version 26.30 excluded every AI Core host	Fixed - 26.99 now covers ai-core's 26.35 floor.
MINOR PR description's "60-second timeout"	Fixed - the description states 132 s and shows the arithmetic.
MINOR isProse deletes interior content lines	Partly fixed - dropWhile/dropLastWhile saves interior lines; the first and last still go. Reply in thread.
MINOR resolveLlmService can resurrect a cleared service	Partly fixed - teardown() is ordered by serviceLock; forgetLlmService still writes unlocked. Reply in thread.
NITPICK withContext(Dispatchers.IO) no-op	Fixed - removed, with a comment saying why.

This round - all MINOR, nothing above it:

MINOR

• SpeechToTextPlugin.kt:509 - an inline-closed fence absorbs the model's trailing explanation into the file
• SpeechToTextPlugin.kt:286 - the new busy guard removes the only recovery from a stuck RECORDING state
• SpeechToTextPlugin.kt:557 (thread reply) - a multi-word fence info string still reaches the file
• SpeechToTextPlugin.kt:530 (thread reply) - dropProse still drops a real first or last line, silently
• SpeechToTextPlugin.kt:182 (thread reply) - forgetLlmService writes outside serviceLock

MINOR is by definition safe to merge, so this review is a COMMENT.

One candidate dropped. The asymmetry between the unguarded addPluginLifecycleListener at line 116 and the runCatching around removePluginLifecycleListener at line 204 does not survive checking: the asymmetry is reasoned - teardown() runs on paths where add never ran - and min_ide_version 26.29 is exactly what keeps that call off a host without the method.

Verification. Toolchain versions is green at head and the branch compiles, but per CLAUDE.md neither is verification for this plugin, and nothing here was exercised on a device this round. The fence and prose paths are only really checkable by speaking into a real editor against a real backend. This repo has no written approve/request-changes rule - CLAUDE.md governs, and it speaks to device verification rather than merge gating - so the default severity rule applied.

[itsaky-adfa]

Approving, which lifts my CHANGES_REQUESTED of 2026-09-14. Both findings it named are fixed at 96067fb: the fence gate is now language-agnostic rather than allowlist-gated, and min_ide_version is 26.29.

The five findings from this round are all MINOR - safe to merge, worth a look before or after:

• SpeechToTextPlugin.kt:509 - an inline-closed fence absorbs the model's trailing explanation into the file. A one-line suggestion is attached to that comment.
• SpeechToTextPlugin.kt:286 - the new busy guard leaves no recovery from a stuck RECORDING state.
• Three reopened threads: the multi-word fence info string, dropProse silently dropping a real first or last line, and forgetLlmService writing outside serviceLock.

Nothing here was exercised on a device. Per CLAUDE.md that is not verification for this plugin, so the fence and prose paths are still worth one run against a real backend - a .dart or .md file is the interesting case.