ADFA-4589 | Add suite of AI-powered developer plugins for Code on the Go#41
ADFA-4589 | Add suite of AI-powered developer plugins for Code on the Go#41jatezzz wants to merge 17 commits into
Conversation
![claude[bot]](https://avatars.githubusercontent.com/in/1236702?s=60&v=4){kind=small}
There was a problem hiding this comment.
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.
e1a6598 to
4986735
Compare
|
Might be too early to properly review this. They all need version ranges, html documentation, tier 1/2/3 help, and README.md. CoroutineScope(Dispatchers.IO) and never gets cancel() in dispose(). @volatile static pluginContext never cleared, native model never unloaded, llama executor never shut down. ai-assistant has security problems
We also need the new plugin api jar otherwise these won't build. |
c9ce9b8 to
9f7c87e
Compare
|
@hal-eisen-adfa Thanks for the feedback, I've addressed all the concers, attached the proof that the new plugins build with the new Code On the Go plugin api. 
https://github.com/appdevforall/plugin-examples/actions/runs/29523333464 |
|
Almost everything is now fixed. One big gap that wasn't addressed yet: LLamaAndroid needs to expose a shutdown that calls (runLoop as ExecutorCoroutineDispatcher).close() / executor.shutdown(), invoked from AiCorePlugin.dispose(). |
|
Added day and night icons for AI plugins 
|
|
Update theme and model validation on fresh installs document_5114081575834749100.mp4 |
hal-eisen-adfa
left a comment
hal-eisen-adfa
left a comment
There was a problem hiding this comment.
Requesting changes. Findings from an xhigh code review of the AI-plugin suite that live in this PR's diff, ranked most-severe first. (Two additional items scoped to the stacked API-key PR #49 are posted there.)
Also worth a cleanup pass (non-blocking): CodeChunker.simpleChunk has an unreachable "single line too large" else branch (the first line is always added, so currentChunk is never empty), and reconcileOverlaps has an empty if body / unused nextChunk and just returns its input unchanged.
Inline ghost-text AI code completions; restored onto main's split layout.
Voice-to-code speech input; restored onto main's split layout.
Semantic code search via embeddings; restored onto main's split layout.
… Downloads Stream the selected content:// URI into a private cache file; drop the Downloads fallback that silently loaded the wrong model (ADFA-4388).
…shing Add GgufModelInspector + UserFeedback; classify the .gguf and throw IncompatibleModelException before native decode SIGABRTs the IDE (ADFA-4388).
….5-flash listModels now queries the live v1beta ListModels catalog (generateContent-only); default bumped to gemini-2.5-flash and retired models migrated on load.
Readable settings title; real model name shown and persisted across reopen; and the chat backend indicator + routing now follow the selected backend, refreshing when the settings dialog closes.
…ggestion-plugin and vector-search-plugin names
- Add version ranges, README/docs, and tiered help coverage for AI plugins - Cancel plugin coroutine scopes during disposal - Move Gemini listModels API key from URL query to x-goog-api-key header - Document Gemini API key storage tradeoff and remove misleading encryption note - Update AI plugin review status with verification results
LLamaAndroid's single-thread run-loop executor was never shut down, so the non-daemon Llm-RunLoop thread outlived the plugin and pinned its classloader. Expose LLamaAndroid.shutdown() (closes the ExecutorCoroutineDispatcher) and call it from LocalLlmBackend.close() after the model unload completes. Regenerate the prebuilt llama AAR so the new API is visible to the plugin build.
… light/dark theme
- Agent tab now shows an error bubble + Snackbar ("Open Settings...") when no local model or Gemini key is configured, instead of silently doing nothing.
- Theme UI to the IDE day/night setting: route inflation through PluginFragmentHelper.getPluginInflater, make PluginTheme a Material3 DayNight theme, add values-night/ colors, and replace hardcoded hex with @color refs.
- Fix AI Settings toolbar contrast (colorOnPrimary title/icon on the primary bar).
llama run-loop now recreates after per-plugin dispose, embedding-dim guard, synced suggestion cache, capped file walk, fallback-safe Gemini model migration and vector search indexes in the background (bounded wait)
3d33b7a to
b939ff5
Compare
- Resolve backend via "auto" in ai-core - Migrate Gemini to REST instead of the google-genai SDK to fix OkHttp 3.x conflict - Implement context-aware suggestion cache key - Add STT raw-transcript fallback and vector-search roots - Refactor code: remove unused code and permissions - Fix version numbers and update plugins documentation
… suite Refresh manifests, README/HTML docs, and day/night icons for ai-core, ai-assistant, code-suggestions, speech-to-text, and vector-search; update PathGuard and search handler.
2 participants
Description
Adds three new AI plugins and hardens the existing
ai-core/ai-assistant(both already onmain):ai-core'sgetEmbeddings(), with a lexical fallback).All three share
ai-core'sLlmInferenceService(on-device llama.cpp and/or Gemini) instead of bundling their own models.Details
Stability & fixes to the existing plugins:
SIGABRTthe IDE (ADFA-4388); load the user-selected model instead of the newest.ggufin Downloads; shut down theLlm-RunLoopexecutor on dispose to stop the classloader leak; move Gemini API key to thex-goog-api-keyheader.gemini-2.5-flash, retiregemini-1.5-flash); fix settings-dialog UX and backend selection/routing.document_5073463760878832078.mp4
Ticket
ADFA-4589
Observation
llama.cppis a submodule atai-core/subprojects/llama.cpp(private fork, branchandroidide-custom, needs SSH):git submodule update --init ai-core/subprojects/llama.cpp. A prebuiltai-core/libs/v8/llama-v8-release.aaris committed, so building llama.cpp from source isn't required just to compile.libs/*.jarare CI-managed and intentionally not committed here, but vector-search needsLlmInferenceService.getEmbeddings(), so the refreshedplugin-api.jarmust be in place (CI /scripts/update-libs.sh) before building.Notes: