feat: AI-powered features (translation, summary, reformulation, suggested replies)

[sk7n4k3d]

Summary

Full integration of AI-powered features using a local LLM (Ollama) via OpenAI-compatible API. Privacy-first: all AI processing happens on the user's own infrastructure, no data sent to third-party services.

Features

• Auto-translation — Async background translation of timeline messages with LRU memory + disk cache, reply quote translation, "Show original" toggle, translated room list previews
• Conversation summary — Summarize last 50 messages via toolbar menu, bottom sheet with copy button
• Message reformulation — Long-press context menu in composer with 4 styles (Formal, Casual, Concise, Fix Grammar)
• Suggested replies — 2-3 contextual reply chips above the composer
• Notification summary — Auto-summarize missed notifications after 5min+ absence, leverages translation cache

Architecture

• All features use TranslationService.complete() → single OpenAI-compatible API endpoint
• Shared translation cache (5000 entries LRU + JSON disk persistence)
• Rate limiter (3 concurrent requests max)
• Debounced timeline cache invalidation (200ms batching)
• Dedicated settings page with per-feature toggles
• Configurable API URL, key, model, and target language

Files changed (38 files, +2668 lines)

features/ai/
    reformulate/ReformulationBottomSheet.kt
    summary/ConversationSummaryBottomSheet.kt
    suggestedreply/SuggestedReplyService.kt, SuggestedReplyView.kt
    notificationsummary/NotificationSummaryService.kt, NotificationSummaryBottomSheet.kt
features/translation/
    TranslationService.kt, TranslateConfig.kt, TranslateApi.kt
    TranslationCache.kt, TranslationRateLimiter.kt
    TimelineTranslationManager.kt, VectorSettingsTranslationFragment.kt

Key design decisions

• Privacy-first: No cloud AI services — works with Ollama, llama.cpp, or any OpenAI-compatible local endpoint
• Cache-based room list previews: No extra API calls, reuses translations already in cache
• Reply quote translation: Extracts quoted text from Matrix reply fallback, translates separately, injects into blockquote HTML
• Debounced invalidation: Groups multiple translation completions into a single Epoxy model rebuild

Test plan

• Enable translation in Settings → Translation & AI
• Configure Ollama URL and model
• Verify auto-translation of messages in a room timeline
• Verify "Show original ▾" toggle below translated messages
• Verify translated previews in room list
• Test conversation summary via ⋮ menu
• Test reformulation via long-press in composer
• Verify suggested reply chips above composer
• Test notification summary after 5min+ absence

🤖 Generated with Claude Code

[CLAassistant]

Thank you for your submission! We really appreciate it. Like many open source projects, we ask that you sign our Contributor License Agreement before we can accept your contribution.
_{You have signed the CLA already but the status is still pending? Let us recheck it.}