Skip to content

Roadmap and Changelog

Puzzak edited this page Nov 20, 2026 · 1 revision

Roadmap & Changelog

Synced with v1.8.9+26112001 (2026-11-20). For the full historical log from 1.0.0, see CHANGELOG.md and ROADMAP.md.


Recent Changelog (1.8.0 – 1.8.9)

v1.8.9 (2026-11-20)

  • ML Kit GenAI 1.0.0-beta5 Integration: Integrated model.warmup(), PromptPrefix caching, native multi-image Content ImageParts (replacing vertical bitmap stitching), countTokens / getTokenLimit, streaming FunctionCall capture, topK / seed sampling, non-deprecated getAvailableFeatures(), and fixed GenAiException.ErrorCode BUSY (9) vs AICORE_INCOMPATIBLE (-101) classification.
  • Lifecycle Sync & Voice Extension Roadmap: Added AppLifecycleState.resumed sync for dismissed Assistant Overlay chats (pending_assistant_chats) and updated mesh_proposals.json / ROADMAP.md for standalone Speech Services (TinyASR + TTS) (page.puzzak.speech).
  • Simplified Chinese (zh-Hans) Parity: Registered assets/translations/zh-Hans.json in pubspec.yaml and completed PR #36 localization across built-in prompts, resources, extensions, UI pages, and Assistant Overlay labels.
  • Kotlin Lifecycle, SAF I/O & Bitmap Memory Fixes: Fixed FlutterLocalAiPlugin.dispose() coroutine cancellation, MainActivity headless handler Activity leak, FileAccessPlugin off-main-thread SAF I/O, and bitmap scaling/recycling.
  • Sliding Window, Cutoff Trimming & Performance: Added token-budgeted sliding window context management, unpunctuated multi-sentence cutoff trimming, and removed the 500ms per-token delay, per-turn background syncRemoteDocs() HTTP calls, and background ChatsPage full-history JSON re-parsing.

v1.8.8 (2026-11-16)

  • Bare-Minimum Default System Prompt & Runtime Self-Knowledge: Slimmed system_default.md to 2 sentences refreshed from assets and expanded query_aicore_documentation (about_paios, paios_info, get_paios_info) with offline PAIOS self-knowledge.
  • Simplified Chinese (zh-Hans) & Script-Aware Locale Detection: Added zh-Hans alongside zh with Hans/CN/SG/MY vs Hant/TW/HK/MO auto-detection.
  • AICore ErrorCode 606 Diagnostics: Added isAiCoreFeatureNotFoundError OEM provisioning diagnostics in Onboarding and Model Settings.
  • Monochrome Icon & Keystore Fallback: Restored Android 13+ <monochrome> adaptive icon layer and hasReleaseKeystore debug-signing fallback in Gradle.
  • Earlier-Sentence Loop Deduplication & trimTrailingIncompleteSentence: Stripped >= 24-char earlier-sentence loops, raised maxNullReasonContinuations to 3, and trimmed dangling tails on final commits.
  • UTF-8 .md Prompt Sync & Storage Sanitization: Fixed utf8.encode exports, preserved Unicode CJK/Cyrillic SAF filenames (nameToFilename), triggered immediate import on folder selection, and sanitized legacy https://puzzak.page/PAIOS links.

v1.8.7 (2026-11-16)

  • Truncated inspect_screen Repair & Fallback: Added string/brace-aware truncated tool JSON closure repair (109-char inspect_screen drafts), raised MAX_TOOL_JSON_CONTINUATIONS to 2, added auto-continue to runToolFreeFallbackAttempt, and preserved screen TTL across aborted attempts.
  • Nano v4 Native <ctrl42>call:... Function Calling: Parsed <ctrl\d+>call:tool{...}<ctrl\d+> tokens (m3x_mesh 1.1.5), scrubbed control tags from thought bubbles (stripControlToolTags), and promoted thought-trapped tool calls.
  • Synthesis Participle & Negation STOP Rescue: Added trailing transitive -ing participle and negation STOP cutoff rescue during synthesis.

v1.8.6 (2026-11-16)

  • m3x_chat ChatThoughtsBlock Smart Auto-Scroll: Added ScrollController auto-follow with manual scroll-up pause (24dp threshold).
  • Overlay Tool JSON Hardening & Extension Registry Sync: Consolidated tool JSON repair into PaiosModelHelper, compacted schemas, synced toolRegistry with disabled modules, and added onboarding ready-state distinctions + composer draft isolation.

v1.8.5 (2026-11-16)

  • On-Demand inspect_screen (3-Turn TTL): Replaced passive force-attached overlay screenshots with lazy inspect_screen tool inspection and silent 3-turn drop/recycle.
  • Mid-Clause STOP Rescue: Added 1-shot STOP rescue for preposition/determiner mid-clause cutoffs ("with advanced").

v1.8.4 (2026-11-16)

  • Cancel & Stream Lifecycle Hardening: _streamGenerationId++ invalidation on cancel, deadlock fixes, 240-char overlap deduplication, 28-char mid-word backoff, stutter loop termination, and Turn-1 tool-free fallback (includeTools: false).

v1.8.3 (2026-11-16)

  • m3x_chat Integration, Multi-Model Downloader & v3 Continuation: Standardized ChatMessageBubble, author attribution + tap-to-reveal timestamps, native ComboStatuses model selector, parallel variantDownloadJobs downloader with speed/progress, single-stable-model onboarding gate, and mojibake-immune Dictionary.

v1.8.0 – v1.8.2 (2026-11-15)

  • Nano v4 & Thinking Mode (1.8.0): Shipped Stable (v3) / Preview (v4-full/v4-fast) switcher and collapsible Thought process bubbles.
  • Hardening (1.8.1 – 1.8.2): Fixed Candidate.FinishReason constants, anti-echo tool notes, startup timeouts, responsive narrow-viewport layout (elements.dart), and overlay ErrorCode 4 inset/chrome filtering.

Project Roadmap Overview

Phase Highlights & Current Status
Phase 1 — AI Core - Basic (13/16) Shipped: v1.0–v1.8.9 multi-chat, AICore v3/v4 & beta5 controls, 23-page 3-tier SWR Docs Explorer, dual-branch onboarding, Hive storage, GitHub explorer, slim system_default.md. Remaining: contributor credits page, monetization, final polish.
Phase 2 — AI Core - Improved (11/13) Shipped: 6-turn agentic tools, multimodality, System Assistant Overlay + inspect_screen, modular extensions, Nano v4 Thinking Mode, m3x_chat, cutoff rescue, nightly CI/CD. Remaining: Context expansion, Ephemeral Short-Term Memory buffer.
Phase 3 — Model-agnosticizm (0/4) Planned: Custom local models (e.g. Gemma 2B), remote hardware connectivity (Ollama bridge) with automatic local failover, and standalone Speech Services (TinyASR + TTS) M3x Mesh companion (page.puzzak.speech).
Phase 4 / 4.5 / 4.6 / 4.7 — Future Planned: Device/network file system access, scripting, extensions marketplace, live voice chat, on-device genai-summarization context compacting, Markdown transcript local RAG, ML Kit neural STT/TTS dashboard, and Google Gemini Cloud Handoff (ACTION_SEND).

Related Pages: Home · Architecture & Model Variants · Building from Source

Clone this wiki locally