neomage 1.4.2

Agnostic AI agent engine for Flutter. API clients for AI Providers and Ollama with tool execution, skills, MCP, and a Gemini Live realtime voice/audio WebSocket client.

1.4.2 — 2026-05-27

• File Picker Compatibility & Platform Support Fix:
  • Upgraded file_picker dependency to ^11.0.2 to resolve compilation errors and restore complete platform support (iOS, Android, Web, Windows, macOS, Linux) on pub.dev.
  • Migrated FilePicker.platform.pickFiles calls to the new static FilePicker.pickFiles API to conform with modern versions of the package.

1.4.1 — 2026-05-25

• Pub.dev Maintenance and Score Optimization:
  • Resolved Flutter SDK deprecation warnings for cacheExtent in ListView and onReorder in ReorderableListView by using inline analyzer ignore statements to maintain backward and forward compatibility across older and newer Flutter versions (such as Framework 3.41.5 stable vs newer channel pre-releases).
  • Updated dependency constraints in pubspec.yaml for llamadart (supporting ">=0.5.4 <0.7.0") and file_picker (supporting ">=10.3.10 <12.0.0") to resolve pub.dev analyzer warnings and support the latest stable releases, achieving 160/160 pub points.

1.4.0 — 2026-05-21

• saia_core Dependency Update: Upgraded saia_core dependency from local path reference to version ^1.0.1 hosted on pub.dev. This allows neomage to consume the fully localization-ready, tested, and high-scoring release of saia_core.
• Swarm Orchestration (Multi-Agent Swarm): Introduced SwarmOrchestrator (lib/utils/swarm/swarm_orchestrator.dart) supporting coordinated agent roles (coordinator, researcher, implementer, reviewer, tester, planner).
  • Directed Acyclic Graph (DAG) scheduling via DependencyGraph with cycle detection using a DFS color-marking algorithm.
  • Topological Sorting (topologicalSort) to resolve execution chains in real-time.
  • Work-Stealing Scheduler to dynamically allocate pending tasks to idle agent threads, maximizing concurrency and throughput.
  • Message Bus: Direct/Broadcast messaging between agents using a pub/sub MessageBus with sharedMemory.
• Advanced Context Compaction: Consolidated CompactionService and QueuedCompactionEngine (lib/data/compact/):
  • Microcompaction: In-memory cheap strip-down that replaces large tool outputs (e.g. from grep, glob, or read_file) with lightweight markers ([Old tool result content cleared]).
  • Auto-Compaction: Automatic context-window protection that triggers when projected token usage crosses configurable thresholds.
  • Partial Compaction: A cache-friendly dual-direction compacting algorithm (PartialCompactDirection.from and PartialCompactDirection.upTo) that preserves Prompt Caching headers on LLM API endpoints.
  • PTL Truncation: Graceful truncateHeadForPtlRetry to auto-recover when API calls return Prompt Too Long errors.
• Offline Sovereignty & Local AI Execution:
  • Offline TTS (Sherpa ONNX): Integrated SherpaOnnxSaiaTtsProvider option under neom_tts leveraging VITS/Piper models in isolates for native offline voice synthesis with sub-millisecond local latency.
  • In-process Local Inference: Embedded LocalLlamaProvider wrapping a native llamadart runtime to execute GGUF models (e.g., Llama 3.2 3B or Qwen 2.5 Coder) via FFI.
  • Hardware Profiler: Integrated HardwareProfiler to dynamically detect hardware tiers and configure GPU layer offloading.
  • JavaScript Sandbox: Consolidated JsSandboxTool leveraging isolated QuickJS runtime to safely compute deterministic algorithms or parsing code without bash/host side effects.
• IDE Bridge & Developer Integration: Consolidated the IdeBridgeServer WebSocket server to establish a bi-directional JSON-RPC edit bridge with VS Code, JetBrains, and command line editors, exposing high-level commands like remote file navigation (openFile), split diffs (showDiff), compilation diagnostics (getDiagnostics), and atomic edits (applyEdit).
• Terminal & Bash Security Suite: Upgraded the bash_security.dart parser to subject terminal command strings to an extensive static validator utilizing 23 security rules (checks for IFS injection, quote-obfuscated flags, process substitution blocks $(...), locale hacks, and restricted directory targets).
• MCP (Model Context Protocol) Client Integration: Built-in support to connect to stdio/websocket MCP servers, proxy-registering remote tools directly into the global ToolRegistry with safety checks and read-only annotations.

1.3.0 — 2026-04-28

• GeminiRealtimeClient + GeminiRealtimeEvent (lib/realtime/): pure-Dart client for the Gemini Live (BidiGenerateContent) WebSocket API. Streams audio bidirectionally over a single socket so apps can build real-time voice conversations without TTS round-trips.
  • Audio formats: input PCM 16-bit 16 kHz mono LE; output PCM 16-bit 24 kHz mono LE. Caller wires its own mic (record, Web Audio API, etc.) and speaker (flutter_pcm_sound, etc.); the client only does bytes-in/bytes-out so neomage stays platform-agnostic.
  • Sealed GeminiRealtimeEvent hierarchy: GeminiSetupComplete, GeminiAudioOut, GeminiTextDelta, GeminiTurnComplete, GeminiInterrupted, GeminiRealtimeError. Pattern-match in consumers — no untyped JSON needed.
  • Test-first: 14 unit tests via an injectable WebSocketChannel fake. Covers setup envelopes, system instructions, base64 PCM encoding, text turns, server-content dispatch (audio + text + interrupt + turn-complete + error).
• New direct dependency: web_socket_channel: ^3.0.0.

1.2.0 — 2026-04-27

• Switch neom_ollama to hosted dep: was path: ../neom_modules/ai/ neom_ollama during local development; now consumes neom_ollama: ^1.2.0 from pub.dev. This unblocks the package for flutter pub publish (path deps are forbidden on pub.dev) and brings in:
  • Hardware profiling (HardwareProfile.detect()).
  • Thinking trace parsing (Qwen3, DeepSeek-R1, QwQ, gpt-oss).
  • PlainTextToolCallParser — recovers tool calls embedded as text (LMStudio bracket form, Hermes / Qwen <tool_call> tag, fenced JSON, bare-JSON last-resort) for local models that don't always populate structured tool_calls.
• Skills localization redesign: replace the skills/es/ folder of 348 duplicate-with-disclaimer markdowns with a single skill_meta_es.json catalog. All LLM-facing skill bodies now load from skills/en/ only; the Spanish-facing UI descriptions (short taglines, MX market order) live in the meta catalog. Saves ~1.5 MB in the asset bundle and removes a source of drift between locales.
• Delete 54 assets/skills/es/... entries from pubspec.yaml flutter assets section.
• Add skill_meta_es.json asset declaration (54 categories + 164 curated skill descriptions with mxOrder relevance hints for the MX market).
• Fix .gitignore: add pubspec.lock + pubspec_overrides.yaml (library packages should not track these) and remove a malformed line that had fused a comment with a pattern.

1.1.0 — 2026-04-16

• Add skills locale folders: skills/es/ and skills/en/ with 348 skills each
• Add 108 asset declarations in pubspec.yaml for locale-specific skill folders

1.0.0

Neomage: multi-provider AI agent engine for Flutter, restructured as a pub.dev package.

Package (lib/)

• Multi-provider API clients: Gemini (native), OpenAI, Anthropic, DeepSeek, Qwen, Ollama — all via a unified ApiProvider interface
• Streaming: SSE parser conforming to W3C EventSource spec, typed stream events (TextDelta, ThinkingDelta, ToolUseStart, etc.)
• Agentic tool system: 31 built-in tools (Bash, FileRead/Write/Edit, Grep, Glob, WebSearch, Agent, SendMessage, TodoWrite, etc.)
• Bash security: 20+ validation checks — command substitution, IFS injection, obfuscated flags, dangerous variables, redirection analysis
• Query engine: Multi-turn agentic conversation loop with automatic tool execution
• Context compaction: Token-aware conversation summarization with configurable strategies
• Skills framework: 283 loadable markdown skills across 40+ categories
• Personality system: 10 modular personality files (Identity, Cognition, Agency, Memory, etc.) assembled into dynamic system prompts
• MCP client: Model Context Protocol support for external tool servers
• Retry with backoff: Configurable exponential backoff with 429/529 awareness
• Domain models: Message, ContentBlock (sealed), ToolDefinition, TokenUsage, branded IDs (SessionId, AgentId)
• Error classification: ApiError with typed categories (rateLimited, authenticationError, promptTooLong, etc.)
• Platform abstraction: dart:io stubs for web compatibility
• Sint framework: State management, DI, navigation via Sint + SintSentinel

Example App (example/)

• Full AI coding assistant: 8 screens (chat, onboarding, settings, Ollama setup, MCP panel, doctor, session browser, splash)
• 30+ widgets (input bar, message renderer, markdown preview, diff view, command palette, permission dialogs, etc.)
• Vim mode, keybinding system, dark/light Material 3 theme
• Skills browser with search, categories, and two loading modes (session visible / context silent)
• Localization: English + Spanish
• Cross-platform: macOS, Web, Linux, Windows, iOS, Android

Infrastructure

• Package/app split: reusable logic in lib/, full app in example/
• Barrel exports with show/hide for clean API surface
• 0 analysis errors, pub.dev dry-run validated

0.1.0

• Initial development release
• Multi-provider support: Gemini, OpenAI, Anthropic, DeepSeek, Qwen, Ollama
• Streaming chat, agentic tool execution, MCP client
• Cross-platform Flutter app