vantura 1.2.0

A Stateful Agentic AI Framework for Flutter enabling on-device LLM agents with ReAct loops, state recovery, custom tools, and seamless coordination.

Changelog

All notable changes to the Vantura SDK will be documented in this file.

The format is based on Keep a Changelog, and this project adheres to Semantic Versioning.

---

1.2.0 - 2026-08-24

Added

• OpenRouter Integration: Native support for OpenRouter via the new OpenRouterClient. Includes site-attribution headers (HTTP-Referer, X-Title) for improved rankings and visibility.
• Reasoning Token Support: Advanced support for "Thinking" and "Reasoning" tokens from models like DeepSeek R1, Claude 3.7, Gemini 2.0 Thinking, and stealth models. These tokens are now captured and streamed to the UI in real-time.
• Debugger Agent: Added a specialized debugger agent in the example app to demonstrate the ApiTestTool for live network diagnostics and API verification.
• UI Error Banners: Implemented automated error propagation from agents to the UI, displaying actionable SnackBar alerts for tool failures or LLM API errors.

Fixed

• Memory Summarization Edge Case: Resolved a critical bug in VanturaMemory where tool call sequences (Assistant -> Tool) were broken during context summarization. Introduced adaptive keepCount to preserve recent context.
• SSE Stream Integrity: Fixed a duplication bug in AnthropicClient streaming where the final response was re-yielded at the end of the stream.
• System Prompt Continuity: Fixed Gemini and Anthropic client adapters to correctly concatenate multiple system messages (persona + historical summaries) into a single unified instruction block.
• Riverpod Lifecycle Safety: Hardened the example app with mounted checks to prevent crashes when async streams complete after a screen has been disposed.

Changed

• Multi-Agent Coordination: Expanded the default AgentCoordinator setup in Orbit to 4 specialist agents (Concierge, Manager, Analyst, Debugger).
• Dependency Upgrades: Upgraded all core and example dependencies to latest versions, resolving critical build issues in the jni and riverpod packages.

1.1.1-dev - 2026-05-27

Changed

• Dependency Updates:
  • device_info_plus: ^12.0.0 → ^13.1.0 (major version bump for latest platform APIs).
  • flutter_dotenv: ^5.1.0 → ^6.0.1 (dev dependency, updated for Dart 3.11 compatibility).
• Project Hygiene: Added grand-report.md to .gitignore and .pubignore to keep internal reports out of version control and published packages.

Example App (Orbit v0.4.0)

• Migrated from single VanturaAgent to AgentCoordinator with multi-agent routing (Concierge, Manager, Analyst).
• Added InventoryJanitor autonomous background worker and LogsScreen observability console.
• Overhauled Dashboard with SliverAppBar, Outfit typography, and refined dark palette.
• Updated example dependencies (device_info_plus → ^13.1.0, sqlcipher_flutter_libs → ^0.7.0+eol).

1.1.0 - 2026-03-07

Added

• VanturaSecurity: A dedicated privacy and security layer for the framework. Features automatic PII detection and redaction (Emails, Phone Numbers) to prevent sensitive data from reaching LLM providers.
• Enhanced Observability: Added internal hooks for fine-grained logging of reasoning steps and tool execution status, supporting real-time log monitoring in consuming apps.
• Worker Redaction: Extended VanturaWorker with built-in security proxies to ensure PII redaction works seamlessly in background Isolates.

Changed

• Multi-Client Resiliency: Finalized onRetry implementation for AnthropicClient and GeminiClient, ensuring better recovery from network transient errors and rate limits.
• Navigation Intelligence: Improved NavigationTool compatibility with complex nested routing and dynamic parameter passing.

Fixed

• Isolate Data Type Conversion: Fixed a critical issue where _Map<String, dynamic> (internal SDK types) caused type errors when being sent across Isolate boundaries in VanturaWorker.
• Reasoning Loop Reliability: Resolved an edge case where certain streaming responses could occasionally swallow the final token if usage data arrived simultaneously.

1.0.0 - 2026-02-28

Added

• Stable 1.0.0 Release: First major stable, production-ready version of Vantura.
• Multi-Provider LLM Support: Introduced LlmClient interface alongside fully native AnthropicClient (Claude 3.5/3.7+) and GeminiClient (Google Gemini 1.5/2.0+) wrappers, complementing the existing OpenAI-compatible VanturaClient.
• Agent Checkpointing: Implemented AgentStateCheckpoint serialization, allowing VanturaAgent to seamlessly resumeFrom() interrupted ReAct loops via the VanturaPersistence layer.
• Conditional Confirmation: Expanded tool logic with dynamic requiresConfirmationFor(args), allowing agents to skip UI confirmation dialogs for low-risk actions automatically.
• Unified Main Library: Created lib/vantura.dart as the primary standard entry point for the package.
• VanturaException Hierarchy: Integrated standardized exception types (VanturaToolException, VanturaCancellationException, VanturaIterationException, etc.) for robust error handling.
• VanturaClient Transparency: Added onRetry callback to clients for better observability during network retries.

Changed

• Hardened Reasoning Loop: Refactored VanturaAgent ReAct loop with improved scoping, iteration control, and standardized failure state updates.
• Dependency Optimization: Removed google_fonts from the core framework to drastically lighten the compiled application footprint.
• Documentation Coverage: Achieved 100% dartdoc coverage for all core public classes and methods.
• Polished Examples: Updated example/example.dart and the Orbit "Business Management" suite to showcase multi-provider instantiation and dynamic confirmation logic.
• Lint Optimization: Resolved all static analysis issues and major pub.dev lint warnings.

0.1.3 - 2026-02-24

Added

• Official Example App: Complete Flutter business management app demonstrating client management, inventory tracking, invoicing, financial ledger, and analytics with advanced tool features.
• Advanced Tool Documentation: Added sections for confirmation flows, complex parameter schemas (arrays, enums, nested objects), and business logic integration.
• Enhanced Examples Section: Expanded README.md Examples section with detailed feature descriptions.

Changed

• Documentation Improvements: Polished code examples, updated example repository URLs, and ensured consistency across all .md files.
• Package Description: Extended pubspec.yaml description to meet pub.dev requirements.

Notes

• Improved package presentation and developer experience with comprehensive examples and documentation.

0.1.2 - 2026-02-23

Added

• Example: Created a standard example/example.dart and example/README.md for better package presentation.
• Enhanced Intelligence: Finalized persistence for tool_calls and tool_results in VanturaMemory, enabling consistent multi-turn reasoning.

Changed

• Security & Privacy Hardening:
  • Implemented recursive redaction of sensitive keys (API keys, Authorization tokens) in logs.
  • Added VanturaLoggerOptions to disable logging of sensitive content (prompts/responses) by default.
  • Added VanturaLogLevel to filter SDK logs.
  • Hardened ApiTestTool with hostname blacklisting (SSRF protection) and strict output truncation.
  • Added SDK-level guardrails to prevent instruction overriding and prompt injection.
• Robustness:
  • Added execution timeouts to VanturaTool (default 30s) to prevent agent hangs.
  • Implemented robust JSON extraction to handle LLM conversational filler and markdown blocks.
  • Added finishReason to VanturaResponse for better observability.
• Dependency Updates: Modernized constraints for connectivity_plus (v7) and device_info_plus (v12).
• Documentation: SIGNIFICANT improvements to API documentation (dartdoc) and a completely rewritten high-end README guide.
• Project Structure: Moved repository marketing assets to root and cleaned up .pubignore.

0.1.1 - 2026-02-22

Changed

• Documentation Upgrade: Completely rewritten "marketing-first" README with premium visuals, use-cases, and comparison tables.
• Package Polishing: Standardized LICENSE to BSD 3-Clause and added .pubignore for a leaner package size.
• Bug Fixes: Regenerated test mocks to ensure internal test suite compatibility.

0.1.0 - 2026-02-22

Added

• Core Agent Loop: VanturaAgent with ReAct-style reasoning (up to 10 iterations), supporting both run() (blocking) and runStreaming() (SSE token-by-token).
• HTTP Client: VanturaClient with connection pooling, 3-retry exponential backoff, and automatic HTTP 429 rate-limit handling.
• Memory System: VanturaMemory with short-term (recent messages) and long-term (LLM-summarized) memory, plus automatic pruning.
• Persistence Interface: Abstract VanturaPersistence — plug in any storage backend (SQLite, Hive, Firestore, etc.) without SDK modification.
• Tool Framework: Generic VanturaTool<T> base class with type-safe argument parsing, JSON Schema definitions via SchemaHelper, and human-in-the-loop confirmation support.
• Built-in Tools: CalculatorTool, NetworkConnectivityTool, DeviceInfoTool, ApiTestTool.
• Multi-Agent Coordination: AgentCoordinator class for routing conversations between specialist agents with automatic transfer_to_agent tool injection.
• Request Cancellation: CancellationToken for aborting generation mid-request at both the HTTP and reasoning-loop levels.
• Token Usage Tracking: TokenUsage object returned in VanturaResponse with promptTokens, completionTokens, and totalTokens.
• Error Callbacks: onToolError, onAgentFailure, and onWarning hooks on VanturaAgent for structured error reporting and telemetry.
• UI State Management: VanturaState (ChangeNotifier) for syncing agent status (isRunning, currentStep, errorMessage) with Flutter UI.
• Markdown Renderer: Zero-dependency MarkdownText Flutter widget supporting headers, bold, italic, inline code, bullet lists, and horizontal rules.
• Logger Interface: Abstract VanturaLogger with a default SimpleVanturaLogger (colored console output).

Notes

• First public release as an extracted, standalone Flutter package.
• Compatible with any OpenAI-compatible API provider (Groq, OpenAI, Ollama, Together AI, etc.).