flutter_ai_assistant 0.2.0

Drop-in AI assistant for Flutter apps. Reads UI via Semantics tree, executes multi-step tasks autonomously, and supports Gemini (Firebase AI Logic), Claude, and OpenAI.

0.2.0

Added

• FirebaseAiProvider — new recommended Gemini provider built on firebase_ai (Firebase AI Logic). Routes calls through Firebase so the Gemini API key never ships in the app binary, supports Firebase App Check (Play Integrity / App Attest / reCAPTCHA Enterprise) for per-request platform attestation, and unlocks implicit prompt caching automatically on Gemini 2.5+ models.
• FirebaseAiProvider.streamMessage(...) for incremental streaming responses in custom chat UIs (the ReAct agent loop still uses non-streaming sendMessage).
• FirebaseAiProvider.generateStructured(...) — one-shot prompt with responseSchema JSON output for direct use outside the agent.
• ThinkingConfig parameter on FirebaseAiProvider to expose Gemini 2.5 thinking budgets.
• bin/generate.dart learnt a Vertex AI + Application Default Credentials auth path. Pass --project=YOUR_GCP_PROJECT after running gcloud auth application-default login; no long-lived GEMINI_API_KEY needed in .env or CI variables.

Changed

• Default Gemini model bumped to gemini-2.5-flash (was gemini-2.0-flash).
• README: new "Recommended: FirebaseAiProvider" section documenting three security postures (with App Check, Firebase without App Check, legacy raw key) and a "Migrating from GeminiProvider" subsection.

Deprecated

• GeminiProvider is now @Deprecated. Upstream google_generative_ai was archived by Google in December 2025. GeminiProvider will be removed in v0.3.0; migrate to FirebaseAiProvider.
• bin/generate.dart --api-key / --env paths emit a deprecation warning recommending --project= + ADC.

Security

• The migration path eliminates the need to ship a Gemini API key inside the app binary. Existing apps that have shipped GeminiProvider(apiKey: ...) should rotate the key in Google AI Studio after upgrading — keys already in distributed binaries remain extractable.

Platform support

• firebase_ai does not ship Linux or Windows desktop bindings. Apps that target those platforms and used GeminiProvider should either pin to ^0.1.x or switch to ClaudeProvider / OpenAiProvider. Android, iOS, Web, and macOS are fully supported.

Breaking

• GeminiProvider now produces an analyzer deprecation warning at every instantiation site. The class still works for one minor cycle.
• Default Gemini model identifier changed (gemini-2.0-flash → gemini-2.5-flash). Override the model parameter to keep the previous model.

0.1.2

Fixed

• Suppress deprecated hasFlag/flags warnings for newer Flutter SDK compatibility.
• Resize demo GIFs in README to 300px side-by-side layout.

0.1.1

Fixed

• Dart formatter compliance across all source files (50/50 static analysis).
• Lint issues in semantics_walker.dart (curly_braces_in_flow_control_structures).
• Unresolved dartdoc reference [enableLogging] in AiLogger.

0.1.0

• Initial release.
• ReAct agent loop with multi-step task execution (tap, type, scroll, navigate).
• Semantics-based screen reading — no instrumentation needed.
• LLM providers: Gemini, Claude, OpenAI.
• Voice input (speech-to-text) with multi-locale support.
• Voice output (text-to-speech) with language auto-detection.
• Rich chat overlay with action feed, handoff mode, and response popups.
• Custom tool registration for app-specific business logic.
• App manifest generation for structured screen descriptions.
• Analytics event system via onEvent callback.