flutter_ai_provider_gemini 0.1.13

Google Gemini LlmProvider for flutter_ai: streams text, thinking, function calls, and Google Search grounding citations as flutter_ai_core AiStreamEvents.

Changelog

0.1.13

• Formatting-only: reflow gemini_event_parser.dart to dart format style (no behavior change). Restores the pub.dev static-analysis score.

0.1.12

• Fix: a mid-stream error payload is now surfaced as a StreamErrorEvent, and a blocked prompt (promptFeedback.blockReason) finishes as content-filtered — instead of both being ignored and rendered as an empty, successful message.
• Fix: tool results delivered across consecutive tool messages (one result per message) are no longer dropped — they coalesce into a single functionResponse turn in call order.
• Fix: reasoningEffort now also sets thinkingConfig.includeThoughts, so the paid thinking is actually surfaced as ReasoningDeltas instead of being invisible.

0.1.11

• Map AiRequestOptions.reasoningEffort to generationConfig.thinkingConfig. thinkingBudget. Requires flutter_ai_core ^0.1.13.

0.1.10

• Fix (Web): the default HTTP client now streams token-by-token on Flutter Web. http.Client() resolves to the XHR-backed BrowserClient on the web, which buffers the entire response body before the stream emits — silently degrading streaming to all-at-once. The default is now a fetch-based client (via a conditional import) that reads the response ReadableStream incrementally. Native platforms are unchanged. Inject your own client to override.

0.1.9

• Fix: raise the flutter_ai_core lower bound to ^0.1.11 — the parser emits AiUsage (added in core 0.1.3) and later APIs, so the old ^0.1.0 bound let dependency downgrades resolve a core that couldn't compile.
• Docs: shortened the pubspec description into pub.dev's 60–180 character window.

0.1.8

• Docs: refreshed the README listing with a hero image, screenshot gallery, and badges (consistent across the package family). No code changes.

0.1.7

• Implements EmbeddingProvider (embed → batchEmbedContents, default text-embedding-004) and TokenCounter (countTokens → the countTokens endpoint, reusing the request content mapping).

0.1.6

• Tool correlation: functionResponse parts are now emitted in the same order as their turn's functionCalls and keyed by name, so two calls to the same tool in one turn line up with their own results (Gemini matches by name+position, having no id channel). Synthesized call ids are order-distinct.
• Declares supported platforms: (all 6).

0.1.5

• Throws typed LlmExceptions (auth/rate-limit/server/request) on HTTP errors instead of a generic Exception; retries 408/409 too.

0.1.4

• A mid-stream stall surfaces a message-scoped StreamErrorEvent instead of also finalizing (which masked the timeout).
• Asserts a non-empty apiKey with an actionable message.

0.1.3

• Structured output: maps AiRequestOptions.responseFormat to generationConfig.responseSchema + responseMimeType: application/json.

0.1.2

• Reports token usage: parses usageMetadata (prompt/candidates/total, cached content + thoughts tokens) into AiUsage on MessageFinished.

0.1.1

• Docs: added a "Buy me a coffee" (Ko-fi) support section to the README. No code changes.

0.1.0

Initial release.

• GeminiProvider — an LlmProvider for the native Gemini API (models/{model}:streamGenerateContent), with an injectable HTTP client and a configurable default model (gemini-2.5-flash).
• Supports Google Search grounding (enableGrounding): grounded answers stream their web sources back as SourcePart citations.
• Maps conversations to Gemini's wire format: system → systemInstruction, assistant tool calls → functionCall, tool results → functionResponse (function name recovered from the matching call). Streams text, thinking, function calls, citations, and finish reasons as AiStreamEvents.
• GeminiEventParser — the chunk→event mapping, unit-tested against recorded SSE.
• Robustness: configurable connect + idle timeout (a stalled stream surfaces a StreamErrorEvent instead of hanging); a wrong-shape chunk emits a StreamErrorEvent instead of crashing the stream; close() only closes a client it created; retry backoff is now capped and jittered. When function tools and enableGrounding are both set, grounding is omitted (the API rejects combining them) rather than 400-ing.
• Re-exports flutter_ai_core.

The mapping is unit-tested against recorded SSE chunks; it has not been run against the live Gemini API in this release.