flutter_ai_core 0.1.13

Dependency-free Dart foundation for AI chat: message models, a streaming MessageProcessor, and the provider and renderer contracts the flutter_ai family builds on.

Changelog

0.1.13

• ReasoningEffort (minimal/low/medium/high) + AiRequestOptions.reasoningEffort: a provider-neutral knob for how hard a reasoning model should think. Exposes budgetTokens (a canonical effort→budget heuristic) and openAiValue for providers to map onto their native control. Additive and opt-in.

0.1.12

• Docs: shortened the pubspec description into pub.dev's 60–180 character window so it renders in full in search results. No code changes.

0.1.11

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

0.1.10

• New AI primitives (opt-in, additive):
  • EmbeddingProvider / AiEmbedding and TokenCounter capability interfaces a provider may implement (check with provider is ...).
  • GenerateObject extension on LlmProvider: generateObject returns a decoded Map constrained to an AiResponseFormat; streamObject yields the evolving partial object as it streams (via JsonAccumulator).

0.1.9

• validateJsonSchema: a tiny, dependency-free validator for the JSON-Schema subset LLM tool declarations use (type, properties, required, items, enum, additionalProperties: false, numeric/string/array bounds, union types). Returns human-readable violation messages. UseChatController uses it to validate tool-call args before execution.

0.1.8

• Perf: streaming text/reasoning deltas accumulate into a per-part StringBuffer and materialize the String lazily, instead of last.text + delta reallocating the whole answer on every token (was quadratic on long responses — the hottest path in the stack). Observably identical: TextPart.text/ReasoningPart.text still return a plain String.
• AiUsage.cacheCreationTokens: carries prompt-cache write tokens (a subset of inputTokens) distinctly; estimateCost bills them at cacheWritePer1M (defaulting to 1.25 * inputPer1M) so cache writes aren't billed at the base input rate.
• Declares supported platforms: (all 6).

0.1.7

• Typed errors: LlmException hierarchy (LlmAuthException, LlmRateLimitException, LlmServerException, LlmRequestException) + a llmExceptionFor mapper, surfaced on StreamErrorEvent.error so hosts can branch on the failure type instead of string-matching.

0.1.6

• ReasoningPart / ReasoningDelta gain an optional signature (preserved and replayed so providers like Anthropic accept thinking blocks on tool rounds).
• MessageProcessor keeps the last good partial tool-call args instead of clobbering them to {} mid-stream.

0.1.5

• AiRequestOptions.cachePrompt: hint that the stable prompt prefix (system + tools) should be cached. Anthropic applies cache_control; OpenAI/Gemini cache automatically (no-op).

0.1.4

• AiResponseFormat (+ AiRequestOptions.responseFormat): request structured output constrained to a JSON schema. Providers route it to their native mechanism; the assistant's text is the JSON object.

0.1.3

• AiUsage model (input/output/cached/reasoning/total tokens) with + to accumulate and estimateCost(...) for cost from per-million prices. Carried on MessageFinished and stored on the completed AiMessage; the processor applies it on finish.

0.1.2

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

0.1.1

Bug fixes in MessageProcessor:

• Zero-argument tool calls (a ToolCallReady with no streamed arguments) now resolve to empty args + inputAvailable instead of being marked errored.
• A ToolResultReceived whose messageId differs from the call's message (the normal case — results arrive in a separate tool-role message) now correctly advances the original call to outputAvailable.
• A tool-scoped StreamErrorEvent (with toolCallId) now marks only that call errored and lets generation continue, instead of failing the whole message — matching UseChatController.
• Doc fix: corrected a stale reference to flutter_markdown_plus.
• JsonAccumulator no longer surfaces an unterminated trailing number/keyword (e.g. 1234 from {"n": 1234) as a complete value — a literal must be delimiter-terminated, preserving the "a partial is always a prefix" contract.
• MessageProcessor resolves a tool result to its owning call by scanning the conversation when the in-memory map misses (after reset()/rehydration).
• deepHash uses order-independent hashing for maps (better distribution).

0.1.0

Initial release.

• Models: AiConversation, AiMessage, AiMessageStatus, AiRole, FinishReason, and the sealed AiPart hierarchy (TextPart, ReasoningPart, ToolCallPart, ToolResultPart, FilePart, SourcePart, DataPart) with manual JSON serialization and value equality.
• Streaming: sealed AiStreamEvent set, MessageProcessor reducer with granular MutationResults, and the tolerant JsonAccumulator for partial tool-call arguments.
• Contracts: LlmProvider, TextRenderer, AiRequestOptions, ToolDefinition.
• Zero runtime dependencies (dart:core + dart:convert only).