flutter_ai_provider_openai 0.1.11

OpenAI-compatible LlmProvider for flutter_ai: streams Chat Completions (text, tool calls, finish reasons) as flutter_ai_core AiStreamEvents over an injectable HTTP client.

Changelog

0.1.11

• Fix: maxOutputTokens now maps to max_completion_tokens instead of the deprecated max_tokens, which reasoning models (o-series, gpt-5) reject with a
  400. A caller targeting an older OpenAI-compatible endpoint can still pass max_tokens via extra.
• Fix: a mid-stream {"error": ...} chunk is now surfaced as a StreamErrorEvent instead of being ignored (which left an empty, apparently successful message).

0.1.10

• Map AiRequestOptions.reasoningEffort to OpenAI's reasoning_effort. An explicit reasoning_effort in extra still takes precedence. Requires flutter_ai_core ^0.1.13.

0.1.9

• 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.

0.1.8

• 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.

0.1.7

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

0.1.6

• Implements EmbeddingProvider: embed(inputs, {model}) POSTs /embeddings (default text-embedding-3-small) and returns AiEmbedding vectors.
• 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

• Readies tool calls on any finish reason (and in finalize()), so calls can't hang in inputStreaming.
• A mid-stream stall now surfaces a message-scoped StreamErrorEvent instead of also finalizing a terminal event that masked the timeout.
• Asserts a non-empty apiKey with an actionable message.

0.1.3

• Structured output: maps AiRequestOptions.responseFormat to a json_schema response_format (with strict).

0.1.2

• Reports token usage: sets stream_options.include_usage and parses the trailing usage chunk (prompt/completion/total, cached + reasoning token details) 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.

• OpenAiProvider — an LlmProvider for the OpenAI Chat Completions API (or any OpenAI-compatible endpoint via a custom base URL), with an injectable HTTP client and a configurable default model.
• Maps conversations (system/user/assistant/tool messages, assistant tool calls, tool results) into the request, and streams text, tool calls, and finish reasons back as AiStreamEvents.
• OpenAiChunkParser — 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.
• Re-exports flutter_ai_core.

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