flutter_ai_provider_anthropic 0.1.12

Anthropic (Claude) LlmProvider for flutter_ai: streams the Messages API (text, extended thinking, tool use) as flutter_ai_core AiStreamEvents.

Changelog

0.1.12

• Fix: reasoningEffort now emits adaptive thinking ({type: adaptive}) on Claude 4.6+ models — including the default claude-opus-4-8, which rejects the legacy budget_tokens shape with a 400. Claude 3.7 and 4.0–4.5 continue to use the budgeted shape. An explicit thinking block in extra still takes precedence.
• Fix: a mid-stream error event (e.g. overloaded_error) is no longer overwritten by a synthetic successful finish — the message now settles as errored.
• Fix: setting both responseFormat and reasoningEffort no longer sends an invalid forced-tool-choice-plus-thinking request (a guaranteed 400). Thinking is dropped when structured output is requested.

0.1.11

• Map AiRequestOptions.reasoningEffort to extended thinking (thinking.budget_tokens): raises max_tokens above the budget when needed and drops temperature (the API rejects both together). An explicit thinking block in extra takes precedence. 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

• Cost accuracy: cache_creation_input_tokens now map to AiUsage.cacheCreationTokens (billed at the ~1.25x write rate) instead of being folded into base input and billed wrong.
• Declares supported platforms: (all 6).

0.1.6

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

0.1.5

• Replays signed thinking blocks before tool_use in the assistant turn, so extended thinking + tools no longer 400 on Claude 4.x.
• 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.

0.1.4

• Prompt caching: when AiRequestOptions.cachePrompt is set, marks the system prompt and the last tool with cache_control: ephemeral (caches the stable prefix for ~90% cheaper repeat input).

0.1.3

• Structured output: maps AiRequestOptions.responseFormat to a forced tool whose input is the schema; its streamed input is surfaced as the JSON answer text and the turn finishes as stop.

0.1.2

• Reports token usage: accumulates input (incl. cache read/creation) from message_start and output from message_delta 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.

• AnthropicProvider — an LlmProvider for the Anthropic Messages API (POST /v1/messages), with an injectable HTTP client, a configurable default model (claude-opus-4-8) and max_tokens.
• Maps conversations into the request: system messages fold into the top-level system field, assistant tool calls become tool_use blocks, and tool results become tool_result blocks. Streams text, extended thinking, tool calls, and finish reasons back as AiStreamEvents.
• AnthropicEventParser — the SSE-event→event mapping, unit-tested against recorded events.
• Robustness: configurable connect + idle timeout (a stalled stream surfaces a StreamErrorEvent instead of hanging); a wrong-shape event emits a StreamErrorEvent instead of crashing the stream; close() only closes a client it created; retry backoff is now capped and jittered; adjacent same-role turns are merged so the API's strict alternation isn't violated.
• Re-exports flutter_ai_core.

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