akashi_anthropic 0.3.0

Anthropic (Claude) provider adapter for the Akashi agent framework: wraps anthropic_sdk_dart behind Akashi's LanguageModel contract, normalizing text/thinking/tool_use blocks.

Changelog

0.3.0

• Coordinated 0.3.0 release (tracks akashi 0.3.0).
• AnthropicProvider.close() releases the shared HTTP client (a no-op when an external client was injected).
• Complete FinishReason mapping: tool_use → toolCalls and refusal → contentFilter, instead of collapsing everything but max_tokens to stop.
• Streaming now honors request.cancel, stopping the upstream drain on cancellation.
• Uses the shared partsToText / encodeToolOutput helpers from akashi core.

0.2.0

Initial release — Anthropic (Claude) provider adapter over anthropic_sdk_dart.

• AnthropicProvider + ClaudeModel (LanguageModel, StructuredOutputCapable with toolMode/promptOnly).
• Content-block streaming normalization: text → TextDeltaPart, thinking → ReasoningDeltaPart (carrying the block signature), tool_use → ToolCallStartPart + input_json ToolCallDeltaPart.
• Leading SystemMessages map to the top-level system field; ToolChoice → Anthropic tool_choice; a default max_tokens is supplied when unset.
• Offline tests against a mock HTTP client; a runnable example/.

Note: reasoning signatures are surfaced on the way out, but anthropic_sdk_dart 5.x exposes no thinking input block, so prior thinking is not re-sent on subsequent turns.