Interleaved reasoning_content lost for non-Claude models on Anthropic SDK

[llc1123]

Description

When non-Claude models (e.g. kimi-k2.5) are configured via the Anthropic SDK (@ai-sdk/anthropic) with interleaved reasoning capability, reasoning content from previous assistant turns is silently dropped in subsequent requests.

The root cause is that the Anthropic SDK only knows how to serialize thinking blocks — it has no concept of a top-level reasoning_content field on assistant messages. So when opencode replays prior assistant messages that contain reasoning parts, the SDK discards them and the upstream API never receives the reasoning history.

Steps to reproduce

1. Configure a non-Claude model (e.g. kimi-k2.5) using @ai-sdk/anthropic provider
2. Enable thinking/reasoning
3. Have a multi-turn conversation where the model produces reasoning content
4. Observe that on subsequent turns, the prior reasoning content is missing from the request sent to the API

[github-actions]

This issue might be a duplicate of existing issues. Please check:

• HuggingFace provider: reasoning_details not stripped from conversation history for Kimi-K2.5 #13002: Similar issue with Kimi-K2.5 reasoning content not being properly handled in conversation history (HuggingFace provider, reasoning_details field causing 400 errors on subsequent turns)
• DeepSeek V3.2 : reasoning_content not cleared from message history on new turns, causing excess token usage and violating API spec #5577: Related issue where reasoning_content from prior turns accumulates in context for DeepSeek models

The root cause described here (Anthropic SDK only serializing thinking blocks and dropping reasoning_content) may be a distinct variant, but the underlying problem domain overlaps significantly.