Deprecated Accessor Footgun in AI SDK v7 (www.lightningjar.com)

A significant oversight in the AI SDK v7 has been identified, affecting how multi-turn tool calls are handled by smaller language models. During benchmarking, it was discovered that the SDK's management of conversation history was flawed: it only retained the final output of a tool call, erasing critical context that models needed to execute follow-up instructions. This led to confusion in model responses, where they appeared to comply with user requests but failed to execute essential edits. In tests, models like Gemini 3.5 Flash and Claude Haiku 4.5 exhibited drastically reduced performance due to this hidden history issue, showcasing success rates plummeting to as low as 3.8% on multi-turn tasks. The situation underscores the importance of careful history management in dialogue systems, especially when transitioning between SDK versions. The findings reveal that the flaw lies not in the model's capabilities themselves but in how external scripts communicate with them. This bug highlights broader implications for the AI/ML community, supporting the need for stringent audits and tailored regression tests to ensure robust model performance. As multi-turn evaluations proliferate, this incident serves as a cautionary tale about reliance on traditional architectures and the critical need for transparency in tool interactions.