Hasty Briefsbeta

Bilingual

Deprecated Accessor Footgun in AI SDK v7

a day ago
  • #AI Benchmarking
  • #Conversation History
  • #Tool Calling
  • A bug in conversation history construction caused small models to fail multi-turn tool calling.
  • The bug erased the model's own tool calls and results from its history, making follow-up tasks impossible.
  • Correcting the history by including tool calls and results restored 100% success rates across all models.
  • Frontier models masked the defect by performing well despite incomplete history, hiding the issue.
  • Prompt mitigations interacted unpredictably with the broken history, sometimes causing catastrophic failures.
  • Auditing conversation histories and adding regression tests are essential to avoid similar bugs.
  • Whole-artifact rewrite designs are immune to this failure class as they have no hidden tool channel.
  • Published benchmark results were corrected after identifying the bug, showing tools and rewrite at parity.