Deprecated Accessor Trap in AI SDK v7 Silently Erases Agent Memory

A benchmark testing multi-turn tool-calling agents kept showing an odd failure: a model would execute a tool call correctly in one turn, then respond with a bare "DONE" when asked to repeat an equivalent edit later, without calling any tool. The initial explanation blamed small-model fragility in multi-turn tasks, but inconsistent follow-up results, including one prompt mitigation crashing a top model's success rate from 100% to 7.5%, pushed the team to inspect the raw message arrays being sent to the model.

The root cause turned out to be a library behavior: in AI SDK v7, result.response.messages silently changed meaning to return only the final step's text in a multi-step tool run, while the accumulated tool calls and results moved to a separate accessor (result.responseMessages or result.steps). This same field had been the documented pattern for capturing full history in v5, so the code still typechecked and ran, just with a different meaning, flagged only by an editor-visible deprecation note invisible to CI. As a result, every multi-turn conversation quietly lost the model's own tool activity, so from the model's perspective there was no prior tool call at all, making "DONE" the only coherent continuation.

A controlled A/B test across 2,160 cells confirmed the mechanism directly: with the buggy history, follow-up success ranged from 5% to 100% depending on the model, but with corrected history every model and configuration hit 100%. Frontier models mask the defect well because they act confidently on the current instruction regardless of context quality, which means the bug can lie dormant until a cheaper, smaller model is swapped in and the failure surfaces as a silent, non-throwing behavior. The recommended fix is a simple audit: dump the actual message array sent on each turn and check whether tool-call and tool-result parts from prior turns are actually present.