One hidden SDK default, two very different benchmarks

The tool-history footgun

Study Overview

The Manufactured Gap

The original +33pp multi-turn gap was manufactured by conversation histories that omitted the model's own tool calls (the AI SDK v5→v7 response.messages trap).

The Correction

Same tasks, same prompts; only history construction differs. Under corrected history, all 2,160 Study G cells pass.

Charts & Tables

Chart G.1: Reference-Edit Success by Model, Hidden vs. Corrected Tool History