Does Scolding Silence the Search Tool? AdvTool-Traps, a Matched-Fork Benchmark of Action-Space Sycophancy under Hostile User Feedback in Tool-Augmented Large Language Model Agents

Data, code, prompts, retained Task C trajectory provenance, final frozen judge-specific per-item workbooks and automated-judge verdicts, and de-identified aggregate human-audit vote tallies supporting the AdvTool-Traps matched-fork study of tool-use degradation under hostile user feedback. Superseded run files that were not used in the reported analyses are not included. Component licenses are documented in NOTICE.txt: original data, labels, and figures are CC BY 4.0; code, prompts, and FreshQA-derived materials are Apache License 2.0.