Summary

• guide Box text reads to use representation signals before uncertain get_file_content and Deep Research fetch reads
• stop retrying Box text reads after the selected file reports no markdown or extracted-text representation
• carry known file size into preview selection so get_file_preview is avoided for files already known to exceed Box's 3 MB limit
• bump the Box plugin version for the skill update

Why

The Box invalid-argument review for UTC May 14-20 found get_file_content text-representation misses as a clean skill-guidance target: 490 errors, or 13.5% of Box errors in that window. The residual-tail pass also found the same explicit missing-representation message in 37 Box fetch error-log rows; this PR folds that exact subset into the same text-read guidance.

A separate residual preview pass found the 3 MB preview limit dominates residual get_file_preview invalid arguments: 40 size-cap log shapes among 53 residual preview errors.

The broader fetch bucket remains mixed: Box also returns one error shape that conflates not-found, non-file, and non-extractable-text cases. This PR does not claim to fix that mixed shape. Generic Box AI invalids, metadata schema lookup, and metadata search also did not justify more skill text because they were already described by the MCP surface/provider error or opaque without more Box-side argument validation detail.

Validation

• git diff --check -- plugins/box
• node plugins/plugin-eval/scripts/plugin-eval.js analyze plugins/box --format markdown (reports existing Box plugin budget/structure warnings; no failures)