What it requires. A generation whose length limit is low enough that answers are cut off, with no indication to the reader that anything was lost.
Why it matters. A response cut off at max_tokens presents as a model failure rather than a limit, and costs real time to diagnose, because a truncated response is a well-formed one - it ends, it reads, and nothing marks the point where it stopped. It is the same failure class as AG-020 and worth keeping separate because the fix is a number rather than a schema.
How to fix it. Raise the limit to fit the longest legitimate answer, or have the template return a structured result whose completeness can be checked.
Source. Root-cause record, max_tokens truncation, August 2026
Is this rule worth checking?
Votes and comments are published here. We read them, and we publish what we change with the reasoning — a vote does not move a rule on its own.